text stringlengths 1.23k 293k | tokens float64 290 66.5k | created stringdate 1-01-01 00:00:00 2024-12-01 00:00:00 | fields listlengths 1 6 |
|---|---|---|---|
USING OF MULTILAYER NEURAL NETWORKS FOR THE SOLVING SYSTEMS OF DIFFERENTIAL EQUATIONS
The article considers the study of methods for numerical solution of systems of differential equations using neural networks. To achieve this goal, the following interdependent tasks were solved: an overview of industries that need to solve systems of differential equations, as well as implemented a method of solving systems of differential equations using neural networks. It is shown that different types of systems of differential equations can be solved by a single method, which requires only the problem of loss function for optimization, which is directly created from differential equations and does not require solving equations for the highest derivative. The solution of differential equations’ system using a multilayer neural networks is the functions given in analytical form, which can be differentiated or integrated analytically. In the course of this work, an improved form of construction of a test solution of systems of differential equations was found, which satisfies the initial conditions for construction, but has less impact on the solution error at a distance from the initial conditions compared to the form of such solution. The way has also been found to modify the calculation of the loss function for cases when the solution process stops at the local minimum, which will be caused by the high dependence of the subsequent values of the functions on the accuracy of finding the previous values. Among the results, it can be noted that the solution of differential equations’ system using artificial neural networks may be more accurate than classical numerical methods for solving differential equations, but usually takes much longer to achieve similar results on small problems. The main advantage of using neural networks to solve differential equations` system is that the solution is in analytical form and can be found not only for individual values of parameters of equations, but also for all values of parameters in a limited range of values.
Ключевые слова: системы дифференциальных уравнений, искусственные нейронные сети, многослойная нейронная сеть, многочисленные методы, метод градиентного спуска, функция погрешности решения Вісник Національного технічного університету «ХПІ». Серія: Системний 82 аналіз, управління та інформаційні технології, № 2 (6)'2021 Introduction. Differential equations and their systems are widely used in mathematical modeling to describe a variety of real processes: physical, environmental, biological, and other. Solving some equations in partial derivatives in cases that allow the separation of variables is also reduced to problems for ordinary differential equations. These are, as a rule, boundary value problems (problems of natural oscillations of elastic beams and plates, determination of the spectrum of natural values of particle energy in spherically symmetric fields, etc.). In addition, higher-order differential equations lead to the solution of systems of differential equations. It is known that solutions of differential equations and their systems can be found analytically or numerically. Finding analytical solutions is a very time consuming process, and in most cases impossible. Therefore, at present, traditional numerical methods are widely used to solve differential equations and their systems, among which the most well-known are Runge -Kutta methods, finite-difference methods, prediction and correction methods [1,2].
The general problem of classical numerical methods is the need to choose their parameters to ensure a compromise between computational costs and the accuracy of the result. Therefore, in this work it is forbidden to use artificial multilayer neural networks, where, in contrast to classical methods, the solution is presented in analytical form, from which you can repeatedly take derivatives [3,4]. Solutions are stored as neural network parameters, which requires much less memory than storing a solution as a discrete array in traditional numerical methods [5,6]. The method is also universal and can therefore be used to solve different types of differential equations and their systems, both ordinary and partial derivatives [7][8].
The main advantage of using neural networks to solve differential equations' systems is that the solution is in analytical form and can be found not only for individual values of parameters of equations, but also for all values of parameters in a limited range of values.
A review of the literature showed the relevance of the problem and the feasibility of creating software. Therefore, the aim of this article is to solve systems of differential equations using a multilayer neural network.
The article's objective is to study the methods of numerical solution of ordinary differential equations` systems and to develop software for their solution using multilayer neural networks.
To solve this problem, the solution is presented in the form [3,5]: whereneural network function with parameters and input values .
In this case, the initial conditions are not satisfied by the creation and therefore are studied gradually during the learning of the neural network.
The construction of the solution of differential equations' systems can be written in a form that satisfies the initial conditions from the beginning: where ( ) is a function that satisfies the initial conditions in advance; Z( )function what construct as the points corresponding which are equal to zero to the coordinates of the initial conditions ( , )output of backforward neural network with input the weights .
The task of a building function ( ) is reduced to the task of the function that takes a certain values in the given points, and can take any value at all other points. To find the function, for example, an interpolation polynomic of Lagrange can be used in this case that looks like: where ( ) -basic polynomials are determined by the formula: To reduce the influence of the shape of the error of the approximation of the solution, we write the expression for Z( ) in the form: A multilayer neural network of direct propagation is chosen as the structure of the neural network for solving differential equations' systems. The number of layers and the number of neurons in each layer are chosen based on the structure of the problem and the complexity of the form of the solution. These parameters are chosen after the experiments, because it is impossible to know in advance the optimal parameters of the neural network structure for each task.
The description of a multilayer neural network. An artificial neural network is a structure that consists of a large number of processor elements, each of which has local memory and can interact with other elements [3,4,6,9,12]. This interaction takes place through communication channels in order to transmit data that can be interpreted in any way. Processor elements independently in time process the local data arriving to them through input channels. Changing the parameters of the algorithms of such processing depends only on the characteristics of the data. If we consider an artificial Вісник Національного технічного університету «ХПІ». Серія: Системний аналіз, управління та інформаційні технології, № 2 (6)'2021 83 neural network as an environment for information processing, then it can be set by defining the elements of this environment and the rules of their interaction. Multilayer artificial neural networks can be considered as a serial connection of single-layer artificial neural networks of direct propagation. The structure of weights in these networks is organized in such a way that more complex classes are processed on layers of highlevel neurons by combining and intersecting simple classes, which are formed at lower levels of artificial neural networks. There is strong evidence that two-layer artificial neural networks are able to recognize any class of convex shape, provided that it is possible to use a sufficient number of hidden layer neurons, and the weights are adjusted accordingly [8][9]. Artificial neural networks of direct propagation with several hidden layers are potentially capable of recognizing classes of arbitrary shape. Therefore, setting the problem on artificial neural networks of direct propagation includes determining the minimum possible number of neurons in the hidden layer and choosing an effective method of adjusting the weights. To date, both of these problems are not trivial. To explain the basic principles of building teaching methods with the teacher we will consider a two-layer artificial neural network. The zero layer of this network performs the auxiliary function of signal branching and does not contain neurons. For this reason, his work does not lead to modification of the input vector. The last layer of artificial neural networks is called the source layer. All layers located between zero and source are hidden layers with nonlinear activation function of neurons. In this example, we will consider one hidden layer with m neurons that use the hyperbolic tangent as an activation function.
It consists of neurons that are simultaneously able to receive the input vector of signals = ( 1 , … , , … , ). To reproduce the elements of this vector use special devices, which are shown to the left of the neurons. These devices do not perform information processing, so they are not considered a layer of the neural network. According to the model of a formal neuron, each of its input signals is multiplied by a weighting factor , wherethe current vector element number , аthe current neuron number. All weights of a single-layer neural network form a matrix of weights Then the vector of arguments is defined as the product of = and the vector of output signals is the vector of values of activation functions: The name of the networks indicates that they have a dedicated direction of propagation of signals that move from the input through one or more hidden layers to the output layer. It is easy to see that a multilayer neural network can be obtained by cascading single-layer networks with matrices of weights 1 , 2 , … , , where is the number of layers of the neural network. If the multilayer neural network is linear, then for activation functions it can be reduced to the equivalent single-layer with a matrix of weights = 1 * 2 * … * . This means that the formation of such structures makes sense if nonlinear activation functions in neurons are used.
The gradient descent method for artificial neural networks.
The idea of the gradient descent method is to sequentially change the parameters of the artificial neural network in a direction that reduces the target function [5]. Since the function is differentiated by each of the parameters, it is possible to calculate the gradient vector. Moving in the direction of the negative gradient for each of the parameters, we find the local minima of the objective function. The change in the parameter is expressed by the formula: This algorithm is called a batch-type algorithm, because to determine the magnitude of the step of changing the parameter, it is necessary to process the entire training sample.
The parameter is searched as (7): The step of changing the weights of the source layer is equal to (8): The step of changing the weights of the hidden layer: But in fact, the correct calculation of values at points closer to the initial conditions is much more important than the calculation at points further away. To correct the optimization to take into account the influence of the values of the functions in the previous points on the values of the functions in the following points, the calculation of the loss function was modified to give the greatest weight to points closer to the initial conditions, keeping the sum of the loss function.
where ( )the loss function; argument of the required function; points at which optimization is performed; the number of points at which optimization is performed.
The main results of the work. The Python and R programming languages were used for perform this work, the TensorFlow library was chosen as the machine learning library with neural network learning support, and the PyCharm environment was used as the integrated development environment.
The final result of solving the problem is shown in fig. 10-11. The obtained optimization result in a form that satisfies the initial conditions for construction: 10000/10000-18s-loss: 3.3917• 10 −07 -rmse: The solution has a root mean square error 1.3370 • 10 −04 in comparison with the solution of the implicit Runge -Kutta's method of the 4th order with the step 10 −03 . Conclusions. The system (12) is difficult to solve with neural networks and could not be solved without additional changes to the loss function, regardless of the form of solution. The applied modification can be used in other cases, when the solution of differential equations by optimization methods coincides to the local minimum. When solving the system (11), the accuracy of the reproduction of the initial conditions had a significant effect on the whole solution. The error of the solution in the basic form was 4.59 times higher than the error of the solution in the form with satisfaction of the initial conditions for construction. Based on the results, we can say that the choice of the form of the solution and the construction of the loss function depends on the differential equations system and the needs of the problem to be solved. Some differential equations require special forms of construction of the loss function to be solved. Fig. 10 The final solution of the differential equations system (12) Fig. 11. The final solution error`s function and the redistributed solution error function of the differential equations system (12) | 3,084.2 | 2021-12-28T00:00:00.000 | [
"Computer Science",
"Mathematics"
] |
The p-Laplacian in thin channels with locally periodic roughness and different scales
In this work we analyse the asymptotic behaviour of the solutions of the p-Laplacian equation with homogeneous Neumann boundary conditions posed in bounded thin domains as Rε=(x,y)∈R2:x∈(0,1) and 0 0. We take a smooth function G:(0,1)×R↦R , L-periodic in the second variable, which allows us to consider locally periodic oscillations at the upper boundary. The thin domain situation is established passing to the limit in the solutions as the positive parameter ɛ goes to zero and we determine the limit regime for three case: α < 1, α = 1 and α > 1.
Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Introduction
Partial differential equations on thin domains (domains in which the size in some directions is much larger than the size in others) appear naturally in biological systems and industrial applications [13,14,24]. In most of the applications, the boundary of those domains is not perfectly flat and one can see irregularities. Then, the influence of such boundary distortions might not be neglected because its effect on the effective equation of the considered system, even far from the rough boundary, can be meaningful [1,9,11]. This motivates researchers to employ different homogenization techniques and try to determine the effective flow behaviour on a lower-dimensional domain which captures the influence of the geometry, roughness and thickness of the perturbed domain on the solutions of such singular boundary value problems. The obtained equations are then suitable for numerical simulations and provide rigorous justification of various natural phenomenon seen in such complex systems.
A simple manner to consider such irregularities is to study domains of type Q ε = (x, y) ∈ R 2 : x ∈ (0, 1) and 0 < y < εg x ε α for ε > 0, where g is a positive, bounded and periodic function satisfying some regularity hypothesis and ε > 0 is a small parameter which goes to zero. Thereby, in the limit ε → 0, the open set Q ε degenerates to the unit interval presenting oscillatory behaviour on the upper boundary (see for instance [1,3,5,[18][19][20][21][22] where similar approach are performed). The periodic rough boundary considered above is certainly a first step, but usually not enough, since most of the irregularities present in real applications are not periodic. In this work we are interested in the following family of rough thin domains R ε = (x, y) ∈ R 2 : x ∈ (0, 1) and 0 < y < εG ε (x) for ε > 0, (1.1) where for some parameter α > 0 with function G satisfying the conditions given by hypothesis (H) set in section 2. This kind of domain perturbation is called in the literature locally periodic thin domain and it is illustrated in figure 1 below.
As an example, one can consider G ε (x) = a(x) + b(x)g(x/ε α ) where a, b : (0, 1) → R are C 1 -piecewise positive functions and g : R → R is a L-periodic function of class C 1 also positive. This includes the case where a, b are positive constants recovering the perturbed regions discussed for instance in [3,5]. Notice that in the case in which α = 0, we also recover the open sets considered in [13] where evolution equations on thin domains without roughness were studied. We observe that the hypothesis (H) considered here is as general as possible for our framework.
In a unified way, we treat the three cases of roughness that can be modelled by the parameter α > 0. We analyse our boundary value problem for 0 < α < 1, α = 1 and α > 1, which represents weak, resonant and strong harshness on the upper boundary respectively. In each case, we have a different effective equation featuring the roughness induced effects on the perturbed model for small values of the parameter ε.
Several references treat issues related to the effect of thickness and rough boundaries on the feature of the solutions of partial differential equations. Indeed, thin structures with oscillating boundaries appear in many fields of science: fluid dynamics (lubrication), solid mechanics (thin rods, plates or shells) or even physiology (blood circulation). Therefore, analysing the asymptotic behaviour of models set on thin structures understanding how the geometry and the roughness affect the problem is of considerable current interest in applied science. In these directions, let us mention [7,10,15,21] and references therein.
In this paper, we are interested in analysing the asymptotic behaviour of the solutions of a p-Laplacian equation given by where η ε is the unit outward normal vector to the boundary ∂R ε , 1 < p < ∞ with p −1 + p −1 = 1, and Δ p · = div |∇ · | p−2 ∇· , denotes the p-Laplacian differential operator. We also assume f ε ∈ L p (R ε ) is uniformly bounded. Such quasilinear equations play an important role in applications, given the fact that many models cannot be described by linear equations. In this sense, considering the p-Laplacian equation becomes natural. Moreover, the p-Laplacian is strongly related to non-Newtonian fluids, which arise in many applications related to polymer processing, hydrology, food processing, turbulent filtration, glaciology (see e.g. [6,16,17,25]). Here, differently from many works [11,12], we deal also with the case 1 < p < 2, which is the most relevant range of p in applications (e.g. [6]) and, of course, the case p 2.
We improve the results from [3] (where the Laplacian operator in locally periodic thin domains were considered) dealing with the p-Laplacian equation for any p ∈ (1, ∞). Moreover, we are improving our previous results from [2] where the purely periodic case in bidimensional thin regions were studied. It is worth noticing that the techniques developed in [2,3] cannot be directly applied in this case. On the one hand, the results concerning the unfolding operator obtained in [4] do not guarantee strong convergence in L p for the unfolding operator applied on solutions of quasilinear operators. On the other hand, the analysis performed in [3] just works on L 2 -spaces. Our goal here is to overcome this situation. We discretize the oscillating region passing to the limit using uniform estimates on two parameters: one associated to the roughness, and other given by the variable profile of the thin domain. In this way, a continuous dependence property for the solutions with respect to the function G in L p -norms is crucial and it is obtained in theorem 4.1. We point out that these techniques also work for the dimension reduction from three-dimensional thin sets to two-dimensional ones. The main change is in the limit problem. In 3D, we somehow lose the explicit p-Laplacian form, as in the unidimensional limit, but, clearly, the monotonicity of this limit operator is preserved (it will be done in a forthcoming work).
Notice that our work also goes a step further from [23] where the p-Laplacian operator is studied in standard thin domains. Let us emphasize that the standard thin domains were previously introduced and rigorously studied in the paper [13] of Hale and Raugel where the continuity of the family of attractors set by a semilinear parabolic equation in thin domains was considered.
According to [1] and references therein, it is expected that the sequence u ε will converge to a function of just one variable x ∈ (0, 1) satisfying a one-dimensional equation of the same type. Combining boundary perturbation techniques [3][4][5] and monotone operator analysis [17], we identify the effective limit model of (2.1) at ε = 0.
The paper is organized as follows. In section 2 we state the main result of the paper. In section 3, we introduce some notations and state some basic results which will be needed in the sequel. In section 4, we prove the continuous dependence of the solutions in L p -spaces with respect to the function G uniformly in the parameter ε > 0 improving [3, theorem 4.1] from L 2 to L p -spaces. In section 5, we perform the asymptotic analysis of (1.2) in piecewise periodic thin domains (that is, in thin domains set by functions G which are piecewise constants in the first variable x, and L-periodic in the second one). See figure 2 below which illustrates piecewise periodic open sets.
Next, we provide in section 6 the proof of the main result of the paper (namely theorem 2.1) as a consequence of the analysis performed in the previous sections. Finally, we discuss in section 7, the convergence of the resolvent and semigroup associated to the equation (1.2) under the additional assumption p 2. As we will see, it is obtained combining the classical result [8, theorem 4.2] and our main result theorem 2.1. Furthermore, we include an appendix where the dependence of the auxiliary solution v on admissible functions G is analysed.
Hypothesis on function G and the main result
First, recall that the variational formulation of (1.2) is given by Moreover, existence and uniqueness of the solutions are guaranteed by Minty-Browder's theorem setting a family of solutions u ε .
Next, we state the main hypothesis on function G setting the main conditions on our rough thin domain R ε introduced in (1.1).
Let G : (0, 1) × R → R be a function satisfying that there exist a finite number of points is C 1 and such that G, ∂ x G and ∂ y G are uniformly bounded in (ξ i−1 , ξ i ) × R getting limits when we approach ξ i−1 and ξ i . Further, we assume there exist two constants G 0 and G 1 such that and a real number L > 0 such that G(x, y + L) = G(x, y) for all (x, y) ∈ (0, 1) × R. 3 As we will see, the homogenized limit equation is a one-dimensional p-Laplacian equation with variable coefficients q(x) and r(x). It assumes the following form where the homogenized coefficients are given by and We emphasize here the dependence of the function q(x) with respect to the parameter α > 0 and variable x ∈ (0, 1) which generalizes our previous work [2]. The functionf is the weak limit off ε in L p (0, 1) withf ε defined by the family of known forcing terms f ε ∈ L p (R ε ) in the following waŷ |Y * (x)| denotes the Lebesgue measure of the representative cell which also depends on variable x ∈ (0, 1). The function v used to set the homogenized coefficient q(x) in (2.2) is the unique solution of the problem is the space of periodic functions on the horizontal variable y 1 , and ϕ O denotes the average of any function ϕ ∈ L 1 loc (R M ) on measurable sets O ⊂ R M . It is worth noticing that problem (2.3) is well posed for each x ∈ (0, 1), due to Minty-Browder's theorem, and then, the coefficient q(x) is also well defined. Further, q(x) is a positive function setting a well posed homogenized equation. Indeed, since v is the solution of The main result of the paper is the following: satisfiesf ε f weakly in L p (0, 1). Let u ∈ W 1,p (0, 1) be the unique solution of the homogenized equation where the homogenized coefficients q(x) and r(x) depend on the parameter α > 0 and are given by the expression (2.2). Then, As mentioned before, we are improving the results from [3] where the Laplacian operator in locally periodic thin domains were considered. We recover them taking p = 2 in theorem 2.1. Moreover, we also have improved our previous results from [2] where the purely periodic case in bidimensional thin regions were studied to the p-Laplacian operator where constant homogenized coefficients are obtained. Here, since we are considering locally periodic thin domains, variable homogenized coefficients can be produced. The main step in the proof is to pass to the limit in the solutions with the representative cell depending on variable x ∈ (0, 1) assuming different orders of roughness (different values for the parameter α > 0). To do that, we discretize the oscillating thin region passing to the limit using uniform estimates on two parameters: one associated to the roughness, and other given by the variable profile of the thin domain. In this way, a continuous dependence property for the solutions with respect to the function G in L p -norms is crucial and it is shown in theorem 4.1 below.
Basic facts and the unfolding operator
In this section, we introduce some basic facts, definitions and results concerning to the unfolding method making some straightforward adaptations to our propose. First, let us just recall some basic properties to the p-Laplacian which can be found for instance in [17].
Unfolding operator
Here, we present the unfolding operators for thin domains in the purely and locally periodic settings. We rewrite it to our context in order to simplify our proofs. They were first introduced in [4,5] where details and proofs can be found.
3.1.1. The purely periodic unfolding. Let G i : R → R be a L-periodic function, lower semicontinuous satisfying 0 < g 0,i G i (x) g 1,i with g 0,i = min x∈R G i (x) and g 1,i = sup x∈R G i (x) for any i = 1, . . . , N. Now consider the thin region The basic cell associated to R ε i is By we denote the average of ϕ ∈ L 1 loc (R 2 ) for any open measurable set O ⊂ R 2 . We also set functional spaces which are defined by periodic functions in the variable y 1 ∈ (0, L). Namely For each ε > 0 and any x ∈ (ξ i−1 , ξ i ), there exists an integer denoted by x ε L such that We still set Now we can introduce the unfolding operator. In the sequel, we point out its main properties.
Proposition 3.5. The unfolding operator satisfies the following properties: Then, The above result sets several basic and somehow immediate properties of the unfolding operator. Property four will be essential to pass to the limit when dealing with solutions of differential equations since it allow us to transform any integral over the thin sets depending on the parameter ε and function G i into an integral over the fixed set (
Locally periodic unfolding.
Next we set the locally periodic unfolding operator discussing some properties that will be needed in the sequel.
Definition 3.6. We define the locally periodic unfolding operator T lp ε acting on a measurable function ϕ, as the function T lp where · denotes the extension by zero to the whole space.
As in classical periodic homogenization, we have the unfolding operator reflecting two scales. The macroscopic one, denoted by x which gives the position in the interval (0, 1), and the microscopic scale given by (y 1 , y 2 ) which sets the position in the cell (0, L) × (0, G 1 ). However, due to the locally periodic oscillations of the domain R ε , the definition given here differs from the usual ones. In this case, we do not have a fixed cell that describes the domain R ε which makes the extension by zero needed.
Remark 3.1. We point out that the convergence above cannot be improved because of the definition of locally periodic unfolding operator.
A domain dependence result
In this section we analyse how the solutions of (1.2) depends on the function G ε . Let us take satisfying hypothesis (H) and considering the associated thin domains R ε andR ε by Now, let u ε andû ε be the solutions of (1.2) for the domains R ε andR ε respectively with f ε ∈ L p (R 2 ). We have the following result.
Remark 4.1. The important part of this result is that the function ρ(δ) does not depend on ε. As we will see, it only depends on the positive constants G 0 and G 1 .
In order to prove theorem 4.1, we use the fact that u ε andû ε are minimizers of the functionals that is, We will need to evaluate the minimizers plugging them into different functionals. For this, we set the following operators introduced in [3]: and U ⊂ R 2 is an arbitrary open set. We also consider the following norm in W 1,p (U) We can easily see that , (4.6) and as η 0.
Also, we need the following result about the behaviour of the solutions near the oscillating boundary.
Now, let us first assume p 2. We use the notations of corollary 3.1.1 to simplify proofs. By proposition 3.2, (4.2) and (2.1) for ϕ = P 1+η u ε − u ε , we get Putting together (4.7) and (4.8), we obtain Consequently Now, let us analyse the integral: To do this, notice that putting the power p, multiplying by 1/ε, integrating between 0 and εG ε (x) and using that (y/(1 + η), y) ⊂ (εG ε (x)), we get Thus, we have (4.10) Hence, due proposition 3.3, (4.9) and (4.10), one gets On the other hand, we have Hence, due to (4.8), we get and then, Thus, due proposition 3.3 and (4.10), we get for p > 2 that Notice that to the case p > 2, we have mainly estimated the term |x − y| p . Now, for the case 1 < p < 2, we have to estimate (1 + |x| + |y|) p−2 |x − y| 2 in view of propositions 3.1 and 3.2. Indeed, we can argue as in (4.11) and (4.12), to get, for 1 < p < 2 that and Finally, putting together the last inequality and (4.13), we also obtain cη + cη 1/p + cη + cη 1/p p/2 , for 1 < p < 2 finishing the proof. Now, we are in condition to show theorem 4.1.
The piecewise periodic case
Now, we analyse the limit of {u ε } ε>0 assuming the upper boundary of R ε is piecewise periodic. More precisely, we assume G satisfies (H) being independent on the first variable in each interval (ξ i−1 , ξ i ).
We suppose that G satisfies with G i (y + L) = G i (y) for all y ∈ R. Moreover, we assume the function G i (·) is C 1 for all i = 1, . . . , N and there exist 0 Notice that the domain R ε can now be rewritten as See figure 2 which illustrates this piecewise periodic thin domain. Before proving the main result of this section, let us recall an important result proved, for instance, in [21]. It is concerned to the purely periodic thin domain situation. Proposition 5.1. Assume G satisfies the condition (5.1) and let u ε be the solution of (1.
where ∇ y · = ∂ y 1 ·, ∂ y 2 · and v i is the solution of the auxiliar problem If α > 1, then there exists an unique and the scaling operator Π ε :
Proof. It follows from [21, theorems 3.1, 4.1 and 5.3].
Remark 5.1. We point out that the results in [21] are proved in the unit interval. Here, we just rewrite it to (ξ i−1 , ξ i ). The limit problems are stated in the next result. Now, we are in condition to show the following result.
Theorem 5.2. Suppose G satisfies the assumption (5.1) and let u ε be the solution of problem
and u is the unique solution of the problem where q, r : (0, 1) → R are piecewise constant functions such that with the homogenized constants r i and q i given by where Y * i is the basic cell associated to R ε i Y * i = (y 1 , y 2 ) ∈ R 2 : 0 < y 1 < L and 0 < y 2 < G i (y 1 ) , and v i is the solution of the auxiliary problem Also, u is the unique solution of the problem (5.3) with q(x) = q i and r(x) = r i for x ∈ (ξ i−1 , ξ i ), (5.6) where If α > 1, then there exists a unique u ∈ W 1,p (0, 1) such that Furthermore, u is the unique solution of the problem (5.3) with Proof. By (5.2), we can rewrite (2.1) taking into account the partition Hence, we obtain from (5.7) (with test functions ϕ(x, y) = ϕ(x) ∈ W 1,p (0, 1)) and proposition 3.5 that By proposition 5.1, we can pass to the limit in each subinterval (ξ i−1 , ξ i ). If we assume α 1, we obtain which is equivalent to J C Nakasato and M C Pereira for all ϕ ∈ W 1,p (0, 1). For α < 1, proposition 5.1 guarantees Since (p − 1)(p − 1) = 1, (5.9) can be rewritten as ϕ dx, ∀ϕ ∈ W 1,p (0, 1), where the functions u i are given by proposition 5.1. Notice that q i > 0 for each i. Indeed, by (5.5), we can take (v i − y 1 ) ∈ W 1,p #,0 (Y * i ) as a test function in such way that Consequently, we obtain from the Minty-Browder's theorem that the problem (5.11) has a unique solution in W 1,p (0, 1), and then, we can conclude that u ∈ W 1,p (0, 1) proving the theorem for α 1. Now, let us assume α > 1. Then, from (5.7) and proposition 3.5, we obtain that where Π ε is the scaling operator introduced in proposition 5.1. Hence, by proposition 5.1, we can pass to the limit taking test functions ϕ(x, y) = ϕ(x) ∈ W 1,p (0, 1). We obtain where the functions u i are given by proposition 5.1. Thus, ϕ dx, ∀ϕ ∈ W 1,p (0, 1). (5.12) As G 0 > 0, it follows from Minty-Browder's theorem that (5.12) is well posed. Hence, we get that u ∈ W 1,p (0, 1) is the unique solution concluding the proof of the theorem.
The locally periodic case
In this section, we provide the proof of our main result, theorem 2.1.
Proof of theorem 2.1. Using proposition 3.3 and theorem 3.7, there is u 0 ∈ W 1,p (0, 1) such that, up to subsequences, where χ is the characteristic function of (0, 1) × Y * (x). We show that u 0 satisfies the Neumann problem (5.3). To do this, we use a kind of discretization argument on the oscillating thin domains. We first proceed as in [3, theorem 2.3] fixing a parameter δ > 0 in order to set a piecewise periodic function G δ (x, y) satisfying (5.1) and 0 G δ (x, y) − G(x, y) δ in (0, 1) × R.
Let us construct this function. Recall that G is uniformly C 1 in each of the domains (ξ i−1 , ξ i ) × R. Also, it is periodic in the second variable. In particular, for δ > 0 small enough and for a fixed z ∈ (ξ i−1 , ξ i ) we have that there exists a small interval (z − η, z + η) with η depending only on δ such that |G(x, y) − G(z, y)| + |∂ y G(x, y) − ∂ y G(z, y)| < δ/2 for all x ∈ (z − η, z + η) ∩ (ξ i−1 , ξ i ) and for all y ∈ R. This allows us to select a finite number of points: Notice that this construction can be done for all i = 1, . . . , N. In particular, if we rename all the constructed points ξ k i by 0 = z 0 < z 1 < · · · < z m = 1, for some m = m(δ), we get that G δ (x, y) = G δ i (y) for (x, y) ∈ (z i−1 , z i ) × R and i = 1, . . . , m is a piecewise C 1 -function which is L-periodic in the second variable y.
Since ϕ and η are arbitrary, we conclude that u * = u 0 . Finally, let us see that the convergence holds. Notice that Hence, we can argue as in (6.9) getting (6.10) from (6.2), remark 4.2 and (6.6). And then, we conclude the proof of the theorem.
Convergence of the resolvent and semigroups
In this section, we show the convergence of the resolvent and semigroup associated to the p-Laplacian operator given by the equation (1.2) under the additional condition p 2. For that, let us first consider the operator M ε : L p (R ε ) → L p (0, 1) given by Next, let A ε : W 1,p (R ε ) → (W 1,p (R ε )) and A 0 : W 1,p (0, 1) → (W 1,p (0, 1)) be given by We consider the L 2 -realization of A ε and A 0 , that is, , and D(A 0,2 ) = u ∈ W 1,p (0, 1) : A 0 u ∈ L 2 (0, 1) ) , Then, for any p 2, λ > 0 and forcing terms f ε ∈ L 2 (R ε ), we can consider the following problems (I + λA ε )u ε = f ε , (7.2) and (I + λA 0 )u =f , (7.3) which are well posed (existence and uniqueness of solutions) by the Minty-Browder's theorem. Notice that here, we are using the dual products ·, · ε and ·, · 0 from W 1,p (R ε ) and W 1,p (0, 1) respectively to set the equations (7.2) and (7.3). Hence, with the additional conditions | f ε | L 2 (R ε ) uniformly bounded and M ε f ε f weakly in L 2 (0, 1), it follows from theorem 2.1 that the family of solutions defined by (7.2) converges to the solution of (7.3) as ε → 0. Consequently, we obtain the convergence of the resolvent operators defined by the equation (1.2). In fact, we have for any λ > 0 that In the next, let us obtain the convergence of the semigroup associated to the equations (7.2) and (7.3). As we will see, it is a consequence of [8, theorem 4.2, p 120]. First, let us write the resolvent operators convergence in appropriate spaces. For this purpose, we use the unfolding operator. We have where W = (0, 1) × (0, L) × (0, G 1 ) and ·, · is the dual product in W 1,p (W ). Next, Notice that It remains to observe that wich holds due to theorem 2.1. Therefore, thanks to Neveu-Trotter-Kato theorem, the semigroup S ε (t) associated to −B ε satisfies where S(t) is the semigroup associated to −B 0 . We have the following theorem. Theorem 7.1. Assume p 2 and consider the operators A ε and B ε defined respectively by (7.1) and (7.4). Then, (a) For any f ε ∈ L 2 (R ε ) with | f ε | L 2 (R ε ) uniformly bounded and M ε f ε f weakly in L 2 (0, 1), we have
where S(t) is the semigroup associated
with B 0 given by (7.5).
Appendix A
In the proof of the main result, we used q δ → q uniformly to obtain (6.7). Recall that q δ and q are given by (6.4) and (6.5) respectively. Here we prove such convergence. For this sake, let us first set Hence, for anyḠ ∈ A(M), we can consider the problem and we are looking for solutionsv such that (v − y 1 ) ∈ W 1,p #,0 (Y * G ). Now, for anyḠ, G ∈ A(M), let us consider the following transformation The Jacobian matrix for L is with det(JL) = F. Also, we can consider It is not difficult to see that B = L T L. Then, we can use the change of variables given by L to rewrite (A.2) in the region Y * G as Notice that this problem still has unique solutionv ∈ By the coercivity of (A.3), we get which means that the solutions are uniformly bounded by a constant independent onḠ and G. Now, let us compare the solutions of (A.2) forḠ = G and (A.3). We need to analyse Notice that L(1, 0) = (1, 0). We will distribute the terms finding estimative for each one. First, observe that for any test function ϕ ∈ W 1,p #,0 (Y * G ) in (A.3), we have On the other side, if 1 < p < 2, we get from Hölder's inequality, proposition 3.1 and (A.13), that Therefore, for 1 < p < ∞, we have where α = 1/2 if 1 < p < 2 and α = 1/p if p 2. Finally, since L∇v − ∇v L p (Y * G ) + L∇v − ∇v L p (Y * G ) , we conclude by (A.14) and (A.12) that We have the following lemma: Remark A.1. We remark that the result of the lemma above, works in a more general framework, that is, the functions do not need to be in A(M). On the other hand, to perform the discretization of the domain in the locally periodic case, in the previous section, we need the hypothesis of A(M) functions defining the domains. | 7,018.2 | 2022-05-05T00:00:00.000 | [
"Mathematics",
"Physics"
] |
Social Media Based, Data-mining Driven Social Network Analysis (SNA) of Printing Technologies in Fashion Industry
extends previous research through focus on information flows organized around different printing technologies. Overall, social media based, data-mining driven SNA provides fashion researchers with an innovative approach to inductively characterize the rapidly evolving printing market. Interpretation of the network findings suggest a number of practical directions for multiple stakeholders in the fashion industry and demonstrate the effectiveness of data mining within the fashion domain.
Introduction and conceptual framework
Social media is increasingly recognized as a rich data resource through which savvy fashion brands gather market intelligence (Nash, 2019). Social media proliferation impacts academia and practice because these data provide contextual insight. However, as big data, this dynamic supply of online information with millions of social media messages derived from human activities is difficult to analyze using conventional methodologies (Tsou, 2015). This study demonstrates application of data-mining driven Social Network Analysis (SNA) to generate a model of four predominant printing terms: screen printing, heat transfer, sublimation, and digital printing. Grounded in Graph Theory (GT), data-mining driven SNA uses computational techniques to capture, analyze, and depict key indicators for a given phenomenon among large social media datasets (Zhao & Min, 2018). The current research is predicated on Freeman's (1978) seminal GT work which suggests that node centrality impacts can be evaluated through centrality indices (i.e. degree, betweenness, and closeness). Advances in big data analytics facilitate visualization and interpretation of social networks.
Methodology
As a subset of a long-term study tracking the development of digital printing technology, this study focuses specifically on four technology terms that emerged from earlier network analyses (Yu et al., 2020). Crimson Hexagon software was used to harvest data from Twitter for one year (January 2019-20). A total of 3,000 tweets were randomly selected based on the presence of one or more of the following hashtags: #screenprint, #heattransfer, #sublimation and #digitalprint. Using Python, a matrix was developed including the initial network nodes as well as edges between nodes which depict hashtag co-occurrence. A total of 3,742 nodes and 52, 802 edges were identified following noise removal. Gephi software was applied to cluster the nodes and generate the outcome network visualization.
Results and discussion
This study reveals the interrelated indicators among four predominant printing technology termsthrough the visual SNA network (Figure 1). Each printing technology represents the local 2020 Proceedings
Virtual Conference
Page 2 of 3
© 2020 The author(s). Published under a Creative Commons Attribution License
(https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. ITAA Proceedings, #77 -https://itaaonline.org hub of its corresponding cluster. Screen printing (#screenprint), depicted as the blue community, indicates the highest degree & betweenness centrality and features #tshirt, #fashion, #clothing indicating the dominance of these apparel decoration techniques in fashion. The digital printing community identified by the color pink features a set of nodes related to Direct-to-Garment (DTG) printing which refers to digital printing directly to garments (e.g. #dtg, #dtgpint, #dtgprinter). The pink community reveals potential advantages of DTG over screen printing including: extensive color options, low minimum requirement and quick turnaround time (Figure). Further, the pink community features #smallbusiness, #kornit, and #bellacanvas which reflect dominant characteristics of the domain including business scale, leading digital printing equipment and preferred cotton t-shirts brands for DTG. The unique features of sublimation and heat transfer technologies are presented in the yellow and green communities, respectively. Sublimation is a digital dye process (#dyesub, #dye) which is widely used for polyester fabric and hard surfaces (e.g. #swimwear, #banner, #mug). The hashtags #transferpaper and #heatpress reflect the logical bridge linking sublimation and heat transfer terms. The durability of heat transfer is not ideal, thus, it is widely used for event t-shirt printing (#event). The purple community which encompasses art and design related nodes, suggests the printing market draws on inspiration from street (#streetstyle) and pop culture (#popart). Notably, despite the general interconnectivity of these communities, each community suggests relatively independent structures which provide a means for deeper interpretation presented in this section.
Conclusion
The visual network demonstrates a blueprint for identifying and understanding unique printing technologies in the fashion industry and their dominant characteristics. The present study extends previous research through focus on information flows organized around different printing technologies. Overall, social media based, data-mining driven SNA provides fashion researchers with an innovative approach to inductively characterize the rapidly evolving printing market. Interpretation of the network findings suggest a number of practical directions for multiple stakeholders in the fashion industry and demonstrate the effectiveness of data mining within the fashion domain. | 1,069.6 | 2020-12-28T00:00:00.000 | [
"Computer Science",
"Business"
] |
NCOA4-Mediated Ferritinophagy: A Vicious Culprit in COVID-19 Pathogenesis?
Coronavirus disease 2019 (COVID-19) is a global pandemic that has caused widespread loss of life. Notably, in this disease, severe inflammatory reactions characterized by cytokine storms are caused by severe acute respiratory syndrome coronavirus 2. The cytokine storms may promote hyper-ferritinemia which can further intensify the inflammation. Moreover, elevated ferritin levels trigger nuclear receptor coactivator 4 (NCOA4)-mediated ferritinophagy, in which ferritin is degraded and iron is released. Excess iron released from ferritinophagy can promote ferroptosis and cellular damage. Therefore, we propose that NCOA4-mediated ferritinophagy can be targeted to limit the ferroptosis and prevent the multi-organ damage and severity in COVID-19 patients.
INTRODUCTION
Coronavirus disease 2019 , triggered by the novel severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), is a serious health concern and can be profoundly detrimental. In March 2020, the World Health Organization declared it as a pandemic. Although various vaccines have been developed, only limited number of therapeutic drugs is available. A significant proportion of patients with COVID-19 experience severe interstitial pneumonia, possibly resulting in acute respiratory distress syndrome and systemic inflammatory response syndrome.
COVID-19 Systemic Inflammatory Response
COVID-19 systemic inflammatory reaction can be life-threatening due to its hyper-inflammation sustained by a cytokine storm. The cytokine storm is not only one of the earliest and most debilitating symptoms in patients with COVID-19 but also contributes to the disease severity at the later stage (Siddiqi and Mehra, 2020). Once inside the cells, the SARS-CoV-2 RNA is recognized by the pathogen recognition receptors, which trigger a downstream cascade of molecules leading to the activation of transcription factors, such as nuclear factor kappa B and interferon regulatory factor 3 (IRF-3) and the subsequent production of type 1 interferons and several other pro-inflammatory cytokines (e.g., interleukin [IL]-1β and IL-6) (Kawai and Akira, 2010;Conti et al., 2020;Yang et al., 2020). Considerably, elevated plasma levels of plasma pro-inflammatory cytokines, such as IL-6, IL-2, IL-7, IL-10, IFN-γ-induced protein 10 (IP-10), monocyte chemoattractant protein-1 (MCP-1), and macrophage inflammatory protein-1 alpha (MIP-1α), have been observed in some patients (Han et al., 2020;Karki et al., 2021). This expansion of an uncontrolled inflammatory response due to SARS-CoV-2 infection potentially leads to cell death by apoptosis, necrosis, or ferroptosis (a newly discovered type of programmed cell death), resulting in multi-organ damage in patients with COVID-19 (Banchini et al., 2021). Considering most of the acute and chronic deleterious effects of the SARS-CoV-2 infection are the consequences of hyperinflammatory response of the virus, several inflammatory drugs have been used for the treatment of severe COVID-19 (Zhang et al., 2020).
HYPER-FERRITINEMIA IN COVID-19: A PART OF HYPER-FERRITINEMIC SYNDROME SPECTRUM
COVID-19-associated systemic inflammation is thought to be the part of the hyper-ferritinemic syndromes. COVID-19 shares many clinical and laboratory features such as lymphopenia, reduced NK cell numbers and activity, abnormal liver functions, and coagulopathy, with the other hyper-ferritinemic syndromes such as macrophage activation syndrome (MAS) and adult-onset Still's disease (AOSD) (Rosario et al., 2013;Shoenfeld, 2020). Elevated pro-inflammatory cytokine and ferritin levels are the other hallmarks of hyper-ferritinemic syndromes common to COVID-19 (Colafrancesco et al., 2020).
Ferritin levels are elevated in various inflammatory disorders and directly correlated with poor prognosis and severity of the COVID-19 patients (Mehta et al., 2020). Although the source of circulating serum ferritin in COVID-19 has not yet been determined, macrophages contribute considerably to release ferritin via a non-classical secretory pathway (Cohen et al., 2010). A recent in vitro study showed that hepatocytes could actively secrete ferritin (Ghosh et al., 2004). Besides its active secretion in hepatocytes, serum ferritin is also derived by hepatic cell death (Colafrancesco et al., 2020). Furthermore, a key iron-regulatory hormone hepcidin could increase intracellular ferritin levels by sequestering iron in enterocytes and macrophages (Daher et al., 2017). Activated innate immunity and cytokine cascades can accelerate the expression of hepcidin (Drakesmith and Prentice, 2012;Cassat and Skaar, 2013). Interestingly, SARS-CoV-2 can mimic hepcidin by inducing ferroportin blockage and can result in the elevated ferritin levels independent of the inflammatory response (Cavezzi et al., 2020). Overall, patients with COVID-19 experience hyper-ferritinemia through the above mechanism.
HYPER-FERRITINEMIA AMPLIFIES INFLAMMATION AND THE CYTOKINE STORM
The magnitude of the cytokine storm has a positive correlation with the serum ferritin levels. The high circulating ferritin levels in COVID-19 may not only reflect an acute phase response but also display other immunomodulatory functions (Recalcati et al., 2008;Edeas et al., 2020). Ferritin contains two subunits: L and H. H ferritin induces the expression of inflammatory cytokines such as IL-beta and IL-6 and promotes the myeloid and lymphocyte proliferation by stimulating a specific ferritin receptor TIM-2 (Ruscitti et al., 2016). Also, there is compelling evidence that the cytokine storm in hepatic cells is directly correlated with the extracellular ferritin levels (Ruddell et al., 2009). Thus, the mutual promotion of ferritin and inflammatory cytokine levels generates a vicious loop that constantly heightens the inflammatory state.
IRON DYSHOMEOSTASIS IN COVID-19
Ferritin is composed of 24 subunits, which can bind up to 4,500 atoms of iron. This makes ferritin one of the major iron storage proteins in the cells. It incorporates Fe 2+ via ferritin iron pores and further oxidizes Fe 2+ to Fe 3+ by an H subunit inside the ferritin cage, leading to inert deposits of Fe 3+ that are unavailable for intracellular use or generation of reactive oxygen species (ROS) (Philpott, 2018). Ferritin can also release Fe 3+ and reduce it to Fe 2+ (Kidane et al., 2006), indicating that iron is critical in COVID-19 pathogenesis (Bellmann-Weiler et al., 2020;Edeas et al., 2020). Notably, iron overload directly influences several critical manifestations of COVID-19 such as systemic inflammation, hyper-ferritinemia, and immune dysfunction.
COVID-19-associated iron overload can be triggered by the release of the free iron from the damaged hemoglobin and ferritin catabolism. The process of viral infection is facilitated by several SARS-CoV-2 accessory viral proteins, including open reading frames (ORFs)-3, 10, and 8. The virus can interact with hemoglobin through angiotensin-converting enzyme 2 (ACE-2), CD-147, CD-26, and other receptors. The heme on the 1-beta chain of hemoglobin is attacked by ORF-8 and surface glycoprotein binding to porphyrin. Consequently, hemolysis and dysfunctional hemoglobin are engendered by SARS-CoV-2 infection, with or without forming a complex with the released heme (Cavezzi et al., 2020). Damage to hemoglobin causes dissociation of the porphyrins from iron and releases iron into the circulation resulting in the condition of iron overload (Cavezzi et al., 2020).
A retrospective cohort study showed that ferritin could lose a part of its inner iron content, giving rise to extremely high serum levels of free iron (Pretorius and Kell, 2014). Iron deposits derived from the ferritin lead to a poorly labile iron pool, indicating that the abundance of ferritin is a key factor governing iron homeostasis. Conversely, depletion of ferritin results in the release of iron into the labile iron pool, resulting in increased sensitivity to ferroptosis (Stockwell et al., 2020). Several studies have suggested that the selective autophagic turnover of ferritin (ferritinophagy) contributes to ferroptosis in fibroblasts and cancer cells (Hou et al., 2016). Although the role of ferritinophagy in COVID-19 remains unclear, its contribution in the pathological processes of neurodegeneration, cancer, ischemia/reperfusion injury, and urinary tract infections is well established (Tang et al., 2018).
Iron metabolism is objectively related to clinical syndromes during COVID-19 by inducing a series of biological events (Cavezzi et al., 2020). By interacting with molecular oxygen, excess intracellular iron can generate ROS (through Haber-Weiss and Fenton reactions), reactive nitrogen species (RNS), and reactive sulfur species (RSS) Frontiers in Molecular Biosciences | www.frontiersin.org December 2021 | Volume 8 | Article 761793 (Oexle et al., 1999). Through redox damage, mitochondrial dysfunction is favored, resulting in ferroptosis, multiple tissue damage, and subsequent fibrosis (Le Lan et al., 2005). During severe inflammatory conditions, excess iron can deteriorate the inflammatory reaction by inducing a severe pro-coagulant state (Pretorius and Kell, 2014). Moreover, higher iron levels may facilitate virus multiplication in the host cells (Drakesmith and Prentice, 2008). Iron depletion or chelation has been considered as a potential antiviral therapy to protect against excessive inflammatory responses and tissue damage by sequestering iron and preventing oxygen radical formation and lipid peroxidation in patients with COVID-19 (Perricone et al., 2020). In addition, the binding of SARS-CoV-2 to its receptors for entry into host cells can be prevented by lactoferrin, an iron chelator (Rainey et al., 2019;Chang et al., 2020). Recently, a ferroptosis signature has been observed in patients with COVID-19 (Jacobs et al., 2020). Ferroptosis inhibitors, such as ferrostatin-1, have been applied as potential drug candidates for COVID-19. Notably, NCOA4 has been reported to be the cargo receptor of ferritin in ferritinophagy (Dowdle et al., 2014;Mancias et al., 2014). NCOA4 binds to ferritin and delivers it to autophagosomes for ferritin degradation and iron release. ROS, RNS, RSS, and ferroptosis are generated by excess iron, consequently resulting in multiple tissue injury (Supplementary Figure S1). Additionally, Joseph et al. demonstrated that a direct association between a key surface arginine in ferritin and a C-terminal element in NCOA4 is required for the delivery of ferritin to the lysosome via autophagosomes (Mancias et al., 2015). Notably, NCOA4 depletion inhibits the delivery of ferritin to the lysosome, resulting in the disruption of ferritin degradation. The degradation of NCOA4 by HERC2, an E3 ubiquitin ligase, leads to suppressed ferritinophagy and elevated levels of ferritin (Mancias et al., 2015).
Therefore, we hypothesize that cytokine storms caused by SARS-COV-2 infection may promote hyper-ferritinemia which can further intensify the inflammation. Elevated ferritin levels can trigger NCOA4-mediated ferritinophagy and may lead to ferroptosis, cell death, and organ damage. NCOA4-mediated ferritinophagy can be targeted to limit the ferroptosis and, therefore, prevent the multi-organ damage and severity in COVID-19 patients (Figure 1). Further studies should be conducted to confirm the involvement of NCOA4 and ferritinophagy in SARS-CoV-2 infection.
AUTHOR CONTRIBUTIONS
The first draft of the manuscript was written by FJ. SK and HL commented on previous versions of the manuscript.
FUNDING
This work was supported by grants from the Shandong Provincial Natural Science Foundation (ZR2020QC088). | 2,324 | 2021-12-15T00:00:00.000 | [
"Medicine",
"Biology"
] |
Carrier Dynamics of Efficient Triplet Harvesting in AgBiS2/Pentacene Singlet Fission Solar Cells
Abstract Singlet fission is a process by which an organic semiconductor is able to generate two triplet excitons from a single photon. If charges from the triplets can be successfully harvested without heavy losses in energy, then this process can enable a single‐junction solar cell to surpass the Shockley–Queisser limit. While singlet fission processes are commonly observed in several materials, harvesting the resulting triplets is difficult and has been demonstrated with only a few transport materials. Here, transient absorption spectroscopy is used to investigate singlet fission and carrier transfer processes at the AgBiS2/pentacene (AgBiS2/Pc) heterojunction. The successful transfer of triplets from pentacene to AgBiS2 and the transfer of holes from AgBiS2 to pentacene is observed. Further singlet fission in pentacene by modifying the crystallinity of the pentacene layer and have fabricated the first singlet fission AgBiS2/Pc solar cell is enhanced. Singlet fission devices exhibit higher external quantum efficiency compared with the control devices, and thus demonstrating the significant contribution of charges from the singlet fission process.
Introduction
Singlet fission (SF) is a spin-allowed process in which internal conversion occurs from a high-energy excited singlet state to two low-energy triplet states. [1] By reducing thermalization losses, [2] the SF process has been predicted to boost the solar cell DOI: 10.1002/advs.202300177 conversion efficiency, theoretically enabling single junction cells to overcome the ey-Queisser efficiency limit. [3][4] Thus photovoltaic devices using SF and the charge generation and transport mechanisms in these devices have recently attracted significant attention. [5][6] Tayebjee et al. calculated the theoretical limit of the efficiency of SF devices to be 45.9% at 300 K and under AM1.5G illumination, almost 50% higher than the maximum efficiency of a normal single junction solar cell (33.7%). [7] Among several molecules that exhibit SF, the organic semiconductor pentacene (Pc) is one of the most popular candidates for SF solar cells, due to the rapid (< 100 fs), highyield triplet generation observed in Pc under solar-spectrum illumination. [8][9] However, simply using SF chromophores as the solar cell active layer is insufficient to achieve solar cell improvement since the doubled photocurrent would also bring a halved voltage of the cell and thus no benefit to the peak power. [3] Instead a suitable triplet acceptor must be chosen that sets the active layer bandgap, harvests high energy photons as SF triplets while also directly converting lower energy (but still above E g ) photons to normal photocurrent.
In recent years, several attempts have been made to use SF in optoelectronic applications. Ehrler et al. first reported an SF solar cell in 2012 with Pc as the SF chromophore and PbSe nanocrystals (NCs) as the triplet acceptor. [5] Macqueen et al. fabricated a tetracene/Si solar cell in 2018, but the endothermic SF process slowed triplet generation, and the low quantity of triplets reduced the benefit of using the SF mechanism. [10] Einzinger et al. utilized electric-field-effect passivation to achieve an efficient energy transfer of triplets generated in tetracene, further revealed the potential of SF to increase the efficiency of Si solar cells. [11] Lu et al. reported SF sensitized by CsPbBr 3 NCs, where triplets were generated efficiently but energy transfer between the CsPbBr 3 NCs and TIPS-pentacene exhibited low efficiency, which was attributed to poor wave function overlap between the two materials. [12] Guo et al. reported high-speed triplet electron transfer (<1.5 ps) in a TIPS-pentacene/MAPbI 3 heterojunction film, but with much slower hole transport (≈13.8 ns). [13] Several factors still hinder device performance including: the relative absorption and luminescence spectra of the materials, the energy difference between E S1 and 2E T1 , the overlap between the wave functions of the triplet donor and acceptor materials and the stability of chromophore and acceptor material, in addition to other normal device considerations. [3] Although Pc exhibits efficient SF under solar illumination, the low E T1 (0.86 eV) makes it incompatible with common semiconductors, most notably silicon and lead halide perovskites. One unexplored option, AgBiS 2 nanocrystals are actually promising candidates for this role, due to their matching band energies. [14][15][16][17] Besides maintaining good stability, solar cells produced this non-toxic material also exhibit strong light-harvesting ability and relatively high power conversion efficiency (PCE). [14,18] In this work, we investigate carrier transfer between Pc and AgBiS 2 nanocrystals and fabricate devices to demonstrate SF and triplet transfer in these materials. Using transient absorption spectroscopy (TAS) measurements we observe that triplets generated in Pc are efficiently transferred to the AgBiS 2 NC layer. Annealing the Pc layer further promotes the SF process and provides more photocurrent due to an increase in the diffusion length of the triplets. AgBiS 2 /Pc SF solar cells were fabricated with internal quantum efficiency (IQE) of nearly 100%, and up to 160 ± 25% in the Pc layer alone, as determined by transfer matrix optical modelling. [19] These results demonstrate the possibility of boosting device performance by employing SF chromophores in AgBiS 2 /Pc solar cells, and provides new ideas for fabricating high-efficiency nanocrystal solar cells.
Results and Discussion
During photoexcitation, photoabsorption by Pc causes the generation of singlet excitons Equation (1) describes the simplified SF process (Figure 1a) and its reverse process, triplet-triplet annihilation (TTA), where S 0 is the ground state of the chromophore, S 1 and T 1 are the lowest-energy excited singlet and triplet states, respectively, 1 (TT) represents a coupled triplet pair, and k represents the rate constant for each step. [20] AgBiS 2 NCs with an average diameter of ≈5 nm were synthesized by adapting a previously reported hot-injection method ( Figure S1, Supporting Information). [14] AgBiS 2 NCs thin films were formed by spin-coating the NC solution onto the substrate. Pc films were formed by thermal evaporation (see the Supporting Information for details). Figure 1b shows the UV-Vis-NIR absorption spectra of those samples. As a narrow-band semiconductor with an indirect band gap at ≈1.1 eV, AgBiS 2 NCs have a wide absorption feature covering the visible to NIR range (400-1100 nm). In contrast, Pc only shows strong absorption at wavelengths below 700 nm, corresponding to its wider effective band gap at ≈1.8 eV. The composite sample exhibits the characteristics of the two materials, with a wide absorption range and small absorption peaks visible at 584 and 666 nm. Since the photoexcitation from the singlet ground state to a triplet state is spin-forbidden, no corresponding absorption peak can be collected in the UV-Vis-NIR absorption spectra.
According to UV-photoelectron spectroscopy (UPS), the highest occupied molecular orbital (HOMO) energies of AgBiS 2 NCs and Pc are −5.2 and −4.9 eV, respectively ( Figure S2 transferred into the Pc, and the spin-forbidden triplet recombination promotes retention of the separated carriers. The increase in excited electrons in the NC film, up to 200% of the photons absorbed by the Pc layer, permits the total IQE to rise above 100%. TAS was performed to observe charge dynamics in Pc and NC bi-layer films. Figure 2 shows the evolution of the transient absorption (TA) signal for AgBiS 2 , Pc, and AgBiS 2 /Pc samples. A 650 nm pump laser, which can excite both the AgBiS 2 and Pc layers, was employed for all samples with 1 kHz repetition rate to avoid heat accumulation. A broadband ground state bleaching (GSB) signal centered around 1050 nm was observed in the AgBiS 2 sample (Figure 2d), corresponding to the narrow indirect band gap observed in the UV-vis-NIR absorption spectrum (Figure 1b). Meanwhile an apparent photo-induced absorption (PIA) signal can be observed via a longer probe wavelength (>1200 nm). The Pc sample presents two sharp GSB features at 584 and 666 nm, corresponding to-the S 0 → S 1 and S 0 → S 2 electronic transition, respectively (Figure 2e). [21] Two PIA signals at around 620 and 700-1100 nm can be explained by the existence of singlets and triplets in the Pc, and the singlet PIA features shows some overlap in wavelength with the GSB signal. Singlets with high energy and short lifetimes result in a quick decay of the PIA signal at 600 nm, [22] while the low-energy triplets with long lifetime dominate the slow decay of the PIA signal at 700-1100 nm. Their larger overall absorbance of the probe results in a stronger PIA signal than the singlet PIA.
For the AgBiS 2 /Pc composite sample (Figure 2f), both the GSB and the PIA signals of Pc are clearly exhibited. In the NIR region (800-1200 nm), the AgBiS 2 GSB signal is diminished, this can be attributed to the carrier transfer process that the holes gen-erated in AgBiS 2 transfer into the valence band of Pc. The increased electron density in the valence band of AgBiS 2 brings an increased absorption of the probe laser, leading to a suppressed GSB signal. However, at long time delay, the triplets generated in Pc will transfer to the conduction band of AgBiS 2 , leading to an enhanced GSB signal around the band gap. Our results match previous observations of the SF process in Pc and confirm the carrier transfer to AgBiS 2 as shown in Figure 3.
When E S1 > 2E T1 , as in Pc, SF is both energetically and entropically favorable, and this results in a rapid triplet generation process ultrafast timescale, outcompeting the radiative decay of the singlet. [3,23] Figure 3a presents the PIA features of Pc due to singlets and triplets, respectively. Because the singlet with a short lifetime exhibits the PIA feature, and this shows obvious spectral overlap with the GSB of Pc, the singlet PIA signal is collected at the node of the spectra at 1 ps to remove the contribution of the GSB signal. Note that the singlet PIA and the triplet PIA features demonstrate a pair of decreasing and increasing functions, respectively. We use a pair of single exponential functions to fit the decay and growth processes of the two curves, and get the time constants for these two features (Equations (2) and (3)) where A is the amplitude, is the rise or decay time constant, t 0 is the starting time, y 0 is a constant background signal (see Table S1 for fitting parameters, Supporting Information). The existence of SF in Pc can be confirmed by a pair of similarly valued singlet and www.advancedsciencenews.com www.advancedscience.com These two time constants are equivalent within uncertainty while the minor difference between singlet and triplet could be attributed to the TTA process and the recombination of few singlets and holes. Both signals stabilized after 1 ps, demonstrating that the SF here is an ultrafast process with a much smaller timescale than conventional carrier transfer, and triplets in Pc are generated within hundreds of femtoseconds and long-lived thereafter. The kinetics of the singlet PIA peak at 620 nm is also notable ( Figure S3, Supporting Information). The ultrafast signal generated in the Pc sample exhibits a sharp decay with a short time delay, and the curve flattens at around 25% of the peak intensity. This strongly suggest that most of the singlets (≈75%) go on to generate triplets in Pc, and the high yield of triplet charges then dominates carrier transfer processes observed in the SF solar cells.
To study the carrier transfer between the two materials, we observed the dynamics of the triplets and the holes generated in Pc and AgBiS 2 , respectively. In Figure 3b, when probed at 1063 nm, triplets in Pc exhibit a long-lived PIA signal. We note that at early time delay (< 1 ns), the AgBiS 2 sample exhibits a strong GSB signal, corresponding to its small bandgap, while the AgBiS 2 /Pc composite sample exhibits a much weaker GSB signal. In contrast, at long time delay (> 1 ns), the GSB signal of the AgBiS 2 /Pc sample is stronger than that of the AgBiS 2 sample. This can be explained by carrier transfer from Pc to AgBiS 2 . The holes generated in the valence band of AgBiS 2 can transfer into Pc, and the decreased hole density in AgBiS 2 enhances the absorption of the probe laser, leading to a suppressed GSB signal. Meanwhile, the triplets generated in Pc enter the conduction band of the AgBiS 2 via Dexter transfer, and the filled orbitals suppress the photoexcitation, resulting in a stronger GSB signal. [12,24] Triplets will not undergo spin inversion during Dexter transfer, so the GSB signal of the composite sample remains at a stable intensity after long time delay. Overall, the more intense and longer-lived GSB signal of the composite sample demonstrates efficient transfer of triplets from Pc to AgBiS 2 . This process can also be confirmed through further observations, as shown in Figure S4 (Supporting Information). There, when probing at 890 nm, the pure Pc sample will display a strong, long-lived triplet PIA signal. However, in the composite sample, the PIA signal quickly decreases to zero, due to the efficient triplet transfer from Pc to AgBiS 2 . We can also observe enhancement of the AgBiS 2 GSB signal in the composite sample. For the pure AgBiS 2 sample, the GSB signal at 1030 nm will disappear within 5 ns, while in the composite sample, the long-lived triplets from Pc will occupy the conduction band of the AgBiS 2 NCs, which extending the lifetime and intensity of the GSB. Note that the nanocrystal GSB signal in the composite sample at 1030 nm undergoes a sharp decrease at ≈2 ps, which matches the expected timing of the triplet transfer from Pc to AgBiS 2 NCs.
The separation of electrons and holes is a key factor in the performance of photovoltaic devices. A 750 nm pump laser was used to observe hole transfer from AgBiS 2 to Pc ( Figure S5, Supporting Information). The extremely low absorbance of Pc at this wavelength demonstrates that the longer-wavelength pump laser can only excite AgBiS 2 , not Pc, as is consistent with the negligible TA intensity of Pc within the entire detection region. Figure 3c presents the TA spectra with a 690 nm probe, corresponding to the band gap of Pc. Note that the pure AgBiS 2 sample exhibits a positive PIA feature which is attributed to broadening of the exciton spectrum, [25] while the composite sample shows an apparently negative signal at this wavelength, proving the interaction between two layers. We propose that the holes generated in AgBiS 2 tend to transfer into the valence band of Pc, and the increased hole density in Pc leads to an apparent GSB signal. The long-lived negative signal indicates that holes can be stored in the Pc, confirming their efficient separation from the electrons. Thus this bi-layer system can be used in photovoltaic devices for efficiently capturing charges from multiple excitons per photon.
Pc films obtained by thermal evaporation are often accompanied by many grain boundaries. It has been reported that the presence of even a small fraction of the amorphous phase (< 10%) will greatly decrease the triplet diffusion length, for example, from 75 to 14 nm in TIPS-pentacene. [26,27] Here we improved the crystallinity of the Pc film by an annealing treatment, reducing the grain boundaries and thereby the trap density. Figure 4a presents the SEM images of three Pc films with different annealing temperatures (room temperature Pc-RT, 60°C Pc-60, 80°C Pc-80), and the grain size is observed to increase with annealing temperature. AFM images in Figure 4b reveal the same trend. XRD patterns of different Pc films are shown in Figure 4c. All the samples exhibit characteristic XRD peaks corresponding to (001), (002), and (003), and the Pc-80 sample shows the best crystallinity, while annealing above 100°C produces irregular films with small crystallites ( Figure S6, Supporting Information).
The carrier dynamics of the annealed sample Pc-80 and AgBiS 2 /Pc-80 were also investigated by TAS ( Figure S7, Supporting Information), both SF and carrier transfer were observed. Figure 4d presents the PIA features of singlets and triplets in Pc-80, and the two curves exhibit singlet and triplet as 88.1 ± 23.4 and 118.9 ± 36.0 fs, respectively (Table S1, Supporting Information). The smaller compared with those of Pc-RT indicates that the better crystallinity of Pc-80 facilitates the SF process. At longer time delay, the triplets in the annealed samples also exhibit longer lifetimes than those in the samples prepared at room temperature ( Figure S8, Supporting Information). Due to the increased density of triplets generated in Pc-80, carrier transfer between AgBiS 2 and annealed Pc-80 is also enhanced. Figure 4e displays the comparison between the carrier dynamics with 650 nm pump and 970 nm probe for all samples. Both the annealed samples (Pc-80 and AgBiS 2 /Pc-80) exhibit longer triplet lifetimes.
Although the better crystallinity of Pc should bring a faster carrier transfer, we note that the annealed samples unusually show a slower decay at short time delay. This can be explained by the higher density and the longer diffusion length of the triplets in the annealed sample. The increased density of triplets close to the interface can then transfer to AgBiS 2 , leading to a long-lived carrier transfer process, and the carriers generated deeper inside Pc-80 can slowly diffuse to the interface and transfer to AgBiS 2 , further extending the time for PIA signal decay. Overall, the an-nealing process at 80°C effectively improves the crystallinity of the Pc film, enhancing SF. The longer diffusion lengths create a long-lived carrier transfer process between the Pc and the AgBiS 2 nanocrystals. Thus, triplets and holes are efficiently separated into the AgBiS 2 and Pc thin films, respectively, indicating the possibility of fabricating a high-performance SF solar cell.
The AgBiS 2 /Pc heterojunction was then incorporated into photovoltaic devices producing what is, to the best of our knowledge, the first SF enhanced AgBiS 2 solar cells to be reported. The device architecture employed here is an n-i-p device type ITO/ZnO NCs/AgBiS 2 /Pc/MoO x /Ag. The band energy diagram indicates that Pc can be both applied as a hole-transporting material (HTM) and the SF material (Figure 5a). Since the current contribution from SF depends on active layer thickness, [28] the active layer thickness was varied in the devices by depositing one layer (1L) or two layers (2L) of AgBiS 2 . PTB7 is used as the optimized HTM for the control devices based on previously report devices. [29][30] The J-V curve shows that AgBiS 2 /Pc could generates higher short-circuit current (J SC ) in both the 1L and 2L devices, demonstrating the current contribution from Pc (Figure 5b). The SF devices generate a lower open-circuit voltage (V OC ) compared with the control devices, which further results in a decreased PCE. The decrease in V OC is attributed to increased nonradiative recombination at the NC/Pc interface for charges originating from the NC layers. [31] In Figure 5c, the SF devices have significantly higher EQE in the wavelength range 600-700 nm, corresponding to the Pc absorption peak. SF devices using Pc-80 slightly improved the J SC ( Figure S9, Supporting Information). Transfer matrix optical modeling was used to calculate the absorbance fraction in the devices ( Figure S10, Supporting Information), which permits calculation of the IQE value for each device. [19] For control devices, the IQE was calculated by dividing the EQE by the fraction of photons absorbed by AgBiS 2 NCs. For SF devices, the overall IQE was calculated by Equation (4) IQE AgBiS 2 ∕PC = EQE AgBiS 2 ∕PC Abs AgBiS 2 + Abs PC (4) where Abs AgBiS2 and Abs Pc are the absorbed fraction of AgBiS 2 NCs and Pc in the SF device, respectively. Figure 5d shows that Pc has a positive overall IQE contribution for both the 1L and 2L SF devices. The IQE contribution from the Pc component is obtained by isolating the AgBiS 2 EQE contribution in the system ( Figure S11, Supporting Information). In the Pc-1L and Pc-2L samples, the IQE within the pentacene layer reaches a peak of 160 ± 25% and 130 ± 20%, respectively, at 610 nm (Figure 5e), which demonstrates an efficient SF process is happening in both devices. For photons at 500 and 600 nm, the peak electric field intensity is located in the AgBiS 2 layer in both the control and SF devices (Figure 5f), indicating an efficient light harvesting in this structure. At 700 nm, the most intense electric field is inside the ZnO layer, which may be the reason for weak exciton generation observed in the EQE spectrum at those wavelengths. Replacement of PTB7 by Pc also enables significantly more efficient light harvesting in the thin layer AgBiS 2 (≈17 nm). Thicker NC films in this n-i-p architecture obscure the SF process and thus further engineering is needed to realize the full potential of SF harvesting in these devices (Supporting Information).
Conclusion
In conclusion, the carrier transport process in the AgBiS 2 /Pc heterojunction is determined by the band structure, as revealed by transient absorption spectroscopy. SF in pentacene was confirmed to generate triplets on femtosecond scale (≈100 fs), and the transfer of triplet electrons and holes in the Pc/AgBiS 2 heterojunction was observed at various specific pump and probe wavelengths, demonstrating efficient charge separation to generate photocurrent. By increasing the annealing temperature of the Pc layer, we found that Pc thin films with higher crystallinity exhibited stronger SF and longer-lived carrier transport, which was attributed to the increased stability and diffusion length. Finally, several SF solar cells were fabricated, in which the Pc layer exhibited high IQE, up to 160 ± 25% in the Pc layer, as modelled from the experimental EQE. Thus the SF process demonstrated significant contribution to the EQE of devices with thin films of AgBiS 2 nanocrystals. Further engineering to increase the AgBiS 2 thickness in n-i-p and p-i-n architectures, without sacrificing SF absorbance and carrier mobility, is needed to realize the full potential of SF in AgBiS 2 solar cells and increase the PCE. However, we believe that these results suggest a reasonable pathway to eventually surpass the Shockley-Queisser limit and incorporate the SF mechanism into solar cell using nontoxic nanocrystalline materials.
TEM image of AgBiS 2 NCs; UPS of AgBiS 2 and pentacene; TA maps of samples prepared at room temperature with 750 nm pump laser. SEM and AFM images of pentacene film annealed at 100°C; TA maps of sample annealed at 80°C with 650 nm pump www.advancedsciencenews.com www.advancedscience.com laser. The evolution of TA spectra for all sample with 650 nm pump laser; Device data of AgBiS 2 /Pc with different annealing temperature; IQE modeling details; Device data of p-i-n AgBiS 2 SF solar cells; Fitting parameters for TA spectroscopy.
Supporting Information
Supporting Information is available from the Wiley Online Library or from the author. | 5,304.6 | 2023-03-20T00:00:00.000 | [
"Physics",
"Chemistry"
] |
Islamic Theological Views on the Teachers Position and Duties
Some people always underestimate teachers. A person as teachers do not preserve their personality; whereas, from the side of tauhid, teaching profession is a very noble job and should be given a high honor. Why? Because the tasks undertaken by the teachers have the relationship with the existence of what has been done by God, Allah, to all His creators, as teaching, educating, leading, guiding, and preserving. On the other hand, a teacher must have a personality with divine attributes (rabbaniyah), such as love and gentle, honest, virtuous, patient, has a broad and deep knowledge, clean or avoid sin, humble and forgiving, and others. Key Word: Teacher, Theology, Islam.
INTRODUCTION
Education seems to be like an important human"s life demand here in this globalization, advanced knowledge and technology era (IPTEK).
Accordingly, human"s cognitive, affective and psychomotor are built and developed through education. Thus, they would be creative and innovative to live, they are able to face life challenges and obstacles, and they are able to take a part in society"s progress (Sudarwan Danim,1995:95-97).
There are elements for education achievement and process. Here it is a teacher or an educationist. In this respect, a teacher or an educationist takes main position in the process of education due to his or her function to support students" cognitive, affective and psychomotor competence. He or she is in charge of students" physics, intellectuals, spirituals and attitude (Samsul Nizar,2002:41).
The above explanation says that a teacher or an educationist is a respectful and honored one. M. Athiyah alAbrasyi once ever quotes al Ghazali"s notion and says that a teacher or an educationist seems like sun that shines its self and other creations. He or she seems like Kasturi oil that makes its self and others fragrant. By the fact, a teacher or an educationistrules an honoredthe important thing. (M. Athiyah al-Abrasyi,1993:135-136).
In contrary, some people do not put a respect on him or her. They say that entrepreneurs, politicians, doctors and judges are better and more respectful than a teacher or an educationist. The worst is another people might still insult and dishonored a teacher or an educationist. Additionally, some people might honor a teacher or an educationist temporarily and pragmatically.
Unfortunately, there are still bad teachers and educationists. They do not honor themselves. Accordingly, they don"t keep their fames; misbehaved, and does the abuse of power for personal interest.(Moh. Uzer Usman,2002:1).
In conclusion, it is more interesting to study about a teacher or an educationist"s existence Islamic theology based. Accordingly, the study would discuss about what a teacher or an educationist is, what a teacher or an educationist rules. The last, a teacher or an educationist"s position, his or her characters Islamic theology based.
What a teacher or an educationist is
National education system says that a teacher is an educationist that rules a teaching process. Here, primary and elementary levels of study take a teacher, where as institute, university and colleges take lecturer. Talking about the definition of teacher; it has been written in section 1, subsection 1 number 1in 2005about the rules for teachers and lecturers (UU Guru dan Dosen), as a professional educator with the main duties to educate, to teach, to guide, to train, to assess, and to evaluate the students from young learners in the formal institutions, in elementary and secondary schools. Accordingly, a teacher is an educationist, and educationist is someone who educates (WJS. Poerwadarminta,1991:250 Ahmad Tafsir says that an educationist is someone who is in charge of students" entire progress; as affective, cognitive and psychomotor progress(Ahmad Tafsir,1992:74). Moreover, Hadari once ever says that a teacher is someone who handles education and teaching. Here, he or she is responsible for students" growth being adults, not just as sitting and standing infront of classroom to give the lesson to their students, but also as a part of society. He or she should be active, creative and free minded to guide students to be a part of adult society too (Hadari Nawawi,1989:123).
In conclusion, a teacher or an educationist is someone who teaches, educates, and guidesstudents" affective, cognitive, and psychomotor progress.Islamic study says that an educationist is Mu'allim and Muaddib; means a teacher (Abuddin Nata:2001:61). Besides, there is a term Murabbi means to guide and keep(Omar Muhammad al-Thaoumy al-Syaibany,1979:41). Moreover, those of all take relationships with God almighty.
Allah the almighty says "dandiamengajarkankepadaAdamnama-nama (benda-benda) seluruhnya" (Al-baqarah:31). It meansAllahswt says that he teaches Adamthe names of objects.""within the verse, "Allama is in fi'ilMadhi (past form) means taught and Allah almightywho taught. Additionally, AllahAlmightysays that he taught prophet(Yusuf 6, 21, 37).Thus, here, AllahAlmightyis a teacher or an educationist. Finally, in the above explanation stated that a teacher or an educationist seems to be like a person who handles teaching, guidance, reparation, and looking after students" affective, cognitive, and psychomotor progress well.Otherwise, a teacher based on Islamic theology in Alquran means:
Muaddib derives from
1. Allah Almighty (see alquran, al 'alaq, alqalam, almuzammil, almudatsir, allahab, attakwiir and al'alam Thus, the highest educationist of universe allowed Prophet Muhammad to teach and educate humankind. The almighty allowed parents to teach and educate their family, and then allowed scholars and scientists to teach and educate public.
Teacher's existence and responsibilities
Here, teacher"s responsibilities are teaching, guiding, and repairing students" affective, cognitive, and psychomotor progress well. In his respect, a teacher is someonebehind brightfuture of society and nation.The society and nationleadermight be appeared from a good teacher or educationist. It might take goodness for their life future. In contrary, it would take badness for their life future. AbuddinNatasays that there are reasons to honor a teacher as follows: 1. A teacher is a knowledge transferor due to knowledge is a human"s degree support. 2. A teacher is a society and leader"s behavior supporter, due to it is to be a main pillar to support society"s future. 3. A teacher is a leader for society"s life, due to it is to guide public to worship god and prophet well and do the goodness for society themselves, parents and others. Nata:2001:69-70).
Both teacher and prophet"s existence and responsibilities are identical. Prophet leaded human to know their god, taught them aqidah and ibadah (well behaved and worship) guided them into good life of here after.(Nasruddin Razak,1990:142). Islamic theology says that both prophet and educationist are the same in responsibilities. AbuddinNataquoted from AsmaHasaFahmisays: "stand up for and honor a teacher due to him or herself is like a prophet" (Abuddin Nata:2001:68
هستقين صرط
Means, we (Allah almighty) has revealed the real verses, and Allah almighty leads who he likes to have guidance.
Teacher's characters
A teacher should have a high competence and professionalism. Moreover, he or she has to have mahmudah (honored) characteristics beside teaching, guiding, and leading students" progress. The last, a teacher should have god almighty"s characteristics (rabbaniyah).In this respect, Abdurrahman An-Nahlawy states that rabbaniyah (god almighty"s characteristics).Characteristic is the main characteristic that every teacher should have. It means that teacher"s activities like intention, speech, gesture, and study movement should be based on Islamic values.(Abdurrahman An-Nahlawy,1995:171).
Abdul Mujibquotes al Razz"s notion and says that robbanior ilaahi (god almighty"s characteristics) characteristics are personal characteristics taken after the transformation of god almighty"s names and characteristics into someone self and then practice it in real life. In simple words, robbani(god almighty"s characteristics) personality is someone"s personality thatreflctsrabbaniyah (god almighty"s characteristics). (Abdul Mujib,2006:188-189).
Allah almighty says that he has got asmaaulhusna (beautiful names) alhasyr21. And then syekhMaqdisi and allamahkaf "ani says that there are95 god"s names.18 (al asmaulhusnaalmustakhraja min al Qur'anal 'azhimwa al Sunnahwaatthibbi an nabawi).God"s names are his characteristics. Theologically, a part of the should bea part of teacher"s characteristics.
Asma HasanFahmi and Abdurrahman
An-Nahlawyquote imam al ghazali"s notion and say that there is no refund to hope being a teacher. (al-Ghazali,2000:55-58, Asma Hasan Fahmi,1979.It deals with what Allahalmighty does for humankind, that is he never asks humankind to pay back what they get from Allahalmighty.Accordingly, Allahalmighty gives humankind commands to do the worships for themselves only. What they gets based on what they do to get, and all of them are done to know to thank Allah almighty who is the only one who gives.Here, a teacher transfers students knowledge, and then he or she deserved to have students" actions based on what they have got from a teacher. That is a thank to a teacher being students. Thus, all of these are dedicated to students "own bright future life.In fact, it is not an absolute thing that a teacher takes no payment for teaching, due to: 1. A teacher like many there human. 2. He or she lives like many other human live. 3. They all do the same thing for life.
The point is, a payment after teaching is something but it is not everything. Even though the reality says that the payment is not really satisfied and sufficient but a teacher still does his or her existence and responsibilities well.
Theologically, a teacher shouldn"t make teaching payment a main point of life, because it will decrease a teacher"s desires to do the responsibilities well. Otherwise, if the payment is good enough then it means that it is something that a teacher deserves to have.
Finally, a teacher"s responsibilities are like Allahalmighty"s responsibilities. To teach and educate are god almighty"s instructions. Thus, a teacher should ask Allahalmighty a lot for payment better than many other human.
Sincerely, a teacher"s knowledge comes from Allah, Allahalmighty who owns that only. A teacher"s knowledge is temporarilyowned. Somehow like a life, a knowledge is some daytaken back again to Allahalmighty.
The last but not least, a teacher should be responsible for his or her duties to guide students" affective, cognitive, and psychomotor progress. Besides, they have to be robbaniyah; means they get god almighty"s characteristics like educationist.
CONCLUSION
Here, the writer knows that being a teacher and an educationist is honored and respectful. Theologically, a teacher or an educationist should have god almighty"s characteristics (rabbaniyah),ikhlas (sincere), and reliable and responsible. Therefore, having god al might"s characteristics is a necessary one to have being a teacher. Today we might not see dreamt teacher.
If there is a research of theology based towards whether there is or not a dreamt teacher? It is a teacher that has rabbaniyahcharacteristic. The answer is yes, there is still, but it is not huge in number. By the reasons that teacher"s selection tends to ask only for self-knowledge mastery which is proved through certificate. No personal evaluation for god almighty"s characteristic.
All of them might make a teacher or an educationist wise less and irresponsible for duties. Finally, national education relays on a teacher much. A good teacher with rabbaniyah(God almighty"s characteristics) characteristic would make national education progress. It is absolutely obstacles for those who want national education getting better. | 2,520 | 2015-08-02T00:00:00.000 | [
"Philosophy",
"Education"
] |
Gene Expression Profile Induced by Two Different Variants of Street Rabies Virus in Mice
Pathogenicity and pathology of rabies virus (RABV) varies according to the variant, but the mechanisms are not completely known. In this study, gene expression profile in brains of mice experimentally infected with RABV isolated from a human case of dog rabies (V2) or vampire bat-acquired rabies (V3) were analyzed. In total, 138 array probes associated with 120 genes were expressed differentially between mice inoculated with V2 and sham-inoculated control mice at day 10 post-inoculation. A single probe corresponding to an unannotated gene was identified in V3 versus control mice. Gene ontology (GO) analysis revealed that all of the genes upregulated in mice inoculated with V2 RABV were involved in the biological process of immune defense against pathogens. Although both variants are considered pathogenic, inoculation by the same conditions generated different gene expression results, which is likely due to differences in pathogenesis between the dog and bat RABV variants. This study demonstrated the global gene expression in experimental infection due to V3 wild-type RABV, from the vampire bat Desmodus rotundus, an important source of infection for humans, domestic animals and wildlife in Latin America.
Introduction
Rabies is a zoonotic, highly lethal and neglected disease that has been affecting humanity for more than 4000 years. Lyssaviruses, such as rabies virus (RABV) enter the body through wounds or by direct contact with mucosal surfaces. The number of human deaths globally due to dog-mediated rabies is estimated to be in excess of 59,000 cases annually with most deaths (around 98%) occurring in Asia and Africa [1]. In contrast, in some countries of the Americas where canine rabies has been eliminated, bats are responsible for the majority of human cases [2,3].
Invasion of the central nervous system (CNS) by RABV occurs by binding to various neuronal receptors, including acetylcholine, the neuronal cell adhesion molecule, or the neuronal growth factor receptor [4]. Viral replication occurs in the neuronal cell body, reached by retrograde axonal transport. Some bat RABV variants may propagate via sensory nerves due to skin tropism [5]. Although clinical signs of rabies are severe, there is minimal inflammation and neuronal destruction compared to other encephalitic viruses [4,6,7]. Variants of RABV differ in the pathogenicity [4,7], but the mechanisms are not completely known. In a detailed analysis of clinical features in published human cases of dog-(V2) or bat-acquired (V3) rabies, Udow et al. (2013) reported that encephalopathy, hydrophobia and aerophobia were more common in dog-acquired rabies, whereas Viruses 2022, 14, 692 2 of 11 myoclonus, cranial nerve abnormalities and motor and sensory abnormalities were more common in bat-acquired rabies [8].
In a previous study [9], we demonstrated that cytokine and chemokine genes associated with the immune response, measured by reverse transcription and quantitative PCR, were differentially expressed in the CNS of mice inoculated with canine or bat RABV variants. Here, we extend these studies to a microarray analysis of brain tissue from mice inoculated with the same variants at 5-and 10-days post inoculation (d.p.i.), to evaluate the gene expression induced by these viruses.
Virus Strains
Two wild-type RABV variants, isolated from human patients infected by a dog variant (V2) or a vampire bat variant (V3) were used to inoculate mice. Both viruses were obtained at the fourth intracerebral passage in mice.
Experimental Design
Specific pathogen free female C57/BL6 mice, 4 to 6 weeks old, were provided by Cemib [University of Campinas (UNICAMP) animal facility]. The animals were kept in ventilated cabinets with HEPA (High Efficiency Particulate Air) filters and fed irradiated food and sterile water "ad libitum". In this experiment a total of 64 animals were divided into eight groups: three groups for V2, three groups for V3, and two groups left noninfected. Animals were inoculated, intramuscularly (i.m.) in the right hind limb with 100 µL of viral inoculum with a dose of 40 LD 50 (50% lethal dose). For both viruses, the LD 50 was determined by intracranial inoculation of mice. Control groups (n = 16), named as non-inoculated, received only viral diluent via i.m., and inoculated groups (n = 48) were inoculated i.m. with viral inoculum, V2 or V3. For each viral variant, one of these groups was maintained over a 30-day period, named as the 30-day period evaluation group, and mice in the other two groups were euthanized at either 5 or 10 d.p.i. when whole brain tissue was collected. Non-inoculated animals were killed as well, at 5 or 10 d.p.i. to serve as controls for the microarray comparison (Table 1). Animals in all groups were weighed and evaluated daily for the onset of clinical signs of infection such as ruffled fur, hunched back, hypo-/hyper-excitability, paralysis of one/both hind limbs and tetraplegia. In the 30-day period evaluation group, animals were killed when they reached a semicomatose state, which was determined as the humane endpoint. Animals were killed by isoflurane inhalation.
RNA Extraction and RT-qPCR
Brain tissue RNA was extracted with the Invitek®kit (Berlin, Germany) according to the manufacturer's instructions and stored at −80 • C. Before microarray analysis was performed, the presence of RABV in the brain was confirmed by detection of the RABV N protein gene by RT-qPCR as previously described [9]. Validation of the results obtained in the microarray were made by RT-qPCR using Mouse Quantitect ® Primer Assay (Qiagen ® , Hilden, Germany) for specific cytokines and chemokines as previously described [9].
Microarray
Whole brain was homogenized, and total RNA was isolated from brain samples collected at 5 and 10 d.p.i. from infected animals and controls using Trizol (Invitrogen®, Carlsbad, CA, USA) and purified with the RNeasy Mini-kit and RNase Free DNase Set (Qiagen ® , Hilden, Germany). Total RNA was quantified by absorbance at 260 nm. The quality of all RNA preparations was controlled by electrophoresis on agarose gels, and they were also evaluated on an Agilent Bioanalyser instrument (Agilent Technologies, Palo Alto, CA, USA). Only samples with a preserved rRNA ratio (28S/18S), no evidence of RNA degradation, and RNA integrity values ≥8 at the Bioanalyser were used in the microarray hybridization and qPCR.
For the microarray analysis, 1 µg of total RNA was labeled with the Cy3 and Cy5 fluorophores array kit (Affymetrix, Santa Clara, CA, USA). The GeneChip ® Mouse Gene 2.0 ST Array, a universal arrangement of Affymetrix (Affymetrix, Santa Clara, CA, USA) that interrogates 35,420 genes was used. Hybridization signals were quantified using a ScanArray Express (PerkinElmer, Boston, MA, USA) and the images were processed using GenePix version 4.0 (Axon, Union City, CA, USA). The CEL files were processed and normalized with the RMA program, after this step an ANOVA analysis was performed between the groups using the LIMMA package (version 3.14) Differentially expressed genes were identified in infected mice compared to the control group for each RABV variant (FDR corrected p value < 0.05 log2 difference of expression of >|1|). Annotation of genes enriched in specific functions were determined with reference to the Kyoto Encyclopedia of Genes and Genomes (KEGG) and gene ontology (GO), using "NIPA" software (https://github.com/ADAC-UoN/NIPA (accessed on 20 March 2022); version 0.6.7). Raw data and experimental design are available from ArrayExpress (https://www.ebi.ac.uk/arrayexpress/ (accessed on 20 March 2022) under the accession E-MTAB-11414).
Statistical Analysis
Kaplan-Meier survival curves were plotted and Log-Rank (Mantel-Cox), Grehan-Breslow-Wilcoxon and Mann-Whitney tests were performed in GraphPad Prism version 8.1.2 for Windows (GraphPad, La Jolla, CA, USA).
Assessment of Lethality of RABV Variants V2 and V3
Lethality of the viruses was determined by group monitored over a 30-day period post-infection. The incubation period was 12 days and the morbidity period (the time from onset of clinical signs of infection to death) was 5 days for each variant. There was no significant difference in survival nor in the amount of virus in the CNS comparing V2 and V3. Median survival was 16 days for the V2 group and 17.5 days for the V3 group. However, three mice survived with no clinical signs for 30 days in the V3 group (62.5% lethality), whereas all mice were dead at 18 days in the V2 group (lethality 100%) ( Figure 1). Animal weights during the 30-day observation period are available in Supplementary File S1.
Transcriptome Analysis of Mouse Brain RNA
To compare the CNS host response to the different RABV variants, whole transcriptome analysis was performed on brain tissue collected from individual mice from all groups at 5 and 10 d.p.i. All animals inoculated with V2 or V3 were positive for the presence of RABV in the brain.
Analysis of mice inoculated with V2 or V3 compared to their respective controls at 5 d.p.i. showed no different expressed genes (DEGs). In the V3 group at 10 d.p.i., a single array probe for an unannotated gene was differentially expressed compared with the control mice. In contrast, 138 array probes associated with 120 DEGs were detected between
Transcriptome Analysis of Mouse Brain RNA
To compare the CNS host response to the different RABV variants, whole transcriptome analysis was performed on brain tissue collected from individual mice from all groups at 5 and 10 d.p.i. All animals inoculated with V2 or V3 were positive for the presence of RABV in the brain.
Analysis of mice inoculated with V2 or V3 compared to their respective controls at 5 d.p.i. showed no different expressed genes (DEGs). In the V3 group at 10 d.p.i., a single array probe for an unannotated gene was differentially expressed compared with the control mice. In contrast, 138 array probes associated with 120 DEGs were detected between mice inoculated with V2 and the control mice at the same time point (10 d.p.i.). Normalized expression of DEGs within the groups are presented in Supplementary Data (Files S9 and S10). The gene ontology (GO) analysis revealed that all the up-regulated genes were involved in the biological process of immune defense against pathogens ( Figure 2). mice inoculated with V2 and the control mice at the same time point (10 d.p.i.). Normalized expression of DEGs within the groups are presented in Supplementary Data ( Figures S9 & S10). The gene ontology (GO) analysis revealed that all the up-regulated genes were involved in the biological process of immune defense against pathogens ( Figure 2). The most enriched GO biological process term in the up-regulated genes was GO:0002376 (immune system process) with 25 genes in the differentially expressed gene list. The KEGG pathway analysis highlighted that a large number of genes that were up-regulated after infection with V2 were the same as those upregulated in response to Herpes simplex infection (genes Myd88, Irf7, H2-Q7, Casp8, Tlr3, Daxx, C3, H2-Aa, Oas2, Oas3, H2-T22, H2-T23 and Cd74). No enriched KEGG pathways were observed in the down-regulated genes. However, vomeronasal receptor (Vmnr) genes were significantly enriched in each of the GO biological terms of 'signal transduction', 'G protein-coupled receptor signaling pathway', 'response to stimulus', 'response to pheromone' and 'sensory perception of taste' representing the down-regulated genes. More detail of GO enrichment and KEGG pathway analysis is presented in Supplementary Tables S2-S8.
Microarray Results Validation
cell chemotaxis +ve chemotaxis +ve regulation T cell mediated cytotoxicity chemotaxis LPS-mediated signaling pathway response to lipopolysaccharide response to interferon-gamma defense response to protozoan presentation via MHC class Ib proteolysis in cell. protein catabolic process +ve regulation of interleukin-8 production defense response +ve regulation of neutrophil chemotaxis positive regulation of chemokine production response to virus type I IFN signalling pathway defense response to Gram+ bacterium Ig mediated immune response cellular response to interferon-gamma +ve regulation phagocytosis, engulfment cellular response to lipopolysaccharide presentation via MHC class II cellular response to interferon-beta +ve regulation of type I IFN production inflammatory response +ve regulation TNF production immune response response to bacterium defense response to virus innate immune response immune system process sensory perception of taste G protein-coupled receptor signaling pathway signal transduction response to pheromone response to stimulus Enrichment (-log10 value) The most enriched GO biological process term in the up-regulated genes was GO:0002376 (immune system process) with 25 genes in the differentially expressed gene list. The KEGG pathway analysis highlighted that a large number of genes that were up-regulated after infection with V2 were the same as those upregulated in response to Herpes simplex infection (genes Myd88, Irf7, H2-Q7, Casp8, Tlr3, Daxx, C3, H2-Aa, Oas2, Oas3, H2-T22, H2-T23 and Cd74). No enriched KEGG pathways were observed in the down-regulated genes. However, vomeronasal receptor (Vmnr) genes were significantly enriched in each of the GO biological terms of 'signal transduction', 'G protein-coupled receptor signaling pathway', 'response to stimulus', 'response to pheromone' and 'sensory perception of taste' representing the down-regulated genes. More detail of GO enrichment and KEGG pathway analysis is presented in Supplementary Files S2-S8.
In our work, the majority of the up-regulated 120 DEGs have their functions associated with cell adhesion, receptor signaling, peptide antigen binding, double-stranded RNA binding and pathways involved in innate immune responses, cellular immune responses, and response to viruses and bacteria. This is in agreement with similar studies in which mice were infected with RABV. Ubol et al. (2006) inoculated 1-day-old mice by the intracerebral route with a primary RABV isolate (passaged once in mouse brain) from an infected dog [14]. They observed altered gene expression at days 2 (29 genes), 4 (109 genes) and 6 (98 genes) post-inoculation in eight major groups: immune response, metabolism, receptor and transport, growth factors, death-mediated factors, transcription and translation factors, proteases and kinases. The up-regulation of VCAM-1 in the present study demonstrates the participation of this molecule in viral penetration, in which the activation occurred immediately after entry of the virus in the brain and in the absence of clinical signs, corroborating previous works published by our group and others [9,[15][16][17].
Results also demonstrate toll-like receptor-3 (TLR3) activation associated with upregulation of various immune response genes, innate immune responses, upregulated virus responses indicating chemotaxis, inflammatory and antiviral responses activated by TLR3 to CNS RABV infection [18]. The early activation of the defense response was only observed in V2, similar to the results obtained by Zhang et al. (2016), who observed activation of the early innate and adaptive immune response induced by infection by the Flury-Hep RABV (an attenuated strain) as opposed to CVS -11 (highly pathogenic strain) [17]. In work by Sugiura et al. (2011), animals inoculated with a lethal dose of CVS intramuscularly did not show any change in gene expression before the onset of clinical signs [19]. This demonstrates that street and fixed virus strains of RABV have distinct expression profiles, corroborated by results obtained in the current study. Mere clinical evaluation of mice in experimental rabies is not able to detect subtle alteration in physiological parameters and behavior and more precise evaluation methods are required.
Besides the immune response pathways, there is also activation of genes associated with the surface and components of the cell and cell matrix, outer cytoplasmic membrane, lysosome, phagocytic vesicles and extracellular space, suggesting alteration in CNS structures and integrity leading to dysfunction, which is a major factor associated with RABV-induced death [20,21]. Neuronal dysfunction can be observed in the absence of classical clinical signs of rabies due to the decreased expression of genes associated with transduction, response to stimuli and pheromones, and taste perception, indicating centrifugal dissemination via sensory nerves and the functional alteration of the CNS resulting from viral replication [5].
Another fact was activation of the canonical pathway associated with viral myocarditis in the infection caused by V2 a clinical finding already described in human rabies cases [22,23].
Although there were no obvious clinical signs in the mice at 10 d.p.i., there was already down-regulation of Vmnr genes associated with signal transduction, response to stimulation, response to pheromones and taste perception not described in other studies. This may be due to different ages of mice and routes of inoculation used.
Moreover, previous studies have not directly compared wild RABV isolates of canine or bat origin in mice inoculated by the intramuscular route, better mimicking a natural infection. The microarray analysis also showed Vmnr genes expressed differentially compared to all other genes analyzed. The VMN receptors are located in the vomeronasal organ (VNO), a secondary chemosensory system that is enclosed in a cartilaginous capsule on the medial surface of the nasal septum. The VNO has chemosensory neurons responsible for identifying numerous pheromones that activates a cascade of molecules that leads to the expression of socio-sexual behavior with involvement of the limbic system [24,25]. In rabies, the change in limbic system activity, which involves behavioral and emotional changes, has been recognized for millennia, so much so that the word rabies is rooted in the Hindi rhabaze, which means to become violent, as in animals the limbic system is directly influenced by recognition of pheromones [25]. Although the differential expression of so many Vmnr genes needs further studies to determine its real influence on animal rabies, it is already known from inoculation of mice with herpes simplex virus via the intranasal route that neurotropic viruses may invade CNS via vomeronasal sensory neurons [26]. In a study of mice inoculated by intranasal instillation with CVS, olfactory receptor cells were selectively infected [27]. The down-regulation of Vmnr genes in mice inoculated intramuscularly with V2 suggests that this variant may spread to or from the CNS by the olfactory route. This is further supported by the herpes simplex infection pathway being the most significant KEGG pathway identified in this study, but such results are yet to be confirmed by different assays.
Although both RABV variants used in this study are considered pathogenic and the same LD 50 was used to inoculate mice, different lethality between V2 and V3 as well as different gene profiles were observed. There were some apparent differences between DEGs and expression levels of selected cytokines and chemokines determined by RT-qPCR. This may result from the difference in sensitivity between this technique and from individual variations between genes and animals that are more evident in the real-time PCR analysis [10]. In a study of a pathogenic wildtype-silver-haired bat RABV in mice, Wang et al. (2005) had similar results regarding discrepancies between detection of some genes in microarray compared to RT-qPCR [4]. Very low or very high expression levels as well as differences in the probes used in each technique decrease the concordance between those two assays [10]. Another important point that might influence the magnitude of gene expression is the clinical status of animals when sampled. In the current study, none of the animals presented clinical signs when samples were collected at 5 or 10 d.p.i. The same observation was made in a microarray study published by Sugiura et al. (2011) in which mice inoculated intramuscularly with CVS at day 3 showed no illness nor significant changes in signal of brain or spinal cord [19].
Despite the discrepancies, the RT-qPCR results confirm the microarray analysis showing that V2 induces expression of a larger number of the analyzed genes and at a higher intensity, especially at 10 d.p.i., compared with the V3 RABV variant. Similarly, in the study by Wang et al. (2005), mice inoculated with the same lethal dose of a highly pathogenic wildtype-silver-haired bat RABV or an attenuated CVS strain (lab strain) via the intracranial or intramuscular routes, showed that the genes upregulated by the attenuated strain were higher in both number and intensity when compared to the pathogenic bat RABV [4]. With the progression of the infectious process, at 10 d.p.i., there was an apparent increase in gene expression of all cytokines and chemokines compared to 5 d.p.i. The increase in cytokines and chemokines during RABV infection, has been reported by others [28][29][30][31] and is associated with viral pathogenicity [32], which is also in agreement with our results, as higher gene expression is associated with higher lethality [28,32].
The gene expressions of cytokines and chemokines in the V2 and V3 groups were also qualitatively distinct, especially at 10 d.p.i. In the V2 group at 5 d.p.i, a higher expression of IL12 was detected while in the V3 group there was an increase in IFNβ and CCL2. At 10 d.p.i., in the V2 group there was an increase in OAS1 and IFNβ, while in the V3 group there was no increased cytokine or chemokine expression. In contrast to what was observed for the other cytokines, in the V3 group, IFNβ expression was higher at 5 d.p.i compared to 10 d.p.i. after inoculation, when its levels drastically reduced. This same expression profile was observed by [16], probably due to the evasion mechanism that interferes with the phosphorylation of IRF3 and the translocation of STAT 2 to the nucleus, blocking the production of type 1 IFN [16,33]. The increase in IFNβ occurred late in V2 group but was not strong enough to diminish viral replication. Levels of RABV-N were higher at 10 d.p.i. when compared to 5 d.p.i. (p < 0.001). Higher expression of CCL2 was Viruses 2022, 14, 692 9 of 11 observed, at 5 d.p.i. in the V3 group, highlighting the possibility of greater permeability of blood-brain barrier (BBB) induced by this chemokine compared to V2. This is in agreement with the lower relative lethality observed for this variant as the increase in BBB permeability has been associated with greater penetration of effector cells [32,34,35], contributing to viral elimination [34,36].
In conclusion, our results demonstrate that the microarray study of two wild type RABV, V2 and V3, have a gene expression profile intrinsic to the viral variant. In the V2 group, a significant difference of signal compared to control mice was only seen at 10 d.p.i. The majority of the 120 up-regulated DEGs were associated with multiple functions in agreement to similar published studies. The V3 group showed only one differently expressed gene with non-annotated function at 10 d.p.i.. However, RT-q PCR showed that some important cytokines and chemokines were detected at 5 and 10 d.p.i. in the V3 group. This may partially explain the lower lethality rate in mice inoculated with V3 and the differences in gene expression between both wild RABV isolates. Such research not only provides insights into the basic pathology of lyssaviruses but also may prove important towards future treatment modalities. | 5,135.2 | 2022-03-27T00:00:00.000 | [
"Biology"
] |
Governance through Economic Paradigms: Addressing Climate Change by Accounting for Health
Climate change is a major challenge for sustainable development, impacting human health, wellbeing, security, and livelihoods. While the post-2015 development agenda sets out action on climate change as one of the Sustainable Development Goals, there is little provision on how this can be achieved in tandem with the desired economic progress and the required improvements in health and wellbeing. This paper examines synergies and tensions between the goals addressing climate change and economic progress. We identify reductionist approaches in economics, such as ‘externalities’, reliance on the metric of the Gross Domestic Product, positive discount rates, and short-term profit targets as some of the key sources of tensions between these goals. Such reductionist approaches could be addressed by intersectoral governance mechanisms. Health in All Policies, health-sensitive macro-economic progress indicators, and accounting for long-term and non-monetary values are some of the approaches that could be adapted and used in governance for the SDGs. Policy framing of climate change and similar issues should facilitate development of intersectoral governance approaches.
Introduction
Governance for the Millennium Development Goals (MDGs; United Nations [UN], 2015) from 2000-2015 was critiqued for not having fully considered interactions among the goals (Waage et al., 2010). When the Sustainable Development Goals (SDGs; UN Sustainable Development Platform, 2015) were developed for 2015-2030 to succeed the MDGs, efforts were made to emphasize potential interactions among them . The 17 SDGs are supported by 169 targets with numer-ous indicators specified at global, regional, and national levels. Such a framework offers an opportunity to identify and exploit beneficial interactions among the goals. In order to design such governance mechanisms and to ensure their effectiveness, it is essential to examine possible tensions and synergies among the SDGs, thereby learning and applying the lessons from what was rarely achieved for the MDGs.
We examine links between SDG 13 addressing climate change and SDG 8 on economic growth, focusing on accounting for their links with SDG 3 on human health and wellbeing to illustrate how intersectoral governance approaches could benefit governance for the SDGs. This approach is comparatively unique because the interactions among such ostensibly disparate SDGs have rarely been investigated in detail. Most comparative analyses of SDGs thus far (e.g. adopt a broadbrush picture for governance framing, rather than detailed critiques of connections among selected goals. Climate change has been proposed as a major challenge for sustainable development (Intergovernmental Panel on Climate Change [IPCC], 2014;UN, 2015;Worldwatch Institute, 2015). SDG 13 is devoted to climate change: "Take urgent action to combat climate change and its impacts" (UN Sustainable Development Platform, 2015). The specific targets of this goal cover both climate change mitigation (reducing greenhouse gases and increasing their sinks) and climate change adaptation (adjusting to climate change impacts), for this paper collectively termed "climate change action".
The achievement of SDG 13 is challenged by continued pursuit of unsustainable economic progress. SDG 8 sets a target for further economic growth for some countries: "Sustain per capita economic growth in accordance with national circumstances and, in particular, at least 7% gross domestic product growth per annum in the least developed countries" (UN Sustainable Development Platform, 2015). Another target of SDG 8 is set for decoupling economic growth from environmental degradation, as per the 10-year framework of programmes on sustainable consumption and production (UN Sustainable Development Platform, 2015).
So far, there has been no absolute decoupling of economic growth from greenhouse gas emissions (Steinberger, Krausmann, Getzner, Schandl, & West, 2013). This may present a significant challenge to the simultaneous achievement of both effective climate change action and economic growth. Furthermore, some argue that climate change may cause significant harm to the global economy, mainly by disrupting set processes and interfering with established mechanisms for creating economic wealth, as measured by Gross Domestic Product (GDP) (Cole, 2007;Weitzman, 2007).
Population health is rarely explicitly considered in decisions pertaining to economic growth even though it is an implicit part of such determinants of growth as labour productivity and human resources. Similarly, the 10-year framework of programmes on sustainable consumption and production, which is the policy and governance mechanism for decoupling economic growth from environmental degradation, suggested in the SDGs, does not refer to human health (UN, 2012).
Both climate change impacts and unsustainable economic growth are expected to have overall negative impacts on the health of populations, although it is always a balance with some positive impacts seen, often delineated by location and subpopulation. To avoid contradictions among the SDGs, such as on climate change and economic growth, their impacts on health targets spec-ified in SDG 3 could be helpful for determining suitable intersectoral governance approaches.
In this paper, we first explain the key tensions between the goals for economic growth and climate change action. We then provide insight into paradigmatic sources of these tensions. Finally, we illustrate how the link of economic growth with climate change is likely to be mediated through human health and then we discuss ways of recognizing how this and similar links could benefit the design of more sustainable intersectoral governance approaches.
Economic Growth as a Driver of Climate Change
Major tension between SDGs 8 and 13 lies in the continued contribution of key drivers of economic growth in the furthering of climate change.
Economic growth has been closely linked to high levels of fossil fuel use and greenhouse gas emissions, which perpetuate anthropogenic climate change. In 2013-2014, 306 tonnes of carbon dioxide were produced per each USD 1 million of the global GDP (PWC, 2015). Such a rate of greenhouse gas emissions requires a 6.3% rate of reduction in the carbon intensity of global economic growth to achieve the climate change target of atmospheric warming down to 2°C above the pre-industrial average which was set in the Paris Agreement from December 2015 (United Nations Framework Convention on Climate Change [UNFCCC], 2015).
The current global economic system was developed during a period of carbon-intensive rapid economic growth (Hall & Klitgaard, 2011;Henriques, 2011;Kraussman & Haberi, 2002), which in many ways it was designed to facilitate (Demirguc-Kunt & Levine, 2001;Fitzgerald, 2006;Rousseau & Sylla, 2001). This has left the global economy heavily dependent upon the continuation of such growth. Canadell et al. (2007) describe how, from 2000-2006, the carbon emissions required to produce a dollar of global economic activity unit have increased approximately 0.3% per year with Peters et al. (2012) reporting a 0.9% increase for 2010 which they attribute principally to burning fossil fuels and producing cement.
According to the Environmental Kuznets Curve hypothesis, economic growth first creates environmental problems, but later serves to reduce them (Grossman & Krueger, 1995). With respect to climate change, the result should be an absolute decoupling of growth from greenhouse gas emissions; i.e., GDP can increase without increasing greenhouse gas emissions . Although some relative decoupling of economic growth from greenhouse gas emissions has been achieved through enhanced energy efficiency and an increasingly service-based economy, there has not yet been absolute decoupling (Steinberger et al., 2013).
The small relative decoupling of 1.3% annual decrease in the carbon intensity of global economic growth (PWC, 2015) has been more than offset by the high rate of carbon-intensive economic growth. Between 2004 and 2014, global GDP growth of 44% produced an increase in greenhouse gas emissions of 22% (Handrich, Kemfert, Mattes, Pavel, & Thure, 2015). Hence, the pursuit of economic growth, as it is currently generated, does not meet the environmental sustainability criteria in relation to climate change mitigation.
Target 4 of SDG 8 suggests decoupling economic growth from environmental degradation following a framework of programmes on sustainable consumption and production (UN, 2012). Documentation of this framework uses the word "sustainable" without defining or providing criteria of "sustainable". Hence, interpretation of target 8.4 hinges on the definition of "sustainable" and "sustainable economic growth". It might potentially refer to "sustained"; i.e., forever and hence assuming infinite availability and use of carbon-intensive resources for economic growth.
Economic Growth and Climate Change Driving Different Priorities
Another level of complexity augmenting the tension between SDGs 8 and 13 is different levels of vulnerability to climate change impacts of those with different power in making decisions pertaining to economic growth and who may interpret "sustainable economic growth" differently.
Sectors that would benefit most from rapid economic growth tend to have high capacity to protect themselves from the impacts of climate change on their health and wellbeing. For example, Canary Wharf, one of London's financial centres, is located in a zone highly vulnerable to storm surge flooding (Dawson, Hall, Bates, & Nichloss, 2005;Jacob, Gornitz, & Rosenzweig, 2007). Although the Thames Barrier provides some protection, to a large extent its construction facilitated the development of this financial centre due to the perception of it being safe from floods (Ward & Smith, 1998). Under climate change, without changes to the Thames Barrier, the Barrier may be inadequate to prevent a major disaster costing hundreds of billions of pounds (Dawson et al., 2005). Thames Barrier upgrades are being discussed now, for plans covering the rest of the century (Environment Agency London, 2012). Hence, despite the climate-related risks, the financial sector has resources to offset their potential losses through constructing ostensibly protective physical infrastructure, using protective financial services (e.g., insurance), and diversifying assets. To maintain resources of the financial sector for such protection, the preferred interpretation of the term "sustainable economic growth" in the SDG 8 may indeed be "sustained".
By contrast, those who have contributed least to greenhouse gas emissions are those who are starting now to experience adverse impacts from climate change and who are likely to advocate for carbon-neutral or carbon-negative "sustainable economic growth" (Brulle, 2015;Parks & Roberts, 2010). Locations highlighted are indigenous peoples in the Arctic and those living along coasts, such as in Bangladesh, Kiribati, Maldives, and Tuvalu (IPCC, 2014). In risk analysis language, the risk takers are different from the risk makers (Glantz, 1996;Glantz & Jamieson, 2000). Several low-lying island countries organised a 1989 conference highlighting their vulnerability to sea-level rise impacts (Island Vulnerability, 1989), which garnered little action outside of the island states. A generation later, some of the island communities are being forced to deal with climate change related challenges physically (Storlazzi, Elias, & Berkowitz, 2015) and socially (Kelman, 2015)-which is also occurring in some Alaskan communities (Bronen & Chapin III, 2013).
The closed élite circle of financial decision-makers and the technical complexity of the economic decisionmaking tools, alongside weak accountability of the decision-makers, further complicates transparency in, and possible involvement from, sectors with differing priorities regarding "sustainable economic growth".
Sources of Tension: Reductionist Paradigm in Economics
The underlying sources of the above-discussed tensions to a large extent lie in reductionist paradigms of economics. The dominant theory and practice of economics today, including methods for estimating economic growth, do not sufficiently account for the complex interactions of economic activities with outcomes such as climate change and its impacts. The concept of "externalities", reliance on the GDP metric, positive discount rates, and short-term profit targets illustrate how these important links are omitted from economic considerations.
Impacts of greenhouse gas emissions on the atmosphere, and consequently on human health, tend to be treated as "externalities" (Brandt et al., 2010;Matthews & Lave, 2000;Navrud, 2001). The term "externalities" describes the effects of production or consumption of goods and services, whose costs and benefits are not reflected in prices of goods and services provided (OECD, 2003). Hence, greenhouse gas emissions and their impacts tend to be external to cost-benefit calculations over the short-term.
Social and health effects of activities that contribute to economic growth, measured by GDP, are similarly often treated as "externalities". In GDP calculations, war expenditures are judged the same as costs to feed and educate the population. Moreover, after a given level of GDP per capita, additional economic growth tends not to produce increases in wellbeing (Anielski, 2007) or happiness (Inglehart, Foa, Peterson, & Welzel, 2008;Layard, 2003Layard, , 2005Stott, 2012). Hence, GDP can be better characterized as a measure of market-based expenditures, which does not judge whether a given expenditure increases or decreases social welfare.
Another example is the use of positive discount rates. A discount rate is used to calculate how much avoided future damage, e.g., from a flood, would be worth compared to the initial cost of actions needed to avoid the damage. In relation to climate change, the effects of which manifest over the long-term, positive discount rates value future impacts at a fraction of current costs (Beckerman & Hepburn, 2007;DeCanio, de Lavergne, & Palter, 2003). The use of positive discount rates is wellcritiqued in the literature for valuing the present more than the future (Beckerman & Hepburn, 2007;DeCanio et al., 2003). Given positive discount rates, few economic incentives exist to avoid climate change, due to its longterm effects.
Furthermore, short-term profit targets are motivated by shareholder and investor pursuit of immediately optimal financial performance and successful revenue management by businesses, foregoing long-term and nonmonetary value creation (EY Poland, 2014). Most business models do not take into account long-term benefits or consequences of their activities, including climate change, or non-monetary values benefitting human health and wellbeing (Paulson, 2015).
Such reductionist thinking renders some economic, financial, and business models to portray climate change action as a costly and irrational act for stakeholders involved in the production of economic growth at all levels: governments, corporations, investors, producers, and consumers.
Sources for Synergies: Climate Change Impacts on Economic Growth via Health
The reductionist paradigm is not consistent with the current scientific understanding of the links between climate change and economic growth. Considering health impacts of climate change and their further implications for economic growth highlights potential shortcomings of working towards the SDGs without addressing the aforementioned reductionist approaches.
Indirect climate change impacts on health are rarely accounted for in estimates of the economic impacts of climate change. However, recent scientific evidence suggests that these impacts have significant implications for labour productivity and human resources. Higher temperatures are shown to be associated with a decrease in the productivity of those performing heavy labour outdoors and, when air conditioning is not available, indoors (Sahu, Sett, & Kjellstrom, 2013). Furthermore, higher temperatures would lead to fewer hours of physiologically safe temperatures for work in non-air conditioned spaces. In South-East Asia, 15-20% of annual work hours are estimated to be already lost under the current climatic conditions; this loss could double by 2050 under projected climatic change (Kjellstrom, 2015).
The projected climate change related decreases in global food availability would challenge the decline of global child undernutrition rates achieved over past decades (UN, 2015). This may subsequently lead to a rise in the long-term consequences of childhood un-dernutrition, such as lower performance of the immune system (Dercon & Porter, 2014), increased risk of chronic diseases (Black et al., 2013), compromised cognitive development (Ampaabeng & Tan, 2013), and lower economic productivity in adulthood (Dewey & Begum, 2011), all further challenging labour productivity and human resources.
As such effects compound, in addition to the health and welfare of people, production and consequently economic growth could be increasingly afflicted. Labour productivity loss is the most substantial economic loss that the world would face from climate change (DARA & Climate Vulnerable Forum, 2012). Already in 2010, the loss of labour productivity globally was suggested as being equated to the net loss of USD 311 billion (2010 PPP), which is around 0.5% of the global GDP (DARA & Climate Vulnerable Forum, 2012). By 2030, the net loss due to compromised labour productivity is projected to reach USD 2.4 trillion per annum (DARA & Climate Vulnerable Forum, 2012). Knock-on effects from these labour impacts mean that even atmospheric warming by 2°C above the pre-industrial levels is projected to result in a loss of USD 4.2 trillion in the asset management industry from the private sector perspective, which is equivalent to the world's listed value of all oil and gas companies combined and which is the equivalent of Japan's GDP (The Economist Intelligence Unit, 2015).
The link illustrated here emphasizes that population health, which is essential for economic growth, will be (and in some places already is) constrained by climate change. The previously discussed reductionist approaches in economics would leave this link unaddressed. By contrast, integrated intersectoral governance approaches designed on the basis of understanding interactions across the SDGs could provide political space for addressing the complex indirect impacts and could further incentivize synergistic action on climate change across sectors.
Suggestions for Intersectoral Governance Approaches: Beyond Reductionism
Links across the SDGs, such as the indirect impact of climate change on economic growth through its impact on health, emphasize the need to govern SDGs in an integrated manner. We discussed four economic paradigms not conducive to such integrated governance approaches, especially as they impede climate change action: (1) the construct of "externalities", (2) reliance on the metric of GDP, (3) discount rates, and (4) shorttermism. Alternative governance processes can be suggested for each of these paradigms. We provide three examples of existing governance mechanisms that could be used to counter these paradigms by considering the links of SDG 3 with SDGs 8 and 13. We conclude this section with a case study on the framing of climate change and disaster risk reduction in wider policy contexts, illustrating the need for policies to be formulated in a way that facilitates development of such integrated governance mechanisms.
Health in All Policies: To Value "Externalities" and Short-Term Health Co-Benefits
Health in All Policies (HiAP) promoted by the World Health Organisation is "an approach to public policies across sectors that systematically takes into account the health implications of decisions, seeks synergies, and avoids harmful health impacts in order to improve population health and health equity" (World Health Organization, 2014). It draws attention to the consequences of public policies on the determinants of health, aiming to improve policy makers' accountability for health impacts of their decisions (World Health Organization, 2014).
In governance for SDGs, HiAP could be used to incorporate health implications across time scales into costbenefit considerations made by stakeholders from international to individual levels and across sectors. Tools such as the Health Impact Assessment and Health Risk Assessments could be adapted to suit the range of possible interactions across the SDGs and incorporated as a regulatory element of governance for the SDGs (Winkler et al., 2013). These elements could help to counter the paradigm of health implications being treated as "externalities" in day-to-day economic decisions as well as to link health with promoting the "green economy" (Winkler et al., 2013).
HiAP could also be used to develop intersectoral policy structures and to provide space for representatives of the health sector to communicate health implications to decision-makers in other sectors. For example, in a debate on discount rates, health sector representatives could lobby for climate change action in spite of positive discounting by emphasizing the immediate health benefits of many choices in favour of climate change mitigation, such as the positive health consequences of reduced car use, including cleaner air and reduced cardiovascular disease (Watts et al., 2015).
Health-Sensitive Macro-Economic Progress Indicators
The UN Statistical Commission and the Inter-Agency and Expert Group on Sustainable Development Goal Indicators have been coordinating the development of an overarching framework of indicators for monitoring and evaluating progress towards the SDGs. As of 17 December 2015, a list of 229 indicators was compiled in a proposal for the framework (United Nations Economic and Social Council, 2016).
The proposed SDG indicators make nearly three dozen references to the GDP metric, including a target for more growth in the least developed countries (United Nations Economic and Social Council, 2016). None of the references exploits possible synergies or addresses tensions between sources of GDP growth and the SDGs. Si-mon Kuznets, who is credited with developing the GDP measure, never intended GDP to be used as a gauge of general social welfare. Kuznets noted, "Distinctions must be kept in mind between quantity and quality of growth, between costs and returns, and between the short and long term. Goals for more growth should specify more growth of what and for what" (Kuznets, 1962).
The specification of "more growth of what and for what" is limited in the current formulation of the targets and their indicators. Attempts to propose macroeconomic progress metrics as alternatives to GDP, which incorporate health, wellbeing, and other sustainability considerations were made in the past, e.g., the Index of Sustainable Economic Welfare (Daly & Cobb, 1989) and the Genuine Progress Indicator (Talberth, 2007). GDP remains the paramount macro-economic metric, to a large extent due to its simplicity and universality (Costanza, Hart, Posner, & Talberth, 2009). To account for "growth of what and for what" in relation to SDGs, complementary macro-economic progress metrics could be developed reflecting the extent to which economic growth of different countries is aligned with their progress towards the SDGs. Such metrics could be used to monitor whether a country's growth becomes more sustainable and more beneficial for the health of the global population. Criteria of what is more sustainable in this context should be defined on the basis of SDG targets and indicators, taking into account their interactions. Interactions concerning SDG 3 may also engage those who would favour interpreting the term "sustainable economic growth" as "sustained". For example, current contributions of economic growth to population health may secure higher potential for economic growth in the future through the links of good population health with higher future human resources and productivity.
Accounting for Long-Term and Non-Monetary Values
Apart from macro-economic progress indicators and policies, individual participants in the economy and particularly the financial system can be engaged in facilitating progress towards the SDGs through socially responsible investment mechanisms encouraging consideration of long-term and non-monetary values compliant with the SDGs in their financial decisions. Existing mechanisms include positive and negative screening, disinvestment, and shareholder engagement. Often, elements of such mechanisms are already aligned with SDGs such as SDG 3 on health. For example, positive screening often includes health and safety considerations addressing such targets of SDG 3 as exposure to hazardous chemicals and pollutants and prevention of substance abuse (Youssef & Whyte, 2013). Climate change impacts have also been considered in more traditional financial decision-making tools, for example, in the design of the long-term investment portfolio risk management strate-gies (Mercer, 2015) and in the insurance sector (Gurenko, 2006;Xu, 2014).
Further incentives for the focus on long-term and non-monetary value creation in the business sector could also be achieved through managerial innovation; for example, restructuring executive remuneration schemes in a way that increases the proportion of their compensation based on long-term company performance (EY Poland, 2014). Greater focus on the longterm performance of companies, in turn, would allow more time for costumer choice to be reflected in a company's performance metrics. Concurrently, consumers and other stakeholders could be sensitized to the social and environmental impacts of businesses pertinent to the SDGs such as health and its determinants, at their individual and community levels through comprehensive education and communication strategies.
The above-illustrated approaches could be adapted and used in governance for the SDGs. Approaches similar to HiAP could further be used to ensure policy coherence and use of shared policies across sectors (Becerra-Posada, 2015). HiAP is particularly relevant for this purpose as it focuses on the determinants of health, which are mostly governed by sectors other than the health sector, requiring complex integrated governance solutions. Establishment of virtual intersectoral boards and taskforces would be required to identify synergies across the SDGs and to devise as well as implement ways of accounting for such effects in daily policy decisions while monitoring progress towards the SDGs.
Beyond Reductionism: Climate Change in Wider Policy Contexts
Development of intersectoral governance mechanisms requires policy framing that permits and encourages intersectoral links. Currently, climate change in policy is mostly formulated as a somewhat isolated environmental process influenced by humanity. Despite its numerous links with many other policies such as those on health and disaster risk reduction, the policy and political processes of climate change have separated it from many other environmental and policy topics.
In the SDGs, climate change is formulated as a separate goal, SDG 13. A footnote to SDG 13 states "Acknowledging that the United Nations Framework Convention on Climate Change is the primary international, intergovernmental forum for negotiating the global response to climate change" (UN Sustainable Development Platform, 2015). Emphasis on a single forum for negotiations on climate change action may ideologically segregate the issue from other intergovernmental fora that could further facilitate addressing climate change impacts.
A contrast can be made with disaster risk reduction policies. As the agreements for the SDGs and UNFCCC (2015) were shaping up, in March 2015 a voluntary international agreement was signed under UNISDR (United Nations Office for Disaster Risk Reduction) auspices, the Sendai Framework for Disaster Risk Reduction (SFDRR; UNISDR, 2015), also running from 2015-2030. The agreement notes the health and economic benefits of disaster risk reduction, synergising with the discussion here regarding climate change.
For example, the outcome in paragraph 16 of UNISDR (2015, p. 9) is "The substantial reduction of disaster risk and losses in lives, livelihoods and health and in the economic, physical, social, cultural and environmental assets of persons, businesses, communities and countries" specifically noting both health and economics. Paragraph 18 of UNISDR (2015, p. 9) includes disaster risk reduction targets to "Reduce direct disaster economic loss in relation to global gross domestic product" (clause c) and "Substantially reduce disaster damage to critical infrastructure and disruption of basic services, among them health and educational facilities" (clause d). As a result, disaster risk reduction measures aim to help minimize negative health and economic impacts of disasters-and often succeed through saving lives (meaning reduced adverse health outcomes) and saving money, as demonstrated by benefit-cost analyses of disaster risk reduction interventions (Shreve & Kelman, 2014).
Climate change is reasonably integrated into SFDRR across DRR contexts; however, the statement "The climate change issues mentioned in this Framework [SF-DRR] remain within the mandate of the United Nations Framework Convention on Climate Change under the competences of the Parties to the Convention" (UNISDR, 2015, p. 11) distances climate change from the DRR mandate instead of fully integrating it into DRR. Hence, the wording of SFDRR separates governance on climate change action and on disaster risk reduction while the wording of the SDGs segregates global governance on climate change from intergovernmental fora other than the UNFCCC.
Climate change policy integration with disaster risk reduction policies would benefit climate change adaptation efforts. Climate change is an important influence, by affecting several hazards including storms, temperature, precipitation, and infectious disease, sometimes exacerbating the hazards and sometimes diminishing them. As opposed to the policy framing in disaster risk reduction, which ensures connections across all hazards, an isolated focus on climate change may constrain instead of facilitate intersectoral synergies supporting climate change action (UNISDR, 2015).
To move beyond the reductionism of climate change and its separation from other processes, especially for connecting health and economic benefits, considering wider policy contexts is necessary. The SDGs, to a large extent, have mainstreamed disaster risk reduction by mentioning the process in numerous SDGs and targets (UN Sustainable Development Platform, 2015). As such, disaster risk reduction is not a standalone process with its own separate SDG but, rather, is integrated into sustainable development. Climate change, as a single hazard influencer among many, was not accorded similar treatment.
Conclusion
Comprehensiveness of the SDG framework offers an opportunity to exploit interactions across the goals. Apart from synergistic interactions, tensions between some of the goals are likely, as shown by our analyses of SDGs 8 and 13, alongside their links with SDG 3 on health followed by the comparison with disaster risk reduction. The reductionist approaches prevalent in economics, such as "externalities", GDP, positive discount rates, and short-term profit targets are likely to be some of the key sources of possible tension between SDGs 8 and 13. These approaches do poorly in considering the complex links among the SDGs, an example being health impacts of climate change and their further implications for economic growth.
In terms of its theoretical value, this paper provides a conceptual baseline for overcoming reductionist approaches. As discussed in section 6.4., health and economic considerations are frequently interpreted and applied in a reductionist manner. The formulation of climate change in policy contexts is often structured in a reductionist manner. However, opportunities for integrating climate change action with policies in other fields could enhance the effectiveness and efficiency of action on climate change. We have provided theoretical suggestions and examples of how to retain the important components of all three topics without becoming ensnared in reductionist thinking.
In terms of its policy value, this paper suggests the intersectoral governance mechanism of HiAP and the development of complementary economic progress metrics aligned with the SDGs. Considering the health impacts of policies-such as those related to economics, climate change, and wider disaster risk reduction-in order to ensure health benefits while avoiding deleterious health consequences would be an important step forward in sustainable development approaches. Here, the importance of HiAP for SDG 8 is demonstrated, yet the lessons apply to policies related to other SDGs and their interactions.
Suggestions made in this paper also have value for practice, particularly when making investment decisions for financial portfolios or development projects. Alternatives to carbon-intensive and growth-focused investments are provided, suggesting how a health focus could lend itself to paybacks and outcomes which might not match economic goals, but which are nonetheless sound economic decisions by supporting healthy people and communities.
The growing recognition of a wide range of socioeconomic factors influencing human health and wellbeing has facilitated development of intersectoral governance approaches, such as HiAP. These approaches could be adapted and incorporated into governance for the SDGs, especially through comparison and analysis of SDGs beyond the three considered here.
About the Authors
Kristine Belesova is a PhD candidate in Global Public Health and Environmental Epidemiology at the London School of Hygiene and Tropical Medicine (LSHTM). Her research examines possible impacts of climate change and extreme weather on food systems and human health. Kristine is also interested in the design and evaluation of policy and governance strategies for sustainable development. She leads the course "Environmental Change and Global Health Policy" at LSHTM and lectures at University College London and Heidelberg University.
Ilan Kelman is a Reader in Risk, Resilience and Global Health at University College London, UK, and a researcher at the University of Agder, Norway. His overall research interest is linking disasters and health, including the integration of climate change into disaster research and health research. That covers three main areas: (i) disaster diplomacy and health diplomacy; (ii) island sustainability; and (iii) risk education for health and disasters.
Roger Boyd is a retired financial industry executive, where he worked for 25 years. He received an MBA in Finance from Stern School of Business, New York University in the United States and an MA in Integrated Studies from Athabasca University in Canada. Over the past decade he has taken a deepening interest in the way in which modern societies, especially their financial systems, will deal with global threats such as energy constraints and climate change. | 7,250.6 | 2016-12-28T00:00:00.000 | [
"Economics",
"Environmental Science",
"Medicine"
] |
Semistructured Data: The TSIMMIS Experience
In this paper we discuss themanagement of semi-structured data, i.e., data that has irregular or dynamically changing structure. We describe components of the Stanford TSIMMIS Project that help extract semi-structured data from Web pages, that allow the storage and querying of semi-structured data, and that allow its browsing through the World Wide Web. A prototype implementation of the TSIMMIS system as described here is currently installed and running in the database group testbed.
Introduction
At a recent w orkshop on management of semistructured data 15 , the workshop attendants de ned semistructured data as data that does not have a regular and static structure like data found in a relational database but whose schema is dynamic and may contain missing data or types.For example, if we l o o k a t w eather forecasts on the Web, the " elds" and their structure may di er across sites.Even at a single site, some forecasts may be missing information, or may h a ve extra information depending on the geographical location of the a ected region e.g., cities in the Rocky Mountains usually include ski reports in the winter months whereas forecast for tropical resorts do not.However, semistructured data is not just limited to the World Wide Web WWW, but is also found in many other interesting sources including le systems, news wires, electronic mail systems, etc. just to name a few.In addition to occurring natively in the above-mentioned classes of sources, semistructured data is often a "by-product" of the integration process when multiple heterogeneous schemas are involved.In those cases, semistructured rather than "fully structured" data arises because the integrated objects may be based on complimentary, sometimes con icting, and often dynamic information from multiple sources, forcing the integrator to lter, merge, or omit certain elds when performing the integration.
The goal of the Tsimmis project at Stanford 4, 7, 11, 13 is to provide integrated access to a wide variety of heterogeneous data sources e.g., databases, object stores, knowl-edge bases, digital libraries including sources containing semistructured data e.g., WWW, le system.In this paper, we present the Tsimmis approach to managing semistructured data.In particular, we discuss three critical aspects semistructured data management: 1 extracting the intended content from its native source how to get it?, 2 reading the extracted data how to query it?, and 3, exploring the result in a graphical, easy-to-understand manner how t o b r o wse it?.In Tsimmis we h a ve developed components that address all of the above issues and together provide an integrated solution to the problem of managing semistructured data.Several other recent projects have similar goals e.g., Lore 10 , Garlic 3 , Information Manifold 9 , Rufus 14 , but we do not survey them here.
2 Representing Semistructured Data in TSIMMIS For the Tsimmis project we h a ve adopted a simple selfdescribing or tagged object model.Similar models have been in use for years; we call our version the Object Exchange Model, o r Oem 4 .Oemis a exible model that is particularly well suited for representing semistructured data.Data represented in Oem constitutes a graph, with a unique root object at the top and zero or more nested subobjects.The fundamental idea is that all objects, and their subobjects, have labels that describe their meaning.For example, the following object represent s a F ahrenheit temperature of 80 degrees: temp-in-Fahrenheit, int, 8 0 Here, the string temp-in-Fahrenheit" is a human-readable label, int" indicates an integer value, and 80" is the value itself.If we wish to represent a complex object, then each component of the object has its own label.For example, an object representing a set of two temperatures may look like: set-of-temps, complex, f temp-in-Fahrenheit, int, 8 0 temp-in-Celsius, int, 2 0 g Oem is very simple, while providing the expressive p o wer and exibility needed for representing semistructured data from a wide range of heterogeneous sources.Our primary reason for choosing a simple model is to be able to accommodate a wide variety of external data models and to facilitate integration.As pointed out in 2 , a simple model such a s Oem has an advantage over complex models when used for representing and integrating heterogeneous data, since the Figure 1: A section of the HTML source le operations to transform and merge data will be correspondingly simpler.Meanwhile a simple model can still be very powerful: advanced features can be emulated" when they are necessary e.g., subclass superclass relationships, inheritance, etc..For additional information on Oem, please refer to 12 .
3 Extracting Data Continuing with our weather example, let us assume that we h a ve an application that needs to process weather data, such as temperature and forecast, for a given city.As one of its information sources, we w ant t o u s e a W eb site called Intellicast, which reports daily weather data for most major cities across the world.Since this site cannot be queried directly from within another application e.g., What is the forecast for Helsinki for May 7, 1997?" we rst have t o extract the contents of the weather table from the underlying HTML page 1 which is displayed in Figure 1.
The Extraction Process
Our con gurable extraction program parses this HTML page based on the speci cation le shown in Figure 2. The specication le consists of a sequence of commands, each de ning one extraction step.Each command is of the form variables, source, pattern where source speci es the input text to be considered, pattern tells us how to nd the text of interest within the source, and variables are one or more extractor variables that will hold the extracted results.The text in variables can be used as input for subsequent commands.If a variable contains an extracted URL, we can also specify that the URL be followed, and that the linked page be used as further input.
After the last command is executed, some subset of the variables will hold the data of interest.Later we describe how the contents of these variables are packaged into an Oem object.
Looking at Figure 2, we see that the list of commands is placed within the outermost brackets ` ' and ` ', and each command is also delimited by brackets.The extraction process in this example is performed by v e commands.The initial command lines 1-4 fetches the contents of the source le whose URL is given in line 2 into the variable called root.
The `' character in line 3 means that everything in this case the contents of the entire le is to be extracted.After the le has been fetched and its contents are read into root, the extractor will lter out unwanted data such as the HTML markup commands and extra text with the remaining four commands.
The second command lines 5-8 speci es that the result of applying the pattern in line 7 to the source variable root is to be stored in a new variable called temperature.The pattern can be interpreted as follows: discard everything until the rst occurrence of the token h=T Ri `*' means discard in the second table de nition and save the data that is stored between h=T Ri and h=T ABLEi `' means save.The two hT A B L Etokens between the `*' are used as navigational help to identify the correct h=T Ri token since there is no way of specifying a numbered occurrence of a token i.e., discard everything until the third occurrence of h=T Ri.After this step, the variable temperature contains the information that is stored in lines 22 and higher in the source le in Figure 1 up to but not including the subsequent h=T ABLEi token which indicates the end of the temperature table.The underscore at the beginning of the name temperature indicates that this is a temporary variable; its contents will not be included in the resulting Oem object.
The third command lines 9-12 instructs the extractor to split the contents of the temperature variable into chunks" of text, using the string hT RALIGN = lefti lines 22, 30, 38, etc. in Figure 1 as the chunk" delimiter.Note, each c hunk" represents one row in the temperature table.The result of each split is stored in a temporary variable called citytemp.The split operator can only be applied if the input is made up of equally structured pieces with a clearly de ned delimiter separating the individual pieces.If one thinks of extractor variables as lists up until now each list had only one member then the result of the split operator can be viewed as a new list with as many members as there are rows in the temperature table.Thus from now on, when we apply a pattern to a variable, we really mean applying the pattern to each member of the variable, much like the apply operator in Lisp.
In command 4 lines 13-16, the extractor copies the contents of each cell of the temporary array i n to the array city temp starting with the second cell from the beginning.The rst integer in the instruction citytemp 1 : 0 indicates the beginning of the copying since the array index starts at 0, 1 refers to the second cell, the second integer indicates the last cell to be included counting from the end of the array.As a result, we h a ve excluded the rst row of the table which contains the individual column headings.Note, that we could have also ltered out the unwanted row in the second command by specifying an additional *h=TRi condition before the `' in line 7 of After the ve commands have been executed, the variables hold the data of interest.This data is packaged into an Oem object, shown in Figure 3, with a structure that follows the extraction process.Notice that this sample object re ects the structure of our extractor speci cation le.That is, the root object of the Oem answer will have a label root because this was the rst extracted variable.This object will have c hildren objects with label city temp and so on.Notice that the variables temperature and citytemp do not appear in the nal result because they are declared as temporary variables.
Additional Capabilities
In addition to the basic capabilities described in our example, the extractor has components for automatic handling of HTML tables, for conditional parsing, and other services.The extractor can also follow URLs in the process, extracting data from multiple Web pages into a single Oem object.Overall, we believe the extractor provides natural facilities for extracting data, as well as for structuring it in di erent ways into Oem objects.For more details on the extractor, please refer to 8 .
Querying Semistructured Data
In this section we i n troduce the Lorel query language, primarily through examples.Lorel is an extension of OQL and a full speci cation can be found in 1 .Here we highlight those features of the language that have an impact on the novel aspects of the system|features designed specically for handling semistructured data.Many other useful features of Lorel some inherited from OQL and others not that are more standard will not be covered.
Simple LORE Examples
Our rst example query introduces the basic building block of Lorel: the simple path expression, which is a name followed by a sequence of labels.For example, Root:City:Location is a simple path expression.Its semantics consists of the set of objects that can be reached starting with the Root object, following an edge to objects labeled City, then following an edge to objects labeled Location.Range variables can be assigned to path expressions, e.g., Root:City:LocationX" speci es that X ranges over the set of locations.
Continuing with our European weather example, the following example query retrieves the locations of all cities located in England when evaluated over the sample Oem database shown in Figure 4.At a high level, the query execution engine will nd all objects which satisfy the path Root:City:Location and for each of these will check whether the where clause is satis ed.The result of Query 4.1 is shown here: answer complex f location string Southern" location complex f longitude oat -0.167 latitude oat 51.5 g g The database over which this query is evaluated presents a n umber of irregularities, as discussed earlier.A guiding principle in Lorel is that, to write a query, one should not This query will not yield a run-time error if a Location object has a string value or is complex, or if Country objects are single-valued, set-valued, or even absent for some cities.Indeed, the above query will succeed no matter what the actual structure of the database is, and will return an appropriate answer.Of course, this query was written with some obvious knowledge of how the graph is laid out within our database.In Sec.4.2 we discuss how an end user can discover the structure of the database.
Value comparisons are made after two objects have been coerced into comparable types.That is, if two objects do not have the same type then attempts will be made to coerce the values into comparable types before applying the comparison operator.Any t ypes which cannot be coerced for comparison will not return type errors, but will simply evaluate to false.This reinforces our underlying principle that Lorel does not require precise knowledge of the data and is most useful when dealing with semistructured data.
The system will in fact translate all Lorel queries into OQL-like queries for evaluation.This is done for two reasons: rst, Lorel is based on OQL and thus OQL gives us well de ned semantics for our queries, and second it allows a user familiar with OQL to directly enter an OQL query to be evaluated over the semistructured data.In some sense, Lorel can be viewed as shorthand for OQL, however Lorel also introduces generalized p ath expressions not present within OQL.Generalized path expressions o er a richer form of declarative n a vigation" in Oem databases than simple path expressions.Intuitively, the user loosely speci es a desired pattern of labels in the database: one can specify patterns for paths to match sequences of labels, patterns for labels to match sequences of characters, and patterns for atomic values.A combination of these three forms of pattern matching is illustrated in the following example: Here the expression weather is a label pattern that matches all labels ending with the string weather e.g., weather, Todays weather, o r Tomorrows weather.For path patterns, the symbol j" indicates disjunction between two labels, and the symbol ?" is applied to the parenthesized expression to the left and indicates that the label pattern is optional.The complete syntax is based on regular expressions, along with syntactic wildcards such as ", which matches any path of length 0 or more.Finally, grep"rain" speci es that the data value should contain within it the string "rain.The grep operator is similar to the Unix grep command.We also support like, based loosely on the SQL like, and soundex for phonetic matching.In English this query is asking for the names of all cities where the forecast or outlook of the weather contains the word rain".Figure 5: Sample DataGuide strated in our rst example.It is not possible to do so with general path expressions, which require a run-time mechanism.Indeed, note that if the database contains cycles, then a general path expression may match an in nite number of paths in the data.When trying to match a general path expression against the database, we match through a cycle at most once, which appears to be a reasonable simpli cation in practice.We conclude with an example that illustrates advanced features of the language.The following query illustrates subqueries and constructed results.For every city in the database that satis es the bottom where clause, we will select out the name of the city along with the current temperature, but only if the current temperature satis es the WHERE clause.The result is shown below.Notice that each city which provides a binding for the C variable and satis es the where clause appears within the answer.Of particular interest is the fact that Plymouth does not have a current temp eld within the answer.This is ltered out as a result of the subquery appearing within the SELECT clause.Specically, the Plymouth object does not have a subobject labeled Todays weather.
answer complex f city complex f name string "London" current temp integer 7 g city complex f Name string "Plymouth" g g 4.2 Query Formulation with the DataGuide Since our data does not have an explicit schema, query formulation and query optimization are particularly challenging.Without some knowledge of the structure of the underlying database, writing a meaningful Lorel query may be di cult, even when using general path expressions.One may manually browse a database to learn more about its structure, but this approach is unreasonable for very large databases.Further, without information about the structure of the database, the query processor may be forced to perform more work than necessary.F or example, consider Query 4.1 that nds the locations of all cities whose country is England.Even if no cities have a country subobject, the execution engine would still needlessly examine every city i n the database.A DataGuide is a concise and accurate summary of the structure of an Oem database, stored itself as an Oem object.Each possible path expression of a database is encoded exactly once in the DataGuide, and the DataGuide has no path expressions that do not exist in the database.As an example Figure 5 shows a DataGuide for the sample database shown in Figure 4.Note that atomic values are usually not stored within the leaf nodes of the DataGuide since it is primarily concerned with the structure of the database.In typical situations, the DataGuide is significantly smaller than the original database.A DataGuide plays a role similar to metadata in traditional database systems.The DataGuide may be queried or browsed, enabling user interfaces or client applications to examine the structure of the database.Assuming the role of the missing schema, the DataGuide can also guide the query processor.Of course, in relational or object-oriented systems the schema is explicitly created before any data is loaded; here, DataGuides are dynamically generated and maintained over all or part of an existing database.
In 5 , formal de nitions for DataGuides are provided as well as algorithms to build and incrementally maintain DataGuides that support annotations.Also given is a discussion of how DataGuides aid query formulation in practice and their use for query optimization.
5 Browsing OEM Results through MOBIE The idea behind our browsing tool centers around the need for displaying semistructured objects in a way that makes it easy for the user to grasp their structure and explore their contents when viewing the result of a Tsimmis query.Oem results are typically irregular in structure and nested, containing a top-level root object and zero or more subobjects sometimes referred to as children.Each subobject may itself be a nested object.In general, nested objects are structured like trees or graphs if we allow cycles.Anybody who has worked with nested objects before can attest to the fact it becomes increasingly di cult to understand the contents of a nested object the more its structure increases in complexity i.e., the larger the number of subobjects and the deeper the level of nesting.
For this reason, we h a ve built a system that transforms Oem results into a web" of hyperlinked documents that can be viewed using any WWW browser.An object that is selected for viewing is formatted as an HTML document.If the object is a complex object, the document m a y also include hyperlinks pointing to some or all of the object's substructure depending on the user's preferences.If the object is atomic, it will be displayed by itself.In addition, each document always contains a link to the parent object, unless the selected object is the root of the entire structure.The main contribution of our system is that it gives the user the option to decide which information is to be displayed, how much of the chosen information he or she wants to see, and when.Information is presented one screen at a time, allowing the user to browse complex objects, which m a y b e t o o large to view all at once, in a cafeteria-style " pick-andchoose fashion.This approach to browsing nested objects is analogous to how one uses the table of contents to explore the individual chapters of a book.
An important part of the functionality of our browser focuses on the layout of information on individual pages.Since this is a process that depends heavily on each user's individual preferences as well as the data that is being displayed, we h a ve paid careful attention to design a system that is exible enough so that it can be tailored to satisfy many di erent needs.Our goal was to provide users with choices as to how information is to be displayed: from the overall layout of a screen down to the format of an individual object.Initially, the system uses default settings that maximize the amount of information that can be displayed within the given real-estate of the window.The result can then be improved upon by c hanging the values of session variables, which control the document l a yout, the level of nesting per screen, the number of subobjects per level, etc.By default, session variables control the formatting for the complete object hierarchy.H o wever, by using the label names that refer to a particular object in the hierarchy, the scope of session variables can be limited: from the entire hierarchy, t o a s p eci c substructure, to one object.Although customization of the object display m a y be time consuming in certain cases, the state of the session variables can be saved on a per-user basis and re-used during subsequent sessions.
We h a ve implemented a fully functional prototype system called Mobie Multimedia OBject and Information Explorer, which currently provides the graphical interface to Tsimmis data sources.However, Mobie is not limited to browsing only data from Tsimmis but can be tailored for displaying and formatting structured information from any object-based database footnoteOne can either use a translator for converting data into Oem or modify our algorithm to work with other object-based data models.. Since a complete description of our browser is beyond the scope of this paper, we i n vite the user to obtain the details from 6 .Instead, we will brie y demonstrate some of Mobie's functionality using screen snapshots from a sample interaction with a Tsimmis wrapper connected to the Intellicast weather source via the above mentioned Web extractor.We start our description when the result is returned from the database, omitting such details as how to connect to the database server, transmission of the query and its results, etc.When displaying data, we use the following conventions.Object labels are displayed in bold, object values are italicized.
Underlining indicates the existence of a hyperlink.
Sample Screen Snapshot
Let us assume that we h a ve submitted the following Lorel query asking for all cities in Europe where tomorrow's forecast calls for showers: Query 5.1 Lorel SELECT city temp FROM intellicast:i WHERE i.city temp.tomorrowstemp.forecast= "shwrs" Let us also assume that the answer to this query consists of three cities that are displayed together under one root object, labeled answer.Figure 6 shows the answer object as it is displayed in Mobie.Each object labeled city temp is a complex object exhibiting additional substructure underneath: the objects labeled city, city url, country, todays temp, and tomorrows temp.Note that the rst subobject the city of Helsinki has one additional subobject labeled 5-day forecast that is not present in the other results.The city, city url, and country subobjects are atomic meaning they contain no further substructure.In those cases, the value of the object is displayed.If there The todays temp and city url subobjects on the other hand, are complex objects that contain additional subobjects: forecast, high, low, and date.Labels belonging to complex objects are underlined meaning that a hyperlink exists that will take the user to the document containing only those subobjects.Those subobjects are displayed in a similar fashion.Also note, the value of the city url subobjects is a standard URL that is part of the answer and has been activated by Mobie for loading.
Formatting Options
As mentioned before, the user can control the formatting of objects through various control parameters.These parameters are called session variables and can be accessed from the User Defaults menu.Formatting options fall into two categories: Global Settings", which apply to the whole object structure, and Label-Based Settings" for which the scope can be speci ed based on object labels.See 6 for details and other options.
The following parameters are available for controlling global settings: Maximum levels of sub-objects controls the number of visible levels of subobjects for each object that is displayed.Sub-object indentation controls the amount of indentation used for subobjects.The following parameters are available for controlling label-based settings: Layout controls the overall look-and-feel" of the output when it gets displayed in the browser window.The two options currently available are table and list layout.Number of displayed sub-objects controls the number of subobjects that are displayed on a screen.Label size and value size control the length of labels and values respectively.
Using these options one can format data in the way that best suits it.For example, Figure 7 shows some data formatted as a table.Labels are shown on the left side.If the subobject is an atomic object e.g., the subobjects labeled city, city url, country, and 5-day forecast the rst column starting form the left will contain the subobject's value.If the subobject is a complex object, e.g., the subobjects labeled todays temp and tomorrows temp, the rst column will be empty, and subsequent columns will contain the values of its immediate subobjects.In the latter case, the column headings are the labels of the lower-level subobjects.Note, if there are several complex subobjects with di erent substructure, the table will display the union of all possible headings.
As mentioned before, label-based settings apply to objects.In order to format an object, a formatting choice associated with its label must be de ned.Thus it is possible, for example, to display three or more levels of nesting for the root object, and then reduce the number of visible levels to just one when viewing its subobjects.As another example, one can display the part of a result that contains numerical values as a table but leave the part that is mostly textual in list format.
Conclusion
In this paper we h a ve presented an overview of the Tsimmis approach to accessing and managing semistructured data.In particular, we h a ve described how semistructured data can be obtained from Web pages, how it can be manipulated in a database system, and how it can be browsed.We believe that semistructured data exists in many applications, and exible tools like the ones we h a ve described can be very helpful for managing it.
Figure 2 .
The nal command lines 17-20 extracts the individual values from each cell in the city temp array and assigns them into the variables listed in line 17 country, c url, city, etc..
Query 4. 3
Lorel SELECT C.Name, SELECT X FROM C.Todays weather.currenttemp X WHERE X 10 FROM Root.City C WHERE C.Country = "England" | 6,360.2 | 1997-09-02T00:00:00.000 | [
"Computer Science"
] |
Optimal Frequency Control for Virtual Synchronous Generator Based AC Microgrids via Adaptive Dynamic Programming
This paper proposes a novel virtual synchronous generator (VSG) controller for converters in AC microgrids (MGs). Such controller improves the control cost and DC-side energy requirements, while considering the system non-linearity for frequency support. First, the frequency dynamics of the MG, which are analytically studied based on the VSG with a secondary frequency controller, are formulated as a nonlinear state space representation. In this later, the reciprocal of the inertia is modeled as the control input. Correspondingly, a cost function is defined by comprehensively considering the angular deviation, frequency deviation, rate of change of the frequency (RoCoF), and a discount factor, which can retain a tradeoff between the critical frequency bounds and the required control energy. Following, the optimal frequency regulation problem is solved by using an online adaptive dynamic programming (ADP) method, where a single echo state network (ESN) is constructed to approximate the optimal cost function and the optimal control input to significantly reduce the computational burden and improve the real-time computation. Finally, simulation results demonstrate that the frequency response of the system is significantly improved, while also retaining more DC-side energy.
I. INTRODUCTION
M ODERN electricity power system is evolving toward the so-called distributed generation (DG) structure with the exhaustion of traditional fossil energy [1], [2]. DGs, such as solar photovoltaic (PV), wind turbine, storage, and diesel generator units, are connected to form an autonomous microgrid (MG) system. However, unlike traditional synchronous generators (SGs), the lack of inertia provision [3] ability of the DGs and the increasing proportion of renewable energy sources (RES) deteriorate the frequency stability of the MGs. Therefore, certain frequency support capability of the RES along with proper control approaches is necessary for maintaining a safe and stable operation of the power system [4].
One of the most promising control methods under development is a virtual synchronous generator (VSG), which is an inertia emulation technology through applying the mechanical equation and electromagnetic equation of the SG to mitigate the transient system dynamics [5]. The implementation usually relies on the assumption that the infinite power can be generated or absorbed by the generator within a short period, whereas the DC-side capacitor is limited in the real world [6]. This problem is solved in [7] with a distributed virtual power system inertia scheme that regulates the DC-link voltages of power converters, where relatively large capacitor units are aggregated to provide frequency support. In addition, a set of the so-called frequency disturbance-based regulators [8], [9], [10] appear, in which the product symbol of the frequency deviation and rate of change of the frequency (RoCoF), is adopted to instruct whether the unit is in the "acceleration" or "deceleration" stage. The principle is to tentatively decide the size of virtual inertia in a gain tuning manner. Following, similar/improved adaptive inertia control methods are proposed, such as that based on dualadaptivity inertia control to improve the overall performance of power and frequency [11] and that in [12] and [13] where the adaptive virtual inertia is developed by referring to the physical meaning of the SG rotor inertia and power angle curve. Nonetheless, all of the above concepts pay only close attention to overall frequency and power improvement, the control cost and energy resources required for such regulation are ignored.
Therefore, researchers begin to employ the optimizationbased method to cope with the parameters setting for VSG. For example, [14] introduces the particle swarm method to tune the control parameters of the VSG units, which can simultaneously realize smooth transition after disturbances and keep the voltage angle deviation within the allowable range. The parameters design rule of the VSG is developed in [15] based on the modal proximity-based method, which can avoid the adverse effects of VSG on the small signal angular stability. The linear quadratic regulator-based optimization technique is employed to achieve the optimal virtual inertia gain for a single-inverter in [16], which is further extended to a uniform multi-machine frequency model in [17]. The aforementioned optimization technique-based methods have shown outstanding advantages in the parameter design of the VSGs on the premise of meeting the stability of the power system. Nevertheless, these studies are established based on small signal modeling method accompanied by linearization processes, although VSG-based MG model is nonlinear built in general. The power system stability under linear control technology can be affected in the case of large disturbances. Based on these analyses, how to adaptively obtain the optimal virtual inertia gain for VSGs, which uses the originally constructed system model, i.e., linearization-free operation, is still an open problem.
When it comes to the optimal control problem of the nonlinear systems, how to solve a nonlinear partial differential equation, which is the so-called Hamilton-Jacobi-Bellman (HJB) equation, is usually challenging. Confront with this challenge, adaptive dynamic programming (ADP) is proposed to solve the HJB equation iteratively by critic-actor techniques and neural networks (NNs) approximation [18]. ADP has successfully solved many control problems, including optimal tracking control [19], robust control [20] and multi-agent consensus control [21]. Furthermore, some improved ADP-based control strategies have been extended to cope with discretetime nonlinear systems [22], [23]. In particular, ADP is applied to the practical problems of wireless connected vehicles [24], power systems [25] and wastewater treatment [26], which shows great practical application prospects [27].
In this paper, an online ADP method is developed to solve the optimal adaptive virtual inertia controller design, which can realize the real-time control of the frequency. Because the equilibrium point of the angular deviation may not be zero when the system is stabilized, this paper propose a cost function by adding a discount factor. Considering that the AC MG is a large-scale system composed of many state variables, it is difficult to design an appropriate activation function of the polynomial NN. Echo state network (ESN) is a novel improved recurrent NN [28]. In ESN, the activation function is called as reservoir, which is generated randomly. Here, a single ESN is designed to implement the ADP method by approximating the optimal cost function and the optimal control input simultaneously, which can obtain better control performance. In other words, the single ESN not only is easy to construct, but also can effectively reduce the computational burden and improve the real-time computation [29]. The main contributions of this paper are listed as follows.
1) A uniform AC MG frequency dynamic model based on the VSG control with a secondary frequency controller is derived, where a nonlinear state space representation is formulated and the reciprocal of the inertia is modeled as the control input.
2) The frequency control problem is solved in an adaptive optimal manner by introducing a cost function, which is defined by using the angular deviation, frequency deviation, RoCoF, and a discount factor. Unlike the conventional VSG control, the proposed approach can adaptively provide the optimal inertia online while retain a tradeoff between the critical frequency bounds and the required control energy. Then, an online ADP method is designed to obtain the optimal frequency controller from the derived HJB equation.
3) The single ESN is constructed to approximate the optimal cost function and the optimal control policy simultaneously, which can reduce the computational burden and improve the real-time computation. The stability of the closed-loop system is analyzed by using the Lyapunov theorem to guarantee that the MG states and the NN output weights are uniformly ultimately bounded (UUB).
The remainder of this paper is organized as follows. The frequency dynamics of the MG are investigated in Section II. Section III presents the optimal frequency controller design approach based on the ADP. Section IV investigates the stability issue of the AC MG with the adaptive algorithm using the Lyapunov theorem. Simulation results are presented to demonstrate the effectiveness of the proposed control scheme in Section V. The conclusions are drawn in Section VI. Fig. 1 shows the generalized structure of a balanced threephase AC MG, where an AC load and an inverter-coupled DG are simultaneously connected with each AC bus.
II. SYSTEM FREQUENCY DYNAMICS
The output of each inverter is determined by the secondary frequency regulation control of the VSG as follows: Secondary Controller (2) where ω i represents the frequency deviation; K Pi and K Ii represent the proportional and integral coefficients of the secondary frequency regulation, respectively; P g,i represents the variation of the mechanical turbine power; R i and T g,i are the droop gain and time constant of the turbine governor dynamics, respectively; J i and D i are the virtual moment of inertia and the active frequency droop damping coefficient, respectively; P l,i is used to represent load disturbance in the network; P e,i represents the change in the electrical power output of the i-th DG, which is given according to the AC power flow calculation as follows [30]: For simplifying (3), the following is adopted
Then, (3) is rewritten as follows
T represents the angle deviation vector. A δ,1 is a N × 1 column vector with all elements equal to 1; A δ,2 is an identity matrix of size N × N; D δ i is a 1 × N column vector with the element in the i-th row equal to 1 and all the remaining elements 0; n is the number of bus nodes in the AC MG; G ij and B ij are the real part and imaginary part of node admittance matrix; superscript 0 is the initial value of the variable. We assume that the internal emf E i is constant because of the action of the excitation system. Combined with (1) and (2), then Substitute (4) into (5), the second-order differential equation of ω i can be described by Considering a constant step change disturbance P l,i and assigning the following vectors T the complete nonlinear model of the AC MG under study is intrinsically represented aṡ where During the frequency regulation, the inverters' inertia parameters are regulated through a state feedback control. The dynamics-based frequency system model is presented in Fig. 2.
Remark 1: Once (6) is established, a respective mathematical expression of the frequency response in the time domain can be obtained by the inverse Laplace transform, where P l,i is considered as a step-change in the active power balance. Take the second derivative of ω i with respect to time, the maximum RoCoF investigated often occurs at t = 0, and is directly determined by P l,i and J i , which indicates that regulating the inertia could have significant improvements on the overall system dynamics.
III. ADP-BASED OPTIMAL FREQUENCY REGULATION
In this section, an adaptive optimal inertia regulation strategy is designed by using the ADP method, where a single ESN are used to implement it. As shown in Fig. 2, the optimal System frequency dynamics model associated with the optimal adaptive frequency controller.
adaptive frequency controller aims to derive the effective control information from the system data. The states of the AC MGs δ(t), ω(t) and ω(t) are used to train the optimal cost function. Then, the optimal inertia J is derived by using the optimal cost function to control the VSGs. In the following, the design and implementation of the optimal frequency control strategy will be presented.
A. ADP-Based Control Design
The main goal of this work is to obtain an optimal control policy u(x) such that the VSG can own an optimal contribution to frequency response, while can also meet the requirements of DC-side energy. To obtain such a balance, a common cost function is defined as where the utility function r(x, u(x)) is chosen as x T Qx+u T Ru.
Here, Q and R are positive definite symmetric matrices with proper dimensions. As it is shown the utility function is constructed of the quadratic function of the angle deviation δ, frequency deviation ω, RoCoF ω and control effort u. However, the equilibrium point of the angular deviation may not be zero when the system is stabilized, which will cause the defined cost function ϒ to fail to converge, and then cannot be used to solve the optimal controller. To solve this problem, a modified cost function with a discount factor is defined where λ > 0 is a discount factor. The term e −λ(τ −t) is used to guarantee that the cost function V(x) converges. According to the Bellman optimality principle, the Hamiltonian function of the problem can be obtained as where ∇V(x) is the partial derivative of the cost function V(x) relative to x, and V(0) = 0. Then, the optimal cost function V * (x) is written as (11) where ( ) is an admissible control set, and the optimal control policy u * (x) that meets the HJB equation is where ( ) is an admissible control set. Suppose (12) holds, then the optimal frequency control policy for benchmark system (7) can be written as Bringing (13) into (12), then the expression of the HJB equation in terms of ∇V * (x) can be further written as where the HJB equation with the optimal control policy u * (x).
B. Single ESN Implementation of the ADP Method
Because of the nonlinear nature of (14), the HJB equation is almost impossible to be directly solved. Thus, the ADP method is adopted to iteratively solve the approximate cost function V * (x) and the optimal control policy u * (x) based on policy evaluation and policy improvement [31]. To implement the ADP method, the polynomial NN is usually adopted to achieve the aforementioned objectives. However, designing and selecting an appropriate activation function for the polynomial NN is always a challenging task, especially when the dimension of the system states is large. In this paper, the system state vector has 3N dimensions, which is hard to design an activation function to achieve a good approximate accuracy by using the polynomial NN. ESN is a promising novel NN with many advancements, which has the property of universal uniform approximate. To further reduce the training burden, the single ESN is used to approximate the cost function and the control policy. Therefore, a continuous time leaky-integrator ESN is defined as followṡ where ξ ∈ R p and ψ(·) denote the reservoir neurons and the input activation function, respectively. W in ∈ R p×n and W ∈ R p×p are the input and internal connection weight matrices, which are generated randomly. ρ > 0 and α > 0 are constant parameters.
ESN is used to approximate the derivative of the cost function (i.e., costate) ∇V(x). It is reasonable to assume that there exist weights W out so that ∇V * (x) can be restructured as where δ c is the ideal approximation error. Then, bring (16) into (14), we have (17) where G = gR −1 g T , δ HJB is the HJB approximation error, which is Substituting (16) into (13), the optimal controller is obtained as where δ a = − 1 2 R −1 g T δ c . It is assumed that the ESN has the echo state property, which is a condition to guarantee the stability and convergence when the ESN is trained. Besides, assuming that the vector [x; ξ ] has a bound of the constant b xξ , the approximation error δ c is bound as δ c < b δc , and the HJB approximation error satisfies the boundness δ HJB < b HJB .
However, the ideal weights W out of the designed ESN are unknown, so the estimated weightsŴ out are first selected. Then, the estimated costate function is expressed as The estimation error of the ideal weights is estimated as Then the estimated controller is as followŝ Bringing (20) and (22) into (10), the approximate Hamiltonian function is derived as Assuming thatV(x) can be obtained from the integral oḟ ) T (f + gû)dτ , and by using (17), (21) and (23), the Hamiltonian error is formulated as Thus, it is reasonable to define E with e to regulate the critic network output weights, which is By minimizing (25) and considering the stability analysis, the tuning law for the critic ESN output weights is designed aṡ where η is the primary learning rate of the output weights. Further, using the estimation error of the ideal weight (21), the dynamics of the estimation output weights error is derived aṡ Therefore, the output weights of the ESN can be obtained by using the updating law (26). Then the optimal frequency control law can be obtained from (22), which can stabilize the frequency of the benchmark system (7) under the disturbance of loads.
IV. STABILITY ANALYSIS OF THE PROPOSED FREQUENCY SUPPORT SCHEME
The stability of the whole closed-loop AC MG system is provided by constructing a complex Lyapunov function of the MG system states and the ESN output weights.
Theorem 1: Consider the complete nonlinear model of the AC MG (7) with the HJB equation (14), the controller is designed by (22), and the adaptive tuning law of the ESN output weights is given as (26). Then, the entire closed-loop AC MG system state x and the output weight estimation errorW out are stable in the sense of uniformly ultimately bounded (UUB).
Proof: A complex Lyapunov function can be constructed as follows where V * (x) is defined in (11), which is constructed of the AC MG system states and satisfies V * (x) ≥ 0, V * (0) = 0. The derivative of (28) iṡ The first term can be derived aṡ According to HJB approximation error (17), we have Substituting (31) into (30), one haṡ where the specific expansions of matrix tr(Ẇ T out λ −1W out ) are given in the Appendix. Combine (32) and (33), (29) becomeṡ According to the frequency dynamic (7), it is reasonable to assume that the following boundedness conditions hold where b f , b g , b xξ and b HJB are positive constants. So (39) Choosing the appropriate parameters of the AC MG model such that > 0, then (37) becomeṡ When the following condition holds we haveL < 0. It means that if L exceeds a certain bound, one hasL < 0. According to the standard Lyapunov extension Theorem, the MG system states and the ESN output weights are all UUB. Therefore, the closed-loop control of the studied AC MG can be guaranteed stable under the designed adaptive controller (22) and the ESN output weights tuning law (26).
V. SIMULATION RESULTS
Simulations based on MATLAB/Simulink are implemented to illustrate the proposed optimal frequency control method. The system parameters are detailed in Table I Fig. 1.
At the beginning, the AC MG is in the steady state, i.e., P l = 0. At t = 2s, the small load changes P l,1 = 1 p.u. and P l,2 = 4 p.u. are performed. Moreover, the large load changes P l,1 = −10 p.u. and P l,2 = −5 p.u. are also For better showing the control performance, six methods of virtual inertia regulator are applied: (a) the no-feedback control with constant inertia (J 0 = 0.1); (b) the non-adaptive control with large constant inertia (J 0 = 2); (c) the LQR-based control; (d) the adaptive control proposed in [11] (J max = 1.2, J min = 0.5, k g = 1e7); (e) the adaptive control proposed in [12] (J 0 = 0.5, k = 4); and (f) the ADP-based adaptive inertia control proposed in this paper.
A. Under Small Load Changes
The LQR control is selected as u = −kx and the quadratic objective is the same with (9), then the control feedback gain can be solved as For the ADP-based control, the predefined parameters are as follows, Q=I 6×6 , R=I 2×2 , λ = 0.01 and k = 2, where I is the identity matrix. The ESN is set as α = 0.01, ρ = 0.5. The input function ψ is selected as identity function. The size of the reservoir is selected as p = 20, so the output weights are a matrix with the dimension of R 6×26 . The input weight, internal weight and initial output weight are generated randomly within a scale of [0, 1]. During the online training, a persistence of excitation (PE) condition is added into the control input to better obtain the system features. After training, the weights of the ESN are obtained, in which only ten dimensions are presented in Fig. 3 for clarity. Fig. 4 presents the evolution of the system states during the online training. Then, by using the converged ESN weights and (22), the optimal frequency controller design can be obtained. 1) Angle, Frequency, and Inertia Tests: Fig. 5 depicts the waveforms of output angular deviation/difference, operating frequency, RoCoF and the emulated inertia of the two DGs. In the first case, with the no-feedback control, both the returning time and deviating time are relatively short, which means that the system has a very fast response. However, it comes at the cost of having a large frequency deviation and RoCoF. In the second case, with the non-adaptive control, the nadir of arrested frequency and RoCoF are higher; but are lower than that of the no-feedback control, with relatively long returning and deviating time and more oscillation. In the third case, with the LQR-based control, whose control effects are similar to the proposed ADP-based control when encountering small load changes, the performance becomes worse than the ADP-based control in the case of large load changes, because it actually belongs to the linear optimal control method. As observed, the system has a shorter returning time and a longer deviating time than that of the non-adaptive control, which can also be well understood by the curves of the emulated inertia (see Fig. 5(d)). At the beginning of the disturbance, relatively large inertia is provided to restrain the frequency drop, while during the "rebound period", relatively small inertia is presented to speed up the steady-state process.
Following, we further compare the proposed method with the existing adaptive methods recently published in [11] and [12]. The results are shown in Fig. 6. Among them, the nadir of frequency and RoCoF with the method in [11] and [12] are similar to that of the ADP-based adaptive virtual inertial control proposed in this paper, whereas the emulated inertia vary greatly. For the method in [11], the virtual inertia is adjusted to the maximum when the nadir of frequency is reached, reduced until the rebound period, and then recovered to the initial until the stable operating point, which increases the risk of system interference. For the method in [12], considering the limitation of the inertia range, that is, it cannot be negative, an excessively large value of the compensation coefficient k cannot be obtained, which hinders the inertia adjustment in other cases. For the strategy proposed in this paper, when the disturbance occurs, the virtual inertia is adjusted to the maximum, then gradually decreases, and finally returns to a stable value, which well improves the frequency, RoCoF and rebound speed.
2) Active Power and Energy Utilization Tests: Figs. 7 and 8 provide the output active power and the energy utilization of the two DGs. As the two DGs have the same control parameters, the active power responses of six methods are similar (see Fig. 7(a) and 8(a)), and the accurate active power sharing is always achieved. Furthermore, it is worth noting that the active power under the non-adaptive control has some oscillations, as depicted in Fig. 7(a). However, after adopting the proposed ADP-based adaptive inertial control, not only the power oscillation can be improved, but also the output power of the two DGs can be smoothed, which indicates that it has a power oscillation damping function. On the other hand, some insightful conclusions about these six control methods can be drawn from the energy utilization, with the respective energy term calculated as E J = J ωdt. As observed in Figs. 7(b) and 8(b), the energy consumption of the non-adaptive method is the largest, followed by the method in [11], the method in [12], the LQR-based and the ADP-based control, which further reflects the role of the developed control strategy in DC-side energy saving.
B. Under Large Load Changes
The case of AC MG under small load changes has been studied in Section V-A. However, another scenario, such as the fluctuation caused by large load integrations, also happens frequently. Here, the effects of such large disturbances are presented. Specifically, the gain k for the LQR-based control is solved by following the same procedure as before as For the ADP-based control, the weights shown in Fig. 9 converge to the optimal values. During the online training, with the effective PE condition, the system states converge as needed, which is shown in Fig. 10. Then, by using the converged ESN weights and (22), the optimal controller design can be obtained.
From Figs. 11 and 12, it is obvious that the ADP-based optimal frequency control scheme performs better than the no-feedback control, non-adaptive control, LQR-based control, method in [11] control and method in [12] control for such large disturbances scenario. Specifically, it can be seen in Fig. 11 that both no-feedback control and the non-adaptive control remain performing frequency regulation poorly under large load changes. Because the LQR-based control method considers linear systems, the occurrence of large load disturbances may lead to poor improvement in frequency regulation compared with the ADP-based control method. The method in [11] and [12] can work, but in the former, the RoCoF even exceeds the acceptable operating range. Meanwhile, the corresponding waveforms of the real output power and the energy utilization under large load changes are also provided in Figs. 13 and 14, respectively. Similar to the results investigated in Section V-A, the proposed ADP-based control has an excellent power oscillation damping function and saves more DC-side energy simultaneously. Large load disturbances may have a great impact on the stability and security of the system. Thus, the adaptability and robustness of the controller is particularly important and necessary. The design of an ADPbased adaptive inertia controller can well cope with various disturbances and be employed in a real-world application.
VI. CONCLUSION
This paper proposed an optimal frequency regulation method for the AC MG-connected VSG, together with a frequency dynamics model of the later. An ADP-based optimal feedback gain was computed to adaptively adjust the emulated inertia according to the predefined cost function. The angular deviation, frequency deviation, RoCoF, and a discount factor are considered in the design of the cost function, while simultaneously preserving a trade-off between the critical frequency limits and the required control energy. The proposal performance was examined based on two different load changes, namely the small load changes and large load changes. Moreover, the proposed control design was compared to the no-feedback, the non-adaptive, the LQR-based, the method in [11], and the method in [12] in terms of the angular deviation/difference, operating frequency, RoCoF, emulated inertia, active power, and, more importantly, energy cost. The advantages of the proposed control include: 1) Lower frequency deviation and RoCoF, which offers the same advantages of the non-adaptive equivalent; 2) Smoother output power, which offers the same advantages of the LQR-based, the method in [11], and the method in [12] equivalent; 3) Lower energy cost, which offers the same advantages of the no-feedback equivalent; On the whole, the proposed method supports frequency dynamics and promotes real-world application.
APPENDIX
The matrix of equation (33). | 6,389.6 | 2023-01-01T00:00:00.000 | [
"Engineering"
] |
Research on the Inner Mechanism and Practical Path of Digital Technology Empowering Rural Cultural Construction under Rural Revitalization
: Rural cultural revitalization is an important part of rural revitalization, and rural cultural construction is an important task for China to continuously promote the construction of socialist spiritual civilization and achieve the goal of the second century. Digital-enabled rural culture construction is a new trend of cultural development in recent years, which provides many advantages for the inheritance and development of excellent rural traditional culture, the innovation and development of digital culture industry and the promotion of rural social civilization, and has important practical significance for promoting rural revitalization. Aiming at the problems of imperfect digital infrastructure, insufficient rural digital literacy, and cultural content failing to meet the needs of the masses and social development, the research group put forward some measures to solve these problems, such as strengthening the construction of rural digital infrastructure, improving the talent support system of rural cultural construction empowered by digital technology, and optimizing the content and communication space of rural cultural products.
Digital Technology Empowers the Practical Significance of Rural Cultural Construction
Rural culture is the spiritual power and ideological guarantee of rural revitalization in the new era, and it is the inevitable requirement to enhance the national cultural soft power and realize cultural power and socialist modernization. Digital technologies, including artificial intelligence, big data and the Internet, provide favorable conditions for improving the level of rural social civilization, improving the level of rural public cultural services and helping the development of rural excellent traditional culture.
Improve the level of social civilization in rural areas
As a big agricultural country, China has a large rural population and diverse rural communities. Therefore, the promotion of rural civilization is related to the promotion of China's overall social civilization. There are many shortcomings in rural cultural construction, for example, the feudal superstitious culture remains, the consciousness of small farmers still exists, and young people tend to western liberal ideology and culture. However, with the gradual integration of digital technology in rural areas, these problems have been improved to some extent. More and more popular science content began to leap into the eyes of rural residents, and more positive energy articles and video content were deeply loved by rural residents. Excellent cultural works on the Internet also made rural residents have a deeper understanding of socialist core values [2]. At the same time, the relevant judicial organs also crack down on illegal and criminal acts hidden in the countryside through various digital means. Residents also supervise each other through digital technology and actively cooperate with the law enforcement actions of relevant departments. For example, the public security department releases videos of popularization of law through short video platforms, actively popularizes legal knowledge and common sense, encourages villagers to know the law, abide by the law and use it, and unblocks online reporting channels. Every villager is a "camera" of rural public security. In addition, the cultural gap between urban and rural areas is gradually being bridged by digital technology, and excellent cultural products in cities are also being transmitted to thousands of rural households through the Internet and digital media, so that rural residents can also enjoy digital dividends and excellent cultural products, which can further boost rural ideological construction and further enhance rural social civilization.
Improve the level of rural public cultural services
As the society pays more attention to economic construction, it pays less attention to cultural construction. Therefore, the development of cultural undertakings in more villages is even less [3]. At the same time, the tension between the cadres and the masses appeared in some villages. At present, the abundant digital technology means can just improve this situation. On the one hand, the application of digital technology is conducive to continuously improving the participation of rural residents, changing the previous phenomenon that the government independently handles the affairs related to rural cultural construction, and concentrating all the strengths and advantages of the government, rural residents and social organizations to carry out cultural construction, participate in cultural governance and share the fruits of cultural development. Through digital technology to promote the informatization of rural management and construction, village cadres, ordinary rural residents, rural cultural talents with special skills, rural netizens who can skillfully use the Internet, and even wanderers who work outside the home, etc., can learn about the development of their hometown culture while making suggestions according to local conditions through the combination of online and offline. On the other hand, the rise of the Internet and various network platforms, as well as their active application in rural life and rural governance, also make more and more highquality cultural products enter the countryside, especially the advanced socialist culture and Marxist theory can be spread more widely through the Internet platform. The relevant official media also provide red classics and red culture for rural residents through the Internet platform and the accurate push of big data. At the same time, digital technology has also brought an increasingly smooth communication platform for the countryside, and rural residents can communicate with rural government staff and cadres through the platform, and even share their experiences. Especially the youth groups, who are better at using digital devices such as mobile phones and computers, prefer to communicate on the Internet, learn more easily through the Internet, get in touch with more outstanding young people, and learn relevant Marxist theories more easily, so as to make up for the shortage of ideological education for rural youth in the past, and help to cultivate rural youth with faith and ideals, so as to better build the countryside.
Help the development of excellent rural culture
Rural cultural construction is inseparable from the development of cultural industry. Proper use of digital technology can not only greatly improve the quality and competitiveness of cultural products, but also continuously promote the integration and development of rural cultural industry and other industries [4]. First, through the Internet, film and television works, games, live broadcasts, short videos and other channels, cultural content with rural characteristics can be productized, personalized, standardized, etc., and can be provided to villagers in a more effective form, thus increasing the effective supply of excellent rural cultural products. At present, many villages have fully demonstrated their excellent culture by skillfully applying various network platforms, such as live broadcast and video contributions, which not only made rural residents know more about their excellent culture, but also attracted many tourists, scholars and other groups, increased their income for rural residents, and also protected and passed on some excellent cultures [5]. Secondly, digital technology empowers the excellent traditional culture with national characteristics, the local culture with historical connotation, the red revolutionary culture with great commemorative and learning significance, and the unique traditional skills of rural residents to promote the development under the interaction and interaction. Especially driven by digital technology, the excellent rural culture can be more conveniently and quickly integrated with the primary and secondary industries, and further developed into products with local characteristics. These cultural products will also be presented to people all over China and even the world through the huge platform of the Internet.
Digital Technology to Empower the Plight of Rural Cultural Construction
In recent years, although the integration of digital technology and various industries has gradually deepened, especially the rural cultural construction has taken to the fast lane through the "east wind" of digital technology, the rural cultural construction empowered by digital technology is still in the development stage. In the process of promoting the rural cultural construction empowered by digital technology, there are still many problems to be solved.
The digital infrastructure is not perfect
By December 2020, the number of Internet users in China has reached 989 million, with 309 million rural Internet users, and the Internet penetration rate in rural areas is 55.90% [6]. There are obvious deficiencies in the construction of digital cultural infrastructure in rural areas, with a big gap between rural areas and cities, uneven distribution of digital resources, most of which are in the initial stage, and there is no corresponding platform for cultural development. It is impossible to realize the integration of digital technology and culture because of the lack of digital equipment. Therefore, it is impossible to protect and inherit the rural excellent culture that is gradually disappearing, and it is impossible to dig deeper into the inheritance and development of many traditional skills. Without the corresponding technical reserves, it is impossible to package, promote and sell highquality cultural resources through big data, artificial intelligence and other means.
Lack of rural digital literacy
Digital technology empowers rural culture construction, which requires a large number of talents with excellent professional skills and high digital professional quality to form a strong talent team. Using digital technology and various platforms to conduct in-depth research on rural culture, we can find a correct way to meet the development of rural culture. It is reported that the digital literacy gap between urban and rural residents is 37.5%, and the digital literacy score of farmers is only 18.6 points, which is significantly lower than that of other occupational groups and 57% lower than the average of all people (43.6 points) [7]. On the one hand, due to the unbalanced development between urban and rural areas, cities with more high-quality resources attract more and more rural youth to leave the countryside, thus reducing the number of rural groups with certain digital literacy. The left-behind people in the countryside are mostly the elderly and minors. The digital literacy of the elderly in rural areas is generally low, while the use of digital devices by minors mostly stays in entertainment. The use of computers, mobile phones and other devices by other people left behind in the countryside only stays in online shopping and entertainment, and it is difficult to use digital platforms and technologies to explore the excellent rural culture and promote the socialist cultural construction in the countryside. On the other hand, many villages can't introduce young talents with excellent digital literacy, and some of them can't work because of lack of funds and corresponding equipment. Some talents can't give full play to their advantages because of the lack of corresponding top-level design in the local area, and their digital literacy is buried.
Cultural content can not meet the needs of the masses and social development needs
Although the integration and development of digital technology can empower the construction of rural culture, the output works and corresponding products cannot be deeply rooted in the hearts of the people. On the one hand, the exploration of local high-quality cultural resources by some products only stays on the surface, and they don't have a correct understanding of their connotation, and they are only published through popular ways on the Internet. Therefore, it can't touch residents' hearts, meet residents' demand for highquality culture, and even achieve the purpose of promoting rural cultural development, ideological education and cultural inheritance. On the other hand, cultural products are mixed with good and bad. Because of the great development of the network platform, many local short video bloggers and live online celebrity have appeared in the countryside. Among these people, there are many "rural online celebrity" who earn traffic through vulgarity and spoofing. The content of its dissemination also deliberately distorts and spoofs the rural production and life and rural culture, which has a bad influence on the rural cultural development. It not only does not conform to the core values of socialism, but also touches the bottom line of the law.
Digital Technology Empowers the Path of Rural Cultural Construction
Digital empowerment of rural culture construction provides a new model and development opportunities for rural culture revitalization, while developing advanced socialist culture can enhance the national cultural soft power [8]. Therefore, it is necessary to explore a feasible path in rural cultural construction, and provide support and guarantee for digital technology to empower rural cultural construction.
Strengthen the construction of rural digital infrastructure
Perfecting rural digital infrastructure is a solid foundation for effectively promoting rural cultural construction. Build broadband communication network, mobile Internet and digital TV network in rural areas, continuously strengthen the implementation of the project of upgrading information about agriculture, rural areas and farmers into villages and households, continuously strengthen the ability to guarantee basic information about agriculture, rural areas and farmers, promote the construction of national, provincial, municipal and county-level media integration centers (platforms), promote the integration of national cable TV networks and the development of 5G integration, collect and sort out cultural heritage data by categories, build a national cultural big data system, and implement the integration and development of publishing. Pay attention to the planning and construction of rural library and rural electronic library facilities, gradually improve the facilities of rural museums and cultural centers, and make precise docking and policy for the cultural needs of rural residents. It is necessary to promote the "internet plus culture" to take root through the Internet platform, continuously introduce high-quality urban culture, and promote diversified exchanges between urban and rural areas.
Improve the talent support system of digital technology to empower rural cultural construction
It is a strong support for that construction of rural culture with talent and the source of power. Building a talent team with high digital literacy is a powerful guarantee to promote rural cultural construction and an inevitable requirement to promote rural talent revitalization. Therefore, we must pay attention to strengthening the excavation of rural digital talents and continuously introduce relevant talents through all available channels. On the one hand, it is necessary to start with basic education, strengthen information education in rural compulsory education, add information education courses, improve students' proficiency in the use of computers and other equipment, and cultivate students' enthusiasm for digital technology and digital technology to empower rural cultural construction. On the other hand, it is necessary to improve the talent introduction mechanism, formulate and introduce corresponding talent incentive policies, and promote the introduction of high-quality talents. Provide corresponding guarantee for talents, meet the material and spiritual needs of talents, and achieve "introduction and retention". In the process of introducing talents, we should adjust measures to local conditions, introduce talents in a targeted way, make good use of talents with emphasis, assign tasks differently according to local cultural characteristics and talent skills characteristics, refine job contents and positions, achieve accurate matching, and improve talent utilization efficiency.
Optimize the content and communication space of rural cultural products
Through the digital platform, we can deeply explore the excellent rural culture, take its essence and discard its dross, and take pictures, live broadcasts and short videos as carriers to show the high-quality rural culture. At the same time, more attention should be paid to the core of cultural products, which should not only fully reflect the local high-quality rural traditional culture, but also fully reflect the core values of socialism and fully demonstrate the beauty of socialist countryside [10]. Moreover, in the process of empowering rural culture construction with digital technology, we must adhere to the guidance of Xi Jinping Thought on Socialism with Chinese Characteristics for a New Era, and properly integrate socialist core values into the construction of rural cultural products, so that rural residents can have a deeper understanding of the connotation of rural culture through these cultural products, learn nutrients from Socialism with Chinese characteristics culture, and actively promote rural cultural construction to help rural revitalization [11]. At the institutional level, local governments should formulate corresponding management measures, conduct grass-roots supervision with a multi-level network supervision system, smooth the channels of supervision and reporting, and prevent the spread of bad content on the Internet [12]. | 3,668 | 2022-12-26T00:00:00.000 | [
"Economics"
] |
Hemoglobin A1c Levels Predicts Acute Kidney Injury after Coronary Artery Bypass Surgery in Non-Diabetic Patients
Introduction Elevated hemoglobin A1c levels in patients with diabetes mellitus have been known as a risk factor for acute kidney injury after coronary artery bypass grafting. However, the relationship between hemoglobin A1c levels in non-diabetics and acute kidney injury is under debate. We aimed to investigate the association of preoperative hemoglobin A1c levels with acute kidney injury in non-diabetic patients undergoing isolated coronary artery bypass grafting. Methods 202 non-diabetic patients with normal renal function (serum creatinine <1.4 mg/dl) who underwent isolated coronary bypass were analyzed. Hemoglobin A1c level was measured at the baseline examination. Patients were separated into two groups according to preoperative Hemoglobin A1c level. Group 1 consisted of patients with preoperative HbA1c levels of < 5.6% and Group 2 consisted of patients with preoperative HbA1c levels of ≥ 5.6%. Acute kidney injury diagnosis was made by comparing baseline and postoperative serum creatinine to determine the presence of predefined significant change based on the Kidney Disease Improving Global Outcomes (KDIGO) definition. Results Acute kidney injury occurred in 19 (10.5%) patients after surgery. The incidence of acute kidney injury was 3.6% in Group 1 and 16.7% in Group 2. Elevated baseline hemoglobin A1c level was found to be associated with acute kidney injury (P=0.0001). None of the patients became hemodialysis dependent. The cut off value for acute kidney injury in our group of patients was 5.75%. Conclusion Our findings suggest that, in non-diabetics, elevated preoperative hemoglobin A1c level may be associated with acute kidney injury in patients undergoing coronary artery bypass grafting. Prospective randomized studies in larger groups are needed to confirm these results.
INTRODUCTION
Coronary artery bypass grafting (CABG) operations are performed safely and successfully in our country as well as in the rest of the world. Acute kidney injury (AKI), not rarely seen following cardiac surgery, is associated with morbidity, increased health costs, and mortality rates [1,2] .
The risk factors and pathophysiology of AKI following CABG were described in the literature and have been the subject of multiple studies [3,4] .
The incidence of AKI following cardiac surgery has been reported as being 5-30% and renal replacement therapy is required in 1-2% of these patients [5,6] . Hemoglobin A1c (HbA1c) is widely used as a marker of average blood glucose concentrations over the preceding 2 to 3 months and it has advantages over glucose tests [7] . Some evidence indicates that high HbA1c levels prior to surgery are strongly associated with the severity of adverse events after CABG [8] .
HbA1c levels were found to be related to cardiovascular and renal complications following open heart surgery [9] . Multiple factors have been implicated as contributors to postoperative AKI, including advanced age, female gender, presence of diabetes mellitus, chronic kidney disease, extended time between heart catheterization and surgery, aortic cross clamp time, duration of cardiopulmonary bypass (CPB), and blood transfusion following surgery [6] .
However, association of elevated HbA1clevels in non-diabetics with AKI after CABG surgery is under debate. The purpose of this study is to investigate the association of preoperative HbA1c levels in non-diabetics with AKI after isolated CABG.
METHODS
In this study, medical records of 315 open cardiac surgery patients operated in the same center by the same surgical team between June 2012 and July 2014 were investigated consecutively and retrospectively. Patients who underwent isolated CABG with CPB and who were non-diabetic with preoperative serum creatinine levels less than 1.4 mg/dl were included in the study. The number of patients that met that criteria was 202. For descriptive purposes, receiver operating characteristic (ROC) curve analysis was performed to identify the cut-off point with the highest sensitivity and specificity. Patients were grouped according to HbA 1c status: < 5.6% (low HbA 1c group; group 1) and ≥5.6% (high HbA 1c group; group 2). AKI diagnosis was made by comparing baseline and postoperative serum creatinine to determine the presence of predefined significant change based on the Kidney Disease Improving Global Outcomes (KDIGO) definition (increase in serum creatinine by ≥0.3 mg/dl within 48 hours of surgery or increase in serum creatinine to ≥1.5 times baseline within 3 days of cardiac surgery) [10] . AKI diagnosis was based on the highest serum creatinine concentration measured during the first 3 days after surgery compared to the baseline serum creatinine concentration, defined as the last concentration measured before surgery. Urine output was not used to define AKI because it may be influenced by diuretics administered during anesthesia and CPB.
Exclusion criteria included patients who had peripheral arterial disease, moderate to severe valvular heart disease, decompensated congestive heart failure, congenital cardiac disease, cerebrovascular event in the last 30 days, malignancy, endocrinological disorders (hypothyroidism, hyperthyroidism), low hemoglobin levels (≤10 g/dl), acute infections, emergency operations; patients who had previous diagnosis of end-stage renal disease and who were on dialysis; patients who were reoperated due to hemodynamic instability or bleeding; patients who required intra-aortic balloon pump; patients who had acute myocardial infarction and percutaneous coronary intervention in the last 30 days prior to operation; and patients who were operated on beating heart or redo CABG. A total of 113 patients were excluded from study, as shown in Figure 1. Additionally, patients for whom data on serum creatinine levels or urine output were missing were also excluded.
Patients' demographic and clinical data were obtained by using the hospital's software system of records and archives to investigate patient files, epicrisis, operation notes, and laboratory results. Age, gender, smoking history, hypertension, hyperlipidemia, left ventricular ejection fraction (LVEF), preoperative and postoperative laboratory parameters (hemoglobin, fasting blood glucose, creatinine, urea, creatinine
(included) non diabetic patients who had coronary artery bypass graft operation
clearance), perioperative data, duration of CPB and aortic cross clamp, amount of blood products used, and intensive care unit (ICU) and hospital length of stay were recorded. Patients were followed in the ICU in the postoperative period, according to protocols of our institution. Electrocardiography, systemic mean arterial pressure, central venous pressure, arterial blood gases, chest tube drainage, and urine output were monitored. Preoperative and postoperative creatinine clearances and peak creatinine clearance were calculated according to the formulas reported in the literature [11,12] .
Operative Technique
All of the patients were operated with median sternotomy under general anesthesia and CPB with aortic and venous cannulations following systemic heparin administration (300 IU/kg). Activated clotting time (ACT) was maintained at over 450 seconds during the operation. Standard CPB circuit and surgical management were used. Antegrade hypothermic and hyperkalemic blood cardioplegia was applied to all patients. Surgery was performed under moderate systemic hypothermia (28-30ºC). CPB flow was maintained at 2.2-2.5 l/min/m 2 , mean perfusion pressure was maintained between 50 and 80 mmHg, hematocrit level was maintained between 20 to 25% during CPB. For the coronary bypass operations, arterial grafts for left anterior descending artery (LAD) revascularization were preferably harvested from the left internal mammary artery (LIMA) whereas saphenous venous grafts were used for the other vessels. Distal anastomoses were done during aortic cross-clamp period and proximal anastomoses were done on beating heart onto the ascending aorta using a lateral clamp.
Statistical Analysis
Statistical analysis was performed using SPSS version 13.0 (SPSS Inc, Chicago, IL, USA). Normal distribution was evaluated by histogram or Kolmogorov-Smirnov test; homogeneity of distribution was evaluated by 'Levene's test for equality of variance' . Normally distributed data were demonstrated as mean ± standard deviation whereas non-normally distributed data were demonstrated as median (minimum-maximum). Difference between groups was evaluated by 'Student's t test' in normal and homogenous distribution and by 'Mann-Whitney U test' in non-normal distribution. Differences between groups were evaluated by parametric or non-parametric Pearson Chi-Square test or Fisher's Exact test with respect to the distribution. Forward stepwise multivariate logistic regression models were created to identify the independent predictors of postoperative AKI. Variables with a P value less than 0.10 in univariate analyses were included in the multivariate model. The sensitivity and specificity of the independent risk factors to predict postoperative AKI were determined by ROC curve analysis. P value less than 0.05 was accepted as significant. Chi-square test was performed for odds ratio. Continuous variables were described as means (standard deviation) or medians (interquartile range), as appropriate; categorical variables were described as percentage.
RESULTS
The demographic characteristics and clinical data of the patients were summarized in Table 1. There were no differences between the two groups in terms of demographic or clinical data.
Preoperative and postoperative blood analysis and haematological parameters of the patients were summarized in Table 2. First postoperative day creatinine levels (P=0.01), 3 rd postoperative day creatinine levels (P=0.0001), and 7 th postoperative day creatinine levels (P=0.0001) were significantly different between the groups.
Intraoperative and postoperative data of the patients are shown in Table 3. ICU length of stay (P=0.004) was significantly different between the groups.
Postoperative AKI occurred in 4 (3.6%) patients in group 1 and in 15 (16.7%) in group 2, showing a statistically significant difference between the groups (P=0.0002). Mortality in the early postoperative period occurred in 2 (1.8%) patients in group 1 and in 6 (6.7%) in group 2, and there was no statistically significant difference between the groups (P=0.14). Renal replacement therapy in the early postoperative period was required in 4 Sensitivity and specificity of preoperative HbA1c levels to predict AKI in non-diabetic patients after CABG was 79% and 59%, respectively. Positive predictive and negative predictive values were 17% and 96%, respectively. Preoperative HbA1c levels higher than 5.6% had an odds ratio of 5.41 for AKI.
Results of univariate and multivariate regression analyses of preoperative risk factors that may influence the development of AKI after CABG in non-diabetic patients are shown in Table 4. In univariate regression analysis, preoperative creatinine (P=0.001), Kocogulları CU, et al .003) was found to be independently associated with an increased risk for AKI. ROC curve analysis of HbA1c level, which was found to be a risk factor for postoperative AKI occurrence in multivariate regression analysis, is depicted in Figure 2. Cut-off value for HbA1c level was determined as 5.75%, at which sensitivity, specificity of the test, and AUC (area under curve) were calculated as 73.7%, 65%, and 0.76 (95% CI=0.62-0.95, P=0.0001), respectively.
Study Limitations
There are some limitations to our study. This study was carried out at a single center, with a limited number of patients, and it was designed as a retrospective study rather than a randomized trial.
DISCUSSION
AKI, not rarely seen following cardiac surgery, prolongs ICU and hospital length of stay and results in increased health costs and mortality rates [1,2] .
AKI following cardiac surgery is a multifactorial state. Risk factors are advanced age, presence of diabetes mellitus, hypertension, low preoperative glomerular filtration rate (GFR) (<60 ml/min/m 2 ), left ventricular systolic dysfunction (LVEF<35%), preexisting kidney dysfunction, atherosclerosis of the ascending aorta, urgent or emergent surgery following myocardial infarction or percutaneous cardiac intervention, and administration of nephrotoxic agents [13,14] . Intraoperative factors also contribute to the development of AKI during cardiac surgery, such as renal hypoperfusion, nonpulsatile flow, and systemic inflammatory response syndrome due to CPB [15,16] . Demographic data and risk factors for AKI such as hypertension, low ejection fraction and EuroSCORE values were similar in both groups.
Long term survival of patients operated for cardiac surgery is directly proportional to the severity of AKI, which is related to changes in serum creatinine levels [17] . Our results showed that patients with higher preoperative HbA1c levels had higher creatinine levels at 1 st , 3 rd and 7 th postoperative days. Nevertheless, there was no difference in mortality. Instead, there was a significant difference when associated with prolonged ICU stay. AKI following CPB is an important cause of morbidity and mortality [18] . Postoperative AKI requiring renal replacement has an independent effect on morbidity and early mortality. It is reported that the overall mortality due to AKI is 40-80% [11] . In the recent literature, there are several studies regarding early diagnosis of AKI and prevention of the inflammation process that is an accepted cause of AKI [19,20] . In a study by Freeland et al. [6] , blood transfusion was found as an independent risk factor for development of AKI following cardiac surgery. The same study also mentioned that longer aortic cross clamp and CPB times increased the incidence of AKI following cardiac surgery [6] . In a study by Khilji et al. [21] , both CPB and total cross-clamp times have been known as potential risk factors for developing kidney injury. In contrast with the literature, we did not find any significant differences between patients with AKI and without AKI regarding CPB and aortic cross clamp times and usage of blood products.
High mortality and morbidity rates following CABG operations have been reported in several studies in patients with type 2 diabetes mellitus (DM) [22] . In addition, some studies have shown that type 2 DM increases postoperative AKI after CABG [23] . HbA1c level is a parameter used to evaluate long term glycemic control in patients with DM [22] . The American Diabetes Association included HbA1c level in the criteria for diagnosing DM [24] . Normal HbA 1c levels are accepted as 4-6%. Tekumit et al. [25] found that the borderline level of HbA1c was 6.1% for patients undergoing CABG. In their retrospective study, Hudson et al. [26] reported that preoperative HbA 1c levels over 6% were associated with 30-days postoperative mortality and occurrence of AKI in patients without DM who underwent open cardiac surgery.
In our study, patients were grouped according to HbA1c levels, with borderline level being described as 5.6%. Patients with levels higher than 5.6% had significantly higher incidence of AKI, according to KGIDO classification. Our results revealed lower levels of HbA 1c than other studies as a risk factor for AKI [27,28] .
According to the KDIGO 2012 AKI Guideline, cardiac surgery with CPB is a 1B risk factor [28] . Despite the lack of consensus on AKI and Hba1c levels in patients with no known renal disease, HbA 1c over 7% is defined as a Class 1A risk factor for patients with chronic renal disease [28] . The cut-off value for AKI in our group of patients was 5.75%.
Azevedo et al. [29] observed that, in critical illness, there was a significant correlation between blood glucose levels and the incidence of AKI. Halkos et al. [30] found that HbA 1c levels greater than 7% were associated with renal failure. Additionally, Gumus et al. [9] found that elevated levels of HbA 1c were associated with increased renal complications. Likewise, in our study, a relationship was found between high preoperative creatinine, BUN, HbA1c levels and occurrence of postoperative AKI in our study. It was also observed that average HbA1c level in the preoperative period is a predictor of AKI in the early postoperative period following CABG.
CONCLUSION
AKI following cardiac surgery causes multiple postoperative complications and leads to prolonged hospitalization, increased Kocogulları CU, Levels and Acute Kidney Injury costs, and eventually increased mortality rates. Our results suggest that elevated preoperative HbA1c level is associated with postoperative AKI and prolonged ICU stay in non-diabetic patients undergoing CABG. However, further prospective randomized studies are warranted to confirm these results.
Authors' roles & responsibilities
CUK ATK RA CD HP HS OF Conception and study design; execution of operations; analysis and/or data interpretation; statistical analysis; manuscript writing or critical review of its content; final manuscript approval Conception and study design; execution of operations; analysis and/or data interpretation; statistical analysis; manuscript writing or critical review of its content; final manuscript approval Conception and study design; execution of operations; analysis and/or data interpretation; statistical analysis; manuscript writing or critical review of its content; final manuscript approval Conception and study design; execution of operations; analysis and/or data interpretation; statistical analysis; manuscript writing or critical review of its content; final manuscript approval Conception and study design; execution of operations; analysis and/or data interpretation; statistical analysis; manuscript writing or critical review of its content; final manuscript approval Conception and study design; execution of operations; analysis and/or data interpretation; statistical analysis; manuscript writing or critical review of its content; final manuscript approval Conception and study design; execution of operations; analysis and/or data interpretation; statistical analysis; manuscript writing or critical review of its content; final manuscript approval | 3,789.8 | 2017-03-01T00:00:00.000 | [
"Medicine",
"Biology"
] |
Evolution of N/O ratios in galaxies from cosmological hydrodynamical simulations
We study the redshift evolution of the gas-phase O/H and N/O abundances, both (i) for individual ISM regions within single spatially-resolved galaxies and (ii) when dealing with average abundances in the whole ISM of many unresolved galaxies. We make use of a cosmological hydrodynamical simulation including detailed chemical enrichment, which properly takes into account the variety of different stellar nucleosynthetic sources of O and N in galaxies. We identify $33$ galaxies in the simulation, lying within dark matter halos with virial mass in the range $10^{11}\le M_{\text{DM}} \le 10^{13}\,\text{M}_{\odot}$ and reconstruct how they evolved with redshift. For the local and global measurements, the observed increasing trend of N/O at high O/H can be explained, respectively, (i) as the consequence of metallicity gradients which have settled in the galaxy interstellar medium, where the innermost galactic regions have the highest O/H abundances and the highest N/O ratios, and (ii) as the consequence of an underlying average mass-metallicity relation that galaxies obey as they evolve across cosmic epochs, where -- at any redshift -- less massive galaxies have lower average O/H and N/O ratios than the more massive ones. We do not find a strong dependence on the environment. For both local and global relations, the predicted N/O--O/H relation is due to the mostly secondary origin of N in stars. We also predict that the O/H and N/O gradients in the galaxy interstellar medium gradually flatten as functions of redshift, with the average N/O ratios being strictly coupled with the galaxy star formation history. Because N production strongly depends on O abundances, we obtain a universal relation for the N/O--O/H abundance diagram whether we consider average abundances of many unresolved galaxies put together or many abundance measurements within a single spatially-resolved galaxy.
INTRODUCTION
Elemental abundances are widely used in astrophysics to constrain the star formation history (SFH) of galaxies (e.g., Kobayashi 2016). An example of a SFH chemical abundance diagnostic is given by [α/Fe] 1 ; from the observed [α/Fe]-[Fe/H] relations, chemical evolution models have demonstrated that the various constituents of our Galaxy (halo, bulge, thick and thin disc) formed on different typical time scales (see, for example, Chiappini et al. 1997;Grieco et al. 2012;Brusadin et al. 2013;Micali et al. 2013;Spitoni et al. 2016;Grisoni et al. 2017); furthermore, by making use of the [α/Fe] ratio estimated from spectral indices, chemical<EMAIL_ADDRESS>†<EMAIL_ADDRESS>1 By α-elements we usually mean O, Mg, Ne, Si, S, Ca. The square bracket notation for the stellar chemical abundances is defined as follows: [X/Y ] = log(N X /N Y ) − log(N X /N Y ) , where N X and N Y represent the number density of the chemical elements X and Y , respectively. lution models have depicted early-type elliptical galaxies as forming from a short and intense burst of star formation in the past, in agreement with observations (Matteucci 1994;Thomas et al. 2003;Pipino & Matteucci 2004;Taylor & Kobayashi 2015a,b;Kriek et al. 2016;De Masi et al. 2018). The observed [α/Fe]-[Fe/H] diagram can be effectively used as a SFH diagnostic (i) firstly, because αelements and Fe are mostly released on different typical time scales by core-collapse and Type Ia Supernovae (SNe), respectively, and (ii) secondly, because the nucleosynthesis of α-elements in stars is not correlated with their Fe abundance (Kobayashi et al. 2006).
In the star forming disc galaxies, however, it is not possible to measure iron abundances. For this reason, the O/H elemental abundance as well as the N/O 2 abundance ratio are among the most measured metallicity proxies in the interstellar medium (ISM). Current Galactic and extragalactic spectroscopic surveys such as MaNGa 2 For brevity, we use the following notation: N/O ≡ log(N/O)gas and O/H ≡ log(O/H)gas + 12, for the gas-phase chemical abundances. c 2018 The Authors arXiv:1801.07911v2 [astro-ph.GA] 23 Apr 2018 are capable of reaching resolutions which were unimaginable only few decades ago. Large amounts of observational data are constantly being released, challenging theorists to develop models which can explain at the same time the variety of different physical observables nowadays available. One of the most important pieces of information we can extract from extragalactic spectroscopic surveys is the N/O-O/H diagram.
Historically, the N and O abundances have been measured for individual targets (e.g. HII regions or star forming regions) within a number of nearby galaxies (e.g., Garnett 1990;Vila Costas & Edmunds 1993;Izotov & Thuan 1999;Pilyugin et al. 2010;Berg et al. 2016;Magrini et al. 2017); then large-scale spectroscopic surveys have improved the statistics considerably, where chemical abundances have been measured from the integrated galaxy spectra (e.g. Andrews & Martini 2013, and references therein); finally, thanks to multi-object spectrographs and integral field unit (IFU) surveys, now it is possible to resolve the abundance patterns in star forming regions within a large number of galaxies (Pérez-Montero et al. 2016;Sánchez-Menguiano et al. 2016;Belfiore et al. 2017a). This growing amount of observational data has suggested the use of the N/O-O/H relation as an alternative SFH chemical abundance diagnostic of galaxies (Chiappini et al. 2005;Mollá et al. 2006;Vincenzo et al. 2016a). In near future, it will be possible to obtain these elemental abundances in high-redshift galaxies with JWST/NIRSpec and study the redshift evolution of the N/O-O/H relation.
Before studying the redshift evolution, it is important to understand the origin of the observed N/O-O/H relations in the local Universe. The observed relations have been obtained with O/H and N/O abundance measurements both (i) as global average values, measured from the galaxy integrated spectra and hence representative of unresolved galaxies, and (ii) as local abundance measurements in resolved HII or star forming regions within single, spatially-resolved external galaxies; these two cases are conceptually different with respect to each other and may give rise -in principle -to different N/O-O/H relations.
All the chemical elements with atomic number A 12 in the cosmos are synthesised in stellar interiors either during the quiescent phases of hydrostatic burning or through explosive nucleosynthesis during SN explosions (Arnett 1996). If a theoretical model is to make predictions about the chemical abundances coming from the analysis of stellar spectra, the chemical enrichment feedback from star formation activity must be properly included in the theoretical machinery by taking into account the variety of different nucleosynthesis sources which can actually produce a given chemical element X; the different distributions of delay times between the formation of each astrophysical source and its death must also be taken into account (see Pagel 2009 for exhaustive reviews on the subject).
Detailed chemical evolution of galaxies have mostly been studied by making use of one-zone models (e.g. Henry et al. 2000;Chiappini et al. 2005;Vincenzo et al. 2016a, but see also Vangioni et al. 2017, where a one-zone model has been embedded in a cosmological framework), which are based on the so-called instantaneous mixing approximation. However, in a real galaxy, chemical enrichment is inhomogeneous, which is important if we want to constrain the SFH from X/Y abundance ratio diagrams (Kobayashi & Nakasato 2011). Cosmological chemodynamical simulations are nowadays the best tools to shed light on how the SFH took place in different galaxies. These simulations are also key to understanding how chemical elements are synthesised, released and later distributed within galaxies, because they are able to address the large amounts of data which are already available or about to come. In fact, cosmological chemodynamical simulations can provide a unifying picture for the formation and evolution of the many different populations of galaxies in the Universe (see, for example, Maio & Tescari 2015).
An advantage of using chemodynamical simulations is that one can predict both local and global relations for a large sample of simulated galaxies; another advantage is that one can naturally have chemical abundance gradients as functions of the galactocentric distance within the ISM of the simulated galaxies. By using chemical abundance measurements from the Cepheids (Andrievsky et al. 2002;Luck et al. 2003;Luck & Lambert 2011;Korotin et al. 2014;Genovali et al. 2015), planetary nebulae (Maciel & Koppen 1994;Costa et al. 2004;Stanghellini et al. 2006;Gutenkunst et al. 2008) or HII regions (Deharveng et al. 2000;Esteban et al. 2005;Rudolph et al. 2006;Fernández-Martín et al. 2017;Esteban et al. 2017) many observational works have shown, for example, that O/H in our Galaxy steadily diminishes when moving outwards as a function of the galactocentric distance, but radial gradients have been observed by those works also for other chemical elements; furthermore, Belfiore et al. (2017a) have shown that the N/O ratios can vary as functions of both the galactocentric distance and stellar mass, when considering a large sample of nearby galaxies in the MaNGa survey.
Historically, multi-zone chemical evolution models have been constructed to reproduce the observed radial metallicity gradients in the Galactic disc by assuming the so-called "inside-out scenario", according to which the innermost (most metal-rich) Galactic regions assembled on much shorter typical timescales than the outermost (most metal-poor) ones, namely by assuming that the Galaxy formed from the inside out (see, for example, Chiappini et al. 2001;Cescutti et al. 2007;Magrini et al. 2009;Spitoni & Matteucci 2011). Chemical evolution models with inside-out growth of the disc and the star formation efficiency being modulated by the angular velocity of the gas predict a flattening of the radial metallicity gradients as a function of time (Portinari & Chiosi 1999;Boissier & Prantzos 2000); also chemodynamical simulations usually predict a flattening of the radial metallicity gradients as a function of time (see Kobayashi & Nakasato 2011;Pilkington et al. 2012;Gibson et al. 2013 and references therein). Finally, there are chemical evolution models predicting an inversion of the radial metallicity gradients at high redshift, corroborated by some observational findings (see Cresci et al. 2010;Werk et al. 2010;Queyrel et al. 2012;Mott et al. 2013, but also Schönrich & McMillan 2017 for a critical discussion).
In this work, we show the results of our cosmological chemodynamical simulation including the latest stellar nucleosynthesis yields; we investigate both (i) local and (ii) global N/O-O/H relations, i.e. (i) the relations obtained of individual targets within single spatially-resolved galaxies and (ii) the relations obtained with average abundances for the whole ISM of many unresolved galaxies put together. If the predicted relations follow a similar trend in the N/O-O/H diagram, we try to understand the causes of this in galaxies. Moreover, we show our predictions for the redshift evolution of the O/H and N/O radial gradients of a sample of galaxies in our cosmological simulation; finally, we show how the simulated galaxies move in the N/O-O/H, stellar mass-O/H and stellar mass-N/O diagrams as they evolve across cosmic epochs, fully exploiting the predictive power of a cosmological hydrodynamical simulation. We would like to note again that only by making use of chemodynamical simulations can we study both local and global evolution of chemical abundances, and that cosmological simulations allow us to study the effect of environment on the chemical evolution of galaxies as well.
Our work is organised as follows. In Section 2 we summarise the main assumptions of our model and the analysis method of the simulation. In Section 3 we present the results of our study. We first discuss the origin of the local and global N/O-O/H relations for nearby galaxies, and then show the redshift evolution and the environmental dependence in Section 3.4. Finally, in Section 4 we draw our conclusions.
SIMULATION MODEL AND METHODS
Our simulation code is based on the GADGET-3 code (Springel 2005) and relevant baryon physics is included, namely UV background heating, metal-dependent radiative cooling, star formation, thermal stellar feedback, and chemical enrichment from asymptotic giant branch (AGB) stars, core-collapse and Type Ia supernovae (SNe). Therefore, the star formation activity within the ISM of galaxies is affected both by the thermal energetic feedback and by the chemical enrichment of star particles through stellar winds and SN explosions (see Kobayashi 2004;Kobayashi et al. 2007;Kobayashi & Nakasato 2011;Taylor & Kobayashi 2014 for a detailed description of the model).
In summary, we evolve a cubic volume of the standard Λcold dark matter Universe with side 10 Mpc h −1 , periodic boundary conditions, and the cosmological parameters being given by the nine-year Wilkinson Microwave Anisotropy Probe (Hinshaw et al. 2013); Ω0 = 0.28, ΩΛ = 0.72, Ω b = 0.046, H0 = 100 × h = 70 km −1 Mpc, and σ8 = 0.82. The mass resolution of our simulation is MDM ≈ 3.097 × 10 7 h −1 M for the dark matter (DM) component and Mgas = 6.09 × 10 6 h −1 M for the gas fluid. Finally, in our simulation we assume a gravitational softening length gas ≈ 0.84 h −1 kpc, in comoving units.
The initial conditions of our simulation are the same as in Kobayashi et al. (2007), but with updated cosmological parameters and better resolution; in particular, we assume initial conditions giving rise to a standard field at redshift z = 0, with no strong central concentration of galaxies. Our initial conditions are different from those in Taylor & Kobayashi (2014).
Chemical enrichment model
According to their mass and metallicity, stars at their death pollute the ISM of galaxies with different fractions of a given chemical element. We cannot resolve single stars in our simulation, hence we assume that each star particle represents a simple stellar population (SSP) with fixed age and chemical composition. Then we assume that all the embedded stars within each single SSP have a universal mass-spectrum at their birth which follows the Kroupa (2008) initial mass function (IMF), as defined in the stellar mass range 0.01 m 120 M . As each given SSP gets older and older as a function of cosmic time, embedded stars with lower and lower mass enrich the surrounding gas particles with their nucleosynthetic products; the number of dying stars within a given SSP at the time t is given by the assumed IMF and SSP mass, while the enrichment time of a star with mass m and metallicity Z is given by the assumed stellar lifetimes, τ (m, Z); in this work, we assume the stellar lifetimes of Kobayashi (2004), which are both metallicity-and mass-dependent.
In our simulation, the stellar nucleosynthetic yields are the same as in , which include the chemical enrichment of AGB stars and SN explosions. The effect of hypernovae is included in our simulation for stars with mass m 25 M with the following metallicity-dependent hypernova fraction: HN = 0.5, 0.5, 0.4, 0.01, and 0.01 for Z = 0, 0.001, 0.004, 0.02, and 0.05, respectively, which is necessary to match the observed elemental abundances in the Milky Way (Kobayashi & Nakasato 2011). We additionally assume that all stars with mass m 25 M and metallicity Z 0.02 which are not hypernovae end up their lives as failed SNe (Smartt 2009;Müller et al. 2016) and pollute the galaxy ISM only with H, He, C, N and F, which are synthesised in the outermost shells of the SN ejecta; the other chemical elements (including O) are assumed to fall back into the black hole, hence they are not expelled by the star into the surrounding ISM (see also Vincenzo & Kobayashi 2018;Kobayashi et al., in prep.).
We assume that each galaxy SSP distributes thermal energy and stellar nucleosynthetic products to its closest 576 neighbour gas particles (with the smoothing kernel weighting). This value, together with the other parameters specified above, is chosen to match the observed cosmic star formation rate (SFR; Hopkins & Beacom 2006;Madau & Dickinson 2014). Figure 1 shows the predicted cosmic SFR history of our simulation with failed SNe (blue solid line), as compared to the same simulation but with the original yields from Kobayashi et al. (2011, green long dashed line) without failed SNe. There is no significant difference in the cosmic SFRs and in the basic properties of the galaxies such as mass and morphology. The red dotted line in Figure 1 shows the predictions for the cosmic SFR of a similar simulation but with lower resolution (2 × 96 3 particles) than in this work (2 × 128 3 particles); an agreement between the two can be only found by assuming a different number of feedback neighbour particles, NFB. In particular, by increasing the resolution and keeping NFB constant, ISM regions with higher and higher densities can be resolved and the SN feedback affects smaller regions around each given star particle; therefore, to obtain similar results in simulations with higher resolution, NFB should be increased accordingly (Kobayashi et al. 2007).
Although our resolution is good enough to study radial gradients of chemical abundances in galaxies, it is not possible to resolve the small-scale physics within star-forming clouds and SN ejecta in galaxy simulations; for this reason, chemical enrichment is included by computing the contribution from each single star particle, depending on the metallicity. Therefore the chemical feedback can vary as a function of time and location within the galaxy (Kobayashi 2004;Kobayashi & Nakasato 2011). We remark on the fact that the evolution of the elemental abundance ratios in the ISM is mainly driven by the difference in the age and metallicity of the enrichment sources, being less affected by the uncertainty in the ISM metallicity due to the limited resolution.
Since in this paper we focus on the evolution of the N/O ratio in galaxies, we briefly recall here how O and N are synthesised by stars in galaxies (see also Vincenzo et al. 2016a for more details). First of all, both N and O can be produced by massive stars, with mass m > 8 M , dying as corecollapse SNe on short typical time scales after the star formation event ( 30 Myr); in this case, stellar evolutionary models predict N to be mainly produced as a secondary element in massive stars, in the CNO cycle at the expense of C and O nuclei already present in the gas mixture at the stellar birth. One-zone chemical evolution models showed that on its own, the "secondary" N component from massive stars is not sufficient to reproduce the observed N/O plateau in our Galaxy at very low metallicity (Matteucci 1986;Chiappini et al. 2005Chiappini et al. , 2008. Therefore, following the original suggestion of Matteucci (1986), many one-zone models assumed an additional primary N production by massive stars to reproduce the observed N/O plateau at very low metallicity (Pettini et al. 2002(Pettini et al. , 2008Spite et al. 2005;Pilyugin et al. 2010), which is however highly scattered. In our simulation, we do not assume any additional primary N production for massive stars.
Low-and intermediate-mass (LIM) stars, with mass in the range 4 m 8M , are dominant stellar nucleosynthesis sources of N, when experiencing the AGB phase (see, for example, Ventura et al. 2013, 2017 for more details). Most of the nitrogen from AGB stars is secondary and its stellar yields steadily increase as functions of the initial stellar metallicity. Note that, however, there may be also a primary N component that can be important in the chemical evolution of galaxies at very low metallicity, which is predicted when hot-bottom burning occurs in conjunction with the so-called third dredge-up (see also Vincenzo et al. 2016a and reference therein).
Analysis of the simulation
From our cosmological simulation, we create a catalogue of 33 stellar systems at redshift z = 0, all embedded within dark matter (DM) halos with virial mass in the range 10 11 MDM 10 13 M ; we make use of the ROCKSTAR friend-of-friends groupfinding algorithm with adaptive hierarchical refinement to determine all the DM halos in the simulation (Behroozi et al. 2013). The 33 stellar systems of our catalogue span a variety of different star formation histories (SFHs) and consequently show different chem- Figure 2. Our ten reference galaxies when viewed edge-on. Black points correspond to the older stellar populations (> 90 per cent in the cumulative age distribution function); blue points to younger star particles (< 10 per cent in the cumulative age distribution function); finally, red points correspond to intermediate-age stellar populations which lie between 10 and 90 per cent in the cumulative age distribution function. ical evolution histories from their formation epoch. For each stellar system at redshift z = 0 in our catalogue, we retrieve the main physical and chemical properties of all its star and gas particles going back in redshift, by means of a simple searching algorithm (each particle in the simulation is univocally characterised by an ID number). At all redshifts, each galaxy in our catalogue is defined as follows.
(i) At any given time t1 in the past, we fit with Gaussian functions the normalised density-weighted distributions of the x, y and z coordinates of all the gas particles within the galaxy, which have been retrieved from the simulation snapshot at a time shortly after t2 = t1 + ∆t; then we consider in our analysis all the star and gas particles at the time t1 in the simulation which lie within 4σ from the centre of the three Gaussians. Therefore, in the presence of merger events, we choose to follow the stellar system with the highest gas densities.
(ii) If the fitting procedure fails at a given redshift (usually corresponding to high velocity encounters or minor/major mergers), we broaden our criteria and consider at that redshift all the gas and star particles which lie within 20 kpc from the centre of mass of the star particles which have been retrieved from the subsequent simulation time step.
By following the analysis as described above, we can study the evolution of the galaxy physical properties continuously as functions of redshift with an automated algorithm. The small fluctuations in the predicted evolution of the average galaxy properties are mostly due to an imperfect centring on the galaxy main body, particularly associated with merging episodes with other stellar systems.
RESULTS
In this Section, we present our new results from the analysis of the cosmological hydrodynamical simulation described in Section 2. In this work, we select ten representative star forming disc galaxies from our catalogue, so that these reference galaxies have a range of characteristic SFHs. All of our ten reference galaxies lie within DM halos with virial masses in the range 10 11 MDM 10 12 M . In the first part of this Section, we show our predictions for the gasphase O/H-N/O abundance patterns in our reference galaxies. In the second part, we show how the average O/H and N/O abundance ratios evolve with time when considering our entire sample of 33 galaxies.
3.1 Star formation history of the reference disc galaxies Our ten reference galaxies are shown in Figure 2, as viewed edgeon. Different colours in Figure 2 correspond to galaxy stellar populations with different age; in particular, the black points correspond to the older galaxy stellar populations (> 90 per cent in the cumulative age distribution function); blue points to younger star particles (< 10 per cent in the cumulative age distribution function); finally, red points correspond to intermediate-age stellar populations which lie between 10 and 90 per cent in the cumulative age distribution function.
In Figure 3, we show the distribution of the total stellar mass of our ten reference galaxies at redshift z = 0 as a function of the stellar age. When passing from Galaxy 0 to Galaxy 9, the stellar mass growth history becomes more concentrated towards later and later epochs. While most of the galaxies have smooth SFHs, Galaxy 1 undergoes a major merger of two stellar systems with average Figure 3. The stellar mass distribution function of our ten reference galaxies as functions of the stellar birth time. The area under the predicted distribution corresponds to the total galaxy stellar mass at redshift z = 0. From top to bottom, our ten reference galaxies have decreasing average stellar ages. The figure can be read as the fraction of the present-day galaxy stellar mass coming from each given age bin in the past. In the upper corner on the left we report for each panel the total galaxy stellar mass at z = 0. ages peaking at ≈ 9 and 7 Gyr ago. After the onset of star formation at z ∼ 6, Galaxies 2 and 3 had a relatively rapid increase of SFR, while the other galaxies had a slow start with very low SFR at z < 4. Galaxy 9 is the youngest and maintains a high SFR at the present epoch.
O/H-N/O relations within single resolved galaxies
In Figure 4, the predicted gas-phase log(O/H)-log(N/O) relation of individual ISM regions (blue points) in our ten reference galaxies is compared to the observations in the local Universe (Dopita et al. 2016;Belfiore et al. 2017a). We remark on the fact that the majority of the gas particles in our ten reference galaxies lie on a thin with the corresponding 1σ deviation, as derived by dividing our simulated galaxies in many concentric annuli and measuring the average gas-phase abundances within each annulus (in this case we only consider gas particles which lie within three times the galaxy half-mass radius, as computed from the stellar mass radial profile); blue points at higher O/H abundances lie in the inner galactic regions. We compare our simulation with the ob- of our ten reference galaxies, from top to bottom. The black points represent the older stellar populations (> 90 per cent in the cumulative age distribution function) in our reference galaxies; blue points to younger star particles (< 10 per cent in the cumulative age distribution function); finally, red points correspond to intermediate-age stellar populations which lie between 10 and 90 per cent in the cumulative age distribution function. The red dashed line indicates the suggested average lowmetallicity N/O-plateau at log(N/O) ≈ −1.5 dex, as predicted by chemical evolution models with pure primary N production by massive stars. disc at the present time. The black points with the error bars represent the average O/H and N/O abundances with the corresponding 1σ standard deviation as predicted when dividing the galaxy in different concentric annuli of galactocentric distance and computing the average gas-phase O/H and N/O ratios therein; in this case, we only consider gas particles which lie within three times the galaxy half-mass radius, as computed from the stellar mass radial profile.
The predicted N/O-O/H relation within all our reference galaxies qualitatively agrees with the observed N/O and O/H abundance measurements of Belfiore et al. (2017a, pink symbols) from a large sample of spatially-resolved galaxies in the MaNGa survey. Our simulation also nicely follows the observed relation as derived by Dopita et al. (2016, grey line), obtained by compiling data from Milky Way metal-poor stars (Israelian et al. 2004;Spite et al. 2005), resolved HII regions in blue compact galaxies (Izotov & Thuan 1999) and local B stars (Nieva & Przybilla 2012). At low metallicity, although we do not have many points, our predictions give slightly larger N/O, which may be due to the difference in the targets. We note that the observed slopes of N/O vs. O/H of Dopita et al. (2016) and Belfiore et al. (2017a) differ from each other because of well-known uncertainties in the abundance measurements, mostly due to the assumed calibration (see, for example, Kewley & Ellison 2008, but also the discussion in Belfiore et al. 2017a).
In At redshift z = 0, we have only a few low-metallicity components in the gas-phase of the galaxy ISM. However, these are predicted to be common at higher redshift, at the earliest evolutionary stages of the galaxies (see Vincenzo & Kobayashi 2018). At redshift z = 0, the low-metallicity components can be seen more clearly in the oldest galaxy stellar populations; this is shown in Figure 5, where the different galaxy stellar populations are discriminated with different colours in the N/O-O/H diagram according to their formation time, by using the same criteria as in Figure 2.
At low metallicity, the majority of the older stellar population (black dots) show a flat trend as a function of metallicity. We recall here that a flat trend of N/O in our chemodynamical model is caused by inhomogeneous chemical enrichment, where a significant contribution of AGB stars appears at low O/H. Depending on the relative contribution between SNe and AGB stars, the exact value of the N/O ratios in the plateau may vary from galaxy to galaxy according to the galaxy formation time and SFH. At z = 0, this effect of inhomogeneous chemical enrichment is more important at the outskirts of our simulated galaxies because of the low SFRs (see Vincenzo & Kobayashi 2018). On the one hand, massive galaxies that have relatively fast star formation also show very low N/O ratios, below the plateau, for the oldest and metal-poor stellar populations; these values are roughly in agreement with the observations in damped Lyα (DLA) systems (Pettini et al. 2002(Pettini et al. , 2008, which are, however, also highly scattered (Zafar et al. 2014;Vangioni et al. 2017). On the other hand, the oldest and most metal-poor stellar populations in the less massive galaxies have log(N/O) ∼ −1.5 dex, on average. Figure 6. The predicted redshift evolution of the O/H (left panels) and N/O (right panels) gradients in the ISM of a sub-sample of four galaxies, as chosen from our ten reference galaxies.
In Figure 6, we show how the gas-phase log(O/H) and log(N/O) ratios vary as functions of the galactocentric distance and time within a sub-sample of our ten reference galaxies. Different colours correspond to different redshifts when the gradients are computed. Although there is a dispersion in the chemical abundances at any fixed galactocentric distance (particularly for the O/H abundances; see also the horizontal error bars in Figure 4), after the first star formation episode in the galaxy, we predict a flattening of the abundance gradients as a function of time, together with an inside-out growth of the galaxy disc. In particular, in the very early "protogalactic" evolutionary stages, we predict highly scattered and overall flat abundance gradients; then, as the first series of stellar populations form and the galaxy contextually accretes gas from the environment, steep gas-phase abundance gradients develop, which then gradually flatten as a function of time. The predicted flattening of the abundance gradients with time is in agreement with the predictions of previous chemodynamical simulations (e.g., Pilkington et al. 2012). Finally, the average O/H abundances at the centre do not show a strong redshift evolution, which is consistent with the observations in our Galactic bulge (Zoccali et al. 2008).
To explain our prediction of negative radial N/O gradients at redshift z = 0, we recall here that the main N-producers in galaxies are not low-mass stars (see also Section 2.1), with chemical composition reflecting the chemical abundances in the ISM, from which they were born quite recently. The fact that we predict negative radial gas-phase O/H gradients at z = 0 makes the N-producers more metal-rich inside than outside. Since N is mainly synthesised as a secondary element (namely, the N stellar yields increase, on aver- age, as functions of the stellar metallicity), there is a correspondent increase of the average gas-phase N/O ratios when moving towards the inner galactic regions along the disc, where we predict the largest metallicities. In summary, the local O/H-N/O relation in Figure 4 can be explained as the consequence of radial gradients in the disc within the galaxy ISM, as shown in Figure 6.
Global average N/O-O/H relation
In this Section, we show how all 33 galaxies of our catalogue move in the N/O-O/H, mass-O/H and mass-N/O diagrams as a function of their evolutionary time. We focus on the average SFR-weighted log(O/H) and log(N/O) ratios of the whole ISM in the galaxies.
The main results of our analysis are shown in Figure 7, where we have put together all 33 galaxies in our catalogue to show how they evolve in the N/O-O/H diagram; each point represents the SFR-weighted average log(O/H) and log(N/O) ratios in the ISM of each galaxy, and the colour coding indicates the redshift of the galaxy. We find that the galaxies in our catalogue follow tracks in the N/O-O/H diagram which agree with the average Dopita et al. (2016, solid grey line) relation. Although we do not have many points at very low average gas-phase O/H abundances in our reference galaxies at high redshift, the points around log(O/H) ∼ 7.5 dex are in good agreement with the observations in DLA systems (Pettini et al. 2002(Pettini et al. , 2008Zafar et al. 2014), in the halo stars of our Galaxy (Matteucci 1986;Spite et al. 2005), and in irregular dwarf galaxies (e.g. Berg et al. 2016), which exhibit log(N/O) ≈ −1.5 dex with a large scatter around this value (see also Vincenzo & Kobayashi 2018).
In a given redshift interval, we predict that the highest N/O and O/H ratios are seen in the most massive galaxies. Therefore (Abazajian et al. 2009, SDSS DR7); the offset between model and data might be due to (i) the assumed IMF when fitting the observed galaxy spectral energy distribution (SED), which mostly affects the galaxy stellar mass estimates, and/or (ii) the assumed calibration in the chemical abundance measurement, which can strongly affect the O/H abundances (see, for example, Kewley & Ellison 2008;Belfiore et al. 2017a). Our redshift evolution of the stellar mass-O/H relation is slightly weaker than in Taylor & Kobayashi (2016), which is due to the failed SN scenario assumed in this work. We note that the feedback from active galactic nuclei (AGNs) is not included in our simulation, however the effect of AGN feedback has been shown to be negligible for this relation (Taylor & Kobayashi 2015a,b).
Since the galaxy stellar mass strongly correlates with the average galaxy ISM metallicity, there is also a correlation between stellar mass and the average N/O ratio in the ISM, which is mostly due to the secondary behaviour of the N stellar yields from massive stars and AGB stars. In this way, we can explain the observed stellar mass-N/O relation of Andrews & Martini (2013) in the redshift range 0.027 < z < 0.25, although there is an offset between model and data.
Finally, because we weight the global O and N abundances with the SFR, we show that the SFRs of our simulated galaxies The solid lines correspond to the average relation when dividing our catalogue in different stellar mass bins, while the error bars correspond to the 1σ standard deviation. The grey points with error bars in both panels correspond to the average observed abundance measurements of Andrews & Martini (2013) in the redshift range 0.027 < z < 0.25 from the SDSS DR7 (Abazajian et al. 2009). are in good agreement with observations. In Figure 9, the predicted global SFR-M relation at redshift z = 0 in our galaxy catalogue (black points) is compared with the observed data from Cicone et al. (2017, grey points with error bars) and the observed average relation from Belfiore et al. (2017b), which is consistent with the Renzini & Peng (2015) relation. Although the scatter in the predicted average SFRs is large, our simulation qualitatively agrees with observations. Finally, we note that the red sequence cannot be produced without the feedback from AGNs (Taylor & Kobayashi 2015a,b). Figure 7 is primarily driven by the mass-metallicity of galaxies, one may expect some environmental dependence. Although the environmental dependence of mass-metallicity relation is seen in some observational data (Ellison et al. 2009), it is not so clear in other works (e.g., Kacprzak et al. 2015;Pilyugin et al. 2017). In Figure 10, we show the effect of the environment on the evolution of galaxies in the N/O-O/H diagram. As the indicator, we use the distance to the fifth nearest halo identified in our cosmological simulation at z = 0, which represents the large-scale structures of galaxies very well (see Figure 5 of Taylor & Kobayashi 2017). High values of s5 for a given DM halo (which can typically be as high as ≈ 0.9 Mpc) indicate relatively low densities of galaxies in the local environment. The various points in Figure 10 represent the redshift evolutionary tracks in the N/O-O/H diagram as followed by all the galaxies in our catalogue with the colour coding indicating the s5 index. Note that the galaxies are over-plotted in the order of s5.
Since the global N/O-O/H relation in
By visually comparing Figures 7 and 10, it can be seen that the environmental dependence is much weaker than the redshift evolution; all galaxies follow the same N/O-O/H relation, indicating that the galaxy chemical evolution is self-regulated even with different SFHs. There may be a small environmental dependence seen, where the galaxies in the densest regions (with the lowest s5 values) can reach higher average N/O ratios by the present time at any fixed O/H abundance. We also find that these galaxies in the densest regions tend to show a larger scatter of evolutionary tracks in the N/O-O/H diagram. This may indicate that not only the star formation efficiency (i.e., O/H) but also chemical enrichment timescale (i.e., N/O) may be different depending on the environment. We should note, however, that this needs to be studied more with a large volume of simulations.
In contrast, there is a variation in the N/O evolution depending on the galaxy SFH. In Figure 11, we show our predictions for the redshift evolution of the average SFR-weighted log(O/H) + 12 and log(N/O) ratios in the gas-phase of our ten reference galaxies (in Figs. 2 and 3), at redshifts z 2 and for total stellar masses M 10 8 M . First of all, we predict the average N/O ratios to increase, on average, in the galaxy ISM by the present; secondly, galaxies with a relatively smooth SFH, like galaxies 3-7, and 8, exhibit also a smooth increasing trend of the average gas-phase N/O ratio. Sudden bumps in the galaxy stellar mass growth history significantly affect the slope of the predicted N/O evolution, causing similar bumps in the predicted N/O evolution.
There are few observational studies in the literature which systematically attempted to measure N/O and O/H in galaxies at high redshifts; they mostly focused on AGN, gamma ray burst (GRB) or SN host galaxies, by making use of a detailed numerical modelling of the galaxy spectral energy distribution. Examples of such systematic studies are the series of works by Contini (2015Contini ( , 2016Contini ( , 2017aContini ( ,b, 2018, but see also the previous works of the same author), which adopted the SUMA numerical code 3 , taking into account the combined effect of photoionisation and shocks (Contini & Aldrovandi 1983, 1986). In Figure 12, we compare the predictions of our simulation for the redshift evolution of log(O/H) + 12 and log(N/O) (top and bottom panels, respectively) with the measurements in GRB and SN host galaxies (red triangles and blue stars, respectively). Our simulation tends to have lower O/H and thus higher N/O than in observations especially at high redshifts. This is rather odd as it is the opposite from what we show in Figure 8. This observational dataset may not be straight-forwardly comparable to our simulation; some spectra in the catalogue of Contini (2016Contini ( , 2017a were taken in the very early phases after the SN explosion, before the SN ejecta disperse into the ambient ISM; this may eventually contaminate the abundance analysis, by leading to higher measured O/H abundances and hence lower N/O ratios. A similar tendency was reported for DLA systems, where GRB-DLA show higher metallicity than QSO-DLA at high redshifts (Cucchiara et al. 2015). In future, comparisons with unbiased large samples of galaxies will provide a more definitive test of our model predictions.
CONCLUSIONS
In this work, we have demonstrated that our model is capable of reproducing the observed increasing trend of N/O vs. O/H at high metallicity in the nearby star forming galaxies, by introducing failed SNe in our cosmological chemodynamical simulation.
We have constructed a sample of 33 star forming disc galaxies at redshift z = 0, embedded within DM halos with virial mass in the range 10 11 MDM 10 13 M . We have analysed the detailed chemical evolution of the N and O abundances within ten reference galaxies of our catalogue, characterised by well distinct SFHs (Figure 3). We have also shown how all 33 galaxies in our catalogue evolve in the N/O-O/H, stellar mass-O/H and stellar mass-N/O 3 http://wise-obs.tau.ac.il/˜marcel/suma/index.htm Figure 11. The predicted redshift evolution of the average SFR-weighted N/O ratios within our ten reference galaxies in Figs. 2 and 3. Figure 12. The predicted redshift evolution of N/O and O/H of all the galaxies in our catalogue (black circles in the top and bottom panels, respectively), comparing with available observations for SN host (blue stars) and GRB host (red triangles) galaxies (see the main text for more details). diagrams, when considering SFR-weighted average abundances in the whole galaxy ISM. Our main conclusions can be summarised as follows. (ii) The global N/O-O/H relation when dealing with average abundances from the whole galaxy ISM is the consequence of an underlying mass-metallicity relation that galaxies obey as they evolve across the cosmic epochs. In this case, the predicted N/O-O/H relation is an average evolutionary trend which is followed by the chemical evolution tracks of all galaxies at almost any redshift.
(iii) We do not find a strong environmental dependence but find that galaxies follow the same global N/O-O/H relation independent of the environment (s5). However, galaxies in the densest environments at z = 0 show a larger scatter along the relation, and thus can have higher N/O ratios at high O/H, than the galaxies in the least dense environments.
(iv) For both local and global relations, the increasing trend of N/O as a function of O/H is mainly due to the fact that N is mainly produced as secondary element at the expense of the C and O nuclei already present in the stars at their birth; the higher the initial stellar O/H abundance, the larger is the amount of synthesised N produced by stars.
(v) The average N/O ratios increase more rapidly in galaxies having SFHs concentrated at earlier cosmic epochs. Smooth stellar mass growth with time gives rise to smooth monotonic evolution of the average N/O ratios with redshift. Conversely, sudden bumps in the stellar mass growth history may also give rise to similar bumps in the z-N/O evolutionary tracks. Therefore, the redshift evolution of N/O in galaxies could be used to contrain the SFH of disc galaxies.
(vi) We predict that the O/H and N/O gradients in the ISM of galaxies flatten -on average -as functions of time, in agreement with previous studies on the metallicity gradient evolution in disc galaxies (e.g. Kobayashi & Nakasato 2011;Pilkington et al. 2012;Gibson et al. 2013); contextually, we predict also an inside-out growth of the galaxy as a function of time. In the very early "protogalactic" evolutionary stages, we predict highly scattered and overall flat abundance gradients; then, as the first series of stellar populations form, steep gas-phase abundance gradients soon develop and then gradually flatten by the present time. | 9,998.8 | 2018-01-24T00:00:00.000 | [
"Physics"
] |
Design of Planar Multilayer Devices for Optical Filtering Using Surrogate Model Based on Artificial Neural Network
: Artificial intelligence paradigms hold significant potential to advance nanophotonics. This study presents a novel approach to designing a plasmonic absorber using an artificial neural network as a surrogate model in conjunction with a genetic algorithm. The methodology involved numerical simulations of multilayered metal–dielectric plasmonic structures to establish a dataset for training an artificial neural network (ANN). The results demonstrate the proficiency of the trained ANN in predicting reflectance spectra and its ability to generalize intricate relationships between desired performance and geometric configurations, with values of correlation higher than 98% in comparison with ground-truth electromagnetic simulations. Furthermore, the ANN was employed as a surrogate model in a genetic algorithm (GA) loop to achieve target optical behaviors. The proposed methodology provides a powerful means of inverse designing multilayered metal–dielectric devices tailored for visible band wavelength filtering. This research demonstrates that the integration of AI-driven approaches in nanophotonics leads to efficient and effective design strategies.
Introduction
The field of nanophotonics has been profoundly impacted by recent innovations in artificial intelligence (AI) paradigms, which have reshaped the landscape of design methodologies [1,2].The process of designing complex photonic structures in the traditional forward approach, by using numerical methods, is usually computer-intensive.The core objective in nanophotonics is to manipulate light at subwavelength scales, thereby achieving desired behaviors in photonic structures.This manipulation involves finely tuning geometrical configurations to control sophisticated optical phenomena, such as plasmonic resonances and interferences.These phenomena are central to a wide range of applications, from achieving subwavelength focusing, which is essential for high-resolution imaging and sensing, to the development of advanced optical absorbers, applied in photovoltaic cells and optical filtering technologies [3].The ability to precisely tune the subwavelength geometries of these structures enables the manipulation of light in ways that were previously unattainable, opening new frontiers in optical technology.
The process of designing these subwavelength structures and predicting their behavior has been heavily dependent on numerical simulation, which requires the discretization and solution of local fields using Maxwell's equations with finite element methods (FEMs) or the finite difference method (FDM), for example.These methods are powerful in that they allow for a detailed analysis of the electromagnetic behavior within photonic structures, accounting for factors such as material properties, geometric configuration, and local field interaction at the nanoscale.
Optics 2024, 5 122 However, the computational cost represents one of the primary challenges imposed by these traditional numerical methods [1].As the complexity of the structure increases, so does the number of elements that need to be solved in the simulation.Moreover, the size of the parametric search space-the range of variables over which the structure is optimized-can be overwhelming.This results in an exponential increase in computational requirements, making the design process not only time-consuming but also resourceintensive.This challenge is particularly pronounced when exploring a vast design space for the optimal configuration of a photonic structure, where thousands or even millions of potential configurations might need to be evaluated.
In response to these challenges, the integration of AI paradigms into nanophotonics has emerged as a revolutionary approach.AI, particularly deep learning techniques, offers an alternative pathway that can significantly reduce the computational costs of design in nanophotonics [2,4,5].These AI methods can learn from existing data and make predictions or suggest designs, thereby accelerating the design process.For instance, by training AI models on data generated from previous simulations, it is possible to predict the optical behavior of new photonic structures without the need for extensive numerical simulations.This not only speeds up the design process but also opens up new possibilities for exploring more complex and innovative photonic structures.Consequently, AI's role in nanophotonics represents a paradigm shift, moving from traditional computation-intensive methods to more efficient, data-driven inverse design approaches.
The conceptualization of photonic devices through AI-aided inverse design has become a topic of interest in recent research.This approach has the potential to drastically reduce the time required for evaluating and designing complex subwavelength structures, a promise that has been documented in various prior studies.In [6,7], the authors used a neural network (NN) to approximate the light scattering of multilayer nanoparticles.The results showed that NN is a powerful tool for accelerating simulations about interactions with nanoscale structures.In another field of research, several papers have presented solutions for inverse design using deep learning (DL).As shown in [8][9][10], the results demonstrated the efficiency of approaches based on NN in achieving the inverse design of multilayer nanoparticles and multilayer nanofilms, mid-infrared graphene plasmons, and plasmonic nanoantenna.As can also be seen in [11,12], inverse design, facilitated by AI, marks a significant departure from traditional methods, offering a more efficient pathway to achieving optimized photonic structures.
In the current work, we build upon these advancements and present a novel methodology for the design of a plasmonic absorber with applications in optical filtering.Central to our approach is the use of a trained artificial neural network (ANN) as a surrogate model.In conjunction with this, we employ a genetic algorithm (GA), an optimization technique inspired by the principles of natural selection, as our strategic search methodology.This algorithm systematically explores the parametric space to identify configurations that align with predefined performance goals, effectively narrowing down the most promising design options.
In this work, we propose a methodology involving investigating and designing a multilayered metal-dielectric-metal (MDM) absorber, which can be tuned arbitrarily for filtering applications in the visible wavelengths without requiring complex three-dimensional periodicity, as demonstrated in previous works [13][14][15][16][17][18], where pyramidal, cylindrical, or cross shapes were needed to obtain the desired optical response.Designing an MDM absorber with arbitrary tunability in the visible wavelengths is essential for nanophotonics given its wide-ranging practical applications.Precise control over spectral filtering in the visible spectrum is of fundamental importance for optical communication, imaging, and display technologies.Additionally, the miniaturization of photonic devices, especially in the nanoscale regime, benefits from such tunable MDM absorbers.They find utility in energy harvesting, photovoltaics, and optical sensors, where the optimized absorption of specific wavelengths is vital.Furthermore, simplified manufacturing processes and Optics 2024, 5 123 cost efficiencies make these absorbers highly relevant for the practical implementation of nanophotonic devices across various industries.
This paper is structured as follows: In Section 2 we present an in-depth exploration of our design methodology along with preliminary results, setting the stage for a detailed understanding of our approach.In Section 3, we unveil and analyze the heart of our findings, discussing the core aspects of our research.Finally, Section 4 encapsulates our major discoveries while also acknowledging the challenges and limitations we faced.This reflection not only highlights our achievements but also paves the way for future research to build upon and address these challenges.
Design Methodology
For the selected multilayered metal-dielectric plasmonic structure, a dataset of various combinations of geometrical parameters and their respective optical responses was created using numerical simulation.With this curated dataset, an ANN was trained as a surrogate model for the numerical simulation.The ANN was then employed in a GA loop search to achieve an arbitrary target optical behavior.In this section, these steps are further discussed in detail.
The Multilayered Metal-Dielectric Plasmonic Structure and the Dataset
Metal-dielectric-metal (MDM) structures have shown remarkable capabilities in enhancing the local density of electric fields, primarily because of the effect of plasmonic resonance.This phenomenon occurs when light interacts with an MDM structure, leading to the excitation of free electrons within the metal layers.This interaction results in the generation of an electric field that is significantly more intense than the incident light itself.The amplification of the electric field at such localized scales is a key attribute of MDM structures, making them highly suitable for a variety of applications.
The dataset used for training the ANN was created based on structures simulated in the COMSOL 5.2 software.This software employs the finite element method to simulate the optical response of various structures, considering the effects of electromagnetic field excitation and specific material arrangements.Figure 1 provides a schematic representation of the planar MDM that forms the basis of our study.
Optics 2024, 5, FOR PEER REVIEW 3 utility in energy harvesting, photovoltaics, and optical sensors, where the optimized absorption of specific wavelengths is vital.Furthermore, simplified manufacturing processes and cost efficiencies make these absorbers highly relevant for the practical implementation of nanophotonic devices across various industries.This paper is structured as follows: In Section 2 we present an in-depth exploration of our design methodology along with preliminary results, setting the stage for a detailed understanding of our approach.In Section 3, we unveil and analyze the heart of our findings, discussing the core aspects of our research.Finally, Section 4 encapsulates our major discoveries while also acknowledging the challenges and limitations we faced.This reflection not only highlights our achievements but also paves the way for future research to build upon and address these challenges.
Design Methodology
For the selected multilayered metal-dielectric plasmonic structure, a dataset of various combinations of geometrical parameters and their respective optical responses was created using numerical simulation.With this curated dataset, an ANN was trained as a surrogate model for the numerical simulation.The ANN was then employed in a GA loop search to achieve an arbitrary target optical behavior.In this section, these steps are further discussed in detail.
The Multilayered Metal-Dielectric Plasmonic Structure and the Dataset
Metal-dielectric-metal (MDM) structures have shown remarkable capabilities in enhancing the local density of electric fields, primarily because of the effect of plasmonic resonance.This phenomenon occurs when light interacts with an MDM structure, leading to the excitation of free electrons within the metal layers.This interaction results in the generation of an electric field that is significantly more intense than the incident light itself.The amplification of the electric field at such localized scales is a key attribute of MDM structures, making them highly suitable for a variety of applications.
The dataset used for training the ANN was created based on structures simulated in the COMSOL 5.2 software.This software employs the finite element method to simulate the optical response of various structures, considering the effects of electromagnetic field excitation and specific material arrangements.Figure 1 provides a schematic representation of the planar MDM that forms the basis of our study.In the modeling of the device, as illustrated in Figure 1, each segment was set to a height of 100 nm.The primary geometrical input parameter for this configuration was the In the modeling of the device, as illustrated in Figure 1, each segment was set to a height of 100 nm.The primary geometrical input parameter for this configuration was the dielectric filling fractions (fn) of each segment, defining the proportion of dielectric material in each 100 nm segment.The device's design comprised 10 such segments, each uniquely composed of a silver-silica mix [19].Additionally, the direction of electromagnetic excitation is represented by k.This modeling approach is in line with prior studies on similar MDM structures [20], allowing us to explore different design methodologies for these configurations and their implications for device performance.
The interaction of electromagnetic waves with a metal-dielectric interface that results in a phase shift, as described in [21], is a particularly intriguing phenomenon in the context of structures engineered for broadband absorption and filter design.Such structures, when designed with strategically stacked metal-dielectric interfaces, offer a means to effectively control the resonances and, consequently, the overall optical behavior.This stacking arrangement allows for the precise manipulation of the wave interactions at these interfaces, thus enabling the tuning of the structure's optical properties for desired applications.This controlled manipulation finds practical utility in the development of devices for photovoltaic applications [3,22,23], where the efficient absorption and conversion of incident light into electrical energy are of utmost importance.
Furthermore, these engineered interfaces play an important role in the domain of color filters, where the manipulation of resonances facilitates the selective transmission or absorption of specific wavelengths of light.These interactions of electromagnetic waves within these structures may also find application in devices such as optical sensors, signal modulation, and controlled light emission processes.The design of these metal-dielectric interfaces not only contributes to a deeper understanding but also opens avenues for advancing the performance and efficiency of optical devices and energy-harvesting technologies.
The Genetic Algorithm-Generated Dataset
The quality of the dataset used to train an ANN is fundamental for its generalization capabilities.A suitable dataset should include individuals with diverse characteristics in both geometric parameters (input) and optical properties (output).In this study, a genetic algorithm (GA) search routine (see Figure 2) was employed to generate individuals for the dataset.We executed 65 different instances of the GA search to obtain reflectivity spectra in the visible band.A total of 65 target spectra were generated using a Gaussian function.
where σ is the standard deviation, which is correlated with the full width at half maximum (FWHM), and λ p is the central wavelength, which was swept from 380 to 700 nm in 5 nm intervals.The simulated structures from the FEM that came up in the optimization process were added to the dataset.
Optics 2024, 5, FOR PEER REVIEW 5 ones.To enhance diversity in the dataset and, consequently, reduce the computational time required to train the ANNs, redundant structures were filtered out.In this study, pairwise Euclidean distances between each individual in the dataset were calculated, and similar individuals were subsequently removed [24,25].To illustrate the effectiveness of the genetic algorithm (GA) search, the outcomes are highlighted in Figure 3, which shows the optimal solutions for wavelengths of 390 nm, 535 nm, and 615 nm. Figure 3 also shows the field distribution along the proposed structure as a function of the position and the operating wavelength.The electric field is concentrated in specific regions depending on the operating frequency, which ultimately results in absorption.Complementing this, Figure 4 provides a detailed overview of the reflectivity spectra across the entire dataset, encompassing all 25,033 simulated structures, with a total simulation time of almost 37 h.A notable feature in these results is the diagonal profile of the reflectance peaks, which is a direct consequence of the sequential simulation process used during the dataset generation.This process began with a As a paradigm of optimization grounded in the principles of natural evolutionary processes, GAs facilitates the systematic examination of a large number of potential solutions for a given problem.At each iteration of a GA, a mandatory step involves the evaluation of the performance of the device from the execution of an electromagnetic simulation for each individual to obtain its optical responses.The subsequent step involves a comparative anal-ysis of these responses against the desired performance metrics through fitness calculations.The fitness function used in the GA optimization was the RMSE of the FEM-simulated spectrum of a given device under test and the target spectrum (Equation ( 1)) for that GA instance: This dataset generation method inherently yields structures with similar geometries and optical behavior, given that the GA involves crossing over structures to produce new ones.To enhance diversity in the dataset and, consequently, reduce the computational time required to train the ANNs, redundant structures were filtered out.In this study, pairwise Euclidean distances between each individual in the dataset were calculated, and similar individuals were subsequently removed [24,25].
To illustrate the effectiveness of the genetic algorithm (GA) search, the outcomes are highlighted in Figure 3, which shows the optimal solutions for wavelengths of 390 nm, 535 nm, and 615 nm. Figure 3 also shows the field distribution along the proposed structure as a function of the position and the operating wavelength.The electric field is concentrated in specific regions depending on the operating frequency, which ultimately results in absorption.Complementing this, Figure 4 provides a detailed overview of the reflectivity spectra across the entire dataset, encompassing all 25,033 simulated structures, with a total simulation time of almost 37 h.A notable feature in these results is the diagonal profile of the reflectance peaks, which is a direct consequence of the sequential simulation process used during the dataset generation.This process began with a Gaussian-shaped target at a reflectance peak of 380 nm and methodically progressed through to 700 nm in increments of 5 nm, effectively mapping the spectrum's response across this range.
similar individuals were subsequently removed [24,25].To illustrate the effectiveness of the genetic algorithm (GA) search, the outcomes are highlighted in Figure 3, which shows the optimal solutions for wavelengths of 390 nm, 535 nm, and 615 nm. Figure 3 also shows the field distribution along the proposed structure as a function of the position and the operating wavelength.The electric field is concentrated in specific regions depending on the operating frequency, which ultimately results in absorption.Complementing this, Figure 4 provides a detailed overview of the reflectivity spectra across the entire dataset, encompassing all 25,033 simulated structures, with a total simulation time of almost 37 h.A notable feature in these results is the diagonal profile of the reflectance peaks, which is a direct consequence of the sequential simulation process used during the dataset generation.This process began with a Gaussian-shaped target at a reflectance peak of 380 nm and methodically progressed through to 700 nm in increments of 5 nm, effectively mapping the spectrum's response across this range.
The ANN Architecture
The feedforward neural network-a fundamental architecture in ANNs, consists of an input layer, several hidden layers, and an output layer is a simple yet powerful structure that has been proven to be able to predict the optical behavior of photonic devices and find a range of applications as surrogate models [2][3][4]26].The input layer is responsible for receiving input data, which, in our case, is the filling factor vector, and the output layer produces the predicted reflectance spectrum.The hidden layers process input through weighted connections and apply activation functions to produce intermediate representations, which are fed to the output layer, as depicted in Figure 5.The main feature of this approach to machine learning is the ability to progressively extract features from the training data, allowing the network to capture non-trivial inputoutput relationships.The training process involves the generation of an extensive dataset created by FEM simulations, which serve as ground truth for the iterative adjustments to
The ANN Architecture
The feedforward neural network-a fundamental architecture in ANNs, consists of an input layer, several hidden layers, and an output layer is a simple yet powerful structure that has been proven to be able to predict the optical behavior of photonic devices and find a range of applications as surrogate models [2][3][4]26].The input layer is responsible for receiving input data, which, in our case, is the filling factor vector, and the output layer produces the predicted reflectance spectrum.The hidden layers process input through weighted connections and apply activation functions to produce intermediate representations, which are fed to the output layer, as depicted in Figure 5.
The ANN Architecture
The feedforward neural network-a fundamental architecture in ANNs, consists of an input layer, several hidden layers, and an output layer is a simple yet powerful structure that has been proven to be able to predict the optical behavior of photonic devices and find a range of applications as surrogate models [2][3][4]26].The input layer is responsible for receiving input data, which, in our case, is the filling factor vector, and the output layer produces the predicted reflectance spectrum.The hidden layers process input through weighted connections and apply activation functions to produce intermediate representations, which are fed to the output layer, as depicted in Figure 5.The main feature of this approach to machine learning is the ability to progressively extract features from the training data, allowing the network to capture non-trivial inputoutput relationships.The training process involves the generation of an extensive dataset created by FEM simulations, which serve as ground truth for the iterative adjustments to The main feature of this approach to machine learning is the ability to progressively extract features from the training data, allowing the network to capture non-trivial inputoutput relationships.The training process involves the generation of an extensive dataset created by FEM simulations, which serve as ground truth for the iterative adjustments to neuron weights.The primary objective is to refine the network's representation until it accurately mirrors the data distributions within the training set.
The artificial neural network (ANN) was implemented using the Python programming language and the TensorFlow module, a publicly available resource [27].TensorFlow's top-level APIs were leveraged to streamline the process of defining and training the necessary machine learning architecture.
The appropriate size of an ANN is an optimization problem.A heuristic approach is commonly used to tackle this problem, such as systematically sweeping a set of combinations of layer sizes.We defined a configuration with three hidden layers as the basic architecture and tested different combinations of neuron counts in each layer, which are the hyperparameters, within a range of 10 to 300.
Expanding upon the ANN training process, it is essential to highlight the significance of optimizing hyperparameters for model generalization.While the architecture's neuron count plays a crucial role, other hyperparameters, such as the learning rate and batch size, were also systematically explored to achieve the highest predictive accuracy.These hyperparameter adjustments were pivotal in ensuring that the ANN could effectively generalize its predictions beyond the training data, a critical requirement for its role as a surrogate model in the inverse design process.Furthermore, the fine-tuning of hyperparameters was conducted through a combination of manual tuning and automated techniques, such as grid search and random search, to identify the optimal settings for our specific problem.
The dataset was divided into three subsets, which were the training (60%), validation (20%), and testing (20%) subsets.The weights and biases in the ANN were adjusted via backpropagation using the training and validation subsets.The testing subset was used afterward to evaluate the performance of the trained ANN in unseen data, which was achieved by determining the mean correlation coefficient of the entire testing dataset ground truth with the ANN predictions.
In Figure 6, each point represents a trained ANN.Notably, the total neuron count exhibited a negative correlation with the validation error, reaching saturation between 600 and 800 neurons.The best performance of ANN demonstrated a maximum mean correlation coefficient of approximately 0.98, with 239, 211, and 256 neurons for the three hidden layers.
neuron weights.The primary objective is to refine the network's representation until it accurately mirrors the data distributions within the training set.
The artificial neural network (ANN) was implemented using the Python programming language and the TensorFlow module, a publicly available resource [27].TensorFlow's top-level APIs were leveraged to streamline the process of defining and training the necessary machine learning architecture.
The appropriate size of an ANN is an optimization problem.A heuristic approach is commonly used to tackle this problem, such as systematically sweeping a set of combinations of layer sizes.We defined a configuration with three hidden layers as the basic architecture and tested different combinations of neuron counts in each layer, which are the hyperparameters, within a range of 10 to 300.
Expanding upon the ANN training process, it is essential to highlight the significance of optimizing hyperparameters for model generalization.While the architecture's neuron count plays a crucial role, other hyperparameters, such as the learning rate and batch size, were also systematically explored to achieve the highest predictive accuracy.These hyperparameter adjustments were pivotal in ensuring that the ANN could effectively generalize its predictions beyond the training data, a critical requirement for its role as a surrogate model in the inverse design process.Furthermore, the fine-tuning of hyperparameters was conducted through a combination of manual tuning and automated techniques, such as grid search and random search, to identify the optimal settings for our specific problem.
The dataset was divided into three subsets, which were the training (60%), validation (20%), and testing (20%) subsets.The weights and biases in the ANN were adjusted via backpropagation using the training and validation subsets.The testing subset was used afterward to evaluate the performance of the trained ANN in unseen data, which was achieved by determining the mean correlation coefficient of the entire testing dataset ground truth with the ANN predictions.
In Figure 6, each point represents a trained ANN.Notably, the total neuron count exhibited a negative correlation with the validation error, reaching saturation between 600 and 800 neurons.The best performance of ANN demonstrated a maximum mean correlation coefficient of approximately 0.98, with 239, 211, and 256 neurons for the three hidden layers.We calculated the correlation coefficient of the 5007 structures of the testing subset, as shown in Figure 7.We also evaluated how the ANN's predictions behaved with respect to its error for each wavelength, which is a performance metric that could have an uneven distribution.These results indicate that the model ANN suffices for its main purpose: to serve as a surrogate model for the FEM simulation.
We calculated the correlation coefficient of the 5007 structures of the testing subset, as shown in Figure 7.We also evaluated how the ANN's predictions behaved with respect to its error for each wavelength, which is a performance metric that could have an uneven distribution.These results indicate that the model ANN suffices for its main purpose: to serve as a surrogate model for the FEM simulation.
Inverse Design
Our approach to inverse design was founded on the integration of a GA optimization loop with the meticulously implemented ANN, aimed at achieving precise target spectral behaviors.This step in the design process was made possible through the utilization of the PyGAD Python library, an open-source optimization and genetic algorithm implementation [28].In this context, the input of this system represents the desired spectral characteristics of the target, and the output is the geometrical configuration of the MDM.
In our methodology, we integrated a designer-configurable optical behavior into the fitness function, affording the freedom to select and define the desired behavior according to specific objectives.This chosen optical behavior serves as a guiding principle for the algorithm, driving the systematic exploration of parameter configurations with the goal of optimizing the device to closely approximate the specified target behavior.This approach proves exceptionally advantageous in scenarios where the parametric search space size is overwhelming.By incorporating this adaptable fitness function, we provide the means to efficiently manipulate the design space to achieve the desired performance characteristics.
The conventional way of testing the behavior of a given device by executing electromagnetic simulations makes use of substantial computational processing time, which becomes impractical given the size of the multiparametric search space.To avoid hindering the GA's ability to find optimal solutions, a large number of individuals need to be tested, which may take a prohibitively long computer time.We substituted the electromagnetic simulation with the deployment of an artificial neural network (ANN) to overcome this computational bottleneck.
The transition to the ANN framework showcases its remarkable proficiency in efficiently capturing and generalizing the complex connections between the desired optical performance and the geometric parameters defining the target photonic device.By using the ANN, we can speed up the process of evaluating the structures by approximately 51-fold.Table 1 summarizes the duration of the events in the process of simulation, prediction training, and generating the dataset using a conventional PC (Ryzen 1700 with 16 GB of RAM).
Inverse Design
Our approach to inverse design was founded on the integration of a GA optimization loop with the meticulously implemented ANN, aimed at achieving precise target spectral behaviors.This step in the design process was made possible through the utilization of the PyGAD Python library, an open-source optimization and genetic algorithm implementation [28].In this context, the input of this system represents the desired spectral characteristics of the target, and the output is the geometrical configuration of the MDM.
In our methodology, we integrated a designer-configurable optical behavior into the fitness function, affording the freedom to select and define the desired behavior according to specific objectives.This chosen optical behavior serves as a guiding principle for the algorithm, driving the systematic exploration of parameter configurations with the goal of optimizing the device to closely approximate the specified target behavior.This approach proves exceptionally advantageous in scenarios where the parametric search space size is overwhelming.By incorporating this adaptable fitness function, we provide the means to efficiently manipulate the design space to achieve the desired performance characteristics.
The conventional way of testing the behavior of a given device by executing electromagnetic simulations makes use of substantial computational processing time, which becomes impractical given the size of the multiparametric search space.To avoid hindering the GA's ability to find optimal solutions, a large number of individuals need to be tested, which may take a prohibitively long computer time.We substituted the electromagnetic simulation with the deployment of an artificial neural network (ANN) to overcome this computational bottleneck.
The transition to the ANN framework showcases its remarkable proficiency in efficiently capturing and generalizing the complex connections between the desired optical performance and the geometric parameters defining the target photonic device.By using the ANN, we can speed up the process of evaluating the structures by approximately 51-fold.Table 1 summarizes the duration of the events in the process of simulation, prediction training, and generating the dataset using a conventional PC (Ryzen 1700 with 16 GB of RAM).
Results and Discussion
The trained ANN successfully predicted the reflectance spectra of arbitrarily provided geometrical inputs from the testing subset, demonstrating high values of correlation, as shown in Figure 7, between FEM-simulated behavior and ANN prediction.We demonstrate the capabilities of the trained ANN in Figure 8, where FEM and ANN spectra are compared with structures that were not present in the training phase of the ANN.
ANN prediction
104 ms Train ANN 122 s Generate dataset 37 h
Results and Discussion
The trained ANN successfully predicted the reflectance spectra of arbitrarily provided geometrical inputs from the testing subset, demonstrating high values of correlation, as shown in Figure 7, between FEM-simulated behavior and ANN prediction.We demonstrate the capabilities of the trained ANN in Figure 8, where FEM and ANN spectra are compared with structures that were not present in the training phase of the ANN.The trained ANN was effectively utilized as a surrogate model in the GA optimization loop using target spectra that were distinct from those included in the dataset generation, where only Gaussian-shaped target spectra were employed.As illustrated in Figure 9, the GA routine successfully identified optimal structures designed to match notch, bandpass, and double-notch target spectra.Table 2 contains the geometrical parameters for each optimal structure.It is noteworthy that these specific targets were not part of the original dataset used for FEM simulations, highlighting the exceptional ability of the trained ANN to extrapolate and predict novel structural configurations with high accuracy.This remarkable predictive capability underscores the versatility and effectiveness of the ANN as a powerful tool for inverse design in nanophotonics.
For each of the aforementioned target spectra that were not part of the training dataset, we executed a GA optimization loop spanning 100 generations, with a population size of 100 individuals in each generation.The computational time required for predicting the spectra using this setup amounted to approximately 17 min.This starkly contrasts with the alternative approach of relying on traditional FEM simulations for structure evaluation, which would have necessitated around 14.7 h of computation on a personal computer.
It is important to note that this time efficiency is achieved after an initial investment in creating the dataset and training the artificial neural network (ANN).The process of generating the dataset and training the ANN, essential steps in enabling the GA to function effectively, accounted for a substantial part of the total computational time.The trained ANN was effectively utilized as a surrogate model in the GA optimization loop using target spectra that were distinct from those included in the dataset generation, where only Gaussian-shaped target spectra were employed.As illustrated in Figure 9, the GA routine successfully identified optimal structures designed to match notch, bandpass, and double-notch target spectra.Table 2 contains the geometrical parameters for each optimal structure.It is noteworthy that these specific targets were not part of the original dataset used for FEM simulations, highlighting the exceptional ability of the trained ANN to extrapolate and predict novel structural configurations with high accuracy.This remarkable predictive capability underscores the versatility and effectiveness of the ANN as a powerful tool for inverse design in nanophotonics.
EER REVIEW 10
Despite this initial time investment, the efficiency gains become increasingly apparent when applying the trained model to new target spectra.For each of the aforementioned target spectra that were not part of the training dataset, we executed a GA optimization loop spanning 100 generations, with a population size of 100 individuals in each generation.The computational time required for predicting the spectra using this setup amounted to approximately 17 min.This starkly contrasts with the alternative approach of relying on traditional FEM simulations for structure evaluation, which would have necessitated around 14.7 h of computation on a personal computer.
It is important to note that this time efficiency is achieved after an initial investment in creating the dataset and training the artificial neural network (ANN).The process of generating the dataset and training the ANN, essential steps in enabling the GA to function effectively, accounted for a substantial part of the total computational time.Despite this initial time investment, the efficiency gains become increasingly apparent when applying the trained model to new target spectra.
Although it is difficult to establish a comparison between different papers given the applications of deep learning to several optical devices with different properties, it is possible to use architecture complexity and the time simulation of model surrogates to position the results of this paper in the scientific literature.Compared with [6,7,11,12], in our paper, the results show that an ANN-MLP (multilayer perceptron) with three hidden layers presented an excellent correlation between ANN prediction and FEM simulation with a time that is approximately 51 times faster.
Conclusions
The proposed methodology leverages the synergy between a GA and an ANN surrogate model to inverse design a planar multilayer device for optical filtering that can be tuned to wavelengths in the visible band.In this work, we used FEM numerical simulations of MDM plasmonic structures to establish a comprehensive dataset to train the ANN.The results highlight the proficiency of the trained ANN in predicting reflectance spectra and its ability to generalize intricate relationships between desired performance and geometric configurations.The ANN predictions exhibit a high correlation of over 0.98 compared with ground-truth electromagnetic simulation while also achieving a computational speed that is approximately 51 times faster.
This study demonstrates the successful deployment of an ANN as a surrogate model in a GA loop, showcasing its ability to find optimal solutions for devices approximating targeted spectral behaviors.The trained ANN exhibits efficiency in predicting novel structures, as evidenced by the comparison of results based on FEM simulation and ANN-predicted reflectivity spectra for configurations that were not included in the initial training phase.
The focus on planar multilayered metal-dielectric structures is due to their relative simplicity and the feasibility of generating a comprehensive dataset for training the ANN within reasonable computational and time constraints.This focus allowed us to establish and validate our methodology effectively, demonstrating the powerful integration of AI in optimizing optical filtering devices.We fully recognize the potential of our methodology in the design of more complex, non-planar geometries.
Although the surrogate model depends on the generation of database training obtained from numerical simulations, a process with significant computational cost, the advantage of the proposal is to replace the numerical simulations with an inexpensive surrogate model (trained model) that can provide the designer with the possibility of determining the parameters of an optical device in quasi-real time.Therefore, the presented methodology is a promising and effective approach for the inverse design of complex photonic structures Optics 2024, 5 131 with diverse applications in areas such as photovoltaics, color filters, optical sensors, and energy harvesting technologies.
Figure 1 .
Figure 1.Schematic of the MDM absorber.The dielectric filling factors (fn) in the 100 nm high segments are the controlled geometrical parameters.
Figure 1 .
Figure 1.Schematic of the MDM absorber.The dielectric filling factors (fn) in the 100 nm high segments are the controlled geometrical parameters.
Figure 2 .
Figure 2. Flowchart of a GA instance used in the generation of the dataset.A set of 65 runs was executed to generate the full dataset.The GA was configured to run for 100 generations, with the stall set to 10.We used a selection roulette with 40% of the population size, which was set to 50.
Figure 2 .
Figure 2. Flowchart of a GA instance used in the generation of the dataset.A set of 65 runs was executed to generate the full dataset.The GA was configured to run for 100 generations, with the stall set to 10.We used a selection roulette with 40% of the population size, which was set to 50.
Figure 2 .
Figure 2. Flowchart of a GA instance used in the generation of the dataset.A set of 65 runs was executed to generate the full dataset.The GA was configured to run for 100 generations, with the stall set to 10.We used a selection roulette with 40% of the population size, which was set to 50.
Figure 3 .Figure 3 ., 5 126
Figure 3.Comparison of FEM-simulated reflectivity and target spectra of three different MDM structures with target peaks at wavelengths (a) 390 nm, (b) 535 nm, and (c) 615 nm.The electric fieldFigure 3. Comparison of FEM-simulated reflectivity and target spectra of three different MDM structures with target peaks at wavelengths (a) 390 nm, (b) 535 nm, and (c) 615 nm.The electric field distribution along the height of the structures is shown for target peaks at (d) 390 nm, (e) 535 nm, and (f) 615 nm.
Figure 4 .
Figure 4. Reflectivity spectra of the individuals in the dataset.
Figure 5 .
Figure 5. ANN architecture.The geometrical parameters are represented by the fn inputs, which are the dielectric filling factors in each 100 nm unit cell, and the output is the corresponding 65-point reflectivity spectrum.
Figure 4 .
Figure 4. Reflectivity spectra of the individuals in the dataset.
Optics 2024, 5 ,
FOR PEERREVIEW 6 distribution along the height of the structures is shown for target peaks at (d) 390 nm, (e) 535 nm, and (f) 615 nm.
Figure 4 .
Figure 4. Reflectivity spectra of the individuals in the dataset.
Figure 5 .
Figure 5. ANN architecture.The geometrical parameters are represented by the fn inputs, which are the dielectric filling factors in each 100 nm unit cell, and the output is the corresponding 65-point reflectivity spectrum.
Figure 5 .
Figure 5. ANN architecture.The geometrical parameters are represented by the fn inputs, which are the dielectric filling factors in each 100 nm unit cell, and the output is the corresponding 65-point reflectivity spectrum.
Figure 6 .
Figure 6.Validation error as a function of the number of neurons in the hidden layer.The arrow indicates the ANN hyperparameter configuration with the highest mean correlation coefficient, chosen for the surrogate model.
Figure 6 .
Figure 6.Validation error as a function of the number of neurons in the hidden layer.The arrow indicates the ANN hyperparameter configuration with the highest mean correlation coefficient, chosen for the surrogate model.
Figure 7 .
Figure 7. Performance metrics for the chosen ANN model: (a) the distribution of correlation coefficients for each sample reveals a high concentration of above 0.9, with 0.98 as the mean value; (b) the histogram illustrates the correlation coefficients of the samples within the training subset; (c) the average error at each wavelength.
Figure 7 .
Figure 7. Performance metrics for the chosen ANN model: (a) the distribution of correlation coefficients for each sample reveals a high concentration of above 0.9, with 0.98 as the mean value; (b) the histogram illustrates the correlation coefficients of the samples within the training subset; (c) the average error at each wavelength.
Figure 8 .
Figure 8.Comparison of FEM-simulated and ANN-predicted reflectivity spectra for three arbitrarily chosen MDM configurations.High correlation coefficients of (a) 0.9355, (b) 0.9846, and (c) 0.9911 were obtained using arbitrarily chosen geometrical inputs from the testing subset as examples of the capabilities of the method.
Figure 8 .
Figure 8.Comparison of FEM-simulated and ANN-predicted reflectivity spectra for three arbitrarily chosen MDM configurations.High correlation coefficients of (a) 0.9355, (b) 0.9846, and (c) 0.9911 were obtained using arbitrarily chosen geometrical inputs from the testing subset as examples of the capabilities of the method.
Figure 9 .
Figure 9.Target and ANN-predicted spectra of optical behaviors not present in the ground-truth dataset.In (a), the target is a filter with notch at 450 nm.In (b), a bandpass filter from 450 to 600 nm was used as the target.In (c), we used a double-notch-shaped target with trough wavelengths at 500 and 600 nm.
Figure 9 .
Figure 9.Target and ANN-predicted spectra of optical behaviors not present in the ground-truth dataset.In (a), the target is a filter with notch at 450 nm.In (b), a bandpass filter from 450 to 600 nm was used as the target.In (c), we used a double-notch-shaped target with trough wavelengths at 500 and 600 nm.
Table 1 .
Duration of events.
Table 1 .
Duration of events.
Table 2 .
Dielectric filling factors for notch, bandpass, and double-notch-shaped filtering devices.
Table 2 .
Dielectric filling factors for notch, bandpass, and double-notch-shaped filtering devices. | 9,494.8 | 2024-03-01T00:00:00.000 | [
"Physics",
"Engineering",
"Materials Science"
] |
Amyloid Beta Peptide Is Released during Thrombosis in the Skin
While it is known that amyloid beta (Aβ) deposits are found in different tissues of both Alzheimer’s disease (AD) patients and healthy individuals, there remain questions about the physiological role of these deposits, the origin of the Aβ peptide, and the mechanisms of its localization to the tissues. Using immunostaining with specific antibodies, as well as enzyme-linked immunosorbent assay, this study demonstrated Aβ40 peptide accumulation in the skin during local experimental photothrombosis in mice. Specifically, Aβ peptide accumulation was concentrated near the dermal blood vessels in thrombotic skin. It was also studied whether the released peptide affects microorganisms. Application of Aβ40 (4 µM) to the external membrane of yeast cells significantly increased membrane conductance with no visible effect on mouse host cells. The results suggest that Aβ release in the skin is related to skin injury and thrombosis, and occurs along with clotting whenever skin is damaged. These results support the proposition that Aβ release during thrombosis serves as part of a natural defense against infection.
Introduction
In 1906, Dr. Alois Alzheimer showed that amyloid beta (Aβ) peptide is the major constituent of amyloid plaques in the brain, thus implicating these peptides in Alzheimer's disease (AD) [1]. Aβ peptide and its β-sheet conglomerates were found not only in the brain, but also between muscle fibers associated with skeletal muscle myopathy [2,3] and in the eye [4][5][6]. In some cases of idiopathic cardiomyopathy, the presence of Aβ peptide in myocardium was also significantly increased in AD [7]. Aβ protein deposits in the skin were described many years ago, in and around the endothelium of dermal blood vessels in aged AD patients, and were proposed as a marker for this disease [8]. Later, skin accumulation of Aβ around blood vessels was found to be unrelated to the severity of symptoms in AD patients, occurring also in some healthy subjects [9]. Some authors also suggested a relationship between Aβ deposits in the skin and the occurrence of amyotrophic lateral sclerosis, as Aβ was detected in the skin of these patients at a higher level than in healthy individuals [10]. Two decades ago, Joachim et al. (1989) [8] pointed out that the source of Aβ deposits in the skin and other organs was probably a common blood-borne precursor. Recently, we showed that activated blood platelets aggregated in blood clots contribute to the accumulation of Aβ peptide in the brain: Aβ peptide was found in and around brain blood vessels in mice after thrombosis, as revealed by immunostaining, while in thrombocytopenic animals, this release was severely reduced [11].
Skin thrombosis is normally a defense against traumatic injury and/or a response to a number of other septic and aseptic causes that activate platelets in the skin and initiate the coagulation cascade. It is known that, besides coagulation factors, platelets also contain a relatively high concentration of amyloid precursor protein (APP), which mostly accumulates within α-granules, and full-length APP (containing Aβ peptide) is liberated upon platelet degranulation [12][13][14][15][16]. APP is itself a Kunitz-type protease inhibitor, which effectively inhibits chymotrypsin, trypsin, and other proteolytic enzymes [13,17,18], and promotes the activation of coagulation factor XII [19,20]. Therefore, platelet-released APP is likely to play an important role in the hemostasis and temporal stability of the thrombus [19].
Platelets are also the primary source (~90%) of Aβ peptide in human blood [21], while the Aβ peptide variants secreted by platelets are similar to those found in the senile plaques of AD patients [22]. When densely concentrated, platelets secrete mainly Aβ ending at residue 40 (Aβ40) as a final product, while the production of Aβ42 does not depend on platelet concentration [23]. Recently, it was shown that Aβ peptide has strong antibiotic activity against both Gram-negative and Gram-positive bacteria, as well as fungi and viruses [24][25][26][27]. Aβ peptide also protects mice against microbial infection in in vivo experiments [28]. Based on these findings, it was suggested that Aβ is a previously unrecognized antimicrobial agent that normally functions in the innate immune system [11,24,28,29]. However, Aβ peptide is not the only weapon in the platelet arsenal, as other antibacterial peptides were identified long ago [30][31][32][33][34]. Like Aβ, one of these antibacterial peptides (platelet microbicidal protein 1) has a variable length of either 72 or 73 amino acids, and is cleaved from a longer precursor [35].
While it was proposed that Aβ peptide oligomers aggregated into blood clot fibrils entrap microbes [28], we suggested another mechanism related to the plasma membrane pore-forming properties of Aβ peptide [11,29]. It was shown that soluble Aβ peptide oligomers at low concentrations perforate cell membranes by forming tetramer channels penetrable by K + ions, and do so at higher concentrations by forming Ca 2+ -permeable hexameric pores [36][37][38]. An excess of Ca 2+ permeability through these pores induces calcium dyshomeostasis, leading to cell death [39]. The same mechanism is employed by other natural peptide antibiotics with channel-forming activity, such as amphotericin B and nystatin [40,41].
We hypothesize that Aβ peptide (cleaved from the released APP or directly released by platelets) may form a sufficient number of large pores in the microbe plasma membrane to cause the death of the microorganism, thereby acting as part of the body's defense system to prevent infection. This mechanism might be of special importance in the skin, which is the organ forming the initial defensive barrier.
Immunostaining and enzyme-linked immunosorbent assay (ELISA) methods were used in this study for in vivo evaluation of the release of Aβ in mouse skin during local experimental thrombosis. Patch-clamp electrophysiology methods were used to study the conductivity of the fungal cell membrane after the application of Aβ40 peptide. Mouse astrocytes were used as control eukaryotic host cells for the evaluation of the pore-forming activity of Aβ40 in eukaryotic cell plasma membranes.
Aβ Immunoreactivity Was Concentrated in Dermal Blood Vessels a Few Minutes after Thrombosis Activation in the Skin and Became More Diffuse Later
To study the distribution of Aβ peptide in skin after thrombosis, we applied a standard model of photo-induced coagulation to obtain thrombotic clots in skin blood vessels. Mice were injected intraperitoneally with the photoactivated dye Rose Bengal, and the bruised spot became visible after application of intense green laser light to the shaved skin. Immunocytochemical evaluation of coagulated and control skin samples revealed Aβ immunofluorescence in and around blood vessels ( Figure 1), and mainly coinciding with the boundaries of the vessels. Some diffuse green staining of Aβ emanating from the vessels is also visible, as well as small parts of (apparently) other blood vessels. The distribution pattern of Aβ immunofluorescence in thrombotic ( Figure 2A) and control ( Figure 2B) skin was markedly different. The blood vessel stained for Aβ can be distinguished near the hair follicle ( Figure 2A1), while similarly processed control skin has only a very low level of relatively uniform Aβ immunofluorescence that is not visible at the same threshold value ( Figure 2B1). The line segment 1-2 ( Figure 2A1) crosses the blood vessel, and the profile of green fluorescence along this line in Figure 2A2 shows Aβ diffusion from the blood vessel. The distribution of Aβ immunofluorescence along the randomly drawn 1-2 line segments in control skin ( Figure 2B1) shows only a very low level of relatively uniform Aβ immunofluorescence ( Figure 2B2).
The 3D distribution of Aβ, which is mainly concentrated within the large and small blood vessels, but to some extent leaks into the nearby tissue, is shown in Figure S1. Interestingly, erythrocytes were stained as Aβ-positive in a thrombotic skin sample ( Figure S1). It is known that Aβ peptide in blood plasma binds to 98% of all erythrocytes in AD patients, and is a marker of the disease [42]. Moreover, the addition of synthetic Aβ not only marks erythrocyte membranes, but also makes them more elongated [43].
Formation of Blood Clots in the Skin Augments the Quantity of Aβ Peptide in Thrombotic Skin
To determine the levels of mouse-specific Aβ40 peptide in skin patches in control and post-photothrombosis skin a commercially available enzyme-linked immunosorbent assay (ELISA) was used. The final concentrations of Aβ40 in control (415 ± 57 pg/mL) and thrombotic (1733 ± 165.7 pg/mL) skin indicate an~4-fold increase after thrombosis ( Figure 3, n = 4, p < 0.001, t = 8.549, df = 5). The concentration of free mouse Aβ40 peptide determined by enzyme-linked immunosorbent assay (ELISA) in control and thrombotic skin homogenate (pg/mL). Mean ± S.E. and significant differences between groups (*) are shown (p < 0.05).
Application of Aβ40 to the External Membrane of Yeast Cells Visibly Augmented Membrane Conductance
Patch-clamp measurements with electrodes filled with synthetic Aβ40 were used in order to evaluate the effect of Aβ40 on membrane conductance in microorganisms. Yeast cells were selected as a model, because these microorganisms have the cell size and other characteristics that allow for the measurement of external membrane conductance using the patch-clamp technique ( Figure 4B). Fine-tip electrodes (~10-12 MΩ tip resistance) were employed in the cell-attached voltage-clamp configuration without rupturing the cellular membrane. In this configuration, the cell is attached to the electrode with holding potential U, which is changed according to the protocol. The current passing first through the electrode (R-elect = 10 MΩ), then (serially) through the patch of cellular membrane (R-patch), and then through the rest of the membrane (R-mem), is measured. A pore-former would affect only R-patch, while other resistances would stay constant, and an R-patch change would affect the total membrane current I ( Figure 4A). Channel-forming peptides are not anchored, and rapidly diffuse around lipid membranes, forming temporary ion channels [44,45]. The resistance of the membrane patch perforated by a channel-former decreases, leading to a current increase. This resistance, and its changes with time, are calculated by measuring the current at the beginning (after attachment, time 1, Figure 4A) and after the effect is pronounced (time 2, Figure 4A), in which R = U/I, and the resistance change with time (Rtime1/Rtime2) can be calculated.
Four different concentrations of Aβ40 (0.4 nM, 40 nM, 400 nM, and 4 µM) were tested. While nanomolar concentrations of Aβ40 had no pronounced effect, 4 µM Aβ40 produced visible augmentation of the current amplitude from 20 pA to about 400 pA ( Figure 4A, upper trace), which developed at 3-5 min after application. The effect was statistically significant, as 4 min after the electrode containing Aβ40 was attached to the yeast membrane, the resistance was reduced 17.4 ± 3.1-fold (n = 7, p < 0.0001, t = 245.7, df = 12). These changes in current indicate that the yeast membrane resistance was rapidly and significantly affected by the application of Aβ40.
Application of Aβ40 to the External Membrane of Mouse Astrocytes Did Not Significantly Affect Membrane Conductance
A standard patch-clamp was used in cell-attached configuration to study whether Aβ40 affects mouse cell membrane conductance at the same concentration as in yeast cells (4 µM). Astrocytes in a mouse hippocampus brain slice were used as the model. After an electrode containing 4 µM Aβ40 was attached to an astrocyte membrane ( Figure 4A, lower trace), the transverse current in the astrocyte external membrane was not significantly affected 3-5 min after application (n = 8, p = 0.7584, t = 0.3136, df = 14). Thus, at a 4 µM concentration, Aβ40 had no immediate effect on the membrane resistance of astrocytes.
Discussion
This study established a correlation between the release of Aβ and thrombotic processes in mouse skin. Experimental photothrombosis was used, which is a well-established method for inducing rapid coagulation without mechanical damage to the tissue [11,46].
Immunostaining with a MOAB-2 (monoclonal mouse anti-Aβ antibody) confirmed that by 10 min after thrombosis, Aβ was present mainly in and around blood vessels in the skin (Figure 1, Figure 2 and Figure S1). The MOAB-2 antibody was chosen because of the high specificity to Aβ peptide. This antibody was extensively tested, previously, and found to be a pan-specific, high-titer antibody to Aβ residues 1-4, reacting with unaggregated, oligomeric, and fibrillar forms of Aβ42 and unaggregated Aβ40, as well as with aggregated amyloid in plaques [47,48]. The MOAB-2 anti-Aβ antibody did not detect APP or its derivatives, and did not colocalize with antibodies against either the N-or C-terminal of APP [47]. It is a technical challenge to stain the endothelial cells of blood vessels to study colocalization of Aβ and the blood vessel boundaries after thrombosis. The commonly used markers for blood vessel endothelium, CD31, von Willebrand factor (vWF), and CD34, are not suitable for staining endothelial cells in this model of platelet aggregation, because they are upregulated at the membranes of activated platelets in the clot, or are involved in platelet-vessel wall interactions, producing confusion [49][50][51]. For this reason, isolectin B4 (BSI-B4) from Bandeiraea simplicifolia was selected as the most suitable marker for the luminal surfaces of endothelial cells [52][53][54] in thrombotic models. Unlike thrombotic skin ( Figure 2A1), in which the blood vessels were clearly marked with Aβ immunofluorescence, in control skin ( Figure 2B1) there was no visible co-staining of Aβ and blood vessels or other structures. The Aβ immunofluorescence profile declined rapidly away from the blood vessel boundaries (Figure 2A2), while it stayed uniformly low in control skin ( Figure 2B2).
Additionally, the study revealed that experimental thrombosis resulted in a 4-fold elevation of Aβ in skin, reaching 1733 ± 165.7 (pg/mL), which is about 0.4 nM. This concentration is in the range previously reported for Aβ in the cerebrospinal fluid (CSF) of young genetically modified 3xTg-AD mice, which spontaneously develop brain plaques [55]. The Aβ40 concentration in control skin in our experiments was 415 ± 57 (pg/mL), which is close to the range of concentrations (340-550 pg/mL) previously found in the blood plasma of healthy humans [56][57][58]. Measurement of Aβ in blood by ELISA reveals mainly free peptides, while a significant amount of Aβ peptide remains bonded to plasma proteins and lipoproteins [59], as well as to cell membranes [42], and our estimation of Aβ40 in skin has similar limitations. The detection of Aβ peptide generation in the skin is consistent with the hypothesis that a role of Aβ is as a natural defense against infection that accompanies thrombosis. The pore-forming properties of innate Aβ peptide might provide a rapid antimicrobial defense mechanism [11,29]. The current study provides additional support for this hypothesis, indicating that Aβ40 peptide significantly affects the conductivity of fungal membranes. Application of 4 µM Aβ40 to the external wall of yeast cells produced marked augmentation of the recorded current amplitudes ( Figure 4A). Bertl et al. (1998) [60] first described the method for patch-clamping yeast cells without cell wall removal, showing that it forms a standard gigaseal. It is known that the fungal cell membrane is mechanically protected by a cell wall that is impossible to "open" by pressure application only [60,61]. Thus, the cell-attached configuration was used to record membrane currents, enabling this problem to be overcome. In yeast cells, a change in membrane resistance developed 2-5 min after peptide application, and became saturated after reaching a new limit. It should be mentioned also that the recommended Tris [tris(hydroxymethyl)aminomethane] buffer [60] was switched to HEPES [4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid], because Tris buffer effectively blocks Aβ peptide-formed channels [62]. At the same time, the similar patch-clamp configuration and Aβ40 concentration had no effect on the resistance of the mouse astrocyte (host cell) external plasma membrane (Figure 4).
The pore-forming effect of Aβ peptide on brain cell membranes is well known [63,64]. The pronounced effect of Aβ40 in yeast (4 µM), but not in eukaryotic host cells in the current study, may reflect the fact that yeast cells have less cholesterol and more ergosterol in their membranes than mammalian cells [65], and it was previously found that membrane cholesterol reduces Aβ40 channel-forming activity [66].
Interestingly, there is a significant mismatch between Aβ40 peptide concentration found in the skin after thrombosis and its effective antimicrobial concentration. The results of this study demonstrate that, when Aβ40 is released in the skin during thrombosis, the concentration detected (~0.4 nM) is lower than necessary for a significant membrane effect in Saccharomyces yeast (4 µM). It was previously shown that pathologic Candida yeast are more sensitive to Aβ40 peptides (0.78 µg/mL, [24]); however, this was 1200-fold more than the concentration detected in the current study in thrombotic skin. At the same time, other antibacterial peptide species that were identified in platelets were found to be very effective against bacteria when applied in nanomolar concentrations as a thrombin-induced platelet supernatant, while purified proteins were effective in the 10-40 µM range [31,35]. Similarly, one might assume that the sensitivity of microorganisms to the native Aβ peptide would be significantly higher than the synthetic peptide.
Another explanation of this contradiction might be that the Aβ40 concentration that was determined by ELISA in mouse skin was not the local concentration around the damaged blood vessels, but rather the mean concentration in tissue. Aβ40 is mostly concentrated locally around blood vessels (Figure 2A,B), and rapidly declines away from the blood vessel boundary.
Many factors affect the sensitivity of microorganisms to Aβ40. First, the amount of oligomerized Aβ40 peptides depends on the solubilization method and the peptide form used in experiments. Oligomer formation by β-sheeting of monomeric Aβ is the necessary first step in forming Ca 2+ -permeable ion channels [67,68]. As an example, Candida yeast are very sensitive to Aβ peptides (0.78 µg/mL, [24]), but the recommended solubilization of the peptide with NH 4 OH [69] reportedly may lead to a lack of toxic effects at low concentrations [70], which supports the possibility that the solubilization method affects the oligomerization of Aβ.
Second, Aβ40 peptide carriers also reduce the effective concentration for pore formation. It was shown that in blood plasma, Aβ is usually attached to carriers, while unbound Aβ is rare. In blood plasma, Aβ40 was found bound to several proteins (that is, apolipoproteins A-I, A-IV, E and J; α2-macroglobulin; complement factors; immunoglobulins; transthyretin; apoferritin; and serum amyloid P component [71][72][73]), and one carrier is associated with many Aβ peptide molecules [74]. It is possible that if a carrier transports numerous Aβ40 peptide molecules to the cellular membrane and fuses to the membrane, it will drastically augment the local effective concentration of Aβ and result in pore formation. This carrier-based reduction of the bulk-phase Aβ40 concentration necessary for pore formation has been shown previously in artificial membranes [75]. When Aβ was concentrated inside liposomes, it was incorporated into plasma membranes more readily than by direct incorporation from aqueous solution.
We suggest that under natural conditions, blood plasma has liposome-like components, such as tissue-specific apolipoprotein type E (APOE), that facilitate the membrane-damaging action of amyloid-prone proteins and peptides. Under normal conditions, APOE participates in lipid transport and removes cholesterol and triglycerides from peripheral tissues. Many cell types produce APOE, including skin epithelium cells [76][77][78]. APOE was found in skin tissue, and plays an important role in primary localized cutaneous amyloidosis (PLCA) formation [79], which is associated with αβ deposits, as well as with deposits of misfolded amyloid-like keratin peptides [80]. Interestingly, the E4 allele of APOE is significantly increased in frequency in patients with PLCA [81], and similarly, the E4 allele is increased in frequency in AD [82]. Thus, APOE involvement in both types of amyloidosis makes it the main suspect as the amyloid peptide carrier and possible facilitator of the pore-forming effects of these peptides.
Materials and Methods
All procedures involving rodents were conducted in accordance with the National Institutes of Health regulations concerning the use and care of experimental animals. All procedures involving animals were approved by Universidad Central del Caribe Institutional Animal Care and Use Committee (Protocol #035-2017-16-01, 16 January 2017). All efforts were made to minimize suffering. For skin thrombosis experiments, animals were anesthetized with isoflurane (4% for induction, 1.75% for maintenance) using a Matrix Quantiflex VMC Anesthesia Machine for small animals (Midmark Corp., Dayton, OH, USA). After photothrombosis experiments, the anesthetized animals were rapidly decapitated and skin patches harvested. This euthanasia method is consistent with the recommendations of the Panel on Euthanasia of the American Veterinary Medical Association. In total, 12 C57BL/6 mice were used in these experiments.
Photothrombosis Model
In order to induce clot formation in mouse skin, a widely used method of photostimulated thrombosis was employed [11,46]. C57BL/6 mice of both sexes, 8-10 weeks old, were used in experiments. Briefly, prior to surgery, Rose Bengal (cat. #198250; Sigma-Aldrich Chemical Co., St. Louis, MO, USA) was dissolved in a sterile saline solution (7.5 mg/mL). The mice were then anesthetized with isoflurane, followed by peritoneal injection of Rose Bengal (150 µg/g animal weight), which was allowed to diffuse and enter the blood stream for 5 min. The animal skin was shaved and exposed to a laser beam (15 mW at 430 nm) until a visible bruise appeared on the skin (in~10-15 min). The resulting thrombosis was 3-4 mm in diameter. The mice were killed, and control and thrombotic skin samples were harvested and fixed in 5% formaldehyde as soon after thrombosis formation (which takes~10 min) as possible. Interestingly, alone among the many xanthene dyes, Rose Bengal exhibited a strong inhibitory effect on Aβ aggregation upon green photoexcitation, because of its high binding affinity for Aβ, and it also exhibits a remarkable red shift and a strong enhancement of fluorescence emission in the presence of Aβ [83]. Thus, Rose Bengal fluorescence is an indicator of the presence of Aβ.
Preparation of Fungi (Yeast) for Patch-Clamping of Plasma Membrane
A common haploid S. cerevisiae yeast strain (BY4741) was used for patch-clamp experiments. Cells were grown overnight in a small volume of modified liquid YEPD medium (1% yeast extract, 2% peptone, and 5% D-glucose; cat. #Y1375; Sigma-Aldrich, St. Louis, MO, USA), with 30 mL per 125 mL Erlenmeyer flask at 25 • C with rotary shaking (90 rpm). The cells were harvested by centrifugation (500× g for 5 min) of 1 mL of suspension, and were then resuspended in the following buffer: 140 mM NaCl, 5 mM MgCl 2 , 10 mM CaCl 2 , 10 mM HEPES, adjusted to pH 7.5 with KOH. A few drops of this suspension were finally added to the patch-clamp chamber with the same buffer. Patch recording was done according to the method developed for S. cerevisiae by Bertl et al. (1998) [60] with modifications. The most important modification was the use of HEPES buffer instead of Tris, as Tris is a well-known blocker of Aβ channels [62]. Another modification was the use of the cell-attached patch-clamp procedure instead of whole-cell patch-clamp. Cells were visualized using an Olympus infrared microscope equipped with a difference of interference contrast (DIC) system (cat. #BX51WI; Olympus, Japan). Placed in the chamber, the cells sedimented rapidly, and under the microscope, cells (3-4 µm in diameter) were selected for patch-clamping. Borosilicate glass pipettes (O.D., 1.5 mm; I.D., 1.0 mm; World Precision Instruments, Sarasota, FL, USA) were pulled in four steps to a final resistance of 10-12 MΩ for recordings using a P-97 puller (Sutter Instrument Co., Novato, CA, USA). Electrodes were filled with the following solution: 175 mM KCI, 5 mM MgCl 2 , 4 mM ATP, 100 nM Ca 2+ , 1 mM EGTA, and brought to pH 7.0 with KOH. The necessary Aβ concentration was then added to this buffer. Currents were recorded in response to 2 s voltage steps, switching from +100 to −100 mV, with a 2 s holding interval at 0 mV between each pulse. Results are presented as concatenated response curves. Data were filtered at 1 kHz and sampled at 100 Hz during recording, and the presentation data were then filtered at 40 Hz. The pipette potential was not corrected.
Whole-Cell Recordings in Astrocytes
Membrane currents were measured using the single-electrode, whole-cell patch-clamp technique. Cells were visualized using an Olympus infrared microscope equipped with a DIC system (cat. #BX51WI; Olympus, Japan). A piezoelectric micromanipulator (MX7500 with MC-1000 drive; Siskiyou, Inc., Grants Pass, OR, USA) was used for voltage-clamp and current-clamp recording, and allowed the patch clamping of cells. A MultiClamp 700A patch-clamp amplifier with two separate patch-clamp channels paired to a DigiData 1322A interface (Molecular Devices, Inc., Sunnyvale, CA, USA) was used for recording and stimulation. The pClamp 10 software package (Molecular Devices, Inc.) was used for data acquisition and analysis. Borosilicate glass pipettes (O.D., 1.5 mm; I.D., 1.0 mm; World Precision Instruments; Sarasota, FL, USA) were pulled to a final resistance of 7-8 MΩ for recordings in four steps using a P-97 puller (Sutter Instrument Co., Novato, CA, USA). Electrodes were filled with the following solution (in mM): 130 K-gluconate, 10 Na-gluconate, 4 NaCl, 4 phosphocreatine, 0.3 GTP-Na 2 , 4 Mg-ATP, and 10 HEPES, and the pH was adjusted to 7.2 with KOH. Currents in the cell-attached configuration were recorded in response to 2 s voltage steps, switching from +20 to −20 mV, with a 2 s holding interval at 0 mV between each pulse. Data were filtered at 1 KHz and sampled at 100 Hz during recording. The presentation data were then filtered at 40 Hz, and the pipette potential was not corrected.
Preparation of Aβ Peptide Solution
We used lyophilized synthetic Aβ40 peptide trifluoroacetate salt (cat. #A4473; Sigma-Aldrich, St. Louis, MO, USA), and the 40 µM stock solution (~0.18 mg/mL) was prepared from the peptide on the day of experimentation by solubilizing Aβ40 in a buffer similar to electrode internal solution (see Sections 2.2 and 2.3). Aβ stocks were sonicated for 5 min on ice, insoluble peptide aggregates were filtrated (0.2 µm pore filter), and the stock was then kept at +2 • C for storage. For experiments, the stocks were diluted into the same buffer that was used to fill the electrode, and once again, filtrated with a similar filter.
Enzyme-Linked Immunosorbent Assay (ELISA) Measurements
A specialized, ready-to-use, mouse-specific, solid-phase sandwich ELISA kit (cat. #KMB3481; Invitrogen) was used for direct measurement of the amount of Aβ40 peptide in the skin with and without thrombosis. The skin samples were homogenized mechanically, and 100 mg of homogenate was then lysed in guanidine solution (5 M guanidine HCl, 50 mM Tris HCl, pH 8.0). A monoclonal antibody to the NH 2 -terminus of mouse Aβ40 was coated onto the wells of the microtiter strips provided in the kit. Samples, including standards of known Aβ40 content for calibration purposes as well as experimental specimens, were pipetted into the wells. After washing, the rabbit antibody specific to the COOH-terminus of Aβ40 was added and detected with horseradish peroxidase-labelled anti-rabbit antibody. The optical density values at 450 nm were determined using a Wallac 1420 Victor2 Microplate Reader (PerkinElmer Inc., Waltham, MA, USA). A standard curve was used for final determination of the concentration of Aβ40 in the samples, and is presented as picograms of Aβ40 per milliliter of initial homogenate.
Statistics and Measurements
Using GraphPad Prism 7.03 (GraphPad Software, Inc., La Jolla, CA, USA) for calculations, an unpaired t-test was employed to estimate statistical differences. Values were determined to be significantly different if the two-tailed p value was <0.05.
Conclusions
The results of the study suggest that Aβ release in skin is associated with clot formation. Aβ affects membrane conductance in yeast microorganisms, but not in mouse somatic cells. Just as clotting is denoted as a "normal" protection and repair process in damaged skin, so does the liberation of Aβ during this process also lead to protection and repair. Skin-associated release of Aβ may also have implications for the development of treatment methods to cure AD, as patients with compromised clotting, as in thrombophilia, thrombophlebitis, or similar conditions, could have additional inflammation problems from an anti-Aβ vaccine. The development of the ACC-011 AD vaccine (Elan-Wyeth Corp., Dublin, Ireland) was halted due to a patient developing skin lesions, which was identified as a suspected case of inflammation of skin blood vessels [84]. Another implication of systemic generation of Aβ is the possible impact on the development of brain damage in AD. This aspect was directly confirmed recently in a study using constant transfusion of blood between genetically modified animals that developed Aβ plaques in the brain and their wild-type littermates. Aβ originated from transgenic AD model mice entered the circulation and accumulated in the brains of wild-type mice [85].
Generation of Aβ during thrombosis in the skin supports the hypothesis that Aβ release is a natural defense against infection that may accompany skin trauma. In support of an Aβ peptide role in natural antimicrobial defense, this study demonstrated that Aβ directly affects the fungal cell external membrane while not affecting the host cell membrane at the same concentration. This remarkable specificity for microbes, with relatively low toxicity for host cells, was shown previously for many antimicrobial peptides from marine invertebrates [86], and we suggest that there is some similarity to the Aβ40 effect. | 6,713 | 2018-06-01T00:00:00.000 | [
"Medicine",
"Biology"
] |
A Dynamical Modelling of the Epidemic Diseases to Assessing the Rates of Spread of COVID-19 in Saudi Arabia: SEIQR Model
A model of critical epidemic dynamics for the emergence of the new coronavirus COVID-19 is being established in this paper. A new approach to the assessment and control of the COVID-19 epidemic is given with the SEIQR pandemic model. This paper uses real knowledge on the distribution of COVID-19 in Saudi Arabia for mathematical modeling and dynamic analyses. The reproductive number and detailed stability analysis are provided in the SEIQR model dynamics. In a Jacobian method of linearization, we will address the domain of the solution and the equilibrium situation based on the SEIQR model. The equilibrium and its importance have been proven, and a study of the stability of the equilibrium free from diseases has been implemented. The reproduction number was evaluated in accordance with its internal parameters. The Lyapunov theorem of stability has proven the global stability of the current model's equilibrium. The SEIQR model was contrasted by comparing the results based on the SEIQR model with the real COVID-19 spread data in Saudi Arabia. Numerical evaluation and predictions were given. The results indicate that the SEIQR model is a strong model for the study of the spread of epidemics, such as COVID-19. At the end of this work, we implemented an optimum protocol that can quickly stop the spread of COVID-19 among the Saudi populations. The key solution to slowing COVID-19 transmission is to stay home and bring sick persons as far as possible in a remote location or in a safe place. Ultimately, it is vital to offer safe and adequate treatment to ill people, and to avoid them, medications, tones, and nutrients should be provided to non-infected persons.
Introduction
As COVID-19 outbreaks continue, the number of infections steadily increases. This is due to the presence of many factors that increase the complexities of COVID-19 infection and create barriers to disease management. Since scientists and researchers all over the world are trying to establish a vaccine or a cure for the outbreak to control such pandemics in the future, from a medical engineering framework, an infectious disease can be well known and understood through the use of mathematical models. This idea began in 1927. After that, many different mathematical models have been constructed for various diseases and infections. For some essential studies, we refer to [1][2][3][4][5][6][7][8] .
To explain transmission dynamics and estimate domestic and global disease spread based on data recorded from December 31, 2019, to January 28, 2020, Wu et al. 9 have implemented the Susceptible Exposed Infectious Recovered Model (SEIR). They also found that COVID-19 had a fundamental reproductive number of approximately 2.68. Read et al. 10 To determine the scale of the disease outbreak in Wuhan, Iman 12 carried out calculational modeling of the possible epidemic tracks with an emphasis on human-to-human transmissions.
Its findings suggest that controls must be efficiently controlled by well over 60% of the transmission. To analyze and forecast the infectivity of the new coronavirus, Guo et al. 13 developed a deep learning algorithm. They found that two animal hosts of this virus were bats and minks. Most of the models illustrate the significant role of a direct transmission mechanism between humans and humans in the outbreak, as demonstrated by the fact that many individuals infected in the Wuhan area have no interaction and the number of infections has been growing rapidly and spreading across the Chinese provinces and over 20 people 14 . There is a relatively long incubation period in many infected individuals, so they do not show symptoms and have not been aware of their infection for 10-14 days. Over time, the disease can easily be spread by direct exposure to other people. On the other hand, the published models have not, to date, taken into account the environmental position of COVID-19 transmission. Several other modeling studies for the COVID-19 outbreak have also been carried out 6,8,[15][16][17][18][19][20][21][22][23] .
Statistical epidemiology is based on the dynamics of health and disease and related population factors. The presence of a pathogenic microbial agent identifies an infectious disease as a clinically obvious disease. For modeling purposes, four forms of transmission are characterized: straightforward, if the causative disease agent is individual; vector, if the causative agent is transmitted from a vector to a person; natural, if the touch of a pathogen infects the human via the environment; and vertical, if the disease agent is transmitted from mother to child at birth. Airborne and personal diseases are generally known to be transmitted directly where transmission occurs through contact between individuals and others. 24 .
Mathematical modeling of infectious diseases is important and critical with the advent of HIV epidemics. Since then, several models for investigating infectious diseases have been developed, studied, and applied. Mathematical modeling currently applies enormously to public health and mathematics. 3,15,18,23,25 .
In the emergence of HIV epidemics, the mathematical modeling of infectious diseases is significant. Since then, numerous models have been developed, studied, and applied for the investigation of infectious diseases 26 . Farman et al. studied the stability and control of glucose insulin glucagon system in human 27 . Farman et al. 26 discussed the dynamical behavior of fractional-order cancer model with vaccine strategy. Gondim and Machado 28 introduced the optimal quarantine strategies for the COVID-19 pandemic in a population with a discrete age structure. Davies et al. 29 studied the age-dependent effects in the transmission and control of COVID-19 epidemics.
The goal of this paper is to construct a new COVID-19 vital dynamical model that is more applicable to cases in any country through mathematical analysis of the model in question by using a system of similar models with different considerations and new in / outflows between population divisions. In addition, this paper presents a new formula that explores the sensitivity of a reproduction number. The mechanisms of virus transmission by humans are to be discovered. Another aim is to investigate and learn the optimal procedures, controls, and techniques to minimize the outbreak substantially.
Formulation of a coronavirus disease (SEIQR model)
During the spread of COVID-19 in any country, the population can be divided into five vital dynamic subpopulations or five groups, which are represented in Fig. 1 and can be described as follows 4,6,16,23,29-31 : The main group ( ) St is dedicated to the healthy people but who may get the disease population. For certain diseases, the infected person may not become infectious immediately, but the latent phase is not contagious. It takes time for the pathogen to replicate and develop itself in the new host. In general, the exposed (latent) cycle follows the sensitive process 4,6,18,20,30 .
Thus, the group ( )
Et is dedicated to the exposed population or individuals who are infected but not yet infectious.
The group ( ) Itis devoted to the population who are confirmed infected (individuals who have contracted the disease and are now sick with it and infected individuals are also infectious).
The group ( )
Qt is dedicated to the quarantined population (separated from the general population even in their houses).
The group ( )
Rtis defined as the recovered population (individuals who have recovered and cannot contract the COVID-19 again), as in Fig. 1. For any group, the outflow based on the natural death rate is defined by the nonnegative rate 1 d .
The total population size is ( ) Nt, which is defined as 4,6,16,22,23,25,32 : Starting with group ( ) St, we have two outflows; a population flows out to the exposed
dS
The group of exposed ( ) Et has only one inflow , while it has four outflows.
The first outflow is the population that flows out to group ( ) All inflows and outflows are shown in the flowchart in Fig. 1, and the five groups can be converted into equations to formulate the following system of first-order ordinary nonlinear differential equations 4,6,16,22,23,25,32 : Proof: It follows from equation (2) that where I
=
It can be re-written in the following form 3 Thus, we obtain Then, we get In similar manner, it can be shown that ( )
Theorem 2 (the domain of solutions)
All the solutions of the model structure that initiate in 5 + are bounded inside the region Proof: By differentiating both sides of equation (1), we obtain Substituting from the model (2)-(6), we obtain From theorem 1, we have + ; hence, the following inequality is valid: Then, we obtain ( ) ( ) Then, when t →we obtain the solution which completes the proof 3,6,25 .
The number 0 is called the reproduction number (RBN), which takes the form 3,6,25 : Then, if 0 1 the system has a unique endemic equilibrium 8 : individual converted to an exposed individual) 6,8 .
Reproduction number by using the Jacobian matrix
To obtain the reproduction number 0 by using the Jacobian matrix method, we consider that the disease-free equilibrium (DFE) of the SEIQR model is acquired by setting (19)- (23). Hence, we obtain DFE in form 0 The Jacobian matrix of the SEIQR model takes the following form: First, we will linearize the first two equations by using the Jacobian method. The first two equations have a disease-free equilibrium (DFE) situation when 00 Hence, we consider Then, we have By substituting from the equilibrium position, we obtain Hence, the system of nonlinear equations (2) and (3) has been converted to the following linear system 6 : For the complete system at equilibrium, the stability of the disease-free equilibrium (DFE) is given by the Jacobian matrix: By calculating the characteristic equation given by and the remaining roots are the solution to the following equation: The roots of the above equation after inserting 0 will take the forms: The formulas (44) generate the following cases 6 : Thus, from (41), we obtain the following condition of equilibrium: Thus, the condition (45) is the only condition of the equilibrium of the SEIQR model. Therefore, the unique equilibrium condition of the SEIQR model is The reproduction number (RBN) 0 is also unique 6 .
Local sensitivity analysis of RBN
Local sensitivity analysis is a sensitivity analysis that examines the change in the output values that results from a change in one input value (parameter) 6 .
The sensitivity or elasticity of quantity G concerning parameter p is given by: The sensitivity of G concern p is positive if G is increasing concerning p and negative if G is decreasing concerning p.
Applying formula (49) into reproduction number 0 , which takes the form 6 : It means that a 1% increase in each one ( ) 1 1 3 2 2 2 , , , , , dd will produce ( ) where is a parameter that will be determined later, and * 1 S d = .
The equation (59) The second possibility is 1 x ; then, the two terms are non-positive. Thus, Therefore, by the Lyapunov theorem, the disease-free equilibrium is globally asymptotically stable for the system of the SEIQR model in all.
Solutions for the system of the SEIQR model
We assume the initial conditions of the SEIQR system in (38), (39), and (4)-(6) to take the form Consequently, we can obtain the other functions ,, S E I , and Q .
Model verification and predictions
To verify the SEIQR model, we will apply it to the real data regarding the COVID-19 outbreak in Saudi Arabia. COVID-19 has been in Saudi Arabia since March 3, 2020. Cases continued to be discovered in small numbers until the beginning of April, and then the number of detected cases increased. Therefore, we decided in this study to consider April 1, 2020, as the real beginning of the spread of the COVID-19 epidemic in Saudi Arabia.
We used tables of statistics issued from the Saudi Ministry of Health 33 and the daily official statement issued by the ministry as well as Wikipedia 34 , which also depends on the ministry's website and other websites that would announce these statistics.
Another source of these data is the "Saudi Centre for Disease Prevention and Control 35 ." We used the official website of the General Statistics Authority of Saudi Arabia for more information about the kingdom's population, mortality rate, and population growth rate.
To study the spread of COVID-19 in Saudi Arabia before June 13, 2020, we will represent the curve of the number of daily infections and the time series curve of the total number of infections, as shown in Figs. 2 and 3: and reached an accumulated amount of 122,259 infections on June 13, 2020. Therefore, we will use these data through the present SEIQR model to discern whether there is a convergence between the model results and the real data 33-35 .
Applying the SEIQR model to Saudi Arabia data of the spread of Covid-19
According to the official data of Saudi Arabia, we have the following initial data, which are considered the initial conditions of the system based on the SEIQR model, as in Table 1 33-35 : . Some of the other parameters have been calculated, estimated, or assumed, as in Table 2. The estimated data has been calculated by using the most powerful methods, however, is calibration or curve fitting. Curve fitting is the process of identifying the parameters of a curve, or mathematical function, that has the best fit to a series of data points. MAPLE software has been used for the fitting curves and to estimate the parameters we need. Now, we will predict the spread of COVID-19 in Saudi Arabia based on the current data and parameters with the same rates without any change in the procedures and considering that everything will continue as it is. We will illustrate the results of the total number of infections by applying the SEIQR model for the next three months, starting from April 1, 2020, and ending on October 18, 2020. As shown in Fig. 6, the curves and results show whether the number of infections will be reduced and whether the spread of COVID-19 will continue to be unstable. The curves have been established by using the same three values of the two parameters and 0 .
The figure shows that the spread of COVID-19 will continue with an unstable situation without being slowed and that the number of daily infections will rise to very high numbers.
Thus, we will later describe the best practices for this situation (best protocol) to control the spread of the COVID-19.
The other parameters change within its suitable range, making all its significant private effects, even the value of the reproduction number 0 higher or smaller than one.
The current state and how to stop the spread of Covid-19 in Saudi Arabia
Now, we are in the most critical part of the assessment of the current situation and evaluate what needs to take place later in Saudi Arabia to control the COVID-19 spread. Therefore, in this section, we will apply the SEIQR model to analyze the current situation with new initial conditions and different values of the system parameters according to the current state. We will consider June 14, 2020, as the fresh start, and we will renew all the initial conditions in Table I. The number of infections on this day was ( ) 0 4223 I = 33-35 . We will keep the values of the parameters 1 d and 2 d as it is without any change, while the other parameters will take the values in Table 3: Table 3. It is noted in figure 8 that the spreading of COVID-19 in Saudi Arabia passed through its peak point on 16-18 July 2020, which agrees with the actual data; after that, the spread has slowed down and kept this attitude until the current days, and the reproduction number takes the value 0 0.1 1 = which means the situation is stable According to this curve, we can also see that the number of daily infections on October 18, 2020, will be approximately 600 persons/day, and we can predict that the spreading situation will go to a more stable position and better state.
The ideal protocol to halt the spread Covid-19 in Saudi Arabia
To obtain the ideal situation, which can help us break the spread COVID-19 in Saudi Arabia, we must start implementing the following protocols and procedures (see figure 9): 1. Decrease the value of the transmission rate from the susceptible population to infected but not detected by testing the population to be in the following interval 2. Increase the value of the transmission coefficient from an infected population but not detected by testing to a quarantine population 1 to be 1 0.2 , which means expanding the detection work and the need to isolate infected people in compulsory quarantine areas as an example.
3. Increase the value of the transmission coefficient from the confirmed detected population by testing to a quarantine population 2 to be 2 0.01 , which means we must help the confirmed infected population, which they need to be in the quarantine zone. Five measures are included in the optimal procedure, and guidance has been comprehensive in helping delay the spread of COVID-19 in Saudi Arabia. Prevention is safer than recovery, one of the key targets in this procedure.
The main approach to slowing down the transmission of COVID-19 is to remain home and to put sick individuals in a distant location or a protected place as far as possible.
In order to evaluate the reported infections, we need more reliable and effective diagnostic methods. In addition, awareness raising on ways in which the infection can be confirmed, the disease symptoms and ways.
Ultimately, efficient and appropriate care of sick patients must be given, and medicines, tones, and nutrients must be distributed to non-infected individuals to prevent them. | 4,205.6 | 2020-12-18T00:00:00.000 | [
"Mathematics",
"Medicine",
"Environmental Science"
] |
Pythagorean Fuzzy Digraphs and Its Application in Healthcare Center
,
Introduction
Fuzzy set is unlike ordinary set whose elements of a set have membership value. It is found by Zadeh [1] in the year 1965. Graph theory is a mathematical tool to solve network problem and study the relation between objects (node). Rosenfeld [2] explained the notion of graph theory under fuzzy environment. Later, researchers defined the various operations of fuzzy graph and types of fuzzy graph such as complement of a fuzzy graph and regular fuzzy graph [3,4].
Definition of IFS is a generalized fuzzy set whose elements of a set have both membership and nonmembership value and it is introduced by Atanassov [5][6][7]. e concepts of intuitionistic fuzzy relations (IFR) and intuitionistic fuzzy graph (IFG) are the generalizations of fuzzy graphs (FG) and it is developed by Shannon and Atanassov [8]. Operations on IFG and shortest path problem under intuitionistic fuzzy environment were developed by Parvathi et al. [9][10][11] and Karunambigai et al. [12]. Intuitionistic fuzzy graphs of nth type were developed by Davvaz et al. [13]. Intuitionistic fuzzy graphs of nth type are a generalization of intuitionistic fuzzy graph and intuitionistic fuzzy graph of second type. In research, several extensions of IFG can be seen in recent years [14][15][16][17][18][19][20][21][22][23][24].
However, the use of IFSs is identified in many fields; it has some limitations. Limitation of IFSs confines the truth and false membership value, that is, sum of truth and false membership does not exceed 1. Pythagorean fuzzy set is one of the extensions of intuitionistic fuzzy set. PyFS was developed to overcome the limitations in IFS and this theory is introduced by Yager [25][26][27]where a membership grade (μ) and a nonmembership grade (]) with the condition μ 2 + ] 2 ≤ 1. Pythagorean fuzzy number (PyFN) is developed by Zhang and Xu [28] to interpret the dual aspects of an element.
But Zhang and Xu theory failed to address the decisionmaking problem where the membership grade and the nonmembership grade are, respectively, 0.9 and 0.3; but 0.9 2 + 0.3 2 ≤ 1. PyFG was originally studied by Naz et al. [29] as a generalized notion of IFG and application of the proposed notion was also investigated. Akram et al. [30] recently studied the operations of PyFGs and properties of PyFGs. e concept of planar graph under Pythagorean fuzzy environment was developed by Akram et al. [31]. Notion of maximal product of two PyFGs, residue product of two PyFGs, and its properties have been introduced and studied by Akram et al. [32]. He et al. [33] combined Pythagorean 2-tuple linguistic fuzzy set and QUALIFLEX method. is combination is used to evaluate the full quality of operation personnel in engineering field. Zhang et al. [17] combined the novel TODIM along with cumulative prospect theory under 2-tuple linguistic Pythagorean fuzzy sets (2-TLPFS) and also basic definitions and operators of 2-TLPFSs introduced by Zhang et al. [34]. Li et al. [35] proposed a new similarity measure under Pythagorean fuzzy environment and investigated multiple criteria group decision-making problem to prove the feasibility of the proposed method. e objective of our work in this paper is to introduce some operations on Pythagorean fuzzy digraph, decisionmaking algorithm for solving problem using the notion of Pythagorean fuzzy digraph. Finally, we explore the proposed algorithm with a real-life example. is paper may motivate to study various real-life problems using the proposed algorithm.
Basic definitions which are used in our work are presented in Section 2. Definition of Pythagorean fuzzy diagraph and its operations are presented in Section 3. Algorithm for solving decision-making problem using the proposed concept is developed in Section 4. Also, a decisionmaking problem is considered and solved using the developed algorithm in Section 5. Comparative study is presented in Section 6. In Section 7, conclusion is presented and also discussed the future work.
Basic Definitions
is section contributes to present the basic definitions of [1,2,10,28,30] which we used in our work. roughout the paper, let U be the universal set.
An object A � (a, μ A (a)): a ∈ X is called FS over the universe X, where the mapping μ A : X ⟶ [0, 1] is called the membership function of A for each element a ∈ X.
Let V be a nonempty vertex set. Fuzzy graph is denoted by G � (P, E) where the mappings of a fuzzy set P : V ⟶ [0, 1] on V and the relation E : e notation ∧ represents the minimum operator.
Definition 3 (see [5]). An object A � (a, μ A (a), η A (a)): a ∈ X is called IFS over the universe X, where the mapping μ A : X ⟶ [0, 1] and η A : X ⟶ [0, 1] are called the membership function and nonmembership function of A for each element a ∈ X.
Pythagorean fuzzy set is an order pair (μ p (e), η P (e)) whose first element is positive grade value and second element is negative grade value for each element e in U. Positive and negative grade mapping, respectively, is called the refusal membership.
Let V be a nonempty vertex set. PyFG is denoted by G � (P, E) where the mappings of Pythagorean fuzzy set e notation ∧ represents the minimum operator; ∨ represents the maximum operator.
Pythagorean Fuzzy Directed Graphs
Pythagorean fuzzy digraph is defined in this section and also operations of Pythagorean fuzzy number, score function is presented in this section.
Remark 1. As the name implies, PyFDG does not hold a symmetric relation on V, like a PyFG holding on V.
e operations PyFN are defined as follows: Definition 8. Let P � (μ, η) be PyFN. Its score function and accuracy function can be derived by using the following formulas: Let P 1 � (μ 1 , η 1 ), P 2 � (μ 2 , η 2 ) be two PyFNs. Comparisons of two PyFNs are defined as follows: Journal of Mathematics (1) If score function of P 1 is less than the score function P 2 , then P 1 < P 2 (2) If score function of P 1 is equal to the score function P 2 , then P 1 � P 2 (a) If accuracy function of P 1 is less than the accuracy function P 2 , then P 1 < P 2 (b) If accuracy function of P 1 is equal to the accuracy function P 2 , then P 1 � P 2
Algorithm for Pythagorean Fuzzy Directed Graphs
Shortest path and its length of a graph were calculated using many algorithms such as Dijkstra's algorithm and Bellman-Ford algorithm. We proposed a new algorithm for calculating shortest path from node i to node j and its length of a PyFDG. Vertex and edges of a PyFDG are assumed as a PyFN. e proposed algorithm in this section motivates to investigate the real-life problem.
� source node, 2, 3, . . .n � terminus node} and edge set which connect two vertices by an arrow V � {1, 2, 3,. . .n}. e path of the PyFDG is denoted by P ij and it is defined as P ij � weight of the arc which connects node i to node j . e existence of a minimum of one path P 1i in the digraph is supposed for every i in V-{1}. Pythagorean fuzzy distance along the path is defined as d(p) � d ij .
Algorithm.
Step 1: label the source node as p 1 � (0, 1) Step 2: compute Pythagorean fuzzy value p j , where p j is the minimum of direct sum of p i and p ij for j � 2, 3,. . .,n Step 3: identify the minimum p j in step 2 and then label [p j , i] if p j is reached from node i Step 4: find the shortest Pythagorean fuzzy path from source node to j � 2, 3, 4,. . .,n by combining the label [p j , i] calculated in step 3 and its corresponding p i Step 5: choose the path from source node to destination node and its corresponding p i is the shortest Pythagorean fuzzy length
Application of Pythagorean Fuzzy Digraph
Healthcare center is a center for maintaining, improving, and helping individual's health through diagnosis and treatment by professionals. Healthcare centers are depending on medical professionals, psychiatrists, physiotherapists, dentists, and nurses. Healthcare centers are classified into four types, namely, primary healthcare center, secondary healthcare center, tertiary healthcare center, and quaternary healthcare center. Primary healthcare center is a first point of contact by all patients within the region. General practitioner provides treatment to the patients in this center. If the problem is serious then the practitioner in the primary healthcare center recommends visiting secondary healthcare center. Secondary healthcare center is found in emergency unit in the hospital. In this center, professionals provide treatment for severe injury and emergency medical condition and during child birth also. Secondary healthcare center service is for a short-period of time while the primary healthcare center service is for a day. Tertiary healthcare center provides advanced medical treatment. e patients admitted in this center are mostly referred by primary or secondary healthcare center. Professionals working in this center are specialists for cardiac, cancer, plastic surgery, neuro surgery, and more complex illnesses. Quaternary healthcare center is a national health center because these centers are found only in limited regions. is center provides advanced treatment compared to tertiary healthcare center. Medical practitioner in the primary healthcare center attends the patients; if the health issues of the patients are severe, then he will recommend visiting secondary healthcare center. Also, if the problem is in advanced level, then the medical practitioner recommends the patients to visit tertiary healthcare center. If the health issue of a patient became severe and is in advanced level, then the patient may visit quaternary healthcare center. ere are two possibilities to each patient who visits these healthcare centers either he/she gets recovered or is forwarded to next healthcare center. Positive membership refers to the patient if he/she recovers in that center, and negative membership refers to the patient if his/her health worsens in that center; and then the patient is forwarded to another center. e node 1 and 2 represent the primary healthcare centers in two regions. Node 3, 4, and 5 represent secondary healthcare centers. Node 6 and 7 represent the tertiary healthcare centers. Node 8 is the quaternary healthcare center. Patient in one region visits a particular center; if more patients visit a center, then the medical practitioner in that region recommends visiting the another center. After diagnosing the patients, the practitioner recommends to visit another center depending on the severity of the issue. We constructed a Pythagorean fuzzy network for this case which is shown in Figure 2 and arc weight of this network is shown in Table 2. Presume that a patient visits center 1 with an ailment. e proposed algorithm shows the way to visit center 1 to center 8. e nodes in the above PyFDG (see Figure 2) are the healthcare centers and connection between the healthcare centers is denoted by edges whose weights are PyFN. Weight of each edge is given in Table 2.
We applied the proposed algorithm to find the shortest path from healthcare center 1 to center 8.
Healthcare center 1 is assumed as source node p 1 � (0, 1) and distance is labelled as p 1 � [(0, 1), 1] and center 8 is the terminus node. So, Pythagorean fuzzy value of p j , where j � 2, 3, 4, 5, 6, 7, 8, can be obtained in the following iterations. Shortest path of PyFDG is obtained by working backward from healthcare center 8 and including the permanently labelled healthcare centers from which the subsequent label arose. e shortest path of PyFDG is 1⟶3⟶5⟶8, with the length (0.88944, 0.064).
SP from the healthcare center 1 to healthcare center j given in Table 3 and the thick lines in PyFDG indicate the SP from the healthcare center 1 to healthcare center 8 shown in Figure 3.
Comparative Analysis
Advantages and limitations of existing digraph and also Pythagorean digraph are shown in Table 4.
Conclusion
Pythagorean fuzzy set has been applied in many fields to deal with uncertainty. is set has been applied in graph structure to find the shortest path. But Pythagorean fuzzy set is not discussed for digraph. So, Pythagorean fuzzy digraph is defined and operations on PyFDG are studied. Crisp values are Pythagorean fuzzified for calculation and score function is used for Pythagorean defuzzification. A real-life problem is investigated with the help of the proposed algorithm. e advantage of this work is to handle imprecise edge weight when sum of membership and nonmembership of an edge exceeds one. Future work will be investigating the various complex problems using Pythagorean fuzzy diagraph.
Data Availability
No data were used to support this study.
Conflicts of Interest
e authors declare that they have no conflicts of interest.
Type of digraph Advantages Limitations
Classical digraph [33] is is applicable when arc weights are precise is method is not applicable when arc weights are imprecise Fuzzy digraph [34] is concept can be applied for imprecise arc weights Membership degree in an arc is discussed but the nonmembership degree in the same arc is discussed Intuitionistic fuzzy digraph [35] is notion can be applied to the imprecise edge weight involving membership and nonmembership degree is concept fails when sum of membership and nonmembership degree of an edge weight exceeds 1 Pythagorean fuzzy digraph (proposed model) is environment can deal with imprecise edge weight when sum of membership and nonmembership degree of an edge weight exceeds 1 Hesitancy degree of an edge weight is not discussed in this concept Journal of Mathematics 5 | 3,241 | 2021-09-06T00:00:00.000 | [
"Computer Science"
] |
Conformational Flexibility and the Interaction of Huperzine A in the Active Site of Acetylcholinesterase : A Quantum Chemical and Charge Density Study
Huperzine A is an herbal reversible inhibitor of Acetylcholinesterase (AChE). A molecular docking analysis on Huperzine A molecule has been carried out to understand its structure, conformational flexibility, intermolecular interaction and the binding affinity in the active site of AChE enzyme. Further, the charge density distribution of huperzine A molecule (lifted from the active site of AChE) was determined from the high level quantum chemical calculations coupled with charge density analysis. The binding affinity of Huperzine A towards AChE was calculated from the molecular docking; the lowest docked energy is -8.46 kcal/mol. In the active site, huperzine A molecule interacts with acyl binding pocket-Phe330 of AChE, that is, the bicyclo ring group of huperzine A forms an intermolecular interaction with the oxygen atom of main chain of the amino acid residue Phe330 at the distances 3.02 and 3.25 Å respectively. On the other hand, a gas phase study on huperzine A molecule also performed using HF and DFT (B3LYP) methods with the basis set 6-311G**. The molecular structure, conformation, and the charge density distribution of huperzine A molecule in the gas phase have determined using quantum chemical calculations and the charge density analysis. The comparative studies between the gas phase and the active site forms of huperzine A molecule, explicitly reveals the degree of conformational modification and the charge density redistribution of huperzine A when present in the active site. The dipole moment of the molecule in the active site is 6.85 D, which is slightly higher than its gas phase value (5.91 D). The electrostatic potential (ESP) surface of active site molecule clearly shows Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 23 August 2016 doi:10.20944/preprints201608.0198.v1
Introduction
Scheme 1.Chemical structure of Huperzine A (-)-Huperzine A (Scheme 1) is one of the reported inhibitors of AChE [1].It is an alkaloid isolated from the club moss [1][2][3][4][5][6][7][8][9][10] Huperzia serrata and used as a Chinese herbal medicine.The pharmacologic studies shows that it is a potent reversible inhibitor of acetylcholinesterase exhibits strong anti-cholinesterase activity, which markedly increases the efficiency in learning and memory in animals [2,[11][12][13][14][15].It is a natural, highly selective, reversible, slow and potent inhibitor of AChE, which is being used for the treatment of Alzheimer disease [16][17][18][19][20]. Cholinergic neurodegeneration is the major pathological feature of this disease [21,22].Huperzine A has been used to treat this disease, as it has high potency and low toxicity [23,24].However, the inhibitory activity of the molecule is mainly attributed to the interaction of this small molecule with the neighbouring amino acids present in the active site of AChE, the strength of binding and charge distribution.When the small molecule inhibitor present in the active site of AChE, it interacts with the nearby amino acids of the active site, where its energy changes; in consequence of this effect the molecule adopts a new conformation in the active site.The biological effects are often related to electron density distribution of ligand molecule and its orientation, the capability of forming of intermolecular interactions and the electrostatic potential of the molecule in the active site [5].
The binding strength of this molecule largely depends on the electrostatic parameters and the charge complementarities, which leads to conformational change and the charge redistribution of the huperzine A molecule in the active site of the enzyme.The charge distribution allows to understand the electrostatic moments of the molecule in the active site of AChE.Hence, it enables to predict the orientation of the huperzine A inside the active site.
In the active site, as the conformation of the molecule changes, its orientation also changes.
To precisely understand the active site effect on huperzine A molecule and its behavior in the active site, the molecule needs to be compared with the corresponding gas phase structure and the charge density distribution.In view of this, the structure, conformation, charge density distribution and the electrostatic properties of huperzine A in the gas phase is also essential.The comparative study between both forms (gas phase and active site forms) of huperzine A could reveal the degree of structure and the conformation of huperzine A modified when it present in the active site and the charge density distribution as well.For the past two decades, several studies have been performed to understand the interactions between AChE and Huperzine A using experimental and computational methods [25][26][27][28][29]. Notably, a recent study [30] explores the efficacy of huperzine A from the studies of thermodynamic and kinetic aspects, and concurrently, it also reports that the drug efficacy does not linearly correlate with binding affinity [31].The most possible binding pathway of Huperzine A is gorge of TcAChE, which is also confirmed from energy landscape theory [30].The AChE also exhibits a conformational flexibility [32] and the active site of AChE was rearranged by the influence of huperzine A [33,34].
In the present study, a molecular docking analysis on huperzine A molecule has been carried out which allows to understand its conformational flexibity and the intermolecular interactions in the active site of AChE.Further, a quantum chemical calculation and the charge density analysis were performed for the huperzine A molecule lifted from the active site gives its charge density distribution and the electrostatic properties.On the other hand, a gas phase calculation on huperzine A molecule also carried out to compare its conformation and charge density distribution with the above said active site parameters.This comparative study insights the molecular conformational flexibility and the charge density distribution and the electrostatic properties of huperzine A in the active site, where the amino acids are interact with the huperzine A molecule as well.The charge density analysis of huperzine A molecule has been carried out for the molecule lifted from the active site of AChE obtained from the huperzine A-AChE docked complex.The toxicity also determined from the global reactivity descriptors such as electronegativity, electrophilicity and chemical hardness [35][36][37] using density functional theory [38], which provides the essential information about the stability and reactivity of the molecule [39,40].
Molecular docking and Huperzine A-AChE Interactions
The molecular docking of plant derived huperzine A molecule has been docked the active site of AChE, the docked lowest binding energy of the molecule is -8.46 kcal/mol (Table S1).This energy is relatively lower than the docked binding energy of galanthamine and curcumin [56] with AChE; the low value of binding energy explicity indicates that huperzine A lacks some expected interactions with the AChE.2).S2), these large bond twists in the active site are mainly attributed to the intermolecular interactions and the energy of the huperzine A molecule in the active site of AChE.
Charge Density Analysis
The electron density distribution of huperzine A molecule has been studied using the Bader's theory of atoms in molecules (AIM) [46,47].The electron density ρbcp(r), the Laplacian of electron density ) ( The Laplacian of electron density 2 ρbcp(r) of all bonds of (I) and (II) forms of huperzine A have been calculated (Table 2).All the other bonds in the molecule exhibit less ellipticity (Table 2), reveals the spherical electron density distribution.
Atomic Charges, Dipole Moment and Electrostatic Potential
The atomic charges of both forms (I & II) of huperzine A were calculated from Mulliken (MPA) [64] and natural population analysis (NPA) [65] and Bader's AIM analysis [47] (Table S3).Notably, the Bader's AIM charges are found to be higher than the MPA and NPA charges for all atoms.Generally, the carbon atoms, which are attached to the negatively charged atoms, carry high positive charge.The AIM charges of C-atoms of (I) and (II) forms are: C The predicted active site dipole moment of huperzine A molecule is 6.85 Debye, which is found to be higher than the same found in the gas phase (5.91 Debye), the small enhancement of dipole moment of huperzine A in the active site is due to the charge redistribution in the molecule caused by the effect of intermolecular interaction between the two molecules.Figure 5 shows the superimposed form of (I) and (II) forms of huperzine A molecule, which displays the orientation of dipole moment vectors of huperzine A in gas phase (I) and in the active site (II).The deviation of dipole moment vector of (II) with respect to (I) is solely attributed to the conformational change as well as charge redistribution.The net dipole moment of (I) and (II) forms of huperzine A and its x, y, z component values are presented in Table 3. and N-atoms in both forms of the molecule.Figure 8b is the pictorial representation of intermolecular interaction shows some important interactions; in which, particularly the N(2) atom forms hydrogen bonding interaction with Tyr130.As expected, the O(1) atom forms hydrogen bonding interaction with amino acid residue [1,12] Tyr130; further, the same oxygen O(1) also forms electrostatic interaction [49] with Gly117.
Toxicity Analysis
The toxicity of huperzine A molecule have been analyzed from global reactivity descriptors such as electronegativity (χ), electrophilicity (ω) and chemical hardness (η).In both forms, the ionization potential (I=˗EHOMO) and electron affinity (A=˗ELUMO) of huperzine A molecule were calculated from HOMO and LUMO energy values ( The gobal reactivity descriptors have not much varied in the gas phase as well as in the active site.From this analysis, we have distinctly confirmed that the huperzine A molecule has low toxicity, high stability when the molecule present in the active site of AChE [23,67].
Calculations
To precisely understand the effect of intermolecular interaction of huperzine A molecule in the active site, it is essential to compute the parameters like molecular conformation, charge density distribution and the electrostatic properties of the molecule in the gas phase as well as in the active site.Hence, the computational calculations were performed in two parts.In the first part, the gas phase (at T=298.15K,P=1atm) of huperzine A molecule (I) was optimized from HF [41] and DFT (B3LYP) [42,43] methods with the basis set 6-311G** using Gaussion03 software [44,45]; in which the threshold limit for maximum force and displacement of DFT optimization were converged at 0.000004 and 0.000257au respectively.Further, a charge density analysis has been carried out using Bader's theory [46,47] of atoms in molecules (AIM).In the second part, a molecular docking analysis of huperzine A with AChE has been carried out using Autodock software [48,49].
For the docking analysis, the ligand was obtained by separating the same from the complex domain of pdb accession code 1VOT [1] and converted into pdb format, whereas the threedimensional structure of AChE of "Torpedo californica" was prepared by separating AChE form the complex domain of pdb accession code 1QTI [50] obtained from Brookhaven Protein Data Bank.Autodock generated 10 different conformations and their corresponding docked energies (Table S1).PyMOL [51] software was used to view the huperzine A-AChE complex and the intermolecular interactions between the AChE and huperzine A. Further, a single point energy DFT calculation was carried out for the huperzine A molecule (lowest energy conformer) lifted from the huperzine A-AChE docked complex at B3LYP/6-311G** level; and further a charge density analysis has been performed for the wave function obtained from the single point energy calculation.
From the charge density analysis, the bond topological properties, such as total electron density, deformation density, Laplacian of electron density, bond ellipticity at the bond critical points, eigen values (λ1, λ2, λ3) and the bond path were determined from the Bader's theory [46,47] of atoms in molecules implemented in AIMPAC software [52].The AIM charges were calculated from AIMALL software [53].The deformation density of both forms [(I) and (II)] of the molecule were plotted by wfn2plots and XD software [54].A cube file was generated from Gaussian03, which has been used as an input file for the Moliso program suite [55] to generate the electrostatic potential (ESP) map.
Conclusion
The geometrical and topological properties of electron density of the huperzine A molecule in gas phase and the same lifted from the active site have been compared.This comparative study explores the conformational modification, intermolecular interaction, charge density distribution and the electrostatic properties of the molecule in the active site of AChE.The lowest binding energy of huperzine A in the active site of AChE is -8.46 kcal/mol.In the huperzine A, the O(1) atom forms hydrogen bonding interaction with Tyr130 and electrostatic interaction with Gly117, whereas the nitrogen N(2) has hydrogen bonding interaction with Tyr130 only.The carbon C(8) forms electrostatic interaction with the residue Phe330, which shows that the molecule binds at the acyl binding pocket, which leads huperzine A to inhibit AChE.For the molecule lifted out from the active site, it is found that the geometry of ring 1, 1' are modified, whereas in ring 2, the geometry is not much altered, it is almost intact.Further, the charges are redistributed and it is found high in polar bonds.
The dipole moment of the molecule in the active site is 6.85 Debye, which is higher than its
Figure 1 (
a,b) displays the interaction between huperzine A and the nearby aminio acids present in the active site of AChE.Huperzine A is a three ring molecule, in which the rings 1 and 1′ are the two cyclohexene rings called as bicylo[3.3.1]nona-2,6-diene; and the ring 2 is a pyridone ring.Nitrogen of the pyridone ring forms hydrogen bonding interaction with Tyr130 at the distance 3.21 Å.The distance of the piperidinic nitrogen of huperzine A 1 with the centroids of the aromatic ring of Trp84 is 4.49 Å; the longer distance indicates that the molecule does not exhibit strong interaction with Trp84.And further, the carbonyl group of the huperzine A molecule also forms strong hydrogen bonding interaction with the hydroxyl oxygen of the residue Tyr130 at the distance 2.88 Å.An electrostatic interaction is found between the bicyclo ring and the oxygen of main chain of Phe330 at the distances 3.02 and 3.25 Å respectively.Apart from the above strong interactions, a large number of hydrophobic interactions also present with the side chains and the main chain atoms of Gly119 and His440 residues.Huperzine A has three potential hydrogen bonding sites, but only one interactive hydrogen bond is exist between the pyridone oxygen of the ligand and the hydrogen of the hydroxyl group of Tyr130 at the distance 2.88 Å.The nitrogen in the bicycle ring does not form any specific interaction.The atom C(12) of the same ring has hydrophobic interaction with the nitrogen group of the Gly119 at the distance 3.08 Å.Here, huperzine A interacts only with Phe330; this shows that the huperzine A binds only with acyl binding pocket-Phe330 and not with the choline binding site-Trp84.Even at the acyl binding pocket-Phe330, it has a repulsive type of electrostatic interaction with the carbon atoms[49].The nearest neighbours and the short contact distances below 3.3 Å between huperzine A molecule and the residues of AChE in the active site are presented (Table
Figure 2 (
Figure 2(a,b) shows the ball and stick model of gas phase (I) and the active site form and bond ellipticity at the bond critical point (bcp) were determined for both forms [gas phase (I) and active site (II)] of huperzine A molecule.In figure 3, (a,b) shows the deformation density maps of (I) and (II) forms of huperzine A molecule respectively.The topological properties of electron density ρbcp(r) of Preprints (www.preprints.org)| NOT PEER-REVIEWED | Posted: 23 August 2016 doi:10.20944/preprints201608.0198.v1both forms [(I) and (II)] are presented in table 2. The deformation density map allows to visualize the areas of charge accumulation and the lone pair position of atoms in both forms of huperzine A molecule.A critical point search has been performed for all bonds of the molecule, predicted a (3,-1) type of critical point for all bonds, which confirms the covalent sharing bonds are present in the molecule.The electron density ρbcp(r) of C-C bonds of ring1, 1' are found to be almost similar in both forms of molecule (I)/(II) [1.656/1.683*eÅ -3 ]; relatively, these values are much less than the density of C-C bonds of ring 2, the corresponding average values are: 1.941/1.975*eÅ -3 .Despite the charge transfer from the hydrogen atoms to carbon atoms of the ring 2, large charge accumulation is found near the hydrogen atom [58].Invariably, the charge concentration in C(9)−C(10): 2.304/2.295*eÅ -3 , C(6)−C(7): 2.285/2.430*eÅ -3 , C(3)−C(4): 2.177/2.071*eÅ -3 and C(14)−C(15): 2.202/2.230*eÅ -3 bonds are found to be higher than all other C−C bonds in both forms of the molecule.The electron density of high polar C=O bond in gas phase is 2.74 eÅ -3 ; this value has been slightly decreased [2.569* eÅ -3 ] when it present in the active site.The gas phase value is well agree with the reported values [59-61].The ρbcp(r) value of C(10)-C(12) and C(7)-C(8) bonds which are linked to the methyl group are: 1.703 and 1.696 eÅ -3 respectively; whereas in (II), these values are found to be 1.678* and 1.743* eÅ -3 respectively.The electron density of C(5)-N(1) bond in I and II are: 1.792/1.760*eÅ -3 ; these values are relatively less on compared with the C(3)-N(2)[2.081/1.950*Å -3 ] and C(13)-N(2)[1.923/2.091* eÅ -3 ] bonds in the molecule.The difference of charge accumulation in both C−N bonds is attributed to the nature of different environments.The electron density of these C-N bonds is not much affected in the active site, as the difference of density is found to be very small.
Figure S1 displays the Laplacian of electron Preprints (www.preprints.org)| NOT PEER-REVIEWED | Posted: 23 August 2016 doi:10.20944/preprints201608.0198.v1density distribution of huperzine A molecule.The Laplacian for the C−C bonds of rings 1 and 1' are ranges from -11.9 to -14.8/-11.5* to -15.8* eÅ −5 , the average are -13.5/-13.9*eÅ −5 ; these values are notably less on compared with the Laplacian of C−C bonds of ring 2, the corresponding values are -18.35/-19.1*eÅ -5 .The Laplacian of C=C bonds [C(9)−C(10):-23.9/-23.7*eÅ −5 , C(6)−C(7): -23.3/-26.7*eÅ −5 , C(3)−C(4): -21.9/-19.5*eÅ −5 and C(14)−C(15):-22.6/-23.2*eÅ −5 ] of (I)/(II) are found to be high.In the active site, relatively, the Laplacian of C(6)−C(7) bond was increased, whereas in C(3)−C(4) bond, this value has been slightly decreased.The gas phase (I) Laplacian value of the carbonyl C=O bond is -8.3 eÅ −5 , while the corresponding active site (II) value is, -13.5* eÅ −5 ; the increase of negative Laplacian in (II) indicates that the charges of the bond become concentrated in the active site.In (I), the Laplacian value of methyl group connected C(10)-C(12) and C(7)-C(8) bonds are -14.4 and -14.3 eÅ −5 respectively; whereas in (II), the corresponding values are -14.0*and -15.2* eÅ −5 respectively.In both forms, the Laplacian of C−N bonds are found to be moderately negative (Table 2), indicates, the Laplacian values are not much affected when the molecule present in the active site except C(13)−N(2) bond [C(5)−N(1): -16.5/-15.7*,C(3)−N(2): -18.7/ -17.7* and C(13)−N(2): -18.2/-20.3*eÅ −5 ], in which the charges are become concentrated.Figure 5 displays the variation of charge concentration at the bcp's of huperzine A in the active site of AChE.The deformation density signifies the aspherical nature of electron density of atoms in the molecule, which is attributed to the bond formation.The features such as spherical/aspherical nature of electron density of bonds can be qualitatively interpreted in terms of bond ellipticity [61].Bond ellipticity ε [ε=(λ1/λ2)−1] is the measure of degree of planarity or conjugation of electron density at the bcp, where λ1 and λ2 are the Hessian eigen values of electron density at the bcp's [62].The lower values of electron density and the Laplacian of electron density together with high ellipticity are the indicative of the charge Preprints (www.preprints.org)| NOT PEER-REVIEWED | Posted: 23 August 2016 doi:10.20944/preprints201608.0198.v1migration to the neighbouring bonds [63].In the present study, the ellipticity ε of C-C bonds of the pyridone ring of both forms [(I) & (II)] is ~ 0.16; these values are found to be high, when compared with all other C-C bonds in the molecule (figure 4).The ellipticity of C=C bonds is ~0.3; this high ε value indicates the degree of aspherical electron density distribution.
Figure 6 (
Figure 6(a,b) shows the molecular electrostatic potential (ESP) of the huperzine A
Preprints
(www.preprints.org)| NOT PEER-REVIEWED | Posted: 23 August 2016 doi:10.20944/preprints201608.0198.v1gasphase value (5.91 Debye).Large electronegative ESP regions are found at the vicinity of O and N-atoms in I and II forms of the molecule, in which, the oxygen and nitrogen atoms of (II) are exhibit high negative ESP as it is less in the gas phase.The present active site charge density study of huperzine A has been performed for the molecule lifted from the active site of AChE; hence, as expected it does not include the polarization effect.
Table 4
The global hardness of the molecule is 2.34/2.26*respectively; it reveals that the molecule is stabilized in the active site.The calculated electrophilicity index of both forms is 2.44/2.49*respectively; this value indicates that the charge transfer of the molecule is low. | 4,797.8 | 2016-08-23T00:00:00.000 | [
"Chemistry",
"Biology"
] |
Diagnosis of cardiac arrhythmia using Swarm-intelligence based Metaheuristic Techniques: A comparative analysis
INTRODUCTION: Heart diseases are the prominent human disorders that have significantly affected the lifestyle and lives of the victims. Cardiac arrhythmia (heart arrhythmia) is one of the critical heart disorders that reflects the state of heartbeat among individuals. ECG (Electrocardiogram) signals are commonly used in the diagnostic process of this cardiac disorder. OBJECTIVES: In this manuscript, an effort has been made to employ and examine the performance of emerging Swarm Intelligence (SI) techniques in finding an optimal set of features used for cardiac arrhythmia diagnosis. METHODS: A standard benchmark UCI dataset set comprises of 279 attributes and 452 instances have been considered. Five different SI-based meta-heuristic techniques viz. binary Grey-Wolf Optimizer (bGWO), Ant Lion Optimization(ALO), Butterfly optimization algorithm (BOA), Dragonfly Algorithm (DA), and Satin-Bird Optimization (SBO) have been also employed for the same. Additionally, five novel chaotic variants of SBO have been designed to solve the feature selection problem for diagnosing a cardiac arrhythmia. Different performance metrics like accuracy, fitness value, optimal set of features and execution time have been computed. CONCLUSION: It has been observed from the experimentation that in terms of accuracy and fitness value of cardiac arrhythmia, the SBO outperformed other SI algorithms viz. bGWO, DA, BOA, and ALO. Additionally, BOA and ALO seem to be the best fit when the emphasis is on dimension size only.
Introduction
Heart diseases are the pre-eminent and critical human global problems that have significantly affected individuals and families across the globe.Heart attack, coronary artery disease, hypertrophic cardiomyopathy, arrhythmia, congenital heart disease, heart failure, stroke are some of the major cardiac human disorders [1].Cardiac arrhythmia is a heart disease related to the heartbeat-rate or heart rhythm.The malfunctioning of heart electrical impulses leads to this human disease.A cardiac arrhythmia victim may suffer from the too slow or too fast heartbeat.Tachycardia, Bradycardia, Supraventricular arrhythmias, Ventricular arrhythmia are the major types of arrhythmias.Tachycardia occurs when the heart-beat or heart rhythm is more than 100bpm(beats per minute).Bradycardia occurs when the heart-beat or heart rhythm is less than 60bpm.Tachycardia and Bradycardia refer to the fast and slow heart rhythms respectively [2].In some cases, the severe arrhythmia causes the dizziness, chest pain, fainting, breathlessness, fatigue, prolific sweating etc. [3].In extreme conditions, it may become life-threatening and cause to cardiac-arrest and stroke [4].Human diseases like coronary artery disorder, high blood-pressure, thyroid, diabetes, electrolyte imbalance, stress and overuse of drugs, smoking may lead to this kind of human disorder.
Swarm Intelligence is a term fabricated by Beni and Wang in the late 1980s [5].SI is based on five major principles viz.proximity principle, quality principle, the principle of diverse response, the principle of stability, and the principle of adaptability [6].SI is a nature-inspired phenomenon that entirely relies upon the collective social behaviour of different species viz.insects, a flock of birds, school of fishes, mammals etc.It has been witnessed that various computing methods like Naïve Bayes, Random-Forest, Decision-Tree, SVM etc., deep learning techniques(CNN, RBM, RNN etc.), swarm intelligence techniques(Ant-Colony, Honey-Bee, Firefly Algorithm, Grey Wolf Optimization Algorithm etc.), and fuzzy logic have been used in the prognosis of a wide range of human disorders like diabetes [7][8], cancer [9][10], psychological [11][12], neurological [13] [14] and heart-related problems [15][16].
Features play an important role in early and precise disease diagnosis.The feature is an attribute that represents important characteristics of the dataset.Feature selection is a pre-eminent step of data pre-processing that helps in extracting a subset of features from the native data set.It is a process of uprooting the significant features related to an optimization problem [17].Feature selection aims to extract the best features and to lessen the duplicate features.The feature selection process has been significantly used in diverse fields viz.bioinformatics, image processing, remote sensing, intrusion detection, text processing, disease diagnosis, etc. [18] The feature selection methods are categorized as wrapper, filter, and hybrid methods [19].Filter method selects the features or attributes using statistical approaches.This method is independent of the classification algorithm and has a reasonable cost.Wrapper method is highly complex.It gives the best set of features by deciding which features or attributes are to be added or removed from the dataset.Hybrid methods amalgamate the best traits of the wrapper and filter methods.Information gain, Gini-index, chi-square, corelation, gain ratio are some of the most commonly used attribute selection measures [20].It is found that very few authors have worked on the use of feature selection for cardiac arrhythmia.The work is novel as no one has explored the performance of emerging SI techniques((bGWO, BOA, DA, ALO, SBO, CSBO_1, CSBO_2, CSBO_3, CSBO_4, and CSBO_5) for finding the optimal set of features to solve cardiac arrhythmia classification problems.Here, SI-based techniques have been employed to determine the optimal set of features used to diagnose cardiac arrhythmia.Different performance metrics viz.accuracy, number of dimensions, execution time, and fitness value have been computed and analyzed.
In the rest of the manuscript, related works and methodology used in this research work have been presented in the second and third section.The fourth section presents the results and discussions.Finally, conclusion and future work have been depicted in the fifth section.
Related Works
Nowadays, the stupendous volume of data is accessible for every province that needs to be vigilantly and effectively mined.Feature selection plays a significant role in extracting meaningful information from mountains of data using a minimal subset of attributes.Feature selection is one of the challenging optimization problems that assist in selecting optimal features from a dataset so that a better predictive rate can be achieved.In other words, it is an extraction process that eradicates the irrelevant and redundant features for a better understanding of data sets [21].
Different deterministic and stochastic techniques have been engaged to untangle the feature selection problem.Filter and wrapper are two basic approaches to the same.Despite these techniques, several meta-heuristic algorithms have also been effectively used to get an optimal set of features.In the last few decades, more than a hundred SI-based meta-heuristics techniques have been designed.The major aim of this study is to present a comprehensive analysis of SI-based meta-heuristic techniques used for feature selection problems.From Table 1, it has been revealed that swarm intelligence approaches have been successively employed in the field of disease diagnosis.However, no one has explicitly determined the performance of bGWO, DA, BOA, ALO, SBO, and chaotic variants in solving the cardiac arrhythmia classification problem.
This section of the manuscript will briefly present the basic principles of five swarm-based meta-heuristics viz.butterfly optimization, dragonfly algorithm, grey wolf optimization, ant lion optimization algorithm, satin bird optimization, and chaotic maps(Chebyshev map, Circle map, Sinusoidal map, Gauss map and Tent map).
Butterfly Optimization Algorithm:
BOA relies on the foraging strategy of butterflies.Butterflies possess different senses that help them for their survival.These species make use of their sense of smell to look for the food(nectar).Usually, all the butterflies produce some fragrance which has an intensity associated with it.This intensity corresponds to some fitness value.Butterflies with low intensity get attracted towards the butterflies with high-intensity value [51].In BOA butterflies propagate the information by sensing the fragrance.Butterflies move in a random direction if they are not capable of sensing the fragrance otherwise they move towards the best species if the fragrance is sensed and hence results in global search.In global search the movement of butterflies towards the fittest butterfly is represented by the following equation: where x i t denotes the solution vector x i of ith butterfly in iteration t, g * denotes the currently found best solution among the solutions of the current iteration.f i represents the fragrance of ith butterfly and r is the random number in[0,1].In local search the equation used is: where, x j t and x k t represents the j th and k th butterflies respectively of the same swarm.
Dragonfly Algorithm:
DA is a novel optimization approach proposed by Seyedali Mirjalili in 2015 [52].DA algorithm is based on the social behaviour of dragonflies.Around 3000 distinct species of dragonflies exist in the world.Migration and hunting are the major objectives of dragonflies.Based on these objectives dragonflies possess dynamic and static swarm behaviour respectively [52][53] In static swarm dragonflies create small groups and fly over different directions which is one of the prime concerns of exploration.Whereas, in dynamic swarm dragonflies makes larger groups and fly in one particular direction.This process is called exploitation.Separation, alignment, cohesion, attraction and distraction are the major factors used for position updation of dragonflies in a swarm.These factors are mathematically represented by the following equations [52][53][54]: Separation is evaluated by using the following equation: ) where X and X j represent the positions of current and jth neighbouring individuals and N denotes the number of neighbouring individuals.Alignment is calculated as: Cohesion is calculated as: The attraction of the dragonflies towards the food source is calculated as: ) where X + represents the position of the food source.The distraction of the dragonflies outwards their enemies is calculated as: where, X − represents the position of the enemy.
Grey Wolf Optimization:
GWO propounded in 2014 by Sayedali Mirjalili et al. imitates the hunting nature and leadership behaviour of grey wolves.Generally, the grey wolves come from the Canidae family[17] [55].Based on the hierarchy grey wolves are categorized as alpha(α), beta(β), delta(δ) and omega(ω).Amongst all alphas are at the top of the hierarchy as they are dominant ones.Alphas take all decisions and the rest of the wolves obey them.Beta wolves help alpha in decision making.Delta wolf assists both alpha and beta wolves.Omega wolves are the lowest-ranked and follow the instructions of α, β, δ.In the hunting process the grey wolves surround their prey which is mathematically represented by the following equations [55] [56][57]: where t denotes the current iteration, X p ����⃗ and X � �⃗ represents the position vectors of prey and grey wolf respectively.A � �⃗ and C �⃗ are the coefficient vectors and are calculated as : where r 1 and r 2 are the random vectors in [0,1] and components of a �⃗ are linearly decreased from 2 to 0 throughout iterations.
Figure 4. Positions updation in GWO[55]
In GWO the wolves generally hunt their prey in a swarm of 5 to 12 called pack.The hunting process is lead by alpha, beta, and delta wolves.However, alpha wolves are considered as the best candidate solutions, beta and delta wolves have finer knowledge about the potential position of the prey.So, the three best solutions achieved so far are saved and the other search agents are required to update their locations according to the location of the best search agents.This process is implemented by using the following equations: In the binary version of grey wolf optimization called bGWO, the solution space is restricted to{0,1}.
Satin-bowerbird Optimization:
Satin bowerbird is a novel optimization algorithm that is propounded by Moosavi and Bardsiri in 2017 [58][59].The fascination of male bowerbirds towards the female birds for procreation is the key idea behind the success of this meta-heuristic.Satin bowerbirds are the insect and fruit-eating passerines that are natives of mesic forest and rainforest of eastern Australia.Male birds usually build gazebo called the bower and decorate it with full zest to woo the female birds.Male bowerbirds have been observed competing to make their bowers more ravishing than that of others.This intention provokes them to demolish the bowers of their adjacent males.Male bowerbirds mostly use the different coloured viz.yellow, white, purple objects of which are normally seen placed at the doorway of their bower The fitness value is evaluated using the following equations: where f(x i ) represents the value of the cost function of ith bower or ith position.
Ant Lion Optimization Algorithm
Ant Lion Optimizer is a swarm-intelligence based algorithm introduced by Mirjalili [62].Antlions belong to a family of insects called Myrmelentidae [63].ALO imitates the hunting strategy of antlions.Antlions catch their prey specifically ants by digging the cone-shaped pits in the sand.After digging the pit antlions hide at the bottom of the pit and wait for the prey to trap.The size of the pit is based on the level of hunger of antlions and the shape of the moon.It is observed that antlions dig the larger pits as they feel more hungry.The behaviour of antlions and ants is depicted mathematically as given below [62][64][65]: Ants move randomly in search of their food and their random walk is represented by the following equation: ) where cumsum evaluates the cumulative sum, n denotes the maximum number of iterations, t represents the step of random walk and r(t) is the stochastic function defined as follows: rand is a uniformly distributed random number in [0,1].The objective is to the random walks within the search space, so these are normalized using the following equations: where a i and b i denotes the minimum and maximum of random walks of ith variable respectively.c i t and d i t denotes the minimum and maximum of ith variable t th iteration.The trapping process of ants in the pits of antlions is mathematically represented as follows: t represents the position of the selected j th antlion at t th iteration.When the antlions come to know that ants are trapped in the pit they start putting sand outwards the middle of the pit.This process is mathematically formulated as : where I is a ratio.Now, the process of capturing of the ants by the antlions and reconstructing the pits to grab the new preys are described by the following equations: where Antlion j t and Ant i t represents the jth and ith positions of selected antlion and ant at iteration t.
Chaotic Maps 3.7
Chaos theory is a part of mathematics that solves the nonlinear complex problems whose behaviours are implausible to predict.Chaos refers to the havoc behavior of the non-linear dynamic systems.Some of the chaotic maps employed in this manuscript are discussed in Table 2 [66][67]: where a=0.5, b=0.
Results and Discussions
A standard UCI cardiac arrhythmia dataset has been taken for experimentation.There are 452 instances and 279 attributes.Out of 279 attributes, 206 are linear and the rest are nominal.The data has been pre-processed to remove the missing values.The dataset is examined to find a set of optimal features required to determine the state of arrhythmia.The binary output reflects the presence or absence of cardiac arrhythmia.
All the experiments have been conducted over a machine (CPU: i3, RAM: 2GB).All the algorithms have been simulated using MATLAB 2016 environment.Initially, all the search agents have been assigned random locations in the search space.The values of the upper and lower bound have been set to 1 and 0 for the cardiac arrhythmia data set.The numbers of search agents are set to 10.For the classification problems, the solution having the least value of features is considered to be optimal.To overcome the bias in stochastic techniques, each algorithm has been individually executed for twenty different runs and the average of the results have been taken.
The SI-based meta-heuristic algorithms viz.DA, BOA, ALO, SBO, and bGWO have been employed to find an optimal set of features required for cardiac arrhythmia diagnosis.To examine the performance of different SI algorithms various metrics like accuracy, dimension size, fitness value and execution time have been computed and analyzed.The equations of parameters used are mentioned below: ) where M denotes the number of runs and g* represents the optimal solution The parameters for execution have been empirically set.The results obtained using DA, ALO, SBO, BOA, bGWO and chaotic versions of SBO have been statistically examined.The metaheuristic techniques that provide higher classification accuracy found to be more promising than others.However, for fitness value, number of dimension, and execution time the smallest value corresponds to better results.The results obtained during 200 experiments have been presented in Tables 4 and 5.The best values obtained using DA, BOA, ALO, SBO for accuracy, fitness value, number of dimensions and execution time for cardiac arrhythmia dataset presented in Table 4 and Table 5 have been underlined and italicized.Here, Avg, max, min, std corresponds to the average, maximum, minimum and standard deviation of values.accomplished using simple SBO is 68%.SBO is also performing well in terms of fitness values and number of dimensions.Additionally, in terms of execution speed, the performance of CSBO_2 is outstanding.Furthermore, the minimum accuracy of SBO is 5.17%, 5.17%, 5.17%, 7.01% and 5.17%, better than CSBO_1, CSBO_2 CSBO_3 CSBO_4 and CSBO_5 respectively.Likewise the average and maximum rate of accuracy of SBO is (3.22%, 3.03%), (4.91%,7.93%),(3.22%, 3.03%), (4.91%, 3.03%) and (4.91%, 3.03%) respectively better as compared to CSBO_1, CSBO_2 CSBO_3 CSBO_4 and CSBO_5.
Figure 6. Variation of the rate of accuracy and number of dimensions of SI-based meta-heuristics
The rate of accuracy and the number of dimensions obtained using ten distinct SI metaheuristics techniques during twenty different runs (experimentation) are depicted in Figure 6.
Conclusion
Cardiac arrhythmia is one of the critical heartbeat related human disorders which may lead to another censorious heart-related problem, in case it not diagnosed and treated on time.A standard UCI dataset(ECG signals) comprises of 452 individuals has been explored during this research work.In this manuscript, five emerging swarm intelligence based meta-heuristic techniques and chaotic variants of SBO have been employed as the feature selection techniques and their results are compared.Here, five distinct variants of SBO have been created by hybridizing the characteristics of SBO and different chaotic maps viz.circle map, chebyshev map, sinusoidal map, tent and gauss map.It is found that SBO outperformed all SI approaches as well as its chaotic variants in terms of accuracy and fitness value.Furthermore, the minimum accuracy of SBO is 7.01%, 8.92%, 12.96%, 5.17% better than bGWO, DA, BOA and ALO respectively.Likewise, the average and maximum rate of accuracy of SBO are (3.22%,1.49%), (4.91%,4.61%),(6.66%, 6.25%) and (3.22%, 3.03%) respectively better as compared to bGWO, DA, BOA and ALO.In the future, more chaotic functions can be utilized on the cardiac arrhythmia dataset as the feature selection approaches.Additionally, the use of these meta-heuristic techniques in the diagnosis of other cardio disorders may also be explored.
Figure 3 .
Figure 3. Position updation principles ofdragonflies[52] [59][60].Male bower birds use different kinds of materials for building attractive bowers viz.twigs, flowers, brown-coloured shells of snails, feathers, drinking straw, etc.[59][61].Female bowerbirds choose their breeding partner after visiting numerous bowers.Based on the virtue of the lives of satin bowerbirds the SBO algorithm is organized into various stages viz.Generation of a set of Random Bowers, Probability Calculation of Each Member of Population, Elitism, Calculating bower's new Position, and Mutation.The probability and the fitness values are evaluated using the following equations[59] [61]: where nb represents the number of bowers and fit i corresponds to the fitness value of the ith solution.EAI Endorsed Transactions on Pervasive Health and Technology 05 2020 -09 2020 | Volume 6 | Issue 22 | e7
Figure 5 .
Figure 5. Ant's random walk inside antlion's trap[62] 2 and x n ∈ (0,1) EAI Endorsed Transactions on Pervasive Health and Technology 05 2020 -09 2020 | Volume 6 | Issue 22 | e7 map, Sinusoidal map, Gauss map and Tent map), the performance of original SBO found to be the more suitable for classification of the cardiac arrhythmia dataset.The maximum rate of classification accuracy EAI Endorsed Transactions on Pervasive Health and Technology 05 2020 -09 2020 | Volume 6 | Issue 22 | e7
Table 1 . Related works Authors Domain Purpose SI-based Technique Used
[33]da R.H. et al.[33]Breast Cancer Feature Selection and ClassificationANN+Fuzzy Logic+GA
Table 4 .
Statistical Analysis of parametersIt has been found fromTable 4 that the minimum and maximum values of accuracy ranges between 0.54 to 0.68.Likewise, the fitness values lie between 0.32 to 0.46.Moreover, as far as accuracy rate and fitness values of cardiac arrhythmia are concerned the SBO outperformed other SI algorithms viz.bGWO, DA, BOA and ALO.In case only dimension size is of utmost importance, then BOA and ALO algorithms are on priority.The values of execution time are ranging from 22.95 to 152.81.In terms of execution time, the performance of bGWO is on top.
Table 5 .
Statistical Analysis of SBO and its chaotic versions.
Table 5
depicts the comparison of SBO with its chaotic versions.The experimental analysis reveals that in comparison to the distinct chaotic variants designed using five different chaotic functions (Chebyshev map, Circle | 4,678.8 | 2020-09-22T00:00:00.000 | [
"Computer Science"
] |
Cluster based Routing Protocols for Wireless Sensor Networks: An Overview
—Energy consumption of nodes in Wireless Sensor Networks (WSNs) is a very critical issue, particularly in scenarios where the energy of nodes cannot be recharged. Optimal routing approaches play a key role in energy utilization, so there is great importance of energy efficient routing protocols in WSNs. Energy efficient routing protocols in WSNs are categorized into four schemes, namely (i) communication model, (ii) topology based model, (iii) reliable routing, and (iv) network structure. Network structure category is further divided into flat and cluster-based approaches. This work focuses on a subtype of “network structure” scheme known as clustered based routing protocols, which are mainly used in WSNs for reduction in energy consumption. This work reviews and provides an overview of prominent cluster based energy efficient routing protocols on the basiss of some primary performance metrics such as (i) energy efficiency, (ii) algorithm complexity, (iii) scalability, (iv) data delivery delay, and (v) clustering approach. Finally, this work discusses some latest research trends with respect to cluster based energy efficient routing protocols in WSNs.
I. INTRODUCTION
Since the last twenty years, there is a rapid growth with respect to technologies in the field of data communication networks.This technological progress facilitates organizations by providing very easy and secure working environment.Wireless networks enable organizations to get rid of expensive procedure of using cables for the purpose of connecting equipment located at different locations.This motivates organizations to use wireless networks for communication purpose.
From topological perspective, wireless networks are commonly classified into two modes i.e., (i) infrastructure mode, and (ii) ad hoc mode.The former supports communication between nodes through a Base Station (BS), while in ad hoc mode, all nodes can communicate with each other directly without requiring any infrastructure and no node is superior to any other node in the absence of any central entity.
A Wireless Sensor Networks (WSN) is a set of low-cost and small-sized sensor nodes having limited communication range, energy, processing, and storage capacity.From network design perspective, WSNs are classifies into structured and unstructured networks.In the former, the deployment of nodes is made with proper planning while in latter the same is done in an ad hoc manner [1].
A wireless sensor network is a combination of various sensor nodes connected with each other.Physical location where these nodes are deployed is known as a sensor field.Data from any node is transferred to other linked nodes and aggregated at sink node in order to be accessible to end users as shown in Fig. 1 [2].
Each sensor node has four major hardware components i.e., (i) sensing unit, (ii) a processing unit, (iii) transmitting / receiving unit, and (iv) power unit.Each sensor comprises of application dependent two additional components, namely (i) location finder system, and (ii) mobilizer.Sensor and analog to digital converter (ADC) are two sub-parts of sensing unit.Initially, the data is observed by a sensor which is forwarded to ADC for conversion into digital form.Then, digital data is sent to the processing unit, which is usually linked with a storage unit consisting of a small storage capacity.In order to perform assigned activities, sensor units cooperate with each other by using procedures organized by processing unit.The transceiver helps a node in connecting with the network.Power unit, considered as the most important part, provides power to all the remaining units.The power may be provided through solar cells or by using power generator system (refer to Fig. 2) [2].
Sensors used in WSNs have various kinds like acoustics, seismic visual, low sampling magnetic, thermal, radar and infrared.These sensors are capable of monitoring several conditions such as noise level, soil makeup, lightening, vehicular movement, temperature, humidity and pressure etc. WSNs have various application areas i.e., performance monitoring of industrial machines, environmental monitoring, monitoring of health and military battlefield [3].Routing protocols are of great attention because of versatilities lying in network architecture as well as in applications using wireless sensor networks.Traditional routing protocols are not applicable in WSNs; therefore WSNs require routing protocols different than traditional ones.Consequently, many energy-efficient routing protocols were developed for WSNs for providing efficient delivery of to the destination.Keeping in view the nature of application and network architecture, each energy-efficient routing protocol may possess specific features.
This research work provides an overview of existing energy-efficient cluster based routing protocols in the context of WSNs.Moreover, it lists a brief comparison of the studied protocols.The rest of the paper is organized as follows: Section II describes the related work.Section III provides the review of the cluster based hierarchical routing protocols of WSNs.Section IV gives comparative analysis of prominent hierarchical clustering protocols in terms of performance metrics and final Section V concludes the study.
II. RELATED WORK
Many surveys have already been conducted in the area of WSNs energy efficient routing protocols, from different perspectives.However, this paper reviews hierarchal / clusterbased routing protocols in WSNs.
In 2004, Al karaka and A E Kamal [4] surveyed routing techniques in WSNs.This work provided a taxonomy about WSN routing protocols by dividing them into two major categories (i) network structure and (ii ) protocol operation.The network structure is further divided into flat, hierarchal and location-based routing.The location based is further divided into (i) negation based, (ii) multipath based, (iii) query based, (iv) QoS based, and (v) coherent based routing.This work also exposed some design tradeoff between communication overheads and energy saving in some routing protocols paradigm.
In a survey in 2005, Kemal Akkaya et al. [5] classified WSNs routing protocols as data-centric, hierarchal and location-based.Each protocol was placed in one basic category, while a few protocols belonged to more than one class conserving various metrics such as QoS, network flow, and data aggregation.All of the clustering protocols were not discussed in this work.
In 2007, Abbasi et al. [6] presented a taxonomy of different clustering schemes and provided an overview of clustering protocols and algorithms from the perspective of variable convergence time and constant convergence time.Moreover, their study provided a comparison of some popular clustering methods.
In 2008, Deosarkar et al. [7] discussed cluster head (CH) selection techniques on the basis of classification as (i) deterministic, (ii) adaptive, and (iii) combined metric schemes.The authors compared the cost of CH selection from various angles like cluster information, creation, and distribution of clusters.
In 2010, Shio Kumar Singh et al. [8] described cluster based energy efficient routing protocols in WSN.The authors highlighted some pitfall and disadvantages of individual protocols along with some future trends and constraints lying in this area.
In 2012, Xuxun Liu [9] comparatively expressed a better survey on cluster-based energy efficient WSN routing protocols.The author developed a novel taxonomy about clustering methods on WSN rather than detailed clustering attributes.This work analyzed some prominent clustering routing protocols in WSNs and compared them through different approaches as discussed in the taxonomy about the cluster (refer to Fig. 3).The author described three clustering approaches i.e., (i) centralize, (ii) distributed, and (iii) hybrid.Centralize clustering approach is responsible for making clusters and CH selection.Distributed approach allows all cluster nodes to work as CH for the current round.Hybrid approach combines the properties of both centralize and distributed approaches.
In 2013, Nikolaos A. P et al. [10] presented a detailed survey on overall energy efficient WSN routing protocols by dividing them into four main categories on the basis of energy efficiency, nemly (i) network structure, (ii) communication model based, (iii) topology based, and (iv) reliable routing based.The first scheme is further divided flat and cluster based approaches.The second scheme is classified in three subtypes i.e., (i) query based, (ii) non query based/negation based, and (iii) coherent based.The third scheme is further divided into location-based and mobile agent based ones.The fourth scheme is divided into QoS based and multipath based schemes.Fig. 4 presents the complete picture of their division.The present work focuses cluster based routing protocols in detail.
In 2014, Agam Gupta and Anand Nayyar [11] discussed many routing protocols.Traditional routing protocols being used in WSNs lack in load balancing and efficiency of energy.The use of clustering not only improves network life time but also supports load balancing.There are many clusters and each cluster consists of many inter-connected sensor nodes, while one of them works as a cluster head (CH).Each cluster head gathers data from the nodes belonging to the cluster and transfer that data to the BS (refer to Fig. 5).There is intra cluster as well as inter cluster data communication between cluster head and member nodes of the cluster [11].www.ijacsa.thesai.orgIn 2015, Santar Paul Singh and SC Sharma [12] conducted a survey on cluster based energy efficient WSN routing protocols.The authors described taxonomy of WSN routing protocols into five categories, namely (i) initiator of communication, (ii) Path establishment, (iii) Network structure, (iv) protocol operation, and (v) next hope selection.Network structure scheme is Specifically further classified into (i) flat, (ii) cluster based, and (iii) location based.The authors in this survey did not review clustering protocols individually.They classified cluster-based routing protocols into three sub-categories, i.e., (i) block cluster-based, (ii)grid cluster based, and (c) chain cluster based.According to this classification, different clustering protocols lie under these three clustering schemes.In the end, the authors discussed some merits and limitations of some prominent cluster-based routing protocols.
In 2015, Priyanka Sharma and Inderjeet kaur [13] discussed WSNs routing protocols by classifying them into three main categories, i.e., (i) path establishment, (ii) network structure, and (c) protocol operation.First scheme path establishment is further divided into proactive, reactive and hybrid.Second scheme network structure is further divided into flat, hierarchal and location based.The third scheme is further classified into eight sub-types, namely (i) query, (ii) bio-Inspired, (iii) multipath, (iv) negation based, (v) QoS, (vi) non-coherent, (vii) coherent, and (viii) mobility.The authors discussed some metrics, pros, cons, and applications of some clustering protocols lying in above-mentioned categories.
In 2015, Ibrihich Ouafaa et al. [14] discussed and compared some prominent cluster-based routing protocols by classifying them into WSN and ad-hoc categories.The authors also compared these prominent protocols considering some important metrics.
In 2016, Yan et al. [15] classified WSNs routing protocols into data-centric, location-based and hierarchal depending on network structure.In data-centric approach, metadata approach is used by the protocols to sense and transmit information to base station.Hierarchal approach adopts clustering technique which can be made by grouping sensor nodes.The cluster reduces the energy utilization of sensor nodes.Clustering technique is more scalable and is used in a number of various applications.The location-based approach uses position/ location of nodes to route the data intelligently.In 2017, Syed Bilal Hussain Shah et al. [16] conducted a survey on hierarchal routing protocols in WSNs to increase network lifetime and conserve energy.In this survey, they reviewed some of the hierarchal routing protocols like LEACH, LEACH-TLCH, APTEEN, TEEN and proposed new scheme adaptive threshold.They also discussed some limitations of LEACH and some merits of newly proposed adaptive threshold.Adoptive Threshold attained good results as compared to some of the previously discussed schemes.But the authors did not studied or compared all hierarchal/ clustering protocols with the newly proposed scheme.
In 2018, G. Beni and C. Selden Christopher [17] discussed a few cluster-based protocols like LEACH, PEGASIS, HEED, TEEN and APTEEN from the perspective of comparing performance metrics like energy efficiency, cluster stability, delivery delay, and scalability.But the authors did not present a detailed survey on all hierarchal / cluster-based protocols of WSN.
Our work, to the best of our knowledge, is a more comprehensive study covering maximum number of famous hierarchal /cluster based protocols of WSNs with different clustering approaches.
III. CLUSTER-BASED ROUTING PROTOCOLS IN WSNS
This section provides an overview of prominent energy efficient cluster-based hierarchal routing protocols of WSNs.In cluster-based routing protocols, multiple nodes connected with each other in a sensor field make a group having one node among them as a cluster head.Data transformation from that particular cluster to sink node occurs through the cluster head (CH).In this way, the energy of other nodes is saved.So in this classification of protocols, the major aspect is clustering.Fig. 6 depicts prominent cluster based / hierarchal routing protocols in WSNs while details of these protocols are provided in subsequent sub-sections.
A. LEACH
In 2000, Heinzelman et al. [18] proposed one of the famous cluster based routing protocol for WSNs namely "Low Energy Adaptive Clustering Hierarchy" (LEACH).LEACH evenly distributed the load of energy between all sensors of the network using random based rotation of cluster head.In order to make dynamic networks more scalable and robust, LEACH used localize coordination.In LEACH, sensor nodes scattered in field organize themselves to make local clusters, among these sensor nodes one node becomes local base station or cluster head (CH).This CH works as router, transfers the signals from all sensor nodes to the sink.LEACH saves energy due to transfer of data by CH rather than transferring of the data individually by all sensor nodes.Optimal number of nodes considered as cluster head are about five percent of the total nodes.In LEACH, all data processing including "data aggregation" and "data fusion" is held locally in the cluster which results in reduction of energy dissipation as well as increasing life time of the system.Cluster head is changed randomly, so energy dissipation between all nodes becomes balanced.CH changing decision is made by randomly selecting a number between 0 and 1.
If selected number (shown in Eq.1) is below the threshold, node may become the CH for some specific round.
In Eq.1, variable "P" is desired percentage of the CH ( for example .05),variable "r" means current round whereas variable "G" means a set of those nodes which are not selected as CH from last 1/P rounds.
The authors claimed that LEACH reduced 8 times energy as compared to direct communication (DC) and minimum transmission energy (MTE) protocol.LEACH, being a single hop routing protocol, possesses some deployment limits in larger networks where every individual node directly communicates to cluster head and sink.
B. TEEN
In 2001, Arati Manjeshwar and Dharma P. Agarwal [19] proposed "Threshold Sensitive Energy Efficient Sensor Network" (TEEN).According to authors TEEN was the first protocols of its time developed for "Reactive Networks".In reactive networks nodes immediately react to drastic and sudden changes in value of sensed attributes.In this protocol the CH broadcasts to its member nodes, at change time of every cluster, in addition to its attributes.TEEN was a combination of data-centric and hierarchical approach.Number of transmissions from member sensor nodes to CH are reduced in this protocol.
Two kinds of threshold values to ordinary member sensor nodes work in TEEN, i.e., (i) when a cluster is formed, and (ii) when CH broadcasts.The first value is called hard threshold (Ht) and second is called soft threshold (St).
Hard threshold (Ht) is sensed attribute's absolute/ minimum value, at which sensing node should turn on its transmitter for reporting to its CH.Soft threshold (St) is small change occurred in sensed attributes, which causes to dictate the node to switch on the transmitter for the transmission purpose.
TEEN saves large amount of energy by reducing number of transmissions between cluster head and member sensor nodes.
The main shortcoming of TEEN is detection of dead nodes.Another limitation of TEEN is that it is difficult to forecast the reason if node is not sending the data.This can be happened because of two main reasons (i) node is unable to meet threshold value, and (ii) node may be dead.TEEN is suitable for time critical applications and suitable for energy consumption and response time [19].
C. APTEEN
In 2002, Arati Manjeshwar and Dharma P. Agarwal [20] proposed an improvement to overcome deficiencies of TEEN named as "Adaptive Periodic Threshold Sensitive Energy Efficient Sensor Network" (APTEEN).LEACH was considered suitable for proactive networks and TEEN was suitable for reactive networks.In APTEEN, the authors used hybrid network approach which had combined best features of both LEACH and TEEN.APTEEN was suitable for time critical events as well as to obtain data periodically.Simulation results showed that network lifetime and energy dissipation of APTEEN existed between TEEN and LEACH.
D. PEGASIS
In 2002, Stephanie Lindsey et al. [21] proposed a chain based protocol namely "Power-Efficient Gathering in Sensor Information Systems" (PEGASIS) which was an improvement over LEACH.In PEGASIS each node sends and receives data to only nearby neighbor nodes.Data reaches to the base station in turns, due to which energy consumption per round is reduced.All member sensor nodes are connected with each other in such a way that they make a chain.Using greedy algorithm, chain computation may be initialized by broadcasting data from a node or base station to all member sensor nodes, as shown in Fig. 7.
E. HEED
In 2004, Ossama Younis et al. [22] proposed protocol named as Hybrid, Energy-Efficient Distributed Clustering (HEED) which was very excellent clustering based protocol.
In HEED node's residual energy is parameter to elect the CH.Node's degree or density is used as a metric, in selection of a cluster to get power balancing.HEED was mainly an improvement over LEACH.
F. SEP
In 2004, G. Smaragdakis et al. [23] proposed a protocol named as Stable Election Protocol (SEP) for cluster-based Hetero-genius wireless sensor networks.SEP is an improved version of LEACH and works like it.SEP prolongs time interval known as "stability period" of the first node before its death, which is very crucial for such kind of applications where feedback from sensor network is considered very reliable.In SEP cluster head (CH) is elected on the basis of energy as a "parameter".In this protocol a node independently selects itself as CH on the base of its own initial energy.SEP depends upon each node's weighted "election probability" to make CH with respect to each node's "remaining energy".From simulation results, it can be concluded that SEP had longer period stability and greater average throughput as compared to existing clustering based heterogeneous oblivious protocols.The study also showed that SEP is more resilient in advance node's energy efficiency.
G. BCDCP
In 2005, Siva D. Muruganathan et al. [24] proposed Base-Station Controlled Dynamic Clustering Protocol (BCDCP).BCDCP maintains clusters and routing paths by utilizing high energy base station.In this protocol cluster head rotations are performed randomly.Intensive energy tasks are performed in this protocol.In BCDCP main concept is formation of balance clusters where every cluster head serves an equal number of member sensor nodes by avoiding cluster head overload.In overall sensor field there is uniform placement of cluster heads.Data is transferred to the base station by utilizing CHto-CH routing.
H. DWEHC
In 2005, Ping Ding et al. [25] proposed a protocol namely Distributed Weight based Energy Efficient Clustering (DWEHC), in which every node in its enclosure region, first of all, locates its neighbor, calculates weight of itself that depends upon two factors namely, (i) its distance from the neighbor node and (ii) residual energy.In that enclosed region, a node having maximum weight is selected as a CH.Other neighbor nodes become member node under this CH hierarchy.This clustering process finally terminates after seven iterations.This clustering process has no dependency on the size and topology of the network.The authors showed through simulation that this protocol performed well.The performance of this protocol was also analyzed from the perspective of Inter-cluster and intracluster communication.Finally, authors compared DWEHC performance with HEED-AMRP algorithm and concluded that it outperforms HEED-AMRP in "energy consumption" and better "cluster generation" perspectives.www.ijacsa.thesai.org
I. EECS
In 2005 and 2006, Mao Ye et al. [26,27] proposed Energy Efficient Clustering Scheme (EECS) for WSNs.EECS is a clustering algorithm that is more suitable for periodical data gathering applications.EECS is like a LEACH.In this scheme sensor network is divided into many clusters and data communication between CH and BS is single-hop.In this scheme, a node candidate to become a CH competes for a given round to gain the ability to elevate the CH.During this competition, CH candidate nodes broadcast residual energy among their neighboring CH candidates.In this duration, if a node under consideration could not get another node with more residual energy than it, then itself becomes the CH.Cluster formation of EECS is different from LEACH.EECS improves in the capability to LEACH by introducing dynamic size in clusters which are based on "distance" of the cluster from BS.
J. CCS
In 2007, Sung-Min Jung et al. [28] proposed a protocol as Concentric Clustering Scheme (CCS), to improve the performance of PEGASIS.As in PEGASIS, data transmission is redundant because when one node is selected as a head node regardless of the base station it takes data from both side nodes to convey it to BS, so data transmission from head node to BS become redundant.To cope with this issue of redundancy, CCS protocol was proposed.The main concept of CCS is to consider the location of the base station so that the lifetime and performance of WSNs can be prolonged.This became possible to achieve, using CCS.In CCS, WSNs are divided into concentric shaped clusters to give data transmission flow.CCS enhanced performance by using four processes, namely (i) level assignment to each node relevant to base station, (ii) constructing chain in level area using greedy algorithm, (iii) constructing chain between head nodes of each level, and (iv) transferring of data from higher level head node to other lower level head node.CCS saved 35 % energy as compare to PEGASIS.
K. HGMR
In 2008, Dimitrios Koutsonikolas et al. [29] proposed a protocol named as Hierarchical geographic multicast routing (HGMR).HGMR is basically location-based and multicast protocols for WSNs.Impeccably it incorporates innovations in the "locations based" & "multicast" and it optimizes them for WSNs.HGMR performs this by concurrently providing scalability and energy efficiency to the networks of large size.It can be concluded from simulation results that HGMR has combined the strength of two protocols namely Hierarchal Rendezvous Point Multicast (HRPM) and Geographic Multicast Routing (GMR).HGMR protocols decompose multicast groups into the subgroups.It uses GMR's local multicast scheme to forward the data packets with multi branches of a Multicast tree in the single transmission.In HGMR, using mobile Geographic Hashing mechanism, multicast groups can be divided into their subgroups.The deployed area is divided into partitions of various equal sized square shaped subdomains called "Cells" and each cell is consisted of subgroups of members having managable size.There exists an access point (AP) in each cell which is responsible for all members of that cell.A Rendezvous Point (RP) manages all the APs.
L. PANEL
In 2007 and 2010, Levente Buttyan and Peter Schaffer [30,31] proposed a protocol named as Position-based Aggregator Node ELection (PANEL) for WSNs.There exist some other aggregator node election protocols but PANEL has a novelty from them in a sense that it supports such kind of sensor network applications which are asynchronous.The sensor collects the reading information through the base station after some delay.Main motivational factor in the design of PANEL was its support for reliable and consistent application of data storage like TinyPEDS.PANEL deals with load balancing and also supportx intra and intercluster routing by allowing communication between sensor and aggregator, aggregators itself, an aggregator to BS and BS to aggregators.Cluster formation and energy consumption capabilities of PANEL are better than HEED.Following are key merits of PANEL: (1) PANEL is energy efficient ensuring load balancing due to the election of each node as an aggregator (2) Beside synchronous scenes this protocol also supports the asynchronous application.
Following are key limitations of PANEL: (i) The supposition that cluster formation is found/ determined before deployment thus cannot be applied upon WSN dynamics, (ii) it has information about the geographical position of the nodes, that is used to find which node must be an aggregator.In WSNs there is a constraint that geographical position is not always available except some special conditions like hardware and software having GPS feature, and (iii) an assumption about PANEL described by the authors is that within cluster nodes form a subnetwork, due to this there may occur such a situation that nodes within the cluster could not hear the announcement of nodes closest to reference point, and they may elect aggregator to another node.
IV. COMPARATIVE ANALYSIS AND RESEARCH CHALLENGES
Table 1 shows a comparative analysis of various clusterbased protocols (discussed earlier in Section III) on the basis of different performance metrics, namely (i) energy efficiency, (ii) algorithm complexity, (iii) delay in data delivery, (iv) scalability, and (v) clustering approach.A tradeoff was observed in terms of energy efficiency and data delivery delay, i.e., BCDCP is very poor regarding energy efficiency but offers small delay.It was also observed that some protocols perform much better in terms of scalability; however, their performance is lower if other metrics are taken into account i.e., scalability of HGMR is very high while having very poor energy efficiency.It is worth mentioning that almost all selected protocols in this review follow the distributed clustering approach.Algorithm complexity is noticed from very low to very high.
The following research challenges require attention from research community The design of energy efficient cluster based protocols for wireless body area networks for the purpose of improving overall energy efficiency is an interesting domain to explore.www.ijacsa.thesai.org Further investigation is required considering integration of these protocols with technologies such as "Internet of things", "Vehicular Ad hoc Networks" and many others.
Sensor nodes are deployed on vehicles in order to monitor events.Data aggregation is an important issue in VANTEs keeping in view high mobility of vehicular nodes.
Security is one of the main concerns in WSNs due to its operation in open environment which requires serious efforts.For secure data transmission, the existing security approaches cannot be applied in present form due to limited resources of WSNs.Thus, there is need of mechanisms which provide secure data transmission by using less energy resources.
The design of routing protocols in the context of Internet of Things requires attention from research community, an overview of the same is provided in [32].
TABLE I .
COMPARISON BETWEEN PROMINENT CLUSTERING BASED ROUTING PROTOCOLS IN WSNS | 5,978 | 2018-01-01T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
The emergence of magnetic ordering at complex oxide interfaces tuned by defects
Complex oxides show extreme sensitivity to structural distortions and defects, and the intricate balance of competing interactions which emerge at atomically defined interfaces may give rise to unexpected physics. In the interfaces of non-magnetic complex oxides, one of the most intriguing properties is the emergence of magnetism which is sensitive to chemical defects. Particularly, it is unclear which defects are responsible for the emergent magnetic interfaces. Here, we show direct and clear experimental evidence, supported by theoretical explanation, that the B-site cation stoichiometry is crucial for the creation and control of magnetism at the interface between non-magnetic ABO3-perovskite oxides, LaAlO3 and SrTiO3. We find that consecutive defect formation, driven by atomic charge compensation, establishes the formation of robust perpendicular magnetic moments at the interface. Our observations propose a route to tune these emerging magnetoelectric structures, which are strongly coupled at the polar-nonpolar complex oxide interfaces.
I feel that there are really two distinct sets of results reported in this paper. I would support in principle publication of a revised manuscript that sticks to the cation vacancy data and analysis. The second part of the manuscript I do not feel rises to the same level of significance or clarity.
I have a few other comments below that should be addressed: In Figure 1, a diamagnetic/paramagnetic background is subtracted. The origin of this diamagnetic background is not discussed or explained. It too seems to depend on the growth conditions, and is different for the various laser fluences explored. It is not clear why this is referred to as a "correction". It is a measured response that was also reported by Ariando (Ref. 10).
Only a ~3% difference in La and Al atomic ratio is observed at higher laser fluence which goes to a maximum of 10% at lower laser fluencies. The authors claim the samples to be La rich at low laser fluencies and Al rich at high laser fluencies. What is the base line with respect to which the samples are considered to be Al rich or La rich? Also, why is the magnetism later assumed to be predominantly arising due to Al deficiency and not La richness?
In Figure 2a. the authors observed hysteresis loops in 9uc LAO/LAO (001) to be 21% smaller compared to 9uc LAO/STO sample. The authors suggest that this implies that a substantial part of the total moment measured in the Al-deficient LAO/STO system arises from the underlying STO layers. However, the Sr and Ti atomic ratio almost remains unchanged according to Figure 1b. How should the reader interpret the relation between the atomic ratios and the induced magnetism?
In addition to the 9uc LAO/LAO samples, strong hysteresis loop is observed in 9uc STO/STO (001) samples, which is dependent on laser fluencies. The results observed is comparable to that of 9uc LAO/STO at mid-range laser fluencies. Therefore, how is the observed magnetism in the case of 9uc LAO/STO sample attributed to as a property of the LAO/STO interface, due to defects and atomic compensation?
The authors have investigated the magnetic and magnetotransport properties of LAO/STO and STO/STO film/substrates grown using different energy fluences with PLD. The authors claim that variation of B-site occupation affects the magnetism in p-doped films. Specifically, the authors suggest the magnetism comes from a combination of spin-polarization of oxygen (which lacks experimental proof) and a moment on Ti (which is seen in the XMCD). For the reasons discussed below I am not convinced and cannot recommend publication of the manuscript as is. 1) Several of the figures show the magnetism as moment, which is labelled as capital M, in units of micro emu, rather than magnetization. This produces a couple of problems. First, have the moments in Fig 1a been normalized to sample size for all the references? Obviously, the more sample the more measured moment. The second problem is the most alarming to me. Take Fig. 1c-the moment saturates at 0.6 J/cm^2. Here the moment is ~50 x10^-6 emu. From the methods section, the sample's area is 5mm x 5 mm. From Fig 3d, there are Ti ions (these produce some of the moment the authors claim) within a region of the LAO ~2nm thick. These numbers yield a volume of magnetic material ~5x10^-8 cm^3.
Ratioing the moment to the volume I get a magnetization of 1000 emu/cc (emu/cc = kA/m). The magnetization of iron metal is 1740 kA/m. There is something very wrong. Which direction is the field applied? 2) Fig. 1c shows a diminishment of the moment with energy. This correlation is compelling. What I do not find compelling is the assertion that the Al and Ti concentration (and hence the related vacancies) change with energy. To me the Ti concentration appears constant in Fig 1b. In the case of Al, the claim that the concentration changes is based on one data point at 1.8 J/cm^2. Take that one point away and the entire thesis disappears.
3) The authors say the induced moment on Ti is important to the overall moment. Fig. 2a shows data for a moment in a sample that doesn't have any Ti in it. Is the difference between the two curves due to the induced Ti moment, and the rest is spin-polarization on O? Why wasn't XMCD of O measured, as done by others? 4) Fig. 2d shows that the moment decreases with increasing energy in STO film/STO substrate. The magnitude of the moment is 3/5 of the LAO film/STO, so nearly the same. The authors suggest the polar vs. not polar character of LAO/STO induces a reconstruction that promotes O spin polarization and Ti redistribution and its induced moment. In the case of STO/STO this mechanism is lacking, and yet the magnetism is nearly as strong as the LAO/STO case. I sense an inconsistency. Further the lack change of Ti concentration with energy (see Fig. 1b) suggests to me the STO/STO sample is fairly uniform, so what is special about the STO film, when the STO substrate does not show magnetism? The authors say there are no impurities, yet what is the evidence for this claim? In (1) I noted the magnetization seems to be non-sensical. If there is some error related to (1), then what concentration of Fe impurities is needed to produce the magnetism? I propose an alternate hypothesis: Magnetic impurities are deposited into the films, e.g., concomitant bombardment of the chamber, and the amount of impurities increases with energy of the PLD. 5) I would say "atomic charge compensation" not "atomic compensation".
The manuscript 'The emergence of magnetic ordering at complex oxide interfaces tuned by defects' by Park et al. reports on magnetic properties of LAO/STO heterostructures produced with varying LAO cation stoichiometry. The authors suggest a mechanism based on the formation of B-site cation defects, namely, titanium vacancies on the STO side, and TiAl antisite defects on the LAO side of Aldeficient LAO layers to be the main driver of magnetism in the system.
The magnetism in these systems is clearly a very interesting and impactful topic to be addressed in the field. Experimental and theoretical work as well as data evaluatuion is appropriate and the presented work can be reproduced based on the given information. The authors present a number of intriguing experimental data, particularly, systematic stoichiometry-dependence of the magnetic signal, and support their conclusions by DFT calculations.
While I like the idea of the paper and while I am impressed by the reported data, I am more reluctant about the details in the presented argumentation. I am missing a couple of points that should be addressed in order to arrive at a full understanding of the data. Therefore, I can recommend publication of the manuscript only after careful revision.
My main concerns are the following: 1. The authors argue that the magnetism at the LAO/STO interface is p-type in nature and argue that the XMCD data obtained for the O K-edge gives evidence for this. The signal however is very small and the discussion of the data in the SI is not detailed enough. In order to be convincing, the authors should provide field-dependent data, indicating that the XMCD is robust and behaving systematic with field (or at least a field-swap). If confirmed, the discussion should be moved to the main paper -as current state of knowledge (see e.g. Refs. 16,17) is that the interface shows dxy-type magnetism. It is also not clear to me how the XMCD (Fig. S7) can 'verify the magnetization of the anticipated antisite defects' (statement on page 5) as it is lacking the sensitivity to single defect species? 2. The authors should address in more detail, why there is a discrepancy between transport (AHE arises at low temperature) and SQUID data (which indicates strong magnetism up to RT). As the authors argue that the magnetism is very robust also at room temperature, why is there no anomalous Hall effect observed at room temperature?
3. The authors do not address the dynamics of the proposed defect formation process -the diffusion of B-site cations in perovskites is typically very slow hindering the formation of B-site vacancy defects (and often favoring A-site defect formation). Also, the formation energy for B-site vacancies is typically large (e.g. Akthar, Catlow, J. Am. Ceram. Soc. 78, 421 (1995))? It would be important the authors address in more detail how is the Ti moving into the LAO layer to form anti-site defects? Also the charge-state of the TiAl anti-site defect should be discussed -where is the extra electron in the LAO layer going? 4. I think it is very interesting finding that homoepitaxial STO and LAO can both show magnetism when grown in a non-stoichiometric manner, which nicely scales with layer thickness. It is not clear for me (this may be a matter of clarity only), why for LAO/STO the magnetization saturated, as I would always expect to see the thickness-dependent magnetization of the thin film on top of the one at the interface? The authors argue that in the heterostructure parts of the cationic defects are cocompensated by oxygen vacancies. This is a straight forward assumption, but why is it not happening in the homoepitaxial thin films, which following this arguments should not show magnetization? 5. The authors rule out other defect scenarios (oxygen vacancies and Sr vacancies) based on DFT, but they do not show their data (page 4, bottom). This should be done at least in the SI as both statements are surprising -oxygen vacancies have been identified in the literature to be a possible origin of magnetism in STO (see e.g. Refs. 17/18). For Sr vacancies on may expect a similar p-type hole formation as observed in the Ti vacancy case -it is hence interesting to understand in how far they result in different behavior? 6. A quantification for the magnetic data should be given. Based on SQUID, one should estimate the number of anti-site defects at the interface.
Besides these major comments, I have some minor comments the authors should consider: • On page 3, I would recommend to mention the growth pressure in the main text, as it is important for the required plasma-gas interaction responsible for the stoichiometry variation.
• Page 3: The term 'angle-dependent' XPS is misleading as measurements were made only at one angle. I suggest to remove. Also, the information depth should be mentioned for the individual core levels. As the authors argue in the SI, the information depth will be different for each element/core level.
• Page 3: 'reduction of Al' -I would suggest to use 'reduction of concentration of Al' to avoid confusion with a valence-change process.
• Fig. 1c -can the authors comment on why the substrate gives a paramagnetic signal, while for the LAO/STO samples a diamagnetic background is observed?
• The section on transport analysis and the data treatment is mainly following Ref. 40, which should be indicated more clearly in the main text. Also, a reference to Joshua, Nat. Comms. 3, 1129(2012 should be added.
• Fig. 3h -data seems to be not consistent to Fig. 2a/3a? Why is the saturation magnetization in perpendicular field much bigger than anticipated in 2a/3a?
• Fig. 4 -it is very interesting that the non-linearity of normal Hall effect and AHE almost cancel out, resulting in an almost linear Rxy. Can the authors comment how stable the fitting of the data is in this case?
Reply to Reviewers:
Reviewer #1: In their manuscript, the authors make a central claim that B-site cation stoichiometry is crucial for creation and control of magnetism at the LAO/STO interface.
DFT calculations show magnetic moments on the B site cation vacancy, consistent with EELS, magnetometry. The story is pretty convincing, and overall this part of the paper seems like a significant advance.
However, the second part of the paper that tries to connect anomalous and nonlinear Hall effect measurements is much less convincing. A basic question arises: if there is ferromagnetism, why is there no hysteresis in the anomalous Hall effect? This discrepancy is not answered. Instead, there is a lot of repeated discussion of nonlinear/AHE that are well documented in the literature.
Reply to comment:
We appreciate the comments of the Referee #1 for his/her positive feedback regarding our central claim that B-site cation stoichiometry is crucial for creation and control of magnetism at the LAO/STO interfaces. Following the comment, we have conducted additional low-temperature magnetoresistance measurements to address for the hysteretic AHE in a B-site cation deficient LAO/STO sample. Also, we have revised and simplified our manuscript accordingly.
We agree with the Referee's comment about the need to show the hysteretic anomalous Hall effect (AHE) in the system. Indeed, in our previous results, the hysteretic AHE was not clearly observed near the zero-field (field intervals of 20 -125 mT/sec). We have therefore performed additional magnetotransport measurements of a B-site cation deficient LAO/STO sample (grown by Eg = 0.6 J/cm 2 ) with much fine field intervals (field steps: 1 -2 mT/sec). These additional results show a clear hysteretic behaviour in the AHE. The coercive strength (40 -50 mT) of the hysteretic AHE is in line with the HC (40 -45 mT) of the ferromagnetic B-site cation deficient LAO/STO sample as shown below in Fig. S10. This reveals that the weak magnetism at the 2DEG interface can be coherently coupled with, and enhanced by the B-site cation defect-induced magnetism. To support our findings, the new results and statements have been added in the revised manuscript and supplementary information with a new Figure S10.
In the revised manuscript, we have added new statements as: ''Furthermore, a hysteretic AHE feature in the B-site cation-deficient LAO/STO sample was observed near zero-field while sweeping the fields (see Fig. S10). The coercive field (B =~40 -50 mT) of the deconvoluted Rxy (AHE) in the B-site cation deficient LAO/STO is consistent with the magnetic coercive field (HC =~40 -45 mT), determined by SQUID magnetometry. These indicate a coupling effect between the defect-induced mangetic moments and the 2DEG. In addition, the hysterestic AHE of the 2DEG visibly appears up to 5 K. This is in contrast to a strong temperature-dependent hysteretic AHE feature in conventional 2DEG interfaces, usually limited by very low temperatures (T < 2 K) in previous reports [40][41][42][43] . These observations further confirm that the weak magnetism of the 2DEG reported until now can be effectively coupled and enhanced by much stronger B-site cation defect-induced magnetism.'' In the supplementary Information, we have added new statements: ''To clarify the coupling effects between the defect-induced magnetic moments and the 2DEG near zero field, magnetotransport measurements of a B-site cation-deficient LAO/STO sample were carried out with refine field intervals (1 -2 mT/sec) considering the measurement time (/field interval) dependence of the ferromagnetic coercive strength of the system. As seen in Figs. S10b,e, a clear hysteretic Rxy appears at the zero field while sweeping the fields. This corresponds to an increase in the Hall coefficient, RH, towards zero field and a spiky negative RH then occurs when the applied field is reversely switched across the zero field due to the magnetic remanence of the system (Fig. S10f).
The coercive field (B =~40 -50 mT) of the deconvoluted Rxy (AHE) part of the B-site cation deficient LAO/STO is consistent with the magnetic coercive field , determined by SQUID magnetometry (Fig. S10g). In addition, the hysteretic AHE of the 2DEG visibly appears up to 5 K.
These observations further confirm that the weak magnetism of the 2DEG reported until now can be effectively coupled and enhanced by much stronger defect-induced magnetism.'' Figure S10. (a) Linear ordinary Hall effect (LOHE) and anomalous Hall effect (AHE) fitting parts of the experimental 5 K-Hall resistance [Rxy (Exp.)] of a B-site cation deficient LAO/STO sample, grown with Eg =~0.6 J/cm 2 , measured in a magnetic field of 5 T with a field sweep of 1 mT/sec. (b) The remanent Rxy (Exp.) of the sample at the zero field while sweeping the fields. (c) Residual Sshaped Rxy over the applied field and hysteretic AHE in small field range in the deconvoluted AHE part after subtracting the LOHE part (Rxy = R0B) to the total Rxy (Exp.). (d) Nonlinear ordinary Hall effect (NOHE) and AHE fitting parts of the Rxy (Exp.) of the sample. (e) The deconvoluted hysteretic AHE part and the difference of the forward and reverse Rxy (Exp.) while sweeping the fields at 5 K. In a closed field loop, the Rxy (F) was measured in a forward field sweep (+5 T to -5 T) and the Rxy (R) was measured in a reverse field sweep (-5 T to +5 T). (f) The hysteretic AHE part of the RH (Exp.) of the sample, measured in the field range of 5 T. The hysteretic AHE part visibly appears in the applied field below 1 T. Correspondingly, a spiky negative RH occurs while the applied fields cross over the zero field and the negative RH remains up to B~40 -50 mT due to the magnetic remanence and the corresponding remanent Rxy at the zero-field. (g) A 5 K out-of-plane magnetic hysteresis loop (in the upper panel) and the deconvoluted AHE part (in the lower panel) of the Rxy of the sample. Dashed lines indicate the coercive field (HC =~45 mT) of the magnetic moment and Rxy while sweeping the applied fields.
I feel that there are really two distinct sets of results reported in this paper. I would support in principle publication of a revised manuscript that sticks to the cation vacancy data and analysis. The second part of the manuscript I do not feel rises to the same level of significance or clarity.
Reply to comment: We thank the reviewer for his/her comments and we therefore tried to make our arguments clearer. Our findings show that interfacial magnetism in LAO/STO system can be effectively tuned by the distribution of B-site cation defects. We have further showed that the defect-induced interfacial magnetism can be significantly coupled with the interface conductivity, primarily created by the presence of oxygen vacancies. Hence, our results provide clear evidence and methodology of designing defect assembly and tuning the interfacial magnetism.
I have a few other comments below that should be addressed: In Figure 1, a diamagnetic/paramagnetic background is subtracted. The origin of this diamagnetic background is not discussed or explained. It too seems to depend on the growth conditions, and is different for the various laser fluences explored. It is not clear why this is referred to as a "correction". It is a measured response that was also reported by Ariando (Ref. 10).
Reply to comment:
To eliminate any external artifacts on the magnetic signals of the SUQID magnetometer samples, we conducted preliminary background measurements without any sample. Any diamagnetic and paramagnetic background components of the samples are therefore mainly due to instrumental accessories (such as the sample holder), see Fig. R1. This is why we used the word ''corrected'', when describing the procedure in the manuscript. To further clarify this, we added this new data into the revised supplementary information (S3) as follows: Figure R 1. (a) 5 K magnetic hysteresis loop of the sample holder used in the SQUID magnetometer, applying fields of +5 and -5 T. A paramagnetic (PM) background component was defined by subtracting the negative slope of the diamagnetic (DM) component. (b) 5 K magnetic hysteresis loop of an annealed TiO2-terminated STO substrate, mounted on the SQUID sample holder by low-temperature glue. (c) A hysteresis loop in the range of + and -0.3 T for the STO substrate. After subtracting the DM and PM background components, a real magnetic hysteresis loop of the STO substrate which shows no ferromagnetism is observed with a moment of~1 -2 x 10 -7 emu, which is not comparable with the measured moments of the LAO/STO samples.
Only a~3% difference in La and Al atomic ratio is observed at higher laser fluence which goes to a maximum of 10% at lower laser fluencies. The authors claim the samples to be La rich at low laser fluencies and Al rich at high laser fluencies. What is the base line with respect to which the samples are considered to be Al rich or La rich?
Reply to comment: We thank the reviewer for rising this question. In our work, we have taken the cation-stoichiometric ABO3 perovskite (50 % A-site and 50 % B-site cations) base-line. To clarify this, we have added the following statement in the revised manuscript: ''Cation stoichiometry of ABO3 perovskite with 50 % A-and 50 % B-site cations is taken as the baseline of the cation-stoichiometry in the grown films.'' Also, why is the magnetism later assumed to be predominantly arising due to Al deficiency and not La richness?
Reply to comment: In ABO3 perovskites, the formation of La interstitials in LaAlO3 is typically energetically unfavourable with respect to the ionic radius of the constituent elements and their thermodynamics, other secondary phases such as superstructures could be formed breaking the AO-and-BO2 plane-altered perovskite cell structures with a higher off-stoichiometry. Also, from our experimental work (atomic resolution-STEM) and theoretical calculations, the B-site cation deficient LaAlO3 upper layer clearly sustains an ABO3-type perovskite structure. The Al-vacancy is the main reason for the interfacial magnetism in this system. Therefore, we show that Al-deficiency of the LaAlO3 film composition creates the interfacial magnetism.
In Figure 2a. the authors observed hysteresis loops in 9uc LAO/LAO (001) to be 21% smaller compared to 9uc LAO/STO sample. The authors suggest that this implies that a substantial part of the total moment measured in the Al-deficient LAO/STO system arises from the underlying STO layers. However, the Sr and Ti atomic ratio almost remains unchanged according to Figure 1b. How should the reader interpret the relation between the atomic ratios and the induced magnetism?
Reply to comment: We appreciate the comment of the reviewer and to avoid any misunderstanding, we corrected the text in accordance with reviewer's comments: For the second question, the change in the Ti/Sr ratios in the LaAlO3 films seems to be small when compared to that of the Al/La (Fig. 1b). However, when the Al deficiency is considered, there is indeed a significant change in the Ti/Sr as a function of the stoichiometricity of the LaAlO3 layer.~3.6 % Ti atomic ratio can compensate approximately 22.2 % of the 16.2 %-Al deficiency in the LaAlO3 layer of the sample, grown by a laser fluence of 0.6 J/cm 2 .
Growth of the LaAlO3 upper layer with an Al-deficient composition initiates the creation of Al vacancies (VAl) near the interface, as confirmed by DFT. This subsequently leads to the formation of substitutional Ti (TiAl antisite defects) in the Al-deficient LAO layer and Ti vacancies (VTi) in the STO.
In addition to the 9uc LAO/LAO samples, strong hysteresis loop is observed in 9uc STO/STO (001) samples, which is dependent on laser fluencies. The results observed is comparable to that of 9uc LAO/STO at mid-range laser fluencies. Therefore, how is the observed magnetism in the case of 9uc LAO/STO sample attributed to as a property of the LAO/STO interface, due to defects and atomic compensation?
Reply to comment: For the case of oxidized B-site cation deficient LAO/STO sample, grown by a laser energy fluence of 0.6 J/cm 2 , a Ti atomic ratio of~3.6 % is found in the Bsite cation stoichiometry of the LAO film layer with~16.2 % Al deficiency (determined by XPS), which can compensate approximately 22.2 % of the Al deficiency. Our experimental results and DFT calculations show that individual cation defects, both VAl and TiAl in LAO and a VTi in STO, separately exist and can induce local moments, 0.87 μB/VAl, 0.9 μB/TiAl, and 0.48 μB/VTi, respectively. Assuming the same amount (~3.6 %) of B-site cation vacancies in the STO giving~3.6 % Ti out-diffusion to the LAO upper layer, 46.0 % (~22 μemu), 37.5 % (~18 μemu), and 16.5 % (~8 μemu) moments of the LAO/STO system can be separately addressed from VAl and TiAl in LAO, and VTi in STO. This could reflect the discrepancy in the total magnetic moments of the heteroepitaxial LAO/STO and homoepitaxial LAO/LAO samples, presented in Fig. 2a in the main manuscript.
We agree with the Reviewer's comment that the measured moments of the homoepitaxial STO(001)/STO(001) samples [as well as LAO(001)/LAO(001) sample] cannot be quantitatively compared with the LAO/STO case due to the different kinetics of the arriving species during PLD deposition and the resultant defect density -this could be interesting subject for a future work. Moreover, the measured moments could originate mainly from B-site cation vacancy defects because of the same charge valence of any compensating ions in the STO lattices. However, it is clear that the variation in the magnetic moments of the STO films strongly depend on the laser-energy-fluence during film growth, very similar to the case of heteroepitaxial LAO/STO. This provides direct evidence into the creation of ferromagnetic moments, driven by B-site cation deficiency in STO. So, this convincingly supports that the total magnetic moment of the B-site cation deficient LAO/STO system should be combined with any induced moments in the STO side as a consequence of Ti redistribution.
Reviewer #2 (Remarks to the Author):
The authors have investigated the magnetic and magnetotransport properties of LAO/STO and STO/STO film/substrates grown using different energy fluences with PLD. The authors claim that variation of B-site occupation affects the magnetism in p-doped films. Specifically, the authors suggest the magnetism comes from a combination of spin-polarization of oxygen (which lacks experimental proof) and a moment on Ti (which is seen in the XMCD). For the reasons discussed below I am not convinced and cannot recommend publication of the manuscript as is.
Reply to comment:
We appreciate the comments of Reviewer #2 about our findings and we have tried to comply with his/her comments point-by-point in the revised manuscript which provide further information to support our observations. We do believe that his/her comments have helped for improving our manuscript.
1) Several of the figures show the magnetism as moment, which is labelled as capital M, in units of micro emu, rather than magnetization. This produces a couple of problems. First, have the moments in Fig 1a been normalized to sample size for all the references?
Obviously, the more sample the more measured moment.
Reply to comment:
We thank the reviewer for his/her comments. We have now corrected the symbol of the moment units to the conventional unit, ''m'', in the revised manuscript to avoid any confusion for readers.
All the samples for the presented references in Fig. 1a [except samples in UNSW (Ref. 26 & 27), no information on the area/volume of STO substrate] are~5 × 5 mm 2 size and the LAO layer thicknesses are highlighted showing no dependence of layer thickness, but it provides clear evidence that there is a strong PLD-growth-parameter dependence of the interfacial magnetism in the LAO/STO system. To clarify this, we added the following statement in the revised manuscript: ''The observed trend shows that relatively weak interfacial magnetism (m ≤~2 μemu) commonly occurs with a low oxygen-pressure (Pg ≤ 10 -4 mbar) growth atmosphere and a high-laser-fluence (Eg
independent of the film thickness and size of the grown samples.''
The second problem is the most alarming to me. Take Fig. 1c-the moment saturates at 0.6 J/cm^2. Here the moment is~50 x10^-6 emu. From the methods section, the sample's area is 5mm x 5 mm. From Fig 3d, there are Ti ions (these produce some of the moment the authors claim) within a region of the LAO~2nm thick. These numbers yield a volume of magnetic material~5x10^-8 cm^3. Ratioing the moment to the volume I get a magnetization of 1000 emu/cc (emu/cc = kA/m). The magnetization of iron metal is 1740 kA/m. There is something very wrong. Which direction is the field applied?
Reply to comment: Our work demonstrates that the total magnetic moment of the Bsite cation deficient LAO/STO system is not solely from the LAO upper layer. It is due to a combined effect and is a result of atomic and/or charge compensating effect across the interface. So, we cannot quantify the total moment (/magnetization) only considering the LAO film layer thickness (Fig. 3d). To avoid inaccurate data evaluation, in this work, we intentionally present the absolute magnetic moment of all the measured samples rather than evaluating the magnetization as the magnetic properties are not solely from the grown film layers. Also, if magnetic impurities would be the main source for the interfacial magnetism, a strong laser fluence dependence and oxygen environment would not appear with a constant magnetic response.
This work presents the magnetic properties of the grown LAO/STO samples, measured along the in-plane direction (Fig. 1c,Figs. 2a,d,and Fig. 3a). We then show that the induced cation defect moments are preferentially aligned along the out-of-plane direction (Fig. 4e), resulting in an enlargement of the total magnetic moment of the B-site cation deficient LAO/STO system. Finally, we combine the magnetic and electronic properties of the system to deliver an emergent magnetoelectric coupling effect at the interface. Fig 1b. In the case of Al, the claim that the concentration changes is based on one data point at 1.8 J/cm^2. Take that one point away and the entire thesis disappears.
2) Fig. 1c shows a diminishment of the moment with energy. This correlation is compelling. What I do not find compelling is the assertion that the Al and Ti concentration (and hence the related vacancies) change with energy. To me the Ti concentration appears constant in
Reply to comment: Our work demonstrates that the Al-deficient LAO layer, grown on TiO2-terminated STO initiates the formation of magnetic moments to the neighbouring oxygen atoms. Importantly, VAl defects tend to reside near the interface, resulting in atomic compensation via Ti out-diffusion, which has been directly observed in our work by atomic resolution STEM and confirmed by DFT.
Al deficiency largely varies from 16.2 %, 15.3 %, 14.0 %, and 12.5 % over a laser fluence range of 0.6 -1.2 J/cm 2 . In contrast, Ti atomic ratios change from 3.60±0.02%, 3.51±0.03%, 3.40±0.05% to 3.25±0.05% within the limited thickness range (up to the maximum first five LAO unit cells). The results suggest that saturation in the Ti atomic concentration of the LAO layer with a higher laser fluence, however, overall, it indeed does show a large variation when the relative moment for each defect amount in the samples is considered as illustrated in Table R1.
Given that the Ti concentration equals the amount of Ti vacancies in the STO as a consequence of Ti out-diffusion, all the defect ratios (VAl, TiAl, and VTi) are approximated from the Al deficiency (50 %-Al%) with respect to a stoichiometric ABO3 perovskite and the Ti concentration (VTi) of the LAO films (STO). The relative moment of each defect compared to the experimentally measured total moment can be estimated by combining the ratio of their theoretical local moments (0.87 μB/VAl, 0.9 μB/TiA l, and 0.48 μB/VTi), as determined by DFT. It should be mentioned that the relative moment are indeed overestimated if they are exclusively associated with the individual defect. It is impossible to assign these moments separately with a high inaccuracy as the total moment is delocalised over the neighbouring ions in the system. Hence, the total magnetic moment of the B-site deficient LAO/STO system is associated with the combined defect-induced effects across the interface. In addition, the effect of the B-site cation defect assembly on the interfacial magnetism can be further controlled by their interactions with oxygen vacancies in a given oxygenreduced condition. Therefore, our findings not only show a cation defect-induced interfacial magnetism with a clear cation stoichiometry dependence but also it deepens our fundamental understanding of the magnetism in LAO/STO interfaces.
3) The authors say the induced moment on Ti is important to the overall moment. Fig. 2a shows
data for a moment in a sample that doesn't have any Ti in it. Is the difference between the two curves due to the induced Ti moment, and the rest is spin-polarization on O? Why wasn't XMCD of O measured, as done by others?
Reply to comment: The O K XMCD spectra of a B-site cation deficient LAO(001)/LAO(001) sample which has no TiAl defects were recorded using the total electron yield method (by measuring the sample drain current). Unfortunately, the fullyoxidized B-site cation deficient 9uc-LAO layers, grown on LAO(001) substrate were too insulating that charging prevented a reliable measurement of the O K-edge XMCD.
Instead, to clarify oxygen spin-polarization nature in the B-site cation-deficient LAO/STO system, we performed additional XMCD measurements. Figure R2 shows the O K-edge XMCD spectra for an oxidized B-site cation-deficient LAO/STO sample, grown by a laser fluence of 0.6 J/cm 2 , measured in normal X-ray incidence to the film surface. A (2013)] since our data shows a polarization of the oxygen states mixed with the Ti eg bands. We conclude that the two XMCD contributions at 529 and 532-533 eV reflect the polarization of the oxygen bands induced by both of TiAl 4+ and VTi vacancies. The XMCD signal centred at 535.5 eV occurs in a region dominated by O 2p states hybridized with La 5d, Al 3s and Al 3p empty bands and could be associated to VAl defects. Figure R2. (a) A schematic of X-ray magnetic circular dichroism (XMCD) measurement protocol. Total electron yield (TEY) detection is obtained through excitation using circularly polarized light (ρ + /ρ -) with normal incidence while applying an external magnetic field along the out-of-plane <001> directions. (b) 20 K X-ray absorption spectra at the O K-edge of a B-site cation deficient 9uc-LAO/STO sample, grown by Eg = 0.6 J/cm 2 . The energy positions of the O-hybridized bands are estimated by the theoretical electronic structure calculations of the bulk LAO and STO. (c) O K-edge XMCD spectra, measured by reversing the applied fields, H = +6 and -6 T. Fig. 2d Reply to comment: In the the manuscript, we clarified all of the possible defects and their contributions to the magnetism in the B-site cation deficient LAO/STO system by comparing the LAO/STO, LAO/LAO, and STO/STO samples. We cannot make a quantitative comparison of the measured moments between the heteroepitaxial LAO/STO and homoepitaxial STO/STO and LAO/LAO samples due to different atomic/ionic kinetics during film growth and vacancy defect density. This is an interesting topic but outside of scope of this paper and further work is required. Moreover, no antisite defects form in the homoepitaxial samples due to the same charge valence of the substituted elements in the appropriate cation sites. However, a strong laser-fluence dependence of the magnetic moment in the STO(001)/STO (001) Reply to comment: First, we observe an opposite behavior as proposed by the reviewer: the magnetic moment decreases with increasing laser fluence (Fig. 1c). In any case, there is no connection between laser fluence and the concentration of impurities as it is supposed to influence ablation from target only. A presence of such high amount of magnetic impurities in the single crystalline LaAlO3 target is not realistic. Moreover, we have not observed any traces of foreign magnetic elements in the XPS spectra. These considerations exclude possibility that magnetism in our films is due to external impurities. The magnetization saturates at about 9uc-thick LAO, which reveals that it is indeed the interfacial effect. If it is due to the impurities, it should just increase with the volume of the films.
5) I would say "atomic charge compensation" not "atomic compensation".
Reply to comment: We agree with the Referee's suggestion and revised the manuscript accordingly.
Reviewer #3 (Remarks to the Author): The manuscript 'The emergence of magnetic ordering at complex oxide interfaces tuned by defects' by Park et al. reports on magnetic properties of LAO/STO heterostructures produced with varying LAO cation stoichiometry. The authors suggest a mechanism based on the formation of B-site cation defects, namely, titanium vacancies on the STO side, and TiAl antisite defects on the LAO side of Al-deficient LAO layers to be the main driver of magnetism in the system.
The magnetism in these systems is clearly a very interesting and impactful topic to be addressed in the field. Experimental and theoretical work as well as data evaluatuion is appropriate and the presented work can be reproduced based on the given information. The authors present a number of intriguing experimental data, particularly, systematic stoichiometry-dependence of the magnetic signal, and support their conclusions by DFT calculations.
While I like the idea of the paper and while I am impressed by the reported data, I am more reluctant about the details in the presented argumentation. I am missing a couple of points that should be addressed in order to arrive at a full understanding of the data. Therefore, I can recommend publication of the manuscript only after careful revision.
Reply to comment: We highly appreciate Reviewer #3 for his positive comments and suggestion and the interest on our work. Following his/her questions and comments, we provide our point-by-point reply blow and have revised the manuscript accordingly.
My main concerns are the following: 1. The authors argue that the magnetism at the LAO/STO interface is p-type in nature and argue that the XMCD data obtained for the O K-edge gives evidence for this. The signal however is very small and the discussion of the data in the SI is not detailed enough. In order to be convincing, the authors should provide field-dependent data, indicating that the XMCD is robust and behaving systematic with field (or at least a field-swap). If confirmed, the discussion should be moved to the main paper -as current state of knowledge (see e.g. Refs. 16,17) is that the interface shows dxy-type magnetism. It is also not clear to me how the XMCD (Fig. S7) can 'verify the magnetization of the anticipated anti-site defects' (statement on page 5) as it is lacking the sensitivity to single defect species?
Reply to comment: Following referee's suggestion, the O K-edge XMCD measurements of the oxidized B-site cation deficient LAO/STO were repeated (with much higher statistics than the data reported in the previous version of the manuscript) at 20 K at both B = +6 and −6 T applied fields. The XAS spectra were collected in groups of eight or octets (ρ+ ρ− ρ− ρ+ ρ− ρ+ ρ+ ρ−, where ρ+ and ρ− indicate photon spin parallel or antiparallel to the applied field, respectively) in order to minimize the effect of any time dependence in the X-ray beam on the measured spectra. In total, the XMCD was obtained as the difference between 40 ρ+ and 40 ρ− spectra, i.e. 5 times more than what is commonly acquired at BOREAS of ALBA for standard ferromagnetic materials. We have added the following text and new data plots (Figs. 4a,b, The O K-edge XMCD measurements were carried out at the maximum magnetic field available in BOREAS beamline in order to maximize the dichroic signal. A systematic study of the magnetization hysteresis loop at different magnetic fields through O K-edge XMCD technique would need very long impractical acquisition times (several days) in order to achieve sufficient reliability in the experimental data for an XMCD signal smaller than 1% of the maximum intensity of the K-edge XAS. Finally, Ti L2,3 XMCD is element sensitive. Our Ti L2,3 XMCD of oxidized B-site cation deficient LAO/STO shows that TiAl antisite defect has a 4+ valence.
The authors should address in more detail, why there is a discrepancy between transport (AHE arises at low temperature) and SQUID data (which indicates strong magnetism up to RT). As the authors argue that the magnetism is very robust also at room temperature, why is there no anomalous Hall effect observed at room temperature?
Reply to comment: The B-site cation defect-induced magnetism in the samples is responsible for the SQUID data, while the AHE originates from the 2DEG at the interfaces. The AHE originates from magnetism at the 2DEG in the LAO/STO system, which has a strong temperature-dependent characteristic [Stornaiuolo et al., Nat. Mater., 15 278 (2016) and Gan et al. Adv. Mater., 31, 1805970 (2019)]. Moreover, the cation defect-induced magnetism is much stronger than the magnetism at the 2DEG interface, leading to a magnetic proximity effect around the interface. Hence, three critical factors need to be considered in order to understand the observed strong low-temperature AHE in the B-site cation-deficient LAO/STO system: (i) a temperature-independent defectinduced magnetism; (ii) a temperature-dependent weak magnetism of the 2DEG; (iii) and the enhanced magnetoelectric coupling effect between the 2DEG and the B-site cation defect-induced magnetic moments only at low temperatures via a magnetic proximity effect. To better communicate these observations, we have added the following statements to the revised manuscript and supplementary information.
In the revised manuscript, we have added new statements as: ''Furthermore, a hysteretic AHE feature in the B-site cation-deficient LAO/STO sample was observed near zero-field while sweeping the fields (see Fig. S10 is reversely switched across the zero field due to the magnetic remanence of the system (Fig. S10f).
The coercive field (B =~40 -50 mT) of the deconvoluted Rxy (AHE) part of the B-site cation deficient
LAO/STO is consistent with the magnetic coercive field (HC =~40 -45 mT), determined by SQUID magnetometry (Fig. S10g). In addition, the hysteretic AHE of the 2DEG visibly appears up to 5 K.
These observations further confirm that the weak magnetism of the 2DEG reported until now can be effectively coupled and enhanced by much stronger defect-induced magnetism.'' Reply to comment: To address this, we need to understand the difference between thermodynamic equilibrium growth condition and the non-equilibrium conditions for solid-state oxide thin film formation. Any artificially designed epitaxial complex oxide films are usually made in non-equilibrium states and/or in quasi-states. This is one of the most significant advantages of using PLD for designing/controlling atomically defined thin film layers on underlying substrates by adjusting material's stoichiometry to trigger exotic physical phenomena, compared to conventional properties in bulk materials. Together with this growth feature, the stoichiometry and related properties of the designed LAO/STO thin films can differ from those of bulk materials.
The authors do not address the dynamics of the proposed defect formation process -the diffusion of B-site cations in perovskites is typically very slow hindering the formation of Bsite vacancy defects (and often favoring A-site defect formation
Akthar et al. [J. Am. Ceram. Soc. 78, 421 (1995)] computed the defect formation energies and diffusion activation energies in bulk STO at the equilibrium conditions and T = 0 K. As the result, the computed Ti activation energy of~11.6 eV is significantly overestimated and, hence, it cannot be directly discussed with the context of our results, especially because Ti diffusion goes into the LAO overlayers across the interface.
Meanwhile, we have found that Ti out-diffusion is significantly limited to a maximum of 5uc of the LAO upper layer in the most B-site cation deficient sample in this work. The Ti diffusion can be effectively hindered by the reduction of the Al vacancy concentration of the LAO layer. This is what we have directly observed in our experimental results as a function of laser fluence. An overall dynamic picture of the consecutive defect assembly is presented in Fig. 4d in the main text of our manuscript.
I think it is very interesting finding that homoepitaxial STO and LAO can both show magnetism when grown in a non-stoichiometric manner, which nicely scales with layer thickness. It is not clear for me (this may be a matter of clarity only), why for LAO/STO the magnetization saturated, as I would always expect to see the thickness-dependent magnetization of the thin film on top of the one at the interface? The authors argue that in the heterostructure parts of the cationic defects are co-compensated by oxygen vacancies. This is a straight forward assumption, but why is it not happening in the homoepitaxial thin films, which following this arguments should not show magnetization?
Reply to comment: We thank the reviewer for appreciating the results. The total magnetic moment of the B-site deficient LAO/STO system stems from the combined defect assembly effect across the interface. As we proposed in the manuscript, by compensating defect formation, the defect level of B-site cation vacancies in LAO can be saturated according to the amphoteric defect theory. Furthermore, an important point made by the referee is that cation diffusion is usually slow. So, Ti out-diffusion to the upper film layer is limited by the LAO layers(e.g. the maximum first 5ucs in the B-site cation deficient LAO/STO sample, grown by~0.6 J/cm 2 laser fluence), as also observed by STEM.
We agree with the referee's comment, and we also expect a similar saturation behaviour in the magnetic moment of homoepitaxial LAO or STO films which would appear with thickness. However, the magnetism is mainly due to B-site cation vacancies as the same valence charge of substitutional cations are always observed around the interfaces.
The authors rule out other defect scenarios (oxygen vacancies and Sr vacancies) based on DFT, but they do not show their data (page 4, bottom). This should be done at least in the SI as both statements are surprising -oxygen vacancies have been identified in the literature to be a possible origin of magnetism in STO (see e.g. Refs. 17/18). For Sr vacancies on may expect a similar p-type hole formation as observed in the Ti vacancy case -it is hence interesting to understand in how far they result in different behavior?
Reply to comment: We did try to search for any oxygen vacancy effect in the magnetism of STO, but this was not the case.
Firstly, DFT shows that Vo-induced magnetism in STO strongly depends on the concentration of Vo for the magnetization of Ti dxy orbitals as shown in Figs (2015)]. However, the Vo concentration in a modelled structure by authors of this paper seems unrealistically high. Also, practically existence of oxygen vacancies in our sample could not be the case as all of the magnetic B-site cation deficient LAO/STO samples were oxidized after film growth (except non-oxidised conductive samples which were used for magnetotransport measurements).
Secondly, the formation of La vacancy defects in the LAO film layer of magnetic B-site cation deficient LAO/STO system can be disregarded, based on our experimental results (e.g. Al/La atomic ratios).
Finally, an STO supercell model structure with a Sr vacancy was calculated by DFT. Unlike the B-site cation vacancy effect, there is indeed a very weak A-site cation vacancy effect (m < 0.07 μB/VSr) on the creation of magnetism in STO (Fig. 3c) although the defective system shows a p-type characteristic. This could be due to much less charge rearrangement of the neighbouring oxygens and cations near the Sr vacancy due to the following factors: (i) only two electrons can contribute to neighbouring oxygens at the Sr vacancy sites, compared with four for the Ti (4+) cation vacancy; (ii) the Sr site has 12 nearest oxygens at relatively larger distances of 2.76 Å, whereas Ti has 6 oxygen neighbours at significantly shorter distances of 1.95 Å. Therefore, the Sr vacancy (or Asite cation vacancies in ABO3 perovskites) would not dramatically change the electronic structure. This eventually results in a magnetic moment at each oxygen atom of less than 0.001 μB.
A quantification for the magnetic data should be given. Based on SQUID, one should estimate the number of anti-site defects at the interface.
Reply to comment: We thank the reviewer for his/her valuable comment, however a quantification analysis is very challenging for assigning specific magnetic moments to each defect in the B-site cation deficient system as the total sample magnetization originates from a combined magnetic effect. Also, the entire defective layer across the interface (not only the film layer) needs to be considered. For example, single ion pictures are often described by assigning a moment per Mn ion in the transition metal-based magnetic material system (e.g. La1-xSrxMnO3 films), but it is not accurate as magnetization is often delocalised over different ions.
We cannot determine practically specifically how much effective moment is from each Bsite cation defect with a gradient defect distribution and a moment spread across the interface (as seen in STEM) although their moments have been determined by DFT. Also, the total moment is distributed over all of the neighbouring ions. A magnetic moment of around 1.8 μB/uc for the most B-site cation defective LAO/STO system (grown by~0.6 J/cm 2 laser fluence) could be deduced by considering the depth of defect distribution (about 16uc thickness) and experimental total magnetic moments. When the magnetic moment of 9uc-LAO upper layers is considered by combining the Ti atomic ratio and estimated defect ratios of VAl, TiAl, and VTi in the B-site of the perovskite cell, it could be further assigned to be 2.6, 1.6, 1.1, 0.97, and 0.14 μB/uc with the laser fluence of 0.6 -1.8 J/cm 2 . Obviously, this moment density variation indicates it is not simply a magnetic impurity effect (e.g. 3.6 μB/Fe). Moreover, an estimated concentration of the out-diffused Ti atoms is~6.2 × 10 20 /cm 3 in the most B-site defective LAO layer. The corresponding TiAl defect magnetic moment from the portion of SQUID moment (210 emu/cm 3 ) is deduced to be 36.4 μB/TiAl. The deduced moment per single defect is unrealistic. Thus, it is noted that each B-site cation defect cannot solely contribute to each portion of the measured moment with respect to B-site cation defect ratios in LAO layer. This means it is difficult to assign moments to single defects, which is not reliable with quantitative inaccuracy. In other words, the obtained large total magnetic moments should be interpreted as collective/combined magnetic moments as a result of a collective magnetic interactions in the B-site cation defect assembly. One of our ongoing projects is to address these issues. Therefore, at this stage, our work only demonstrates a qualitative picture to describe the effect of B-site cation defect assembly across the interface combining experimental results, supported by theoretical model calculations.
Besides these major comments, I have some minor comments the authors should consider: • On page 3, I would recommend to mention the growth pressure in the main text, as it is important for the required plasma-gas interaction responsible for the stoichiometry variation. Reply to comment: We agree with the reviewer and according to the referee's suggestion, we have added statements in the revised manuscript as:
''The laser fluency (energy per unit area) and plume-background gas interaction significantly affect the transfer of ablated species to the surface of underlying substrates during deposition, resulting from the non-congruent ablation and/or mass-dependent variations in the atom and ion trajectories through the laser plume.''
• Page 3: The term 'angle-dependent' XPS is misleading as measurements were made only at one angle. I suggest to remove. Also, the information depth should be mentioned for the individual core levels. As the authors argue in the SI, the information depth will be different for each element/core level. Reply to comment: We agree with this, so the term, ''angle-dependent XPS'' has been corrected to be ''surface-sensitive XPS'' in the revised manuscript. In addition, to inform the inelastic mean-free path of photoelectrons for constituent element core levels of LAO and STO, we have added the details in the revised manuscript and the Method section as: '' Figure 1b shows ''i.e. the reduction of the concentration of Al while increasing of La (Fig. 1c).''
• Fig. 1c -can the authors comment on why the substrate gives a paramagnetic signal, while for the LAO/STO samples a diamagnetic background is observed?
Reply to comment: Diamagnetic and paramagnetic background components in the measured magnetic data are from an instrumental accessory (sample holder) in the SQUID magnetometer, used in this work. Therefore, we have corrected the data again as Fig. 1c in the revised manuscript.
• The section on transport analysis and the data treatment is mainly following Ref. 40, which should be indicated more clearly in the main text. Also, a reference to Joshua, Nat. Comms. 3, 1129(2012 should be added. Reply to comment: We agree with Referee's suggestion, so the suggested paper is added to the list of references in the revised manuscript: Fig.2 were grown with a laser energy fluence of~0.6 J/cm 2 , while for the thickness-dependent series, the films were grown using a laser energy fluence of~0.7 J/cm 2 . As for the second question, the magnetic moments of the ordered cation defects along the out-of-plane of the B-site cation-deficient LAO/STO sample could be more collective as a result of the consecutive defect formation across the interface. • Fig. 4 Fig. R4. Here, the non-linearity of the normal Hall effect for the two-carrier case occurs at high magnetic field (after a critical field), leading to a Lorentz-shaped RH curve. However, the non-linearity of AHE occurs at low magnetic field, leading to a hump around zero field if a non-hysteretic feature appears (Fig. R4). The deconvoluted hysteretic AHE part and the difference of the forward and reverse Rxy (Exp.) while sweeping the fields at 5 K. Rxy(F) was measured in a forward field sweep (+5 T to -5 T) and Rxy(R) was measured in a reverse field sweep (-5 T to +5 T). (f) The hysteretic AHE part of the RH (Exp.) of the sample, measured in the field range of 5 T. The hysteretic AHE part visibly appears in the applied field below 1 T. Correspondingly, a spiky negative RH occurs while the applied fields cross over the zero field and the negative RH remains up to B~40 -50 mT. (g) A 5 K out-of-plane magnetic hysteresis loop (in the upper panel) and the deconvoluted AHE part (in the lower panel) of the Rxy of the sample. Dashed lines indicate the coercive field (HC =~45 mT) of the magnetic moment and Rxy while sweeping the applied fields.
• Page 6: It is not clear to me, why a proximity effect of the hysteretic out-of-plane magnetization in the LAO layer on the 2DEG does not yield a hysteresis in the observed AHE?
Reply to comment: To address this, we have performed additional low temperature magnetotransport measurements of a B-site cation deficient LAO/STO sample (Eg = 0.6 J/cm 2 ) with small field intervals (5 -20 mT) which are much smaller than the magnetic coercive field (HC = 40 -45 mT). Our new results visibly show a hysteretic AHE of the measured Rxy near the zero field while sweeping the field in a closed loop. Therefore, we have added the new results in the revised supplementary information (Figs. S10): ''To clarify the coupling effects between the defect-induced magnetic moments and the 2DEG near zero field, magnetotransport measurements of a B-site cation-deficient LAO/STO sample were carried out with refine field intervals (1 -2 mT/sec) considering the measurement time (/field interval) dependence of the ferromagnetic coercive strength of the system. As seen in Figs. S10b,e, a clear hysteretic Rxy appears at the zero field while sweeping the fields. This corresponds to an increase in the Hall coefficient, RH, towards zero field and a spiky negative RH then occurs when the applied field is reversely switched across the zero field due to the magnetic remanence of the system (Fig. S10f).
The coercive field (B =~40 -50 mT) of the deconvoluted Rxy (AHE) part of the B-site cation deficient
LAO/STO is consistent with the magnetic coercive field (HC =~40 -45 mT), determined by SQUID magnetometry (Fig. S10g). In addition, the hysteretic AHE of the 2DEG visibly appears up to 5 K.
These observations further confirm that the weak magnetism of the 2DEG reported until now can be effectively coupled and enhanced by much stronger defect-induced magnetism.'' • XPS survey scans in SI -there are two core levels labeled with La 3d. Reply to comment: In the revised manuscript, ''La 3d'' is corrected to be ''La 4d'' at the binding energy of approximately 99 -110 eV. I am still lost by the arguments made by the authors concerning the "second" part of the manuscript that claims to make a non-trivial connection between the magnetism due to the B-site cation defects (which are firmly established) and any magnetic properties in the STO. I feel the need to reply in-line to the brief response of the authors: "I feel that there are really two distinct sets of results reported in this paper. I would support in principle publication of a revised manuscript that sticks to the cation vacancy data and analysis. The second part of the manuscript I do not feel rises to the same level of significance or clarity.
Reply to comment: We thank the reviewer for his/her comments and we therefore tried to make our arguments clearer. Our findings show that interfacial magnetism in LAO/STO system can be effectively tuned by the distribution of B-site cation defects.
On this we are in agreement.
We have further showed that the defect-induced interfacial magnetism can be significantly coupled with the interface conductivity, primarily created by the presence of oxygen vacancies. Hence, our results provide clear evidence and methodology of designing defect assembly and tuning the interfacial magnetism." Here I am still not convinced. I was pleased to see the new measurements confirming existence of hysteresis in the AHE. It is not clear to me, however, that it is anything other than the contribution of the magnetic field from external magnetic moments that were previously identified. It is not clear what the authors are referring to as "weak magnetism". This measurement basically confirms that there is no intrinsic hysteretic (ferromagnetic or ferrimagnetic) behavior in the STO. The hysteretic behavior is no more surprising than if any other ferromagnetic layer were placed in contact with the top surface and switched hysteretically.
The existence of oxygen vacancies and their impact on transport in the STO layer are a completely separate matter. There is absolutely nothing novel here concerning the use of oxygen vacancies to tune the electron density in LAO/STO. The final sentence of the author's response is also an overreach.
There is no evidence that the effect of the magnetic moments on the STO is any different from the externally applied magnetic field.
It was my recommendation previously that the authors scale back their discussion to focus on the magnetic moments that they were able to create and control using oxygen pressure and laser fluence. I would have recommended publication of that manuscript. Here, these interesting results are mixed with overreaching claims that are not supported by the data. Unfortunately, I cannot recommend the manuscript for publication in its present form.
Personally, I think the observations of a) the stoichiometry dependence of magnetism in LAO and STO and b) the oxygen dicroism in STO (including the clear field swap presented in the revised manuscript) are most striking and deserve publication. As for the LAO/STO interface, the most striking result is the observation of the hysteretic anomalous Hall effect, revealed in the revised version of the manuscript. At the same time this hysteresis raises some new questions as listed below. I can therefore only recommend publication after another revision.
-The hysteretic AHE is highly interesting and could be worth to be moved into the main paper. The authors should comment, however, why the magnetic moment causing the proximity effect seems to be out-of-plane? This is counter-intuitive, as one may expect any moment to align in-plane given the thin layer thicknesses and the 2D-character of the interface.
-It may be technical, but it is unclear from the experimental point of view how the data for forward and backward field sweep was obtained. As described the samples were measured in van der Pauw configuration. Were the data averaged over different geometries to avoid misalignment voltages? And how does the potential data treatment affect the hysteretic Hall signal, which is symmetric in field and could be averaged out through averaging. Further, it is unclear how the authors technically determined the coercive field from the data.
-Finally, the model applied to describe the AHE in the RH data (elaborated in the SI) is symmetric in B, while RH now has asymmetric contributions due to the observed hysteresis. Can the authors comment why the model is still valid to be used?
-O-K-edge XMCD: While I appreciate the clear field swap shown in the new figure 4, I would like to ask why the feature 535eV cannot be related to oxygen-Sr hybrids? In my understanding, the total electron yield in XAS is not similarly surface sensitive as e.g. XPS, so they spectra may contain significant contributions from the substrate in this region. Can the authors clarify?
-I also think the statement on the Ti XMCD originating from Ti_Al defect sites should be rephrased. Obviously, XMCD is element sensitive. Hence, the authors observe XMCD originating from Ti with no doubt. Their conclusion the respective Ti ions are located in the LAO layer forming anti-sites however cannot be made based on the XMCD measurement alone. The authors should clarify this to avoid misunderstanding.
-As for the defect dynamics, I see the argument made by the authors, but I would suggest to comment on this in the paper. It is not obvious that anti-site defect can be formed under the nonequilibrium conditions of PLD, and should hence be mentioned in the manuscript.
-Finally, a quantitative estimation of a magnetic moment per defect -be it anti-site, cation vacancy, or something else should be made. I understand that exact values for each potentially involved magnetic center cannot be determined, but it should be made plausible that the measured total moment is consistent with the discussed (superimposed) mechanisms. A similar plausibility argument was also requested by reviewer #1.
Reviewer #1 (Remarks to the Author): We thank the reviewer 1 for his/her additional comments and concerns to improve our manuscript.
I am still lost by the arguments made by the authors concerning the "second" part of the manuscript that claims to make a non-trivial connection between the magnetism due to the B-site cation defects (which are firmly established) and any magnetic properties in the STO. I feel the need to reply in-line to the brief response of the authors: "I feel that there are really two distinct sets of results reported in this paper. I would support in principle publication of a revised manuscript that sticks to the cation vacancy data and analysis. The second part of the manuscript I do not feel rises to the same level of significance or clarity.
Reply to comment: We thank the reviewer for his/her comments and we therefore tried to make our arguments clearer. Our findings show that interfacial magnetism in LAO/STO system can be effectively tuned by the distribution of B-site cation defects.
On this we are in agreement.
We have further showed that the defect-induced interfacial magnetism can be significantly coupled with the interface conductivity, primarily created by the presence of oxygen vacancies. Hence, our results provide clear evidence and methodology of designing defect assembly and tuning the interfacial magnetism." Here I am still not convinced. I was pleased to see the new measurements confirming existence of hysteresis in the AHE. It is not clear to me, however, that it is anything other than the contribution of the magnetic field from external magnetic moments that were previously identified. It is not clear what the authors are referring to as "weak magnetism". This measurement basically confirms that there is no intrinsic hysteretic (ferromagnetic or ferrimagnetic) behavior in the STO. The hysteretic behavior is no more surprising than if any other ferromagnetic layer were placed in contact with the top surface and switched hysteretically. The existence of oxygen vacancies and their impact on transport in the STO layer are a completely separate matter. There is absolutely nothing novel here concerning the use of oxygen vacancies to tune the electron density in LAO/STO. The final sentence of the author's response is also an overreach. There is no evidence that the effect of the magnetic moments on the STO is any different from the externally applied magnetic field.
Reply to the reviewer comments: Since the reviewer is still puzzled by the "second" part we will try making our argument clearer by introducing also the "first" part of the discussion. The overall magnetism in our LAO/STO samples are large with the Bsite cation deficient LAO overlayer. The magnetic properties are created by tuning the atomic charge compensation process which occur across the interface. In the 1 st case (samples fabricated and annealed after the deposition in high oxygen partial pressure in the chamber) the magnetic moments come from the contribution of the defects at LAO layer VAl and TiAl as well as defects created at the STO interfaces, VTi. This magnetic contribution of the B-site cation vacancy has been qualitatively verified and presented in Figs, 2a-c. Note, as expected, no evidence for oxygen vacancies were observed (samples are nonconductive).
In the 2 nd case (samples fabricated and cooled after the deposition in lower oxygen partial pressure) we believe again that the magnetic contribution (lower than the 1 st case) comes from the same sources, defects at LAO layer VAl and TiAl as well as defect created at the STO interfaces, VTi. However, here we indeed observed oxygen vacancies (samples show conductivity) however, their contribution to the magnetism is in a "negative" way as they compensate with the p-type defects, thus reducing the magnetism. This magnetic contribution of the B-site cation vacancies has been qualitatively verified and presented in Figs, 2d-f.
The first set of results agree well with the observation reported by Ariando's group [Ariando, et al. Nat. Commun. 2, 188 (2011)] where enhanced magnetism is observed in the case of samples treated in high oxygen partial pressure as compared to lower oxygen partial pressure where the magnetism is reduced. The B-site cation defectinduced magnetic moments dominantly gives rise to polarizing the 2DEG via a magnetic proximity effect, resulting in the observed large AHE. Furthermore, we emphasize here that the observed large magnetoelectric coupling effect can be driven by the large magnetic moment of B-site cation defects in pure LAO/STO system without employing external magnetic elements/layers.
Regarding the weak magnetism at the 2DEG-interface, we recognized that there is miscommunication in the manuscript, which we have now revised.
''At low temperatures (≤ 50 K), the motion of the confined dxy electrons couples with the preferential out-of-plane magnetic moments''.
The out-of-plane magnetic moments are induced by the B-site cation defects, it does not mean that 2DEG itself has out-of-plane magnetism. This have been corrected in the manuscript: ''At low temperatures (≤ 50 K), the motion of the confined dxy electrons couples with the preferential out-of-plane magnetic moments of the B-site cation defects, resulting from a magnetic proximity effect.'' It was my recommendation previously that the authors scale back their discussion to focus on the magnetic moments that they were able to create and control using oxygen pressure and laser fluence. I would have recommended publication of that manuscript. Here, these interesting results are mixed with overreaching claims that are not supported by the data. Unfortunately, I cannot recommend the manuscript for publication in its present form.
Reply to the comments: In principle, we could agree with the reviewer's idea of separating our manuscript into two parts. However, we also need to balance the concerns of other the reviewers who are highly interested in the corresponding large magnetoelectric coupling effect around the interface, induced by the B-site cation defect-associated magnetism. In fact, the other two reviewers do not recommend separating the paper into two part. The magneto-transport data of the 2DEG-interface provide an additional experimental proof for the existence of the B-site cation defects with magnetic nature. We do believe that these findings will be interesting for readers with a broader scope of research, such as the other two reviewers.
To try to comply with the reviewer request/wish, we have shortened the second part in the revised manuscript and remove some of the discussion and results to the supplementary information for better illustrating the completion of the manuscript.
We thank the reviewer 3 for his/her additional comments for improving our manuscript.
Personally, I think the observations of a) the stoichiometry dependence of magnetism in LAO and STO and b) the oxygen dicroism in STO (including the clear field swap presented in the revised manuscript) are most striking and deserve publication. As for the LAO/STO interface, the most striking result is the observation of the hysteretic anomalous Hall effect, revealed in the revised version of the manuscript. At the same time this hysteresis raises some new questions as listed below. I can therefore only recommend publication after another revision.
Reply to comments: We thank the reviewer for acknowledging the results of the paper and for recommending it for publication after this revision.
-The hysteretic AHE is highly interesting and could be worth to be moved into the main paper. The authors should comment, however, why the magnetic moment causing the proximity effect seems to be out-of-plane? This is counter-intuitive, as one may expect any moment to align in-plane given the thin layer thicknesses and the 2D-character of the interface.
Reply to comments: We thanks the reviewer for the good suggestion and we have moved the hysteretic AHE part into the main text. Also, we have revised the second part of the manuscript and supplementary information to deliver a clearer message for readers.
The main question raised by the reviewer is why the cation defects create preferential out-of-plane magnetic moments as experimentally observed in SQUID, XMCD, and magneto-transport measurements. The magnetic anisotropy of LAO/STO system is often describe using a magnetic model including the oxygen vacancies in TiO2 interfacial layer. For instance, magnetic Ti 3+ ions appear on a regular sublattice after removing each second oxygen from that layer (~50% oxygen vacancy concentration), which can interact ferromagnetically along the in-plane, thus their magnetic moments could align to the in-plane. In thin films with cubic symmetry, the shape anisotropy effect usually forces magnetic moments to be in the in-plane. In our case, the situation is different, the magnetic LAO thin layer is tetragonally compressed in the B-site cation deficient in the LAO/STO system, which can induce the uniaxial magnetic anisotropy as seen in our measurements. Furthermore, the cation defect-induced magnetic configurations in the B-site cation deficient LAO/STO play their own role and forms the preferential out-of-plane magnetic axis below: In the manuscript, we have suggested a possible mechanism for the preferential outof-plane magnetic moment of the B-site cation defects in the system with an exchange interaction of electrons around the consecutive cation defects, dipole-dipole interactions, and a planar orbital magnetism. In brief, these defects are characterised by unpaired p-electrons of O, which surround an Al or Ti vacancy, or by partially occupied Ti 3d electrons in the LAO environment. These electrons are "almost" free to move within a confined space (cavity) around the defects. These cavities are magnetic due to the unpaired O p-orbitals or Ti 3d-electrons and interact magnetically with each other. Using the magnetic force theorem we estimated the exchange interaction between the cavities. In the LAO/STO interfaces this interaction leads to a ferromagnetic order. The magnetic interaction within a cavity results in a strong effective magnetic field (Weiss field), which forces almost free electrons to move inside the cavity -it is an orbital motion perpendicular to the field direction. This motion leads to a large orbital magnetic moment, which depends on the cavity size and shape, the number of free electrons within the cavity and their spin magnetic moments, the exchange interaction between the moments, and the temperature. For example, the total magnetic moment depends strongly on the cavity size, which could reflect the magnetic moment variation with the B-site cation defect clustering in the system. We believe that this model explains two important facts of our experiments: (i) an enhanced magnetisation due to the orbital motion and the (ii) dependency of the magnetisation on the applied field direction. We also believe that the cavity shape is not spherical along the interface -the asymmetry is due to strong relaxations in the interface vicinity: the cavities are extended along the interface and are squeezed in the direction perpendicular to the interface. Therefore, the total magnetic moments are larger for the case if the applied magnetic field is parallel to Z-axis. However, the theoretical description and details are beyond the scope of this work and can be a subject of following work. More details of our model can be found in the manuscript [1], which will be submitted soon. -It may be technical, but it is unclear from the experimental point of view how the data for forward and backward field sweep was obtained. As described the samples were measured in van der Pauw configuration. Were the data averaged over different geometries to avoid misalignment voltages? And how does the potential data treatment affect the hysteretic Hall signal, which is symmetric in field and could be averaged out through averaging. Further, it is unclear how the authors technically determined the coercive field from the data.
Reply to comments: To provide additional detailed information on the AHE in the samples, we performed four-point probe Hall effect measurements in relatively low fields as schematically illustrated below. The samples were measured in PPMS using external electronics (Keithely 2400 and Keithely 2000). The magnetic fields were applied, perpendicular to the xy-plane of the samples and Hall resistance were measured in a five-loop sequence of the magnetic field (e.g. 0 → +7 T to -7 T → + 7 T to 0 T).
The main message is that the observed hysteresis in the measurements is not attributed from a geometrical contribution or data averaging. These asymmetry will not produce hysteresis, they maybe only contribute non-linear and asymmetric effects relative to the zero field. So, it is a pure AHE intrinsic contribution of our sample. Data symmetrization could also not affect the hysteretic AHE part. Note that the data recorded in the high-field (B ≥ 13 T) and the low-field (B ≤ 7 T) were used for the twoband analysis and AHE, respectively. We only subtracted the OHE part from the raw data to extract the AHE part -for the case of hysteretic AHE, we did not employ the non-hysteretic AHE equation (in SI9) to describe the deconvoluted hysteretic AHE part.
Importantly, the Rxy difference in the forward and reverse field sweeps provides a solid evidence for the existence of the AHE hysteresis as presented in Figs. 5g,h of the revised manuscript. This technical information for new measurements has been included in the method section of the revised manuscript.
As for the second question, the coercive field (BC) of the hysteresis of the deconvoluted Rxy (AHE) part were determined by using a conventional way, BC = (|+BC|+|-BC|)/2.
-Finally, the model applied to describe the AHE in the RH data (elaborated in the SI) is symmetric in B, while RH now has asymmetric contributions due to the observed hysteresis. Can the authors comment why the model is still valid to be used?
Reply to comments: As we already indicated in our previously revised supplementary information 9: "Note that this functional expression is only valid to describe a non-hysteretic AHE, used for fittings of the symmetized non-hysteretic Hall resistance data (Fig. S9).''. Therefore, in the previous manuscript, we presented only the hysteretic Rxy and RH (AHE) parts after subtracting OHE contribution to the experimental Rxy data (recorded from the fine-field interval measurements) without modelling. For more clarity, we have further revised the supplementary information S9 and Fig. S10.
-O-K-edge XMCD: While I appreciate the clear field swap shown in the new figure 4, I would like to ask why the feature 535eV cannot be related to oxygen-Sr hybrids? In my understanding, the total electron yield in XAS is not similarly surface sensitive as e.g. XPS, so they spectra may contain significant contributions from the substrate in this region. Can the authors clarify?
Reply to comments: Since TEY-detected XAS collects all kind of secondary electrons, XAS and XPS have a nearly identical probing depth when the Auger emission is very small, which typically occurs for medium to heavy atoms [Frazer et al., Surface Science 537, 161 (2003)]. As for the first question, the reviewer is right in the sense that there is certainly a contribution from the O 2p-Sr 4d hybrids to the XAS at around 534-537 eV, but the contribution is minor when compared to that of the O bands hybridized with La 5d, Al 3s and Al 3p. For example, in our O K XAS spectrum of 9uc-LAO/STO, the characteristic peak of Sr 4d at 536.5 eV cannot be discerned, instead, there is a broad minimum in the intensity. This is because the probing depth of TEY at these energies is only ~4 -5 nm and the signal for the Sr 4d is hidden by the dominant signals originating from the LAO top layer.
-I also think the statement on the Ti XMCD originating from Ti Al defect sites should be rephrased. Obviously, XMCD is element sensitive. Hence, the authors observe XMCD originating from Ti with no doubt. Their conclusion the respective Ti ions are located in the LAO layer forming anti-sites however cannot be made based on the XMCD measurement alone. The authors should clarify this to avoid misunderstanding.
Reply to comments: We thank the reviewer and agree with his/her suggestion. Regarding the point on the Ti-Al antisite defects, we refer to the following text in the supplementary S7: "Intriguingly, the Ti L2,3-edge XMCD measurements exhibit prominent dichroism (Δρ of the total absorption signal) at a photon energy of 457.5 eV, as illustrated in Fig. S7c, while a broad and asymmetric dichroism feature in the O K-edge was observed at RT (Fig. S7d). The 457.5 eV sharp dichroism reflects the in-plane magnetization of the t2g orbitals of the Ti 4+ ions, with respect to the anisotropic linear dichroism (Fig. S7a). This is consistent with our DFT results for the magnetization of TiAl antisite defect in LAO, i.e. the dxy orbital magnetization of TiAl in LAO (Figs. 3c,d in the main text). No spectral features related to Ti 3+ ions can be observed and the spectral shape of our XMCD spectrum is very similar to the pure Ti 4+ spectrum reported by Saluzzo et al. [S7]. This observation also confirms that there is minor or no impact by Ti 3+ on the magnetism of the oxidized B-site cation deficient interface." The referee is completely right. From the Ti L2,3 XMCD, we can only say that the measured XMCD is a characteristic of Ti 4+ and there is no magnetic signature associated with Ti 3+ . Ti L2,3 XMCD cannot distinguish between Ti 4+ in STO and TiAl 4+ antisite defect in LAO although the Ti signal is barely seen from STO.
To avoid any misunderstanding, we have removed the statement in the revised supplementary information S7.
-As for the defect dynamics, I see the argument made by the authors, but I would suggest to comment on this in the paper. It is not obvious that anti-site defect can be formed under the nonequilibrium conditions of PLD, and should hence be mentioned in the manuscript.
Reply to comments: We would agree with the reviewer comments that a defect compensation process across the interface could be rather limited under nonequilibrium/moderate growth condition in PLD. However, according to our experimental observation (XPS and STEM-EELS), it is clear for Ti out-diffusion behaviour to the Al-deficient LAO overlayers keeping an ABO3 perovskite structure as other research groups have also observed similar Ti diffusion and cation exchange behaviours at similar interfaces [Chen, et al. Nat. Commun. 4, 1371(2013]. No segregation of Ti/TiO2 was observed in the Al-deficient LAO overlayer. Although all the films were grown in thermodynamically non-equilibrium states/partial relaxation, such a cation exchange near the interfaces are commonly observed. This is strongly dependent on the cation stoichiometry of oxide overlayers and the interface roughness [NaKagawa, et al. Nat. Mater. 5, 204 (2006) This anti-site defect formation in the B-site cation deficient system could be further promoted by longer relaxation (e.g. longer cooling-process and post-annealing) after film growth. Thus, it would be interesting to undertake a further detailed investigation on a microscopic view of Ti diffusion behaviour and defect clustering with different relaxation process after film growth, but this is beyond the scope for the main message of the present work.
Following the reviewer's comments, we have modified the manuscript accordingly: ''Our findings indicate that the formation of interfacial magnetism stems from the atomic charge compensation across the interface region between the B-site cation deficient LAO and STO. Such a cation exchange/intermixing is commonly observed at a complex oxide interface, which can be strongly dependent on the stoichiometry of oxide overlayers [Ref. 23,Ref. 24,Ref. 44]. Our results indicate that the Al vacancy sites in the LAO overlayer play a crucial role in initiating the creation of secondary magnetic components and consecutive defect alignment.'' -Finally, a quantitative estimation of a magnetic moment per defect -be it anti-site, cation vacancy, or something else should be made. I understand that exact values for each potentially involved magnetic center cannot be determined, but it should be made plausible that the measured total moment is consistent with the discussed (superimposed) mechanisms. A similar plausibility argument was also requested by reviewer #1.
Reply to comments: Experimentally it is still challenging to quantify a magnetic moment per single defect in the B-site cation deficient LAO/STO system. Therefore, we have proposed a possible mechanism for the generation of the observed large total magnetic moment of the B-site cation defect clusters in the B-site cation deficient LAO/STO system. The proposed model for a large collective orbital moment could reflect the experimentally observed large magnetic moment of the samples, strongly dependent on the B-site cation deficiency and defect assembly in the LAO/STO system as already proposed in the manuscript: ''In other words, the total magnetic response of the B-site cation-deficient heterostructure results from an extended defect configuration at the interface, instead of isolated defects or individual layer contributions. Furthermore, the preferential orientation of the magnetic moment of the B-site cation-deficient LAO/STO system is revealed to be the out-of-plane in [001] direction (Fig. 4e) and temperature-independent between 5 K -300 K (Fig. 4f). The observed large ferromagnetism could be related to exchange interaction of the ordered defects with a large spin wave stiffness 35,36 and/or a collective planar orbital magnetism 37 .'' | 19,420.8 | 2020-07-20T00:00:00.000 | [
"Materials Science",
"Physics"
] |
Slow-roll inflation in Palatini $F(R)$ gravity
We study single field slow-roll inflation in the presence of $F(R)$ gravity in the Palatini formulation. In contrast to metric $F(R)$, when rewritten in terms of an auxiliary field and moved to the Einstein frame, Palatini $F(R)$ does not develop a new dynamical degree of freedom. However, it is not possible to solve analytically the constraint equation of the auxiliary field for a general $F(R)$. We propose a method that allows us to circumvent this issue and compute the inflationary observables. We apply this method to test scenarios of the form $F(R) = R + \alpha R^n$ and find that, as in the previously known $n=2$ case, a large $\alpha$ suppresses the tensor-to-scalar ratio $r$. We also find that models with $F(R)$ increasing faster than $R^2$ for large $R$ suffer from numerous problems.
Introduction
Past and recent observations of the cosmic microwave background radiation (CMB) support the flatness and homogeneity of the Universe at large scales. Such properties can be explained by assuming an accelerated expansion during the very early Universe [1][2][3][4]. This inflationary era is also able to generate and preserve the primordial inhomogeneities which generated the subsequent large-scale structure that we observe. In its minimal version, inflation is usually formulated by adding to the Einstein-Hilbert action one scalar field, the inflaton, whose energy density induces the near-exponential expansion.
Recently, the BICEP/Keck Array experiment [5] has reduced even more the available parameter space, disfavoring most of the minimal inflationary models. On the other hand, the maybe most popular inflationary setup, the Starobinsky model [1], still lies in the allowed region. Such a model involves non-minimal gravity and it can be equivalently described by the addition of a R 2 term to the Einstein-Hilbert action or by a scalar field non-minimally coupled to gravity (e.g. [6] and references therein).
However, in the context of non-minimally coupled theories there is more than one choice of the dynamical degrees of freedom. In the common metric formulation, the metric tensor is the only dynamical degree of freedom, while the connection is fixed to be the Levi-Civita one. On the other hand, in the Palatini formulation, both the metric and the connection are independent variables. Their corresponding equations of motion (EoMs) will dictate the eventual relation between the two variables. When the action is linear in the curvature scalar, the two formalisms lead to equivalent theories (i.e. the Levi-Civita connection arises from the solution of one EoM), otherwise the theories are completely different [7] and lead to different phenomenological predictions, as recently investigated in e.g. .
In particular, there is a dramatic difference between the metric and the Palatini formulations when an R 2 term is added to a single scalar field inflationary action. In the metric case, we obtain a bi-field inflationary setup (e.g. [53,54] and references therein), while in the Palatini case we still obtain a single field scenario [19]. In the latter case, it is remarkable that the presence of the R 2 term leaves essentially unchanged all the phenomenological parameters except for the tensor-to-scalar ratio r, which can be arbitrarily lowered by increasing the coupling in front of the R 2 term. Motivated by the results of [19], one wonders if any other F (R) in the Palatini formulation can produce analogous results. However, such models are more complicated to study than F (R) = R + αR 2 . When studying F (R) theories, it is common to use a representation via an auxiliary field and move the problem to the Einstein frame. In Palatini F (R) = R + αR 2 the EoM for the auxiliary field is independent of α and quite simple to solve [19]. Unfortunately, the same does not happen for any generic F (R), where the EoM for the auxiliary field may not be analytically solvable at all. The purpose of our work is to present a method that allows the computation of inflationary predictions even in such a case.
The article is organized as follows. In section 2, we introduce the theory of a single field inflaton in the presence of Palatini F (R) gravity and develop a new method that allows slow-roll computations even when it is not possible to solve exactly the EoM of the auxiliary field. In section 3, we apply our method to test scenarios of the form F (R) = R + αR n with arbitrary n > 1. Finally, in section 4, we check the behaviour of the aforementioned scenarios beyond the slow-roll approximation, in particular in the high R limit. We present our conclusions in section 5. In appendix A, we compare our results to an earlier work [55] that discussed similar models.
Slow-roll computations
We start by considering the following action for a real scalar inflaton φ minimally coupled to a F (R) gravity (in Planck units: M P = 1): where k(φ) > 0 is the non-minimal kinetic function for the inflaton and V (φ) its positive scalar potential 1 . We stress that we are considering the Palatini formulation of gravity by using the notation R(Γ), where R is the curvature scalar and Γ ρ µν is the connection in the 1 The study applies also to theories with φ non-minimally coupled to gravity, if it is possible to perform a frame transformation and cast the action in the form of (2.1) (or equivalently (2.2) e.g. [6,40]). The cases where this is not possible need to be investigated separately. We postpone such a study to a future work.
Jordan frame. As is customary, we rewrite the F (R) term using the auxiliary field ζ where F (ζ) = ∂F/∂ζ. Then, we move the theory to the Einstein frame via the Weyl transformation which leads to the action where the canonically normalized scalar χ is defined by 5) and the full scalar potential is . (2.6) By varying (2.4) with respect to ζ, we get its EoM in the Einstein frame, where we assumed F (R), F (R) = 0. The standard procedure would be now to solve (2.7), determine the solution for the auxiliary field as ζ(φ, ∂ µ φ∂ µ φ) and insert it back into the action (2.4). However, for a generic F (R), we cannot expect (2.7) to be explicitly solvable, even though it should still be satisfied. On the other hand, we might still be able to perform inflationary computations in the slow-roll approximation. Assuming that slow-roll conditions are satisfied (i.e. g µν J ∂ µ φ∂ ν φ V (φ)) 2 , we can write the EoM (2.7) as In the slow-roll approximation we still cannot always provide explicit solutions for ζ for a given F (R). However, it is possible to perform inflationary computations by using the auxiliary field as a computational variable and using (2.8) as a constraint. First of all, by using (2.8), we replace V (φ) with G(ζ) in (2.6), obtaining a scalar potential 3 that depends only on the auxiliary field ζ: (2.10) 2 We will go back to the full evolution of the system in Section 4 and discuss the slow-roll conditions in more detail. 3 Note that U (ζ) is an actual scalar potential only when the kinetic term of φ is negligible like in slow-roll.
We stress that the last result implies that ζ > 0 in the slow-roll regime, since both the potential U (ζ) and F (ζ) (controlling the signs of the Weyl transformation and the kinetic term) should be positive there. It was shown in [19] that F (R) = R + αR 2 gives an asymptotically flat potential, regardless of the initial V (φ). We can easily deduce from (2.10) that no other asymptotic form of F (R) than F (R) ∼ R 2 can give such a result. Let us proceed and perform inflationary computations. We need to compute the first derivative of U (ζ) with respect to χ, the canonically normalized scalar field in the Einstein frame: where we used the chain rule for the derivative of composite functions together with (2.5), and V −1 (G) is the inverse function of V (φ) defined via (2.8). In the end, after using (2.9) for G, we have a function of ζ only. Similarly, for a general function f (ζ), we have: (2.13) This derivative can be explicitly computed as long as V is invertible. This allows us to easily express higher order derivatives: From (2.11) and (2.14) we can derive the slow-roll parameters straightforwardly: Hence, the equations for the CMB observables r, n s , A s read: where A s has to satisfy [56] ln 10 10 A s = 3.044 ± 0.014 (2.20) at the CMB scale. Analogously, the number of e-folds becomes where value at the end of inflation, ζ f , is determined by 4 (ζ f ) = 1. Equation (2.21) determines the auxiliary field value ζ N at the time a given scale leaves the horizon, corresponding to N e . In our examples below, we take N e ∈ [50,60] at CMB. Before proceeding to examples, we will briefly comment on the mandatory requirements of our procedure in the following subsection.
Requirements
First of all we need to satisfy the usual requirements of any non-minimal gravity model, i.e. reproducing Einstein gravity as a low energy limit and having the correct positive sign in the Weyl transformation (2.3) (i.e. F (R) > 0). Then, we need to satisfy the constraint in eq. (2.8). This induces several additional conditions. First, V (φ) needs to be an invertible function, so that we can define the function g(ζ) in eq. (2.13). The function g(ζ) must also be uniquely defined, therefore G(ζ) must be a bijective function, at least in a smaller domain that satisfies eq. (2.8). Last but not least, since V (φ) is positive everywhere, the same must be true also for G(ζ) in the region of validity of the slow-roll approximation, i.e. ζ > 0. As we can see from eq. (2.9), this is not, in general, true for an arbitrary F (R). The problem lies in the fact that G(ζ) is the difference of two generally positive terms F (ζ) and ζF (ζ). Assuming that F (ζ) ∼ ζ n for very large positive ζ, we can see from (2.9) that G(ζ) is positive in this limit only when n ≤ 2. If that is the case, since G is continuous and behaves linearly around 0, G will take all possible positive values ensuring the existence of a ζ that satisfies (2.8) (and also (2.7)). We can easily extend the same reasoning to functions F that do not possess a monomial asymptotic behaviour, leading to the following summary of requirements for successful scenarios: While it is relatively easy to satisfy the first four constraints, the last condition reduces noticeably the number of available F (R) models. However, when (2.26) is not satisfied, it is still possible to perform some inflationary computations if some additional constraints are realized. When (2.26) does not hold but (2.22) and (2.23) do, G(ζ) is a function bounded from above (in the real positive domain) with at least one local maximum. This means Figure 1: Reference plots of G(ζ) (left) and V (φ) = φ 2 (right) for F (R) = R + R n with n = 3/2 (continuous) and n = 5/2 (dashed).
that (2.8) can be satisfied only for the V (φ) values that lie within the upper limit of G(ζ).
In this case, inflationary computations can still make sense if slow-roll is realized within such a region and G(ζ) is treated as an effective description. In Fig. 1, we present a visual example of the issue. In the left panel we show a reference plot of G(ζ) for F (R) = R + R n with n = 3/2 (continuous) and n = 5/2 (dashed), while in the right panel we show a reference plot of V (φ) = φ 2 . As we can see in the right panel, V is covering all positive values in the y-axis as expected. The same happens in the left panel for n = 3/2 i.e. when (2.26) is satisfied. Therefore, a ζ that satisfies (2.8) always exists. On the other hand, for n = 5/2 i.e. when (2.26) is not satisfied, G presents a local maximum and then decreases towards negative values. Therefore, when the V value is high enough, there is no real solution for ζ that can satisfy (2.8). Hence, an effective description is the only available working option. The positivity and bijectivity of G (together with condition (2.22)) can be still realized within the origin and the local maximum of G. Therefore slowroll must be realized within this interval in order to have at least a feasible inflationary model.
In the following section, we will present a numerical study considering test scenarios in both n < 2 and n > 2 configurations. After that, we will comment on the beyond slow-roll behaviour and see in more detail the problems that arise if (2.26) is not satisfied for large ζ.
Test scenarios
In this section we test our method with We consider two different scenarios 5 , n < 2 and n > 2. We stress that, during slow-roll, the positivity of both U (ζ) and F (ζ) implies ζ > 0 and α > 0. Before proceeding, we also check that our procedure reproduces the results of [19] with F (R) = R + αR 2 for any kind of V (φ).
n = 2
We can easily verify that for the F (R) = R + αR 2 we get the same results as in [19]. These results can be cast in the following form: where . . . 0 means that the quantity is computed for α = 0. First of all, we check the scalar potential. Eq. (2.9) becomes in agreement with eq. (3.2). Let us check now the inflationary parameters, starting with r. First of all, we need to evaluate the g function: Therefore, the tensor-to-scalar ratio becomes again in agreement with the result of [19] in eq. (3.3). Analogously we can prove the validity of the remaining results in eq. (3.4).
n < 2
In this subsection we apply our method on the model in (3. .
where we used the hypergeometric function with (q) n the (rising) Pochhammer symbol. Notice that the n = 2 case is also described by eqs. (3.9)-(3.12). We can also derive more readable expressions in the limit |n − 2|α → ∞ (which automatically excludes the n = 2 configuration). In such a limit we can approximate the number of e-folds as N e ∼ n 8m 2 ζ N , (3.14) obtaining We show in Fig. 2 The orange areas represent the 1,2σ allowed regions coming from the latest combination of Planck, BICEP/Keck and BAO data [5]. The numerical results in Fig. 2(a) were obtained by varying the parameter m from m = 6.98 · 10 −6 (N e = 50) and m = 5.82 · 10 −6 (N e = 60) to m = 1.6 · 10 −5 (both N e = 50, 60). This is equivalent to increase the parameter α since the relation between the two parameters is fixed by the amplitude of the power spectrum (cf. eqs. (2.19) and (3.17)), as shown in Fig. 2(d). From Fig. 2(b) we notice that the net effect of the αR n term is to lower the tensor-to-scalar ratio r. As we get closer to n = 2 we see that this effect is enhanced, as expected, approaching the asymptotic value (3.15) for α → ∞. A similar discussion also applies to n s in Fig. 2(c). In this case, n s increases until α is big enough, and then n s approaches the asymptotic value (3.16).
To conclude, we stress that the limits for α → +∞ with n = 2 and n = 2 are two completely different configurations. However, as can be deduced from Fig. 2, keeping n fixed, it is always possible to identify a maximum value for α so that the results between n = 2 and n = 2 are indistinguishable and this maximum value increases with n getting closer to 2. Then, for α values above such a maximum, the results for n = 2 and n = 2 depart, converging towards two different asymptotic configurations. In particular, eqs. (3.10) and (3.11) with n = 2 are asymptotic limits, approached when n is very close to 2, but never reached. Our results show that at a given α 1, a slight variation from n = 2 might completely jeopardize the inflationary predictions of the n = 2 case. This happens when α 1 |n−2| .
n > 2
In this subsection, we test our method with the setup of eq. (3.1) but for n > 2. Such a scenario satisfies the requirements (2.22)-(2.24), but not (2.26). Moreover also (2.25) requires additional care. We start by checking the constraint (2.8): for N e = 50 (dashed) and N e = 60 (dotted). In the same color code, we show the limit values rᾱ (eq. (3.21)) and mᾱ (eq. (3.19)) for N e = 50 (small dot) and N e = 60 (large dot). The scalar amplitude A s is fixed to its observed value. The orange areas represent the 1,2σ allowed regions coming from the latest combination of Planck, BICEP/Keck and BAO data [5].
as an effective theory if slow-roll is realized between 0 and ζ max . In such a domain we also manage to satisfy (2.25).
In the region of validity of the model the inflationary parameters are still given by (3.9)-(3.12). We show in Fig. 3 a more detailed numerical analysis, where we plot r vs. n s (a), r vs. α (b), α vs. n s (c) and m vs. α (d) for n = 3 (blue), n = 5/2 (red), n = 9/4 (green) with N e = 50 (dashed) and N e = 60 (dotted). For reference we also plot the results for n = 2 with N e = 50 (black, dashed) and N e = 60 (black, dotted). The scalar amplitude A s is fixed to its observed value. The orange areas again represent the 1,2σ allowed regions coming from the latest combination of Planck, BICEP/Keck and BAO data [5]. The numerical results were obtained by varying the parameter m in the range 3.95 · 10 −6 < m < 5.82 · 10 −6 (N e = 60) and 4.44 · 10 −6 < m < 6.98 · 10 −6 (N e = 50). Once again the net effect of the αR n term is to lower r, and this effect is more enhanced as n approaches 2. However, the effect on n s and m is the opposite with the respect to the n < 2 case. In fact, now, by increasing α both n s and m are decreasing.
We can also see that, for a given n, there is an upper limit on α. Since slow-roll inflation happens between 0 and ζ max , the possible number of e-folds is bounded from above in a given model. In order to get the required amount of e-folds, we need ζ N < ζ max at N e ∈ [50,60]. However, by increasing α, ζ max decreases and the distance between ζ N and ζ max decreases as well. We can set a rough upper limit for α when ζ N = ζ max . The limit is only rough because η has a pole at ζ = ζ max meaning the loss of validity of the slow-roll approximation. Such a pole is reflected in Fig. 3(c) with the appearance of horizontal asymptotes with n s pointing towards −∞. Therefore the actual upper limitᾱ takes place for ζ N not equal, but slightly smaller than ζ max . Nevertheless, we can still provide useful estimates for the limit values of r, m and α by using ζ N = ζ max . First of all, we impose such a condition on the amplitude of the power spectrum (3.12), obtaining a limit for m 2 : where A s satisfies (2.20). We can now compute the number of e-folds (3.9) till ζ max , obtaining We can use (3.20) as a definition forᾱ. Using the previous results, we obtain the limit value for r rᾱ (n − 2)(ᾱ(n − 2)n) The limit values rᾱ and mᾱ are shown respectively in Fig. 3(b) and Fig. 3(d) for n = 3 (blue), n = 5/2 (red) and n = 9/4 (green). The small (large) dot stands for N e = 50 (60) e-folds. As we can see, the numerical values for r approach closely the limit ones, but cannot reach them because that would imply a violation of the slow-roll approximation. Analogously, mᾱ and the actual limit of m have the same order of magnitude.
Beyond slow-roll approximation
To gain a sense of the global dynamics of our models, it is interesting to solve their evolution numerically without the slow-roll approximation, in particular for the problematic n > 2 case. Starting from the action (2.4), after some manipulations, the full Einstein frame EoMs read:φ These can be used to also deriveḢ Again, solving the constraint equation (4.3) may be problematic, but it can be replaced with its time derivative which, using (4.1) and (4.3), readṡ To solve the full time evolution, one only needs to solve (4.3) once to get the initial condition of ζ; after that, it is simple to follow the time evolution of ζ through (4.6). The constraint (4.3) can later be used to check the accuracy of the result. Let us check the slow-roll limit of the full equations. There, the potential terms dominate over the kinetic ones. The Hubble constraint becomes 3H 2 = U as usual. The constraint (4.3) becomes (2.8), G(ζ) = V (φ), fixing a one-to-one correspondence between φ and ζ and the new canonical field χ, which in this limit can be defined through (2.5). Since in this limit, ∂ ζ U (φ, ζ) = 0 (by construction), we have and the field equation becomes as expected. In practice, the goodness of the slow-roll approximation can be estimated by comparing the extra terms in (4.1), (4.2), and (4.3) to the leading slow-roll terms.
n < 2
In this subsection, we study the time evolution beyond the slow-roll approximation for the test scenario in (3.1) with n < 2. In particular, we consider the following benchmark point 6,7 : n = 3/2 , α = 8710 , m = 1.15 · 10 −5 . (4.9) The corresponding time evolution given by (4.3), (4.6) in the (ζ, φ) plane is depicted in the flow chart of Fig. 4. The darker orange region corresponds to inflation with H < 1. As expected, inflation takes place only when ζ > 0 (cf. eq. (2.10)). The grey areas represent excluded regions either because F (ζ) turns negative andφ diverges 8 (for ζ < −5.9 · 10 −9 ), or because V (φ) < G(ζ) and the constraint equation (for small φ, large ζ). Slow-roll happens at the edge of this region, where V (φ) ≈ G(ζ), as explained above. The blue dot denotes the CMB scale with N e = 50, and has A s = 2.1·10 −9 , n s = 0.967, r = 0.096. In Fig. 4, it was assumed that H > 0 andφ < 0. This is only one branch of the possible solutions. However, from the EoMs and symmetry of the potential V we see that the system stays invariant under the transformationsφ ↔ −φ, φ ↔ −φ (in particular,ζ does not change). Thus, theφ > 0 branch is obtained by mirroring Fig. 4 with respect to the x-axis. The system can only jump from one branch to the other whenφ = 0, that is, at the slow-roll edge of the right hand side grey region. In Fig. 4, trajectories with φ < 0 end up on the lower edge and switch to theφ > 0 branch: the field rolls up the potential, slows down, stops, and turns around, entering slow-roll on the other branch. Trajectories with φ > 0 approach the upper edge, but slow down due to Hubble friction and enter slow-roll right next to the edge on the same branch. This can be seen as sharp turns in the (ζ, φ) trajectories in the top right panel of Fig. 4. Note that two more branches, with H < 0, can be obtained by simply switching the direction of the flow; these can't be reached smoothly from the inflating branches in a spatially flat universe filled only with a scalar field.
Near the end of a slow-roll trajectory, the field approaches the origin with φ = ζ = 0, oscillating around it. This can be seen in the bottom right panel of Fig. 4. Hereφ changes sign repeatedly, and the evolution jumps from one branch to another as the oscillation amplitude dies down due to Hubble friction. Note that ζ < 0 repeatedly during the oscillations. Due to a rescaling symmetry of the action, the stream lines of a figure like 4 exactly describe the dynamics of a family of models where n and α · m 2(n−1) are kept constant [57], up to a linear rescaling of ζ. This rescaling preserves the values of N e , n s , and r, but changes A s . Beyond that, a chart that is qualitatively similar to Fig. 4 can be drawn for any model of the form (3.1) with 1 < n < 2. Remarkably, in these models all trajectories with large enough initial field values eventually end up in the inflationary region and on a slow-roll trajectory.
n > 2
In this subsection we study the time evolution beyond the slow-roll approximation for the test scenario in (3.1) with n > 2. In particular, we consider the following benchmark point: n = 3 , α = 2.32 · 10 14 , m = 6.40 · 10 −6 . (4.10) The corresponding time evolution given by (4.3), (4.6) (theφ < 0, H > 0 branch) in the (ζ, φ) plane is depicted in Fig. 5. This is analogous to Fig. 4, but with the crucial difference that now G(ζ) has a maximum (here at ζ = 3.79 · 10 −8 , corresponding to |φ| = 17.5 on the slow-roll line) and begins to decrease for larger ζ, making the grey region where constraint (4.3) can't be satisfied terminate at ζ = 6.62 · 10 −8 . Note also that for |φ| > 17.5, not all φ-values are allowed, but there's a minimum value of |φ| needed to solve (4.3). This limits the possible initial conditions of the system in terms of the Jordan frame variables.
A new feature emerges for ζ > 3.79·10 −8 , marked by the dashed black line: a divergence ofζ from (4.6), possible here since G (ζ) < 0. The darkest orange region inside this curve is separated from the rest of phase space-as Fig. 5 shows, it either attracts or repels all trajectories around it and can't be dynamically crossed-and while the system inflates there, and is even attracted to a slow-roll trajectory near the grey boundary as shown by the top right panel of Fig. 5, it is driven towards the point (6.62 · 10 −8 , 0) where condition (2.22) is broken and we don't have the usual Einstein-Hilbert low-energy limit for gravity.
We also have G(ζ) < 0 for ζ < 0. The low ζ limit is not given by the condition F (ζ) = 0 as above (here F (ζ) > 0 everywhere), but instead another divergence ofζ (this time caused by F (ζ) < 0), denoted by the left-most dashed line. We cut off Fig. 5 near this line since, once again, it can't be crossed dynamically, so the left hand side is cut off from the observationally allowed trajectories in the vicinity of φ = ζ = 0.
These features are generic for models with n > 2 and signify their unhealthiness. Slowroll can still happen at the edge of the grey area and terminate succesfully near φ = ζ = 0, as depicted by the bottom right panel of Fig. 5, though the maximum number of slow-roll e-folds is limited, as discussed above. However, most trajectories pass the slow-roll region and continue to ζ → ∞, never turning back. In this sense, the slow-roll region is not an attractor of the dynamics on a global scale.
Conclusions
We studied the dynamics of a (minimally coupled) single field inflaton in the presence of Palatini F (R) gravity. Since such a scenario is not always explicitly solvable, we developed a method that allows the computation of the inflationary parameters if certain conditions are satisfied. Apart from the usual requirements of a generic F (R) theory, as reproducing GR in the low energy limit or having attractive gravity, we found one additional constraint to be important: for curvatures going to infinity, either F (R) should not diverge or it should not diverge faster than R 2 . In case this last requirement is not satisfied, the theory exhibits problematic UV behaviour with additional divergences in phase space, though it can treated as an effective theory during slow-roll inflation. Moreover, to successfully apply our procedure, the Jordan frame inflaton potential V (φ) has to be an invertible function of φ.
We applied our method to a test scenario of an inflaton with a canonical kinetic term and a quadratic potential embedded in F (R) = R + αR n gravity. We computed the inflationary predictions for both the n < 2 case, which satisfies all the requirements, and the n > 2 case with problematic behaviour. Both cases share the same effect on the tensorto-scalar ratio r: for α increasing, r decreases, with the effect getting enhanced with n approaching 2. For the other phenomenological parameters, the two cases have opposite behaviours: for α increasing when n < 2 (n > 2), n s and m are increasing (decreasing). Moreover, for n > 2 and given n and N e , α shows an upper limit. We also checked the evolution of the system beyond the slow-roll approximation for both models. We discovered that n < 2 behaves well: all trajectories with large enough initial field values eventually end up in the inflationary region and on a slow-roll trajectory. On the other hand, in the n > 2 case, the slow-roll region is not an attractor of the dynamics on a global scale, another sign of the intrinsic illness of this setup.
We conclude with a remark about the n = 2 scenario. Such a configuration has been quite a powerful tool to adjust inflationary models, reducing r while leaving n s practically unchanged. However, our study proves that in the strong coupling limit α 1, a slight variation from n = 2 can induce a large change in the n s predictions. This might have a dramatic impact in ruling in/out inflationary models, especially in light of the increased precision of future experiments (e.g. Simons Observatory [58], PICO [59], CMB-S4 [60] and LITEBIRD [61]).
Our notation
Notation of S. Bekov et al. [55] φ χ V (φ) = 1 2 m 2 φ 2 U (χ) = 1 2 m 2 Table 1: A dictionary between the notation of this article and that of [55]. Note, in particular, the relation of the time derivatives: we use the Einstein frame time, whereas [55] works in the Jordan frame, leading to a difference related to the conformal factor F (ζ). We have set k(φ) = 1, and worked in Planck units, M P ≡ κ −1 = 1.
solution. We see similar behaviour in our F (R) = R+αR n case, where G becomes negative when the αR n term starts to dominate at large field values, see the discussion in sections 2.1, 3.3, and 4.2. Indeed, in our slow-roll results in section 3.3, slow-roll always happens in the small-field regime where the linear R term is still significant.
In addition to the sign error, the authors of [55] used slow-roll parameters computed in the Jordan frame in the standard CMB formulae (2.17), (2.18). However, these formulae assume metric Einstein-Hilbert gravity with a canonical scalar field, and thus only work with slow-roll parameters computed in the Einstein frame as we did in our sutdy. These problems render the results of [55] invalid and explain the differences between our results and theirs. | 7,644 | 2021-12-22T00:00:00.000 | [
"Computer Science"
] |
Text-Relevant Video Segments and Reading Comprehension of Culturally Unfamiliar Texts with Adult Speakers of English as a Second Language (ESL)
This pilot study is a mixed-methods sequential explanatory study that investigated the effects of watching text-relevant video segments supported by the Schematic Information-Processing (SIP) model of reading comprehension on reading comprehension of adult speakers of English as a Second Language (ESL) in terms of text familiarity and the retention of the content of a culturally unfamiliar text. Twenty homogeneous participants were randomly assigned to two groups: one control group and one experimental group. The results of the quantitative analysis showed that the experimental group outperformed the control group in both the immediate and delayed post-tests. Qualitative findings also highlight the effectiveness of implementing text-relevant video segments supported by the SIP model over traditional text-based reading instruction in terms of reading comprehension and retention of the content of the culturally unfamiliar passage. Limitations of the study and suggestions for future research are discussed in detail.
Introduction
Reading comprehension is the ultimate goal of reading instruction (e.g., Nation, 2019;Smith et al., 2021;Snow, 2002;Spencer & Wagner, 2018). In other words, reading comprehension is central to "academic progress, because it underpins content-area learning in all subjects" (Smith et al., 2021, p. 215). Successful reading comprehension is a complex and multifaceted process in the human mind. In fact, reading comprehension is an active and interactive process (Anderson, 1999;Grabe & Stoller, 2002) between different cultural, linguistic, psycholinguistic, meta-cognitive, and cognitive aspects. Karami et al. (2021) divide the most critical factors affecting reading comprehension into internal and external factors. The individual's linguistic knowledge, background knowledge and experience, vocabulary knowledge, and internal motivation (e.g., Aferbach et al., 2017;Brevik, 2015;Brown, 2017;Fountas & Pinnell, 2001;Hollenbeck, 2011;Karami, 2021;Snow, 2002;Woolley, 2011) are some internal factors while the cultural familiarity of the reader with the text (e.g., Carrell & Eisterhold, 1983;Droop & Verhoeven, 1998;Sabatin, 2013), the context, the text, and the teaching strategy applied by the teacher (Karami, 2021) are some of the external factors. Among all these factors, the teacher's teaching strategy is the only one of which the teacher has full control and can adapt based on their observations. Brevik (2019bBrevik ( , p. 2281 highlights this idea and argues that a deep understanding of a text "requires sustained emphasis on reading comprehension instruction and scaffolded strategy practices." Although reading comprehension has been extensively researched in the literature, some scholars still believe that further investigation is needed (August & Shanahan, 2006;Hwang & Duke, 2020;Samson & Lesaux, 2015). The review of the literature shows that "researchers apply the same set of models" to their research . Some examples of the most consistently applied models are the "Construction-Integration Framework (Kintsch, 1988(Kintsch, , 1998Kintsch & van Dijk, 1978), the Dual-Coding Theory (Clark & Paivio, 1991), the Cognitive Load Theory of Multimedia (Moreno & Mayer, 1999), and the Generative Theory of Multimedia Learning (Mayer, 1997).
This pilot study steps further and paves the way for future studies by taking a new reading comprehension model into consideration. The Schematic Information-Processing (SIP) model of reading comprehension (Karami, 2021) is a new model that provides theoretical support for the utilization of text-relevant video segments in reading instruction to teach culturally unfamiliar texts to English language learners (See Karami, 2021 for detailed discussion).
The purpose of this pilot study is to examine the relationship between text-relevant video segments and reading comprehension of culturally unfamiliar texts supported by the SIP model of reading comprehension (Karami, 2021) with adult speakers of English as a Second Language (ESL). The authors chose a culturally unfamiliar text and paired it with relevant video segments to investigate whether text-relevant video segments can improve the familiarity of the readers with the content of the text, their reading comprehension and retention. The following quantitative (Q1 and Q2) and qualitative (Q3) research questions were addressed:
1)
Is there a significant difference in terms of reading comprehension between two types of reading instruction, watching text-relevant video segments supported by the SIP model of reading comprehension and the traditional text-based reading instruction, in adult speakers of English as a Second Language (ESL)?
2)
Is there a significant difference in terms of the retention of the content between two types of reading instruction, watching text-relevant video segments supported by the SIP model of reading comprehension and the traditional text-based reading instruction, in adult speakers of English as a Second Language (ESL)?
3)
How do text-relevant video segments help adult speakers of English as a Second Language (ESL) comprehend a culturally unfamiliar text?
Review of the Literature
The history of using audio-visual materials, as a supplementary tool, in classrooms goes back to World War II when filmstrips were used to train soldiers (Hovland et al., 1949). Since then, the innovation of technology and the development of new teaching methods have helped educators recognize "the power of audio-visual materials to capture the attention of learners, increase their motivation and enhance their learning experience" (Cruse, 2007, p. 1). Video players and televisions then become the simplest and the most accessible technology in language classrooms in different parts of the world which makes videos and video segments one of the most important teaching tools in classrooms.
"The utilization of videos in reading instruction is a familiar strategy across the curriculum" (Karami, 2021, p. 2) and has been investigated extensively in the past. The findings of previous studies also highlight the idea that videos can promote learning. Concerning second/foreign language pedagogy, for example, the findings show significant impacts of using videos in classrooms (e.g., Arthur, 1999;Canning-Wilson, 2000;Hemei, 1997;Lin, 2016;Wright, 1976). For example, Lin (2016) conducted a study to find the effects of video-based instruction on text comprehension of the second language learners and reported significant beneficial effects of video-based materials on overall comprehension of the text. Some other researchers (Caspi et al., 2005) focused on other areas such as students' preferences and argued that video-based reading instruction can be either promising or unfavourable for students depending on their learning habits and preferences.
An overall review of the articles related to the utilization of audio-visual materials in language learning classrooms, in general, and videos and reading comprehension, in particular, shows the efficacy of this language teaching strategy . In fact, most of the studies have reported the positive effects of using full-length or short videos on various aspects of language learning (Akbulut, 2007;Caspi et al., 2005;Lin, 2016). For example, Webb and Rodgers (2009a) highlighted the benefits of watching videos in the target language and stated that these resources can improve language learning due to the repetition of the target language and viewers' exposure to it. The review also shows that video-based reading instruction was based on the same set of theories such as the Cognitive Theory of Multimedia (Moreno & Mayer, 1999), the Dual-Coding Theory (Clark & Paivio, 1991), the Construction-Integration Model (Kintsch, 1988(Kintsch, , 1998Kintsch & van Dijk, 1978), and the Cognitive Components-Resource Model (Atkinson & Shiffrin, 1971).
This pilot study was based on the guidelines of the Schematic-Information Processing (SIP) model of reading comprehension developed by Karami (2021). According to Karami (2021), the SIP model is a theoretical framework that …provides theoretical support and evidence for the positive effects of using text-relevant video segments to teach culturally unfamiliar texts in second/foreign language classrooms. According to the SIP model, appropriate and well-chosen text-relevant video segments should be played in three stages of reading instruction: pre-reading, while-reading, and post-reading as the text is being read in the classroom. It is important to include text-relevant video segments in all three stages because 'the methodology of inserting the supplementary videos is very important for improvement of teaching' (Ljubojevic et al., 2014, p. 287). It should be mentioned that text-relevant video segments are playing a supportive role in this model which means that they should be considered as supplementary tools in teachers' toolbox. (p. 3)
Purpose of the Study
Although previous studies have investigated the effects of watching videos on reading comprehension, vocabulary acquisition, and other areas of language, no study has investigated the effects of watching text-relevant video segments on reading comprehension of culturally unfamiliar texts in terms of text familiarity and retention of the text. This pilot study fills this gap by interpreting the results of both qualitative and quantitative data and paves the way for future studies. The combination of both will provide "a more complete and synergistic utilization of data" (Wisdom & Creswell, 2013, p. 1).
Participants of the Study
Twenty international graduate students were selected and assigned to two different groups: a control group and an experimental group. The participants were enrolled at a mid-southern land grant institution of almost 28,000 students in the United States of America. Convenience sampling was used since participation was voluntary. Participants of this study were all from Iran. Farsi was their first language. They all had learned English in an English as a Foreign Language (EFL) setting (Iran). Although participants were raised in different cities of Iran with specific microcultures, they all shared the greater Iranian culture.
Homogeneity of the Participants
Each group had 10 participants composed of six males and four females. The participants were randomly assigned to either the experimental group or the control group. The average age of the participants was thirty-three years old (M = 33) and all had learned English as their second language in an EFL setting. Since they were graduate students majoring in different disciplines, they had met the minimum admission requirements of the university such as the Graduate Record Examination (GRE) and an English language proficiency test. The participants of this study had a minimum score of 6.5 on the International English Language Testing System (IELTS), or a minimum score of 79 on the Internet-based Test of English as a Foreign Language (TOEFL).
Participants were selected based on their unfamiliarity with a culturally specific topic, American football. The content of a reading passage that discussed the rules of the game was both culturally and linguistically unfamiliar to the participants because American football is neither played nor televised in Iran. Therefore, participants were most likely homogeneous in terms of their cultural knowledge and experience, as well as their prior knowledge of American football.
A pre-test was administered to ascertain the unfamiliarity of the participants with the topic and the content of the reading passage. The minimum score was 1 out of 20 and the maximum score could be 20 out of 20. The results of the pre-test showed that the control group (M = 1.30, SD = .483) and the experimental group (M = 1.40, SD = .516) had almost the same mean scores and standard deviations of the same low magnitude which means the participants of the study were not familiar with American football and its rules and regulations.
Research Instruments
An article about American Football by Osorio (2011) was chosen and adapted for this study. The short article, three and a half pages, focused on the basic rules and regulations of American football. The article starts with a brief description of American football and moves to some factors and variables that can affect the results of the game.
Two authentic video segments about American football were selected and used as a part of the instruction of the experimental group. The video segments were produced for English native speakers and were chosen from the YouTube channel. One of the video segments was taken from the NFL (National Football League) website (https://www.youtube.com/watch?v=3t6hM5tRlfA), which was one minute and 19 seconds long, and the second video segment was six minutes and 20 seconds long (https://www.youtube.com/watch?v=Ddwp1HyEFRE&vl=tr).
The total number of 11 questions designed to measure the participants' comprehension and retention of the text included ten closed-response questions consisting of fill in the blank, multiple-choice, short answer, and true or elt.ccsenet.org English Language Teaching Vol. 14, No. 10;2021 false with one open-response question. The same set of questions was administered in three stages: pre-test, immediate post-test, and delayed post-test. Figure 1 shows an example of the reading comprehension questions.
Figure 1. An Example of the Reading Comprehension Questions
The reading comprehension questions were carefully designed by the researchers and reviewed by two literacy experts not only to cover different parts of the text but also to include different formats as well. The first ten questions measured recall of the details of the passage; i.e., the basic rules and regulations of American football and how teams gain scores.
The last question was a descriptive question focused on the overall understanding of the text by asking the participants to provide a summary of what they learned from the one-session instruction about American football without concern for grammar, spelling, and punctuation. The purpose of this descriptive question was to compare the experimental group with the control group in terms of their recall and retention.
Qualitative data were collected using a semi-structured interview with two randomly selected participants from each group to answer the third research question. The number of interview questions for the control group was seven and for the experimental group 13. The first seven questions concerned participants' feelings, motivation, perception, background knowledge, and prior experience. Figure 2 shows an example of interview questions designed for both groups.
Figure 2. Examples of Interview Questions Designed for Both Groups
The experimental group had six additional questions about the effects of watching video segments on activating prior knowledge and cultural experiences. Figure 3 shows an example of interview questions designed specifically for the experimental group.
Reliability and Validity of the Instruments
The results of the immediate and delayed post-tests were used to find the test-retest reliability of the tests. Test-retest reliability is used to measure the consistency of the tests over time. Test-retest correlation was reliable and significant (r = .785, p < .05). Figure 4 shows this correlation.
Figure 4. Test-Retest Correlation
The intraclass correlation was also used to avoid overestimate relationships due to the small sample size in this study. The result of the intraclass correlation is also significant (p < .05) and indicates good reliability between the two tests. Table 1 shows the results of the intraclass correlation performed by SPSS 24. Both sets of questions, comprehension and interview questions, were reviewed by two language experts to make sure that they were valid in terms of face validity and content validity. This was done to make sure that questions measured what they were expected to measure.
Variables
Two independent variables and one dependent variable were identified in this study. The independent variables of this study are as follows: • Teaching strategy as the between-subjects factor with two levels: traditional vs. video-based.
• Time as the within-subjects factor with two levels: immediate post-test (IPT) and delayed post-test (DPT).
Reading comprehension is the only dependent variable of this study. Since the study includes two independent variables with two levels of each and one dependent variable, a 2*2 Mixed ANOVA was performed to analyse the data.
Assumptions of 2*2 Mixed ANOVA
Uttley (2019) states that "Parametric tests rely on certain assumptions about the way in which data were collected and the way in which they are distributed" (p. 145). He emphasized four assumptions that must be tested in parametric tests; a) The dependent variable must be measured on an interval scale. To begin with the first assumption, reading comprehension scores were recorded on an interval scale through the administration of tests in such a way that the difference and distance between scores were meaningful. The second assumption was also met because the participants of the study did not influence each other in terms of data collection. Each group received different treatments, and the responses from one participant to another were independent. The assumption of normality was checked through the evaluation of graphical representation, descriptive statistics, and "statistical tests of deviation from normality" (Uttley, 2019, p. 150). The output of the normal Q-Q plot ( Figure 5) for the video-based instruction in immediate post-test shows that the magnitude of the violation is small. While an ANOVA test can be used for small violations (Uttley, 2019), some researchers (Field et al., 2012) also state that an ANOVA is fairly robust to violations of homogeneity of variance in groups that have equal sample sizes such as the equal sample sizes of this study. Schmider et al. (2010) highlighted this and argued that results will still be close to the results of the normal distribution even if there is a minor violation of the assumption of normality. Therefore, the minor violation of the data in the results of the immediate post-test in video-based instruction can be ignored. To check the homogeneity of variance, "Levene's test can be used to test whether this assumption of homogeneity of variance is true" (Uttley, 2019, p. 151). Table 3 shows the results of this test. Since there is no significant result in this test, the data meet this assumption as well; therefore, according to the testing of assumptions, a 2*2 Mixed ANOVA is an appropriate parametric test for this study.
Reliability and Validity of the Instruments
The results of the immediate and delayed post-tests were used to find the test-retest reliability of the tests. Test-retest reliability is used to measure the consistency of the tests over time. The results showed evidence of test-retest reliability. Test-retest correlation is significant with p < .05 and reliable with Pearson's r = .785.
Both sets of questions, comprehension and interview questions, were reviewed by two language experts to make sure that they were valid in terms of face validity, content validity, and construct-related validity to make sure that questions measured what they are expected to measure.
Intervention
The experimental group received one sixty-minute treatment session. The participants in this group read the text about American football and watched two video segments during the instruction. The control group read the same text during the sixty-minute session without any video support but with the same instruction.
To familiarize the participants of the experimental group with the rules and regulations of American football, a video segment (one minute and 19 seconds) was played at the beginning of the sixty-minute session. This video segment served as an introduction to American football including such facts as the number of participants of each team, the length of the game, the aim of the game, and the name of some techniques performed in this game.
Participants in the experimental group were then directed to begin the reading passage. At the appropriate section in the text, some parts of the second video segment were played based on the content of the reading passage and the context of the video. For example, if the paragraph was about the techniques of the game such as interception, that part of the second video was chosen and played about interception. The second video was viewed in its entirety during the reading. When the participants concluded reading the passage, the first and second video segments were played again without stopping to confirm and stabilize the input that they had received during the session.
The control group had a sixty-minute session without exposure to the video segments. This group simply read the passage and received the same instruction as the experimental group without watching the video segments.
Both groups received immediate post-test questions at the end of the treatment session and were not allowed to use the text to answer the reading comprehension questions. One week later, both groups received and answered the same set of questions as the delayed post-test. The participants of the study knew in advance that they would receive three tests as their pre-test, immediate post-test, and delayed post-test due to the consent forms that they had signed before participating in this study. The same tests were administered in three stages including a pre-test, an immediate post-test, and a delayed post-test to measure participants' retention, their over-time progress, and to achieve meaningful results regarding comprehension. Having both groups receive the same test administration and questions, assured there would be no test bias. In addition, participants' answers to the open-ended question showed whether their answers to reading comprehension questions were most likely due to learning or by chance.
Quantitative Results
Statistical Package for Social Science (SPSS) 24 was run to answer the first and second research questions. The results of the descriptive statistics show that the mean score of the control group (M = 12.40, SD = 2.01) is lower than the mean score of the experimental group (M = 15.70, SD = 1.76) in the immediate post-test (IPT). According to the results, the overall performance of the experimental group was better than the overall performance of the control group in terms of the comprehension of the culturally unfamiliar text.
The results of the descriptive statistics also show that the participants of the experimental group retained the content of the reading passage longer than the participants of the control group. The experimental group in the .553 To answer the second research question, a test of Within-Subjects Effects was performed. The results showed a significant difference between participants' retention in the video vs. traditional mode. The results show a statistically significant relationship between post-test scores F(1, 18) = 6.517, p < .05, which indicates that the video-based instruction had significant effects on the results of the post-tests. In other words, a significant relationship between teaching strategy (traditional vs. video-based) and longer retention of the content of the culturally unfamiliar text of adult speakers of English as a Second Language (ESL) is revealed in this study. The interaction effect between post-tests and reading modes was not significant F(1, 18) = 1.197, p > .05, but there is an ordinal interaction between them. Table 5 provides a summary of the SPSS output. The ordinal interaction indicates that both reading the passage and watching video segments were influential strategies for reading comprehension. It is also worth noting that including text-relevant video segments in two stages of reading instruction promoted retention. The group with videos outperformed the group with no video segments in post-test scores indicating the efficacy of including video segments in teaching the culturally unfamiliar text. The results also show a minimal reduction in the scores of the experimental group from immediate to delayed post-test while the scores of the control group were more reduced than the scores of the experimental group. This indicates that including text-relevant video segments in reading instruction helps the retention of the content and the level of retention of the culturally unfamiliar text.
To examine the retention level of the participants, Question 11 (open-response question) was graded holistically on both the immediate and delayed post-tests. The scoring rubric was adapted from a recommended holistic rubric by Kelley and Clausen-Grace (2007). The scores of the rubric ranged from one (the lowest score) to five (the highest score). A "five" indicates that the participant provided a good summary of the text that is comprehensible and accompanied by the correct responses regarding the rules and regulations of American football while a "one" indicates a poor summary that is incomprehensible with little to no understanding of the rules and regulations. Figure 6 shows a comparison of the holistic scores of the two groups of the study in both immediate and delayed post-tests. The results of the open-ended question show that the experimental group had better overall comprehension of the passage, receiving higher scores in comparison to the control group. The mean score of the experimental group is 4.3 (M = 4.3) out of 5 which is 0.8 greater than the mean score of the control group (M = 3.5) indicating that text-relevant video segments were beneficial in the comprehension of the culturally unfamiliar reading passage for this study.
To compare the retention level of these two groups, the results of the delayed post-test were also demonstrated in Figure 7. The results of the inferential statistics show that integrating text-relevant video segments with text can facilitate the comprehension process and improve the comprehension of a culturally unfamiliar text. The improvement in the comprehension of the text points to better retention of the content of the text as well.
Qualitative Results
Interview responses were transcribed, followed by initial coding to highlight words or phrases that were repeated or appeared to stand out. Saldaña (2009, p. 81) states that initial coding "is not necessarily a specific formulaic method." These words and phrases were then reviewed a second time to look for themes. Creswell (2013) states that re-coding of data after the initial cycle of coding also helps the researcher improve the reliability of the findings. Following the guidelines of the thematic analysis outlined by Braun and Clarke (2006), three themes were identified: culturally unfamiliar texts, multiple strategies, and text-relevant videos with deeper levels of processing information.
Culturally Unfamiliar Texts
Interviewees of both the control and the experimental group believe that the text was useful, informative, and easy to read. Interviewees also acknowledged that it was their first time reading and learning about American football.
One of the control group interviewees answered, "Honestly, I had no idea about American football," showing cultural unfamiliarity. The other control group interviewee stated that "there was no overlap between American football and my own culture." One of the interviewees stated that he had seen some players "passing the field… and I was watching the players from outside," showing that he/she was curious about American football.
Interviewees of the experimental group mentioned the importance of videos and their roles in the comprehension of the text. "Basic things are explained very well." One of the interviewees highlighted some features that the videos could clarify such as the number of players.
Since soccer is played in the interviewees' home country, it is expected to see a connection between these two sports in the individual's mind indicating the integration of ideas with prior knowledge and the formation of "a strategically associated memory structure about the content" (Wijekumar et al., 2018(Wijekumar et al., , p. 1974. Two interviewees of the control group reported that they could not make any connection between the text and their prior experience indicating a lack of schema regarding American football, which could hinder the successful comprehension of the text. One of the interviewees of the control group mentioned a lack of connection between American football and his/her prior knowledge and stated that "There is no connection between American football and other sports that I know of." This was confirmed by a second interviewee from the control group saying that "I could not make connections." The interviewee believed that some basic terms were the same as soccer, which is well-known and played in his/her country but was not a very deep connection. The reason for the lack of connection between American football and soccer for these individuals could be that they might know less about soccer, or it could be about reading.
This is also true for the participants of the experimental group. They had no idea and could not make any connections to their background knowledge except for the number of players and some basic terms such as defensive or offensive, which are the same terms used for soccer players in their country. One of the interviewees reported that he could make a connection between soccer and American football while watching video segments because of these similar terms.
One of the interviewees of the experimental group mentioned the role of text-relevant video segments in remembering the answers to the reading comprehension questions and said, "When we watched the video, after that, following the text was kind of easier for me."
Multiple Strategies
Interviewees reported positive effects of reading the text on their understanding of American football. "Despite the efforts of some of my American friends explaining American football to me, this passage really helped me, and I learned a lot of things about it." One of the interviewees from the control group stated that even though he had been to the stadium before and had watched American football in the stadium, he had no clue how American football was played and what the rules were before reading this passage. It is clear from this sentence that only watching video segments is not enough and cannot lead to successful comprehension of the culturally unfamiliar text. The interview responses indicate that the integration of video segments and reading was more beneficial.
One of the interviewees from the experimental group talked about his/her feelings towards American football after reading this text and said, "I have a good feeling about American football right now, and I think I understand all of it." Videos along with reading instruction, as an integrated strategy, can facilitate the comprehension process of the culturally unfamiliar text. Interviewees of the experimental group emphasized that watching the video segments for the second time (while-reading stage) and discussing the rules simultaneously clarified the text.
Text-Relevant Videos with Deeper Levels of Processing Information
Interviewees of the experimental group mentioned the role of text-relevant video segments in understanding the culturally unfamiliar text. They stated that the "video was great. In the video, it was the basic concept about everything." The interviewees of the experimental group mentioned that text-relevant video segments also enhanced their interest and motivation to learn more about American football. One of them stated that "without video, I was not interested to follow, actually, this subject." Concerning the level of interest and the motivation that text-relevant video segments could provide, interviewees reported positive effects. For example, one interviewee from the experimental group mentioned the role of video in enhancing motivation and stated that "video was a great kind of motivation to follow the text." The interviewees of the experimental group also talked about the long-term effects of watching text-relevant video segments and argued that they would like to follow American football, find games on the Internet, or go to the stadium to watch the game. This shows that reading about American football and watching text-relevant video segments promoted their motivation and interest in such a way that they want to continue learning more about it.
With respect to the role of text-relevant video segments and their effectiveness on vocabulary learning, interviewees answered, "Yes, exactly." They believed that text-relevant video segments explained technical words very clearly and clarified their meanings. One of the interviewees of the experimental group could remember some words. The interviewee believes that he/she remembers words because of watching the text-relevant video segments. One of the experimental group interviewees mentioned the role of videos and said, "I mostly remember the video." Interviewees of the experimental group also mentioned the positive role of videos in understanding some technical words that they had not heard before. Control group interviewees mentioned the problems that they had in terms of vocabulary and stated that they needed some extra help such as Google to understand the meaning of those words. Learning words incidentally through watching videos related to the text can provide readers with an opportunity to see the authentic use of the words produced by native speakers in the target culture (Karami, 2019).
Interviewees were also asked to provide researchers with any suggestions that they thought were important. One of the interviewees of the control group mentioned figures and pictures. The interviewee said, "I was reading through the text and then I was like hey, what does it mean. Then I was trying to imagine what it means or how it looks like in reality. I wish you had some pictures in that to show it." Interviewees also believe that a combination of text and related videos can lead to even more successful comprehension although one of them said that "I think video, alone, can help a lot." The findings of the thematic analysis also highlight the effectiveness of the utilization of text-relevant video segments in the three stages of reading instruction. According to the interview responses, watching text-relevant video segments helped readers retrieve their prior knowledge and experience, facilitated the comprehension process, and promoted reading comprehension of the culturally unfamiliar text. Watching text-relevant video segments also enhanced readers' interest and motivation and provided readers with a clear picture and outline of the game.
Limitations and Suggestions
This pilot study investigated the effects of watching text-relevant video segments supported by the Schematic Information-Processing (SIP) model of reading comprehension on reading comprehension of a culturally unfamiliar text with adult speakers of English as a Second/Foreign Language (ESL/EFL). The sample size of this study was small with 20 adult international graduate students of the same nationality. Future studies need to be conducted with larger sample sizes and various linguistic and cultural backgrounds to help the generalizability of the results.
The participants of this study were all adults with the same English language proficiency level. Future studies could be conducted with younger students to examine the effect of watching text-relevant video segments on their reading comprehension. The treatment session was limited to reading one text in one 60-minute session. Future research could extend sessions to multiple texts and sessions with longer periods of time to see the results and delayed effects of the instruction on second/foreign language learners' language proficiency in terms of cultural familiarity, reading comprehension, listening, speaking, vocabulary knowledge, grammar, and students' feelings and perceptions before and after watching the video.
This study did not differentiate between males and females in terms of their cultural knowledge. Therefore, future studies could consider this by comparing a group of males with a group of females to find out whether gender stereotypes affected the participants' understanding about American football or their prior knowledge about soccer and other sports that could have been relevant.
This study compared video-based reading instruction with traditional text-based reading instruction. Future studies could compare other classroom instructions with video-based instruction to see which instruction leads to more successful comprehension of culturally unfamiliar texts.
Discussion and Conclusion
This pilot study addressed three research questions. Concerning the first research question, the findings showed that there is a significant difference between watching text-relevant video segments supported by the Schematic Information-Processing (SIP) model of reading comprehension and the comprehension of a culturally unfamiliar text of adult speakers of English as a Second Language (ESL). The results imply that playing text-relevant video segments in three stages of reading instruction could cover various aspects that are necessary for successful comprehension of a culturally unfamiliar text like retrieving background knowledge, the construction of schemas, and providing scaffolding.
For the second research question, the results showed that there is a significant difference between the comprehension and retention of the content between the two groups. Although the interaction effect between post-tests and reading modes was not significant, an ordinal interaction between them illustrates a relationship. This implies the idea that viewing text-relevant video segments in three stages of reading instruction could help readers repeat and review the content once more, provide an opportunity for them to see the use of words and sentences in real-life situations, and help them remember the content of the text longer.
The third research question focused on the participants' views and addressed how text-relevant video segments could help adult speakers of English as a Second Language (ESL) comprehend a culturally unfamiliar text better. Interview responses highlighted the applicability of text-relevant video segments in reading instruction in such a way that playing text-relevant video segments in three stages is a multiple-layered strategy that can improve deeper levels of information processing.
Since language learning is a process of active interactions between various sources such as linguistic, cultural, conceptual, emotional, and prior experience, researchers and teachers need to have a closer look into reading comprehension models in general and models of video-based instruction to teach reading specifically. For example, Bohn-Gettler (2019) proposed the Process, Emotion, and Task (PET) framework to study the complexity and the influence of readers' emotions on comprehension. Bohn-Gettler (2019, p. 387) stated that "studies must consider the specific comprehension process under investigation, identify the processes likely to be influenced by emotion, and provide hypotheses for how these processes may be influenced by emotion." Playing text-relevant video segments in three stages of reading instruction not only motivates readers but also facilitates the process of reading comprehension and helps them retrieve their prior knowledge and background experience easier and faster.
This study investigated the effects of watching text-relevant video segments supported by the Schematic Information-Processing (SIP) model of reading comprehension on reading comprehension of a culturally unfamiliar text with adult speakers of English as a Second Language (ESL). The results of the quantitative analysis, as well as the findings of the qualitative phase, show that the SIP model can be an effective strategy to teach culturally unfamiliar texts in adult English language learning classrooms. | 8,528.8 | 2021-09-15T00:00:00.000 | [
"Linguistics",
"Education"
] |
A Novel Transcription Mechanism Activated by Ethanol
Background: The regulatory role of ethanol on gene expression has not been fully defined. Results: Ethanol reduces OCT-1 binding to the Slc7a11 promoter. Mutation of the OCT-1 binding motif abolishes ethanol-induced Slc7a11 promoter activity. Conclusion: Ethanol increases Slc7a11 expression by reducing OCT-1 binding to its promoter. Significance: Ethanol up-regulates gene expression by inhibiting the DNA binding activity of transcriptional repressor(s). Solute carrier family 7, member 11 (Slc7a11) is a plasma membrane cystine/glutamate exchanger that provides intracellular cystine to produce glutathione, a major cellular antioxidant. Oxidative and endoplasmic reticulum stresses up-regulate Slc7a11 expression by activation of nuclear factor erythroid 2-related factor 2 and transcription factor 4. This study examined the effect of ethanol on Slc7a11 expression and the underlying mechanism involved. Treatment of mouse hepatic stellate cells with ethanol significantly increased Slc7a11 mRNA and protein levels. Deletion of a 20-bp DNA sequence between −2044 to −2024 upstream of the transcription start site significantly increased basal activity and completely abolished the ethanol-induced activity of the Slc7a11 promoter. This deletion did not affect Slc7a11 promoter activity induced by oxidative or endoplasmic reticulum stress. DNA sequence analysis revealed a binding motif for octamer-binding transcription factor 1 (OCT-1) in the deleted fragment. Mutation of this OCT-1 binding motif resulted in a similar effect as the deletion experiment, i.e. it increased the basal promoter activity and abolished the response to ethanol. Ethanol exposure significantly inhibited OCT-1 binding to the Slc7a11 promoter region, although it did not alter OCT-1 mRNA and protein levels. OCT-1 reportedly functions as either a transcriptional enhancer or repressor, depending on the target genes. Results from this study suggest that OCT-1 functions as a repressor on the Slc7a11 promoter and that ethanol inhibits OCT-1 binding to the Slc7a11 promoter, thereby increasing Slc7a11 expression. Taken together, inhibition of the DNA binding activity of transcriptional repressor OCT-1 is a mechanism by which ethanol up-regulates Slc711 expression.
that activation of hepatic stellate cells is a central event in alcoholic fibrosis and cirrhosis (2). In addition, an increase in oxidative stress has been shown to affect cell proliferation and differentiation (3), and is thought to play a causal role in alcohol-induced liver cirrhosis (4).
Slc7a11 and solute carrier family 3, member 2 (Slc3a2) encode the light chain (xCT) and the heavy chain (4F2hc) of the plasma membrane cystine/glutamate transporter (system X c Ϫ ), respectively (5). The xCT confers substrate specificity and mediates the transport activity, whereas 4F2hc locates the heterodimer on the plasma membrane (5,6). The system X c Ϫ transports one molecule of cystine into cells and releases one molecule of glutamate into the extracellular space (7,8). After uptake by system X c Ϫ , cystine is rapidly reduced to cysteine, which is an amino acid involved in the synthesis of glutathione (8). As the most abundant endogenous antioxidant, glutathione prevents damage to important cellular components caused by reactive oxygen species (9,10). Thus, induction of system X c Ϫ is essential for cellular protection from oxidative stress (11,12).
The mouse Slc7a11 promoter region contains binding motifs for so-called redox-sensitive transcription factors. For example, there are at least four putative elements similar to the electrophile response element, also called the antioxidant response element, in the upstream promoter region of the Slc7a11 gene. The electrophile response element is recognized by the transcription factor nuclear factor E2-related factor 2 (Nrf2). Induction of Nrf2 binding to the electrophile response element located between 141 and 133 nucleotides upstream of the transcription start site (TSS) 2 has been demonstrated to be a mechanism whereby the electrophilic reagent diethyl maleate upregulates Slc7a11 expression (13).
The expression of Slc7a11 can also be up-regulated by endoplasmic reticulum (ER) stress (14,15). The mouse Slc7a11 gene contains a tandem of amino acid response elements that represent binding motifs for activating transcription factor 4 (ATF4). Binding of ATF4 to the amino acid response elements in the * This study was supported, in whole or in part, by National Institutes of Slc7a11 promoter region has been suggested to be responsible for the induction of Slc7a11 expression by amino acid deprivation (14) and by salubrinal (15). In addition, it has been suggested that ATF4 regulates the basal expression of Slc7a11 (15).
Ethanol has been shown to induce gene expression via various mechanisms, including oxidative as well as ER stress (16 -19). In addition, ethanol has been shown to induce Slc7a11 expression in various cell types. For example, Slc7a11 was increased significantly, 1.54-fold, after exposure to 75 mM ethanol for 9 days using HepG2 cells (20). The mechanism(s) responsible for ethanol-induced Slc7a11 has not been defined.
In this study, we report that ethanol exposure significantly increases Slc7a11 expression in mouse hepatic stellate cells (MHSCs) and that octamer-binding transcription factor 1 (OCT-1) is a repressor of Slc7a11 transcription. Ethanol exposure inhibited OCT-1 binding to the Slc7a11 promoter in MHSCs and, therefore, up-regulated Slc7a11 expression.
EXPERIMENTAL PROCEDURES
Isolation and Culture of Mouse Hepatic Stellate Cells-Hepatic stellate cells represent only 5-8% of liver cells. The immortal cell line of MHSCs used in this study was established from H-2K b -tsA58 transgenic mice, as we described previously (21). Briefly, under anesthesia of ketamine hydrochloride, mice were perfused through the portal vein with a perfusion solution containing 137 mM NaCl, 5.4 mM KCl, 0.6 mM NaH 2 PO 4 , 0.8 mM Na 2 HPO 4 , 10 mM HEPES, 4.2 mM NaHCO 3 , 5.5 mM glucose, 3.8 mM CaCl 2 , and 180 mg/liter collagenase (pH 7.4) at 37°C at a flow rate of 7 ml/min for 10 min. The liver was then excised and incubated in the perfusion solution supplemented with 400 mg/liter Pronase E and 20 mg/liter DNase I at 37°C for 20 min. The mixture was then filtered through a mesh (pore size, 150 mm), and centrifuged at 450 ϫ g for 7 min. The supernatant, enriched with stellate, Kupffer, and endothelial cells, was overlaid with a triple-layered density cushion (Gey's balanced salt solution/8.2% Nycodenz/17% Nycodenz) and centrifuged at 1400 ϫ g for 20 min (22). Gey's balanced salt solution contains 120 mM NaCl, 5 mM KCl, 0.84 mM Na 2 HPO 4 , 0.22 mM KH 2 PO 4 , 1.85 mM MgCl 2 , 1.53 mM CaCl 2 , 27 mM NaHCO 3 , and 5.5 mM glucose. Stellate cells in the upper white layer were resuspended in DMEM and cultured in 100-mm dishes at a density of 1 ϫ 10 3 cells/dish. The H-2K b -tsA58 mice express a thermolabile SV40 tumor antigen (tsA58) driven by a mouse major histocompatibility complex H-2K b promoter (23). Cells obtained from these mice continuously divide at 34°C (23). After 5 days of culture, cell colonies were harvested. Quantitative real-time RT-PCR and Western blot analysis were performed for determination of desmin, a marker gene for stellate cells. Desmin-positive cells were cultured in DMEM supplemented with 10% FBS. Cells were cultured and passaged at 34°C and transferred to 37°C before ethanol treatment.
Western Blot Analysis-Quiescent MHSCs in serum-free DMEM were treated with ethanol, DEM, TM, or culture medium alone (control) at the doses indicated in the figure legends for 6 h. For whole cell protein extraction, cells were lysed in M-PER mammalian protein extraction reagent (Thermo Scientific, Rockford, IL). Samples containing 5 g (for -actin detection) or 40 g of protein (for detection of proteins of interest) were resolved on 10% SDS-PAGE gels and then transferred to a PVDF membrane (Millipore, Billerica, MA). Antibodies against -actin, Slc7a11, OCT-1 (Santa Cruz Biotechnology, Inc., Santa Cruz, CA), followed by detection with horseradish peroxidase-conjugated secondary antibodies (Santa Cruz Biotechnology, Inc.), were used as described previously (24). Immunoreactive bands were visualized using ECLplus chemiluminescence reagent (GE Healthcare/Amersham Biosciences, Piscataway, NJ) and analyzed using a GS-700 imaging densitometer (Bio-Rad).
Recombinant Plasmid Construction and Cell Transfection-For functional analysis of the Slc7a11 promoter, four Slc7a11 promoter fragments (P1-P4) were generated by PCR from wildtype C57BL/6J mouse genomic DNA using platinum TaqDNA polymerase (Invitrogen). The 5Ј end of these fragments starts at nucleotide ϩ22 downstream from the TSS of the Slc7a11 gene. The 3Ј end of the P1, P2, P3, and P4 fragments start at nucleotides Ϫ2044, Ϫ2024, Ϫ306, and Ϫ134 upstream from the TSS of the Slc7a11 gene, respectively. The following primers were used for generation of these DNA fragments: the shared reverse, atcccaagcttGCCCTATCATTACACACCAG; P1 forward, atccgctcgagCCTGTCTATGCTCATTCATC; P2 forward, atccgctcgagTGGAGAATTATGTGAGTGGA; P3 forward, atccgctcgagCGGAGCTGCTTAAAGTCACC; and P4 forward, atccgctcgagAATGTTGGCGCTTTCTCAAG.
The recombinant plasmid constructs were transfected into MHSCs using a Lipofectamine 2000 kit (Invitrogen) according to the protocol of the manufacturer. Briefly, MHSCs grown in six-well plates to ϳ60 -70% confluency were incubated with 2 ml/well serum-free DMEM containing 10 l of Lipofectamine 2000 and 4 g of pGL2 plasmid encoding either the wild-type or mutated Slc7a11 promoter fragments. At 6 h post-transfection, cells were replenished with fresh medium containing 10% FBS and then cultured for an additional 24 h. The transfected cells were treated for 12 h with ethanol, DEM, TM, or culture medium alone (control) at the doses indicated in the figure legends prior to reporter gene analysis.
Luciferase Assay-For the luciferase assay, transfected cells were lysed with 100 l of reporter lysis buffer provided by the luciferase assay kit (luciferase assay system, Promega). Lysate (10 l) was incubated in a 96-well plate at room temperature for 2 min with 100 l of luciferase assay reagent (Promega). Luminescence was measured using the BL10000 LumiCount (Packard BioScience, Meriden, CT). The protein level in the lysate was determined using a BCA protein assay kit (Thermo Scientific). The luciferase activity was expressed as the luminescence intensity relative to the protein level.
ChIP-Quiescent MHSCs were treated with 0, 50 mM, 100 mM, or 200 mM ethanol for 4 h. The binding of OCT-1 to the Slc7a11 promoter region was determined by ChIP as described (24). Briefly, ethanol-treated cells were cross-linked with 1% formaldehyde at room temperature and then lysed in 500 l of cell lysis buffer (5 mM PIPES, 85 mM KCl, 0.5% Nonidet P-40, 1 mM PMSF, and protease inhibitor mixture (pH 8.0)). Nuclei were isolated and homogenized in 300 l of nuclear lysis buffer (50 mM Tris-HCl, 1% SDS, 10 mM EDTA, and protease inhibitor mixture (pH 8.1)). The resulting nuclear lysate was sonicated until cross-linked chromatin was sheared to an average length of 0.3ϳ1.0 kb. Supernatant (5 l) was used as an input control. The remaining lysate was diluted 10-fold with ChIP dilution buffer (16.7 mM Tris-HCl, 167 mM NaCl, 0.01% SDS, 1.1% Triton X-100, 1.2 mM EDTA, and protease inhibitor mixture (pH 8.1)), and precleaned with salmon sperm DNA/protein A-agarose (Santa Cruz Biotechnology, Inc.). The precleared sample was incubated with OCT-1 antibody followed by salmon sperm DNA/protein A-agarose. Bound protein-DNA complexes were eluted with a solution containing 0.1 M NaHCO 3 and 1% SDS. Following reversion of the protein-DNA cross-links, DNA fragments in the eluate and input controls were purified using the QIAquick PCR purification kit (Qiagen) and then subjected to quantitative real-time PCR using an iCycler system (Bio-Rad). The PCR primers used were as follows: 5Ј-TTAGTGGT-CAAAGCCTGGTG-3Ј (forward) and 5Ј-CTCTGTTCCATG-TCTCCTGT-3Ј (reverse). A 200-bp DNA fragment containing an OCT-1 binding site of the Slc7a11 promoter region was amplified. The amount of DNA-bound OCT-1 was expressed as a ratio of the PCR product amplified from the immunoprecipitated samples versus the input controls. The binding of OCT-1 to aryl-hydrocarbon receptor (AhR) and NAD(P)H: quinone oxidoreductase 1 (NQO1) were detected by the same ChIP assay. The AhR ChIP PCR primers were as follows: 5Ј-CCTGGTAAATCTTGATGTCT-3Ј (forward) and 5Ј-CTCC-CAACACTTCTTGCAGT-3Ј (reverse). A 140-bp DNA fragment contains a putative OCT-1 binding site (GGCATAATG-TGGC) (M00137) in the AhR promoter. NQO1 ChIP primers were as follows: 5Ј-TCTAAGAGCAGAACGCAGCA-3Ј (forward) and 5Ј-CCTCCTGGGTACAAAATGGA-3Ј (reverse). A 245-bp DNA fragment contains a putative OCT-1 binding site (GGCTGATTATGTA) (M00137) in the NQO1 promoter.
Measurement of Intracellular ATP Levels-The cellular ATP levels were determined using a bioluminescence ATP detection kit (Promega). MHSCs grown in a white-bottom, clear 96-well plate at confluence were incubated at 37°C in serum-free DMEM (100 l/well) with the indicated concentrations of ethanol or culture medium alone as a control for 6 h. In the experiments for studying the cytotoxicity resulting from the combinational treatment of DNA transfection and ethanol exposure, MHSCs were transfected with empty pGL2 plasmid as described above and then treated with the indicated concentrations of ethanol. Thereafter, 100 l of kit reagent was added to each well. After a 10-min incubation, luminescence was quantified using a BL10000 Lumicount luminometer (Packard BioScience, Downers, IL). Cellular ATP contents were expressed as the luminescence intensity (arbitrary unit) generated in the assay. At the end of the experiments, cells were lysed in M-PER mammalian protein extraction reagent (Thermo Scientific). Protein levels in the lysate were determined using a BCA protein assay kit (Thermo Scientific). The levels of ATP were normalized by the protein levels.
Statistical Analysis-For experiments using the microplate reader, the mean value for each experiment was averaged from triplicate wells in the same plate. Data are reported as the mean Ϯ S.E. of the mean for at least five independent experiments. Differences between the control and treatment groups were analyzed by one-way or multiple factor analysis of variance followed by Tukey's post-hoc tests. Statistical significance was considered when p was less than 0.05. Statistix software (Statistix, Tallahassee, FL) was used for statistical analyses.
Ethanol Exposure Up-regulates Slc7a11 Expression in MHSCs-
Data from previous studies using various cell types indicate that Slc7a11 plays a protective role against injuries induced by a variety of factors (26 -29), including those induced by ethanol (20). This report studied the effect of ethanol on Slc7a11 expression in MHSCs. The data in Fig. 1A show that the basal mRNA level of Slc7a11 is relatively low, i.e. about 0.02% of the GAPDH mRNA level, in MHSCs. Ethanol exposure resulted in a dose-dependent increase in Slc7a11 mRNA levels. Specifically, the Slc7a11 mRNA levels in MHSCs treated with 50, 100, and 200 mM ethanol were 2.8-, 5.1-, and 20.0-fold higher, respectively, than that in untreated control cells (Fig. 1A). Similarly, the same doses of ethanol elevated Slc7a11 protein levels in MHSCs 3.6-, 4.9-, and 8.8-fold, respectively (Fig. 1, B and C). These data suggest that ethanol is able to induce Slc7a11 expression at both the mRNA and protein level.
Ethanol Activates the Slc7a11 Promoter via a Mechanism Involving the OCT1-binding Motif-To explore the DNA motif(s) responsible for ethanol-induced Slc7a11 transcription, a series of promoter-reporter constructs were generated by progressive 5Ј deletion of the Slc7a11 promoter ( Fig. 2A). As the data in Fig. 2B show, ethanol exposure significantly increased the reporter gene activity of the P1 construct, which includes 2044 bp upstream of the TSS of the Slc7a11 gene. We also observed that prolongation of the P1 construct by 126 bp (using MluI/BglII as the cloning sites) at the 5Ј end of the Slc7a11 promoter did not significantly alter the basal and ethanol-induced reporter gene activities (data not shown). In contrast, deletion of 20 or more bp from the 5Ј end significantly increased the basal activity while completely abolishing the ethanol-induced activity of the Slc7a11 promoter. Specifically, the basal activities of the reporter gene in cells transfected with the P2, P3, and P4 constructs were 72, 71, and 38% higher, respectively, than those observed in cells transfected with the P1 construct. Ethanol did not induce a significantly higher reporter gene activity in cells transfected with the P2, P3, and P4 constructs. These data suggest that the 20 nucleotides spanning from Ϫ2024 to Ϫ2044 upstream of the TSS are indispensable for ethanol-induced Slc7a11 promoter activity.
A computer database search of the Slc7a11 promoter showed that an OCT1-binding site (ATGCTCAT) located between 2029 and 2040 nucleotides upstream of the TSS is included in the deleted 20-nucleotide promoter region. The Slc7a11 promoter OCT-1 binding site is highly homologous to the octamer motif (ATGCAAAT) found in the human H2B gene (30). We assessed the impact of this DNA motif on Slc7a11 transcription using a mutated promoter-reporter construct (P1-OM), as shown in Fig. 3A. The data in Fig. 3B show that mutation of the OCT1-binding site abolished ethanol-induced activity but sig-nificantly increased the basal activity of the Slc7a11 promoter when compared with that derived from the wild-type Slc7a11 promoter counterpart, P1. These findings are consistent with the data obtained from the deletion experiments (Fig. 2B) and support the view that the OCT1-binding motif plays an inhibitory role on the basal activity of the Slc7a11 promoter and a stimulatory role on ethanol-induced Slc7a11 promoter activity.
Ethanol Activates the Slc7a11 Promoter but Not through Nrf2, ATF4, AP1, and NF-B-A computer database search showed that the eight putative AP1 binding sites in the mouse Slc7a11 promoter are located 46, 140, 298, 429, 529, 549, 685, and 714 nucleotides upstream of the TSS, whereas the putative NF-B binding site is located 438 nucleotides upstream of the TSS ( Fig. 2A) (31). The schematic in Fig. 2A also shows that the mouse Slc7a11 promoter contains a functional Nrf2-binding site and an ATF4-binding site, respectively, which span from 141 to 133 and from 94 to 86 nucleotides upstream of the TSS, respectively. The data in Fig. 2 also show that ethanol did not increase the reporter gene activity in cells transfected with the P2 construct, which contains all the aforementioned binding motifs for Nrf2, ATF4, AP1, and NF-B. Collectively, these data suggest that the induction of Slc7a11 by ethanol is not regulated by Nrf2, ATF4, AP1, and/or NF-B.
The Nrf2 and ATF4 binding motifs have been reported to be responsible for the induction of Slc7a11 expression by both
OCT-1 Represses Slc7a11
oxidative and ER stresses (13)(14)(15). In addition, it has been suggested that the ATF4-binding site controls the basal expression of Slc7a11. This study assessed the function of these DNA motifs using diethyl maleate (an oxidant) and tunicamycin (an ER stress inducer). The data in Fig. 2B show that both diethyl maleate and tunicamycin significantly increased the reporter gene activity by ϳ85% in cells transfected with the P1, P2, and P3 constructs. Thus, as long as the Nrf2-and ATF-binding motifs remained intact, deletion of as much as 1638 nucleotides of the DNA sequence (P1 versus P3 construct) did not significantly affect the induction of Slc7a11 promoter activity by diethyl maleate or tunicamycin.
To confirm the function of the Slc7a11 promoter Nrf2-binding motif in response to oxidative stress, we generated a DNA construct, P4, in which the Nrf2-binding motif was deleted from the Slc7a11 promoter region (Fig. 2A). The data in Fig. 2B show that deletion of the Nrf2 binding motif abolished the induction role of diethyl maleate Slc7a11 promoter activity but did not alter the induction of tunicamycin. These data are consistent with the view that this Nrf2-binding motif is responsible for oxidative stress-induced Slc7a11 transcription.
To study the function of the Slc7a11 promoter ATF4-binding motif, we generated a DNA construct, P1-AM, in which the ATF4-binding motif in the Slc7a11 promoter region was mutated as depicted in Fig. 4A. The data in Fig. 4B show that the basal level of reporter gene activity in cells transfected with the P1-AM construct is 4.5-fold lower than in cells transfected the wild-type Slc7a11 promoter counterpart P1 (Fig. 4B). In addition, tunicamycin did not induce reporter gene activity in cells transfected with the P1-AM construct. These findings are consistent with previous studies showing that the ATF-binding motif controls basal and ER stress-induced Slc7a11 transcription (15). The data in Fig. 4B also indicate that ethanol increased reporter gene activity in cells transfected with the P1-AM construct, although the activity is significantly lower than in cells transfected with the wild-type P1 construct. Thus, mutation of the ATF-binding motif abolished the induction role of tunicamycin in Slc7a11 promoter activity but did not alter the induction of ethanol.
Ethanol Exposure Reduces OCT-1 Binding to the Slc7a11 Promoter but Does Not Alter OCT-1 Expression-Having demonstrated the regulatory role of OCT-1 in ethanol induced Slc7a11 promoter activity, we next studied the effect of ethanol on OCT-1 expression. The data in Fig. 5 show that OCT-1 protein and mRNA levels are comparable in cells with or without ethanol treatment. Thus, ethanol-dependent increases in Slc7a11 expression were not due to ethanol dependent decreases in OCT-1 expression per se.
In contrast to the lack of changes in OCT-1 expression as a function of ethanol concentrations, ethanol exposure reduced OCT-1 binding to the Slc7a11 promoter in a dose-dependent manner (Fig. 6A). As the data show, the amount of OCT-1 bound to the Slc7a11 promoter region was 28.2, 36.8, and 44.6% lower in cells treated with 50, 100, and 200 mM ethanol, respectively, than that seen in untreated control cells. These results suggest that ethanol inhibits OCT-1 binding to the Slc7a11 promoter. Taken together, it is highly likely that OCT-1 functions as a transcription repressor of the Slc7a11 promoter under normal conditions. Ethanol exposure, however, reduces OCT-1 binding to the Slc7a11 promoter and, therefore, increases Slc7a11 expression.
JOURNAL OF BIOLOGICAL CHEMISTRY 14819
Ethanol Reduces OCT-1 DNA Binding in a Sequence-specific Manner-It has been reported that oxidative modification of OCT-1 nonspecifically inhibits its DNA binding activity, i.e. oxidative stress reduces the binding of OCT-1 not only to the octamer motif but also to other OCT-1 binding elements (32). Ethanol is able to induce oxidative stress (16,18). This study, therefore, also monitored the binding of OCT-1 to the promoters of the AhR and the AhR target gene, NQO1. The AhR and NQO1 promoter OCT-1 binding elements contain nucleotides TAATGTGGC and GCTGATTATGT, respectively. They differ greatly from the octamer motif. We reported previously that acute ethanol exposure induced a dose-dependent decrease in AhR mRNA and protein levels in MHSCs (21). In contrast, ethanol did not affect the expression of NQO1 (data not shown). The data in Fig. 6B indicate that treatment of MHSCs with 50 and 100 mM ethanol did not significantly alter the amount of OCT-1 bound to the AhR promoter, whereas 200 mM ethanol diminished OCT-1 binding to the AhR promoter by ϳ25%. In contrast, ethanol induced a dose-dependent increase in OCT-1 binding to the NQO1 promoter (Fig. 6C). These results suggest that the effect of ethanol on the DNA binding activity of OCT-1 varies among genes and that the reduction in OCT-1 binding to the Slc7a11 promoter in the presence of ethanol is a gene-specific effect.
Effect of Ethanol on Cellular ATP Content-Ethanol has been shown to reduce viability and functionality and induce cytotoxicities in liver parenchymal cells (33). However, hepatic stellate cells are relatively resistant to ethanol-induced cytotoxicity as compared with the liver parenchymal cells (34,35). This study determined the effect of ethanol on the metabolic viability of MHSCs by measuring cellular ATP content. The data in Fig. 7 show that ethanol at concentrations of 50, 100, and 200 mM did not reduce the ATP levels in MHSCs incubated for 6 h.
A preliminary study from our laboratory suggested that transfection of MHSCs with PGL2 plasmids reduced ethanolinduced Slc7a11 expression. Specifically, the Slc7a11 mRNA levels increased by 50, 100, and 200 mM ethanol in PGL2-transfected MHSCs were ϳ50% less than in non-transfected cells (data not shown). This study, thus, determined the effect of ethanol on cellular ATP contents in MHSCs transfected with PGL2 plasmids. Fig. 7 shows that the ATP level in PGL2-transfected MHSCs was ϳ10% lower than in non-transfected cells. However, ethanol treatment did not significantly reduce ATP content in the transfected cells. These data, together with the results presented in Figs. 1-6, suggest that the reduced OCT-1 binding to the Slc7a11 promoter region, and the increased Slc7a11 expression and promoter activity in the MHSCs treated with ethanol, are not due to a reduced cell viability.
DISCUSSION
Data from this study demonstrated that ethanol exposure significantly increased Slc7a11 expression at the mRNA and protein level in MHSCs. This finding is consistent with a previous report from our laboratory that ethanol up-regulates Slc7a11 expression in HepG2 cells (20). The main function of the Slc7a11 cystine/glutamate antiporter is to maintain the intracellular level of glutathione and protect cells from oxidative damage (8,12). Thus, up-regulation of Slc7a11 expression might be a compensatory effect in response to ethanol-induced oxidative stress.
Changes in gene expression could result from a changed activity of transcription factors. Indeed, ethanol has been reported to regulate the activity of many transcription factors (17, 18, 36 -39). For example, ethanol has been shown to reduce or increase the binding activity of NF-B to an NF-B oligonucleotide probe in rat pancreatic acinar cells (36) and gastric mucosal epithelia, respectively (37). In addition, it has been suggested that ethanol increases the activity of transcription factors and STAT3, specificity protein-1 (Sp1), and AP1, thereby up-regulating IL-10 expression in human monocytes (38). Moreover, maternal ethanol exposure was found to increase both Nrf2 protein levels and Nrf2-antioxidant response element binding in mouse embryos, resulting in an increase in the expression of a number of Nrf2 target genes (18). Furthermore, ethanol was shown to increase Nrf2 and ATF4 expression (17,18,39). The Slc7a11 promoter contains binding motifs for many transcription factors, including AP1, Nrf2, NF-B, and ATF4 (13,14,31). It has been reported that ATF4 controls the basal expression of Slc7a11 and that activation of Nrf2 (13) and ATF4 (15), in response to oxidative and ER stresses, up-regulates Slc7a11 expression. The data from this study confirmed the role of ATF4 in the basal and ER stress-induced expression of Slc7a11. However, our data clearly indicate that induction of Slc7a11 by ethanol in MHSCs is neither through ATF4 nor AP1, NF-B, or Nrf2 because ethanol could not activate the Slc7a11 promoter-reporter constructs that contained the binding motifs for these transcription factors unless the OCT-1 binding motif was also present.
This report is the first to demonstrate the regulatory role of transcription factor OCT-1 in ethanol-induced Slc7a11 expression. Specifically, we observed that ethanol reduced the amount of OCT-1 bound to the Slc7a11 promoter but did not significantly affect the expression of OCT-1. Deletion or mutation of the OCT-1 binding motif completely abolished ethanol-induced Slc7a11 promoter activity. Previous reports demonstrated that OCT-1 can function as either a transcription enhancer or a repressor, depending on the target genes involved (40 -43). For example, it has been reported that OCT-1 up-regulates the expression of snRNA genes (40). In contrast, binding of OCT-1 to the octamer motif in the mouse -opioid receptor gene negatively modulates its expression (42). Furthermore, OCT-1 can functionally interact with the retinoid X receptor. This interaction interferes with the contact of the retinoid X receptor with the thyroid hormone receptor, which, in turn, reduces the binding of the thyroid hormone receptor/retinoid X receptor to the promoter region of thyroid hormone receptor target genes and, therefore, negatively regulates the expression of these genes (43). Our data clearly demonstrate that OCT-1 functions as a repressor of the Slc7a11 promoter. It is highly likely that binding of OCT-1 to the Slc7a11 promoter constitutively inhibits its transcriptional activity. Ethanol exposure reduces OCT-1 binding to this motif, and, therefore, Slc7a11 expression increases.
The OCT-1 binding motif ATGCTCAT in the mouse Slc7a11 promoter region is highly homologous to the octamer element originally found in the human H2B gene (30). Besides this octamer motif, several variants of DNA elements are able to bind OCT-1 protein (44). Data from this study demonstrated that the effect of ethanol on OCT-1 binding differs among genes. Specifically, ethanol increased OCT-1 binding to the NQO1 promoter in a dose-dependent manner. In the case of the AhR promoter, only a high concentration of ethanol (200 mM) had a significant effect on OCT-1 binding to that promoter. These observations suggest that a reduction in OCT-1 binding to the Slc7a11 promoter is a sequence-specific effect rather than a nonspecific effect of ethanol on the DNA binding activity of OCT-1.
Several kinases and phosphatases have been shown to regulate the binding activity of OCT-1 to the cognate octamer motif (32). For example, PKC has been shown to inhibit the binding of OCT-1 to the octamer motif oligonucleotide probe in vitro (32) and to the promoter in vivo (45). Besides PKC, other protein kinases, such as PKA, have been shown to inhibit the binding of OCT-1 to the octamer motif. In contrast, casein kinase 2 (CK-II) enhances the binding of OCT-1 to this octamer motif (32). In addition, the calcium/calmodulin-activated phosphatase calcineurin is effective in increasing OCT-1-dependent transcription (46). Moreover, phosphatase 2A has been suggested to augment the binding of OCT-1 to the octamer motif (47). Further studies are required to investigate the mechanism underlying the inhibitory effect of ethanol exposure on OCT-1 binding to the Slc7a11 promoter, including the study of the involvement of the above-mentioned kinases and phosphatases in ethanolinduced changes in the OCT-1 DNA binding activity.
In summary, data from this report demonstrated that ethanol exposure up-regulates Slc7a11 expression, increases Slc7a11 promoter activity, and diminished OCT-1 binding to the Slc7a11 promoter. Deletion or mutation of the Slc7a11 promoter OCT-1 binding motif completely abolished ethanol-in- duced Slc7a11 promoter activity. In addition, mutation of the Slc7a11 promoter ATF4 binding motif remarkably reduced the basal level of Slc7a11 promoter. These data suggest that under normal conditions, bound OCT-1 inhibits, whereas ATF4 enhances, Slc7a11 promoter activity. Ethanol exposure removes OCT-1 from the Slc7a11 promoter and, therefore, increases expression of Slc7a11. | 6,778.8 | 2013-04-16T00:00:00.000 | [
"Biology",
"Medicine"
] |
Probing Oral Microbial Functionality – Expression of spxB in Plaque Samples
The Human Oral Microbiome Database (HOMD) provides an extensive collection of genome sequences from oral bacteria. The sequence information is a static snapshot of the microbial potential of the so far sequenced species. A major challenge is to connect the microbial potential encoded in the metagenome to an actual function in the in vivo oral biofilm. In the present study we took a reductionist approach and identified a considerably conserved metabolic gene, spxB to be encoded by a majority of oral streptococci using the HOMD metagenome information. spxB encodes the pyruvate oxidase responsible for the production of growth inhibiting amounts of hydrogen peroxide (H2O2) and has previously been shown as important in the interspecies competition in the oral biofilm. Here we demonstrate a strong correlation of H2O2 production and the presence of the spxB gene in dental plaque. Using Real-Time RT PCR we show that spxB is expressed in freshly isolated human plaque samples from several donors and that the expression is relative constant when followed over time in one individual. This is the first demonstration of an oral community encoded gene expressed in vivo suggesting a functional role of spxB in oral biofilm physiology. This also demonstrates a possible strategy to connect the microbial potential of the metagenome to its functionality in future studies by identifying similar highly conserved genes in the oral microbial community.
Introduction
The Human Oral Microbiome Database (HOMD) contains DNA sequence information from over 1300 genomes [1]. This number is most likely to grow in the future. The wealth of information provided allows the interested researcher to screen the available sequence data for common features and investigate a biological relevant relation to the community structure. For example is a specific gene or a set of genes associated with a healthy community or a predictor of a pathogenic oral biofilm?
One caveat of this approach is that the metagenome only gives a static snapshot of the oral bacterial community potential at the time of sampling [2]. It will not allow determining a dynamic relationship or the functionality of the bacterial community at any given time. While the potential is determined by the metagenome, the functionality of the bacterial community is driven by its metatranscriptome and metaproteome, thus at the RNA and protein level [2,3]. However, investigating the metatranscriptome and the metaproteome comes with certain challenges, including the amount of RNA or protein required to successfully cover the entire metatranscriptome or metaproteome of any given plaque sample. For example detection and quantification of low abundance transcripts by RNA Seq can be challenging [4] since the starting material for RNA isolation, dental plaque, is limited. Furthermore, the amount of species open reading frames would lead to large amounts of diverse sequences reads requiring specific software to adequately manage the data [5][6][7].
Alternatively, the metagenomic information could be used to identify an ecological relevant oral community encoded gene. The expression of this gene in human plaque samples could be determined directly thus probing its functionality in vivo. Ideally several members of the oral microbial biofilm share this gene. We previously identified a high homology of the pyruvate oxidase gene spxB between the oral commensal Streptococcus sanguinis and Streptococcus gordonii [8]. SpxB is an oxido-reductase, catalyzing the conversion of pyruvate to acetyl phosphate, CO 2 and H 2 O 2 under aerobic conditions [9,10]. The gene provides several advantages for the encoding species. Initially, we identified H 2 O 2 as inhibiting substance in the dual species competition with cariogenic Streptococcus mutans [11,12]. Deletion of the spxB gene in both S. sanguinis and S. gordonii rendered them non-competitive against S. mutans demonstrating that SpxB is responsible for competitive H 2 O 2 production [12]. Besides the inhibitory H 2 O 2 action, two additional effects increasing commensals competitiveness are obvious: i) ATP production from acetyl phosphate for energy generation and ii) H 2 O 2 induced release of extracellular DNA (eDNA) [12]. eDNA is a major component of the extracellular polymeric substance (EPS) of biofilms, promoting cell-cell and cell-tooth contact [13,14]. H 2 O 2 induced release of eDNA can also serve in horizontal gene transfer promoting genetic diversity as we demonstrated recently [15]. In addition, the ability to produce inhibitory amounts of H 2 O 2 seems to be limited to oral streptococci [16].
In this pilot study, we demonstrate that spxB is a suitable candidate gene encoded by several important commensal streptococci, identified by homologous sequence search using the HOMD database. Oral streptococcal production of H 2 O 2 as measured with specific indicator plates seems to correlate with the presence of the spxB gene as determined with PCR and spxB specific oligonucleotides. Furthermore, we isolated RNA from freshly isolated human plaque samples and demonstrate that spxB is expressed with Real-Time RT-PCR, suggesting a functional role in the oral biofilm.
Ethics Statement
The Institutional Review Board of University of Oklahoma HSC approved the study protocol for human subjects. IRB protocol # 1934. Participants signed a written consent form following an approved procedure by the IRB.
Bacterial Strains and Growth Conditions
Bacterial strains are listed in Tab. 1. Bacteria were routinely grown aerobically in 5% CO 2 at 37uC overnight in BHI medium (Brain Heart Infusion; Difco, Sparks, MD) or on BHI plates, or as otherwise indicated.
Subjects and Plaque Sampling
The present study sole intention was to determine the feasibility of measuring spxB gene expression in dental plaque samples. Therefore, no subject related data were collected from the 8 volunteers asked to donate plaque samples. From each individual, supragingival dental plaque samples were recovered between 16 and 18 hours after tooth brushing, from interproximal, vestibular and lingual surfaces of all teeth using a dental probe. The collected plaque samples were immediately removed from the dental probe after a visible amount has built up using a sterile tip and resuspended in 1 ml TRIzol (lifetechnologies). The sampling procedure lasted not longer than five minutes. Resuspended plaque samples in TRIzol were immediately frozen at 280uC until further processing the next day. Plaque samples for plating were removed from the teeth using a cotton swap and directly plated onto the H 2 O 2 indicator plates.
RNA Isolation and cDNA Synthesis
To isolate RNA, cells were disrupted three times for 30 seconds each using 0.1 mm zirconia/silica beads (BioSpec, Inc) in a FastPrep FP210 Homogenizer (Thermo Scientific) using the highest speed setting. Total RNA isolation from TRIzol was carried out according to the manufacturer's instructions (Isolation of total RNA using TRIzol, lifetechnologies). RNA samples were treated with Turbo DNase to remove traces of chromosomal DNA after manufacturers recommendations (Ambion). RNeasy Mini-Elute cleanup kit (Qiagen) was used to purify RNA samples after DNase treatment. 2 ml RNA was immediately run on a 1% agarose gel to check for integrity. cDNA was synthesized using qScript TM cDNA synthesis kit (Quanta Biosciences) according to the manufacturer's protocol.
RNA Integrity
RNA integrity was determined using an Agilent 2100 Bioanalyzer (Agilent Technologies, Inc.). The yield of the extracted RNA and intactness of rRNA was assayed from 1 ml RNA after DNase treatment and RNeasy MiniElute cleanup in a 2100 BioAnalyzer using the protocol of the RNA 6000 Nano Lab Chip Kit (Agilent Technologies, Inc.). The chip allows for analyzing 12 samples simultaneously. RNA was therefore collected first and stored at 280uC until all 8 samples were available for the RNA 6000 Nano Lab Chip run.
RT-PCR and Real-Time PCR
RT-PCR was performed as described earlier [17]. Real-time RT PCR was performed to determine specific cDNA copies with the comparative threshold cycle (CT) method using a MyiQ single-color real-time PCR detection system (Bio-Rad) and PerfeCtaTM SYBR H Green SuperMix for iQ TM (Quanta Biosciences). Relative changes in cDNA copies representing differential gene expression were calculated using the DC T method described previously [18]. The 16S rRNA gene was used as the housekeeping reference gene. Oligonucleotides were synthesized by IDTDNA. Oligonucleotide sequences are: 16S rRNA F -59-AAGCAACGCGAAGAACCTTA-39; 16S rRNA R -59-GTCTCGCTAGAGTGCCCAAC-39; universal spxB F -59-CATCATGGGTGACGGTGCAT-39; universal spxB R -59-GCGTTAGGGAAGTCACAACC-39.
PCR
Colony PCR was performed by scraping a small amount of cells from the respective agar plate using a sterile pipet-tip and resuspending the cells in a pre-aliquoted PCR reaction mix. Alternatively, cells were inoculated in 2 ml BHI overnight; 1 ml was transferred into tubes containing lysing matrix B (MP Biomedicals, Solon, OH) and cells were homogenized in a FastPrep FP210 homogenizer (Thermo Scientific) (speed setting of 6.5). After centrifugation for 10 min at 13.200 rpm in a tabletop centrifuge, 2 ml supernatant containing chromosomal DNA was used as PCR template. PCR was performed with a G-Storm GS1 thermocycler (Gene Technologies) according to the manufacturer's protocol. GoTaq-DNA polymerase was obtained from Promega, and oligonucleotides specific for spxB and 16S rRNA were the same as listed above. S. mutans UA159 [32] S. mitis 12261 [33] S. gordonii DL1 [34] S. gordonii V288 [35] S. oralis MC3-1 [36] S. oralis J22 [36] S. sanguinis SK36 [37] S. sanguinis 133-79 [38] S.
Detection of H 2 O 2 Production
Indicator plates for H 2 O 2 production were prepared and used as described [19]. This indicator plates allow for the detection of bacterial H 2 O 2 production resulting in a blue pigment (Prussian blue, ferric ferrocyanide) that forms in the presence of H 2 O 2 . Plates were aerobically cultured in 5% CO 2 at 37uC overnight.
Results
Prevalence of the spxB Gene in the Oral Biofilm Community A BLAST search was performed using the S. gordonii CH1 nucleotide sequence as query against all available oral microbial genomes on the HOMD server. In addition to S. gordonii, the following oral Streptococci encode spxB homologs: Streptococcus sanguinis, Streptococcus mitis, Streptococcus infantis, Streptococcus oralis, Streptococcus oligofermentans and Streptococcus cristatus (Tab. 2). Furthermore, Streptococcus pneumoniae encodes also spxB as reported in the literature [20] and found during the BLAST search, but was omitted since it is usually not found in the dental associated oral biofilm. The homology of the spxB genes encoded by the oral streptococcal community is high ranging from 93% to 97% on the nucleotide level (against S. gordonii CH1) with a core sequence of about 1700 bp when compared to the 2308 bp of S. gordonii CH1. This suggests that the spxB gene is highly conserved among oral Streptococci. Interestingly, no homolog was found in cariogenic Streptococcus mutans confirming our previous observation that S. mutans does not produce competitive amounts of H 2 O 2 [11].
Detection of spxB with spxB Universal Oligonucleotides
The available sequence data for spxB from oral Streptococci was used to design an spxB specific set of oligonucleotides. Initially, the oligonucleotides were tested for their ability to amplify spxB from several Streptococci present in our laboratory culture collection. The H 2 O 2 production potential of the here used streptococci was assessed first by inoculating 10 ml of an overnight culture on specific H 2 O 2 indicator plates. The plates were incubated overnight to allow for growth and H 2 O 2 production. As shown in Fig. 1 A, with the exception of S. mutans and S. salivarius, all other Streptococci produced H 2 O 2 evident from the formation of a blue pigment. Subsequently, cells were removed and chromosomal DNA isolated to perform PCR with 16S rRNA and spxB specific oligonucleotides. The 16S rRNA control showed amplification for all strains tested. A clear correlation was evident between positive spxB amplification in S. mitis, S. gordonii, S. oralis, S. sanguinis, S. infantis and S. parasanguinis and the formation of the blue pigment (Fig. 1B). No significant amplification was observed for all S. mutans strains and S. salivarius, but we did recognize weak bands after longer integration during picture documentation (Fig. 1B).
To determine if the observed faint bands would interfere with our intention to measure spxB expression by giving false positive amplification products, RNA was isolated from S. mutans, S. gordonii and S. salivarius and cDNA synthesized. RT-PCR with spxB oligonucleotides showed only amplification for S. gordonii in the RT-PCR reaction, but no amplification for the no RT control (Fig. 2). 16S rRNA amplification was positive in all three RT-PCR reactions (data not presented). This suggests that the newly designed oligonucleotides are able to amplify spxB from H 2 O 2 positive oral Streptococci and are suitable to be used in gene expression analysis.
PCR Amplification of spxB from H 2 O 2 Positive Plaque Colonies
To further evaluate the spxB amplification potential of the newly designed spxB specific oligonucleotides, dental plaque was collected from five subjects. The plaque samples were inoculated on H 2 O 2 indicator plates to separate single colonies. Fig. 3A represents plaque samples from two subjects and illustrates that colonies with and without blue pigment can be distinguished. Ten blue colonies from each subject were picked and used for colony PCR with the spxB specific oligonucleotides (Fig. 3B). PCR amplification was successful for all colonies in the five subjects. Obvious, however, was the difference in PCR efficiency. While subject 1 had strong bands for all ten colonies, subject 4 showed some heterogeneity in the band intensity (Fig. 3B). In addition, 45 white colonies were subject to PCR amplification with the spxB oligonucleotides. No amplification was observed (Fig. 3C). In summary, the spxB specific oligonucleotides are able to amplify spxB from freshly isolated H 2 O 2 positive plaque bacteria suggesting that the oligonucleotides could be used in general to evaluate the presence of this gene in human plaque samples.
RNA Isolation from Plaque Samples
The goal of this study is to detect the expression of spxB from freshly isolated dental plaque samples to evaluate the in vivo expression of this gene. The major challenge in the detection of gene expression from host-derived biofilms is the isolation of sufficient amounts of high quality RNA [21]. To test feasibility of RNA isolation, plaque was sampled from 11 subjects asked to refrain from tooth brushing in the morning to collect a sufficient amount of plaque in the afternoon. The plaque samples were immediately processed for RNA isolation. Isolated RNA was analyzed and RIN (RNA Integrity Number) determined. Fig. 4A and 4B showing agarose gel-electrophoresis and Bioanalyzer gel visualization. The respective RIN numbers and RNA concentrations are presented in Fig. 4C. Sample 1 and 2 showed very low RIN and were not further used. The remaining samples ranged from RIN 4.8 to RIN 7.8. The RNA concentration range was from 50 ng/ml to 418 ng/ml.
Detection of spxB Expression from Plaque Samples
Initially the expression of spxB in 9 individual plaque samples was compared using 16S rRNA as housekeeping reference. The expression was normalized to the subject with the highest spxB expression relative to the others. The result presented in Fig. 5 demonstrates that there is a high degree of interpersonal variation in spxB expression. The observed fold difference reached about 88 fold when the highest and the lowest expression were compared (subject 8 vs. subject 3). Contamination by chromosomal DNA was excluded by running a no RT control along with a RT-PCR (data not shown).
In addition, the expression of spxB from a single subject on five different time-points was determined to learn how stable the expression is on a day-to-day basis. As shown in Fig. 6, time point 1 and 2 as well as 3 and 4 were taken on two subsequent days, one hour apart while time point 5 was taken several days later. Although slight variations are visible, the expression of spxB in one subject seems to be constant over time. Furthermore spxB expression can be detected repeatable from RNA isolated from human plaque samples.
Discussion
The HOMD contains DNA sequence information from over 1300 genomes (www.homd.org). An eminent question is what kind of information can be extracted from the deposited sequence data which goes beyond the determination of what is present in the oral cavity or what kind of potential metabolic pathways are encoded [3,22]. The next logical step is to move away from the static information of sequence data to determine if the microbial potential of the chromosomal sequences is converted to a functional response in vivo. The here presented study uses the sequence data provided by HOMD to identify an oral microbial community shared gene, spxB and demonstrates its functional expression in freshly isolated human plaque samples.
Several oral commensal Streptococci encode the spxB gene as we identified in this study. Up to 80% of the detected initial colonizers belong to the genus Streptococcus. Some species are even discussed as constant members constituting a core group for initial biofilm formation [23,24]. The spxB gene is therefore highly abundant during initial biofilm formation and might be considered as core gene involved in oral biofilm formation. The production of an antimicrobial substance like H 2 O 2 could therefore be regarded as an important protection mechanism of the initial colonizers of the resident biofilm community against invading and competing species like the extremely H 2 O 2 sensitive S. mutans [11]. More importantly, it might also be a mechanism to shape the colonization process towards a specific species composition. Only species co-evolved with oral streptococci and therefore adapted to withstand H 2 O 2 can integrate or colonize in close proximity to the initial colonizers and extend the developing biofilm community [8]. The gene is expressed in the different subjects suggesting a function in the biofilm mode of growth of oral streptococci. The next step is to design a large-scale study to relate the expression of spxB with an assessment of the oral health status. A positive association of spxB expression with oral health would allow for a better caries risk assessment in the future. Expression studies on infection relevant genes in vivo in humans and animal models have been done before as summarized in [25]. Three in vivo studies are relevant in the context of our study: i) An investigation of in vivo gene expression in the human host was carried out with Streptococcus pyogenes (Group A Streptococcus, GAS) [26]. The author's demonstrated GAS host-pathogen interactions by analyzing the expression of 17 GAS genes in throat swab specimens sampled from 18 pediatric patients with pharyngitis. Several known and putative virulence genes and regulatory genes were highly expressed during infection [26]; and ii) In a study with the periodontal pathogen Porphyromonas gingivalis, Shelbourne et al. demonstrated a clear correlation between periodontal disease status and elevated expression of dnaK and htpG encoding general stress response proteins [27]; iii) while another study established the expression of an uncharacterized gene in P. gingivalis (G1334) as more frequent in diseased sites compared to healthy sites [28]. The importance of the G1334 gene in virulence was confirmed in a mouse abscess model of infection [28]. The here presented expression of a community-encoded core gene of oral biofilm formation and the single-species focus by other groups detecting virulence and stress related gene expression in vivo demonstrate the feasibility to specifically determine the expression of genes of interest in their ecological context.
Limitations of the study became apparent when the RIN where determined. Although all samples where processed after the same protocol, RNA degradation was a problem for some samples. Two RNA samples were severely degraded and not further considered, while others showed varying degrees of degradation. In general, an RIN of 10 would indicate no degradation. RIN greater than five indicates good total RNA quality for reverse transcription [29]. Some of the samples were close to 5 and therefore would be considered not ideal for spxB expression quantification. Nonetheless, we used the total RNA of subject 3 and 4 for cDNA synthesis to determine spxB expression. The respective relative expression of spxB from subject 3 and 4 were among the lower spectrum, however, subject 5 and 10 showed a similar relative expression level albeit higher RIN. Another caveat of this study lies in the fact that the oral microbial diversity among subjects varies [30] and the spxB sequences will have sequence dissimilarities. This will result in nucleotide mis-pairing of the spxB oligonucleotides used for spxB expression. As a consequence, non-ideal PCR amplification conditions will occur due to different primer efficiencies when plaque samples with undefined species composition are used. This might be prevented in the future if more spxB sequences become available to optimize oligonucleotides for amplification. In addition, the main advantage we take in our approach is the wide distribution of a highly conserved gene among the most prominent genus in the oral biofilm. If other similar important genes can be identified needs to be determined.
In summary, we showed in vivo expression of an oral biofilm community encoded gene in its ecological context, thus suggesting an active role of spxB in oral biofilm physiology. This study demonstrates how the HOMD database can be used to determine a dynamic relationship or the functionality of the potential encoded in the metagenome of the oral biofilm. | 4,874.8 | 2014-01-29T00:00:00.000 | [
"Medicine",
"Biology"
] |
Development of low-cost non-contact structural health monitoring system for rotating machinery
Condition monitoring systems are increasingly being employed in industrial applications to improve the availability of equipment to increase the overall equipment efficiency. Condition monitoring of gearboxes, a key element of rotating machines, ensures to continuously reduce and eliminate costs, unscheduled downtime and unexpected breakdowns. This study demonstrates a low-cost microcontroller-based non-contact data acquisition system for condition monitoring of rotating machinery. Experimental validation of the proposed system was carried out by performing examination tests on a gearbox test rig. A user-friendly graphical user interface was also developed which facilitates users to perform signal processing in both real-time and offline mode. The proposed system can perform most of the functions available in complex, stand-alone vibration analysers. The use of a general-purpose PC and standard programing language makes the system simple, economical and adaptable to a variety of problems. The tests show the developed system can perform properly as proposed.
Introduction
The rapid development of engineering and mechanical components, together with the increase in complexity, cost of acquiring equipment and machine downtime, has attracted the attention of researchers towards the development of non-contact structural health monitoring (SHM) systems. Transformation of modern manufacturing technologies from mass production to lean manufacturing opens a new paradigm to implement the SHM for the improvement in safety, reliability, economy and intelligence against catastrophic failure applied in real-time environments [1,2].
Gears, a vital component of the transmission system, contributes to wide range of applications such as industrial, automotive and daily-life applications and productivity of the system is decreased with failure of any of the machine part. Catastrophic failures related to gears are a result of improper operating conditions and loading, thus leading to failure of the whole mechanism. According to the latest survey [3], the premature failure of gears and bearings has a large share of more than 95% in the breakdown of wind turbine gearboxes. Appropriate maintenance strategy can reduce the unplanned stoppages and minimize the number of failures, thereby reducing serious and costly consequences.
Acoustic as a diagnostic tool
Not only are the current conditions monitored by diagnosis tools, but future conditions of machines are also predicted while in operation. Therefore, the information must be collected externally without affecting the working of machine. Various monitoring approaches like direct and indirect techniques can be put on to any application by adaptations and alterations, to accomplish an acceptable level of consequences. Direct techniques, such as vision and optical methods, measure the actual geometric changes in the mechanical component [4]. These techniques are not used in real-time applications because the machine has to be stopped for physical inspection. Indirect techniques viz. vibration, oil analysis, thermography, laser vibrometer, acoustic analysis and motor current signature analysis are easy to implement. However, vibration, thermography and acoustic monitoring are generally practiced as they reflect the status of the machine instantaneously. Researchers have used measurements of vibrations [5][6][7][7][8][9][10][11][12], acoustics [5,[13][14][15][16][17][18][19][20][21][22] and thermography [23][24][25] to estimate the gearbox condition. In contrast to the aforementioned variables, vibration is one of the features of modern mechanical machinery that is now continuously monitored in many significant applications. However, acoustic-based condition monitoring provides some technical advantages over vibration-based methods. Firstly, acoustic sensors can be placed at any convenient location on the periphery of the monitored system, whereas the vibration monitoring technique requires surface contact placement in specific orientation to get accurate and meaningful information. Secondly, acoustic monitoring is more sensitive to vibrating bodies than vibration sensors, and hence provides an opportunity to identify faults at an early stage [2,[20][21][22]26,27]. Successful implementation of non-contact incipient fault diagnosis of fixed-axis gearboxes has been carried out [28][29][30][31][32].
Structural health monitoring
SHM is a procedure that aims to estimate the structure condition, fault detection and problem diagnostics in the machine by means of the evaluation of some measured physical features. Microphones are the most widely used non-contact-type instrument to measure the acoustic emission from mechanical components. An effective SHM system can improve productivity and ensure work-piece quality contributing to a major influence on machining economics. The effectiveness of the SHM system relies on the sensitivity of sensor used, position of sensor mounting, number of sensors to be used and types of sensors to be practiced. Current health monitoring approaches are much able to determine the [33]: -presence of the fault in the mechanical component (level 1); -geometric location of the fault (level 2); -level/severity or type of fault (level 3); and -anticipation of the remaining lifetime (level 4).
The present work develops a non-contact SHM system using a low-cost microcontroller-based data acquisition system. The system used the microphone to measure the acoustic signal of the gearbox. The produced acoustic signal is then analysed using signal analysis graphical user interface (GUI) developed in MATLAB 2016a to determine the health of the machine. Experiments are carried out on a gearbox test-rig to collect data for the analysis. This study proposes to develop a cost-effective data acquisition system for monitoring the gearbox condition. The objectives of this work include the following: (i) to develop an economical non-contact acoustic sensor-based device to monitor the current status of the gearbox for reliable SHM using microcontroller interfaced with MATLAB; (ii) to correlate the vibration signatures with machine health; and (iii) to develop a GUI for easy implementation of SHM.
Proposed data acquisition system structure
The functional information flow of the proposed data acquisition system is shown in figure 1.
Microcontroller-based structural health monitoring data acquisition system
Acoustic data acquired through the data acquisition system was analysed and implemented for gearbox condition monitoring. The hardware system development, software system development, integration and testing of the data acquisition system, and calibration along with the existing data acquisition system is described in the following sections.
Hardware system development
The non-contact data acquisition system employed off-the-shelf equipment in a novel integrated approach. During gearbox operation, real-time acoustic data were collected through a microphone. The raw signal was processed to extract the characteristic features from the gearbox signals. The various equipments that are used for developing a system are described below.
Test rig
A test rig, gearbox diagnostics simulator was used for simulating faults in rotating machinery. The test rig had a feature to simulate industrial drive-trains, especially as an experimental research tool. Figure 2 shows the experimental set-up used for studying gear faults. The set-up has been designed to simulate real working conditions of a gearbox. The detailed specifications of the test rig are listed in table 1. Table 2 summarizes the specifications of the gearbox.
Acoustic sensor
For acoustic signal acquisition a general-purpose microphone (Ahuja GN45) was used. The microphone gave the output in sound pressure level, i.e. Pascals (Pa). The position of the microphone is fixed at 24 cm vertically from the bottom of the set-up near the meshing pair in accordance to the optimal sensor location experimentation conducted by Vanraj et al. [34,35]
Microcontroller-based data acquisition device
A microcontroller-based data acquisition device (Arduino) has been used to acquire data. This device is interfaced to computer program using MATLAB via a USB port. The device enables acquisition of acoustic signals using a low-cost hardware by using a standard high-speed USB port and PC-based analogue and digital input/output data acquisition. A MATLAB microcontroller is associated with the sensor, and brings forth a digital output after burning the program for the microphone. Figure 3 shows the schematic of the non-contact acoustic measurement circuit. Figure 4 shows the circuit diagram of various component integration. Arduino analogue Pin0 is used to collect the audio input signal. To control the audio effects, a potentiometer connected to analogue rsos.royalsocietypublishing.org R. Soc. open sci input 0 was used. Pin 11 on microcontroller board was used as pulse-width-modulation (PWM) audio output for recording the audio signal in MATLAB. interfaced with the hardware system for real-time data acquisition. Figure 5 shows the GUI developed for acoustic signal analysis. The developed GUI facilitates the user to compute frequency spectra, cepstrum analysis, signal editing of vibration, acoustic and velocity, archiving data in various ASCII formats. Both real-time signal processing and offline signal processing can be carried out for SHM of machinery.
Simulated faults
Since the time period to overhaul a new gearbox may vary from six months to 1 year depending upon working conditions, naturally generated faults and their detection becomes difficult. Therefore, the only option left is to study seeded fault trials in gearboxes. Most common gearbox faults can be categorized as: (i) root crack, (ii) surface spalling, and (iii) chipped tooth. The chipped tooth is the rupture of material from the working tip of a gear. Root crack is extremely common in numerous industrial practices which triggers other gear faults and is very difficult to detect at its initial stage [27].
Calibration
In order to calibrate the results obtained using the developed system, a NI-cDAQ-9178 data acquisition system was used to acquire the acoustic signals using a G. gearbox operating conditions. The sampling frequency of the NI-DAQ was 12.8 kHz and 30 k data points were collected for each case. Time-domain and frequency-domain-based comparison was done to check the accuracy of the proposed system. Mostly used statistical parameters [36] were extracted from the signals for each gear running condition.
Results and discussion
Based on the analysis and the theoretical framework provided in the earlier sections, the experiment has been performed on the test rig, and results obtained from the analysis phase in the development of a non-contact SHM system for gearboxes are discussed in this section. A number of experiments have been conducted to check the accuracy in the measurement of the developed system by calibrating it with a standard microphone and data acquisition system.
Time domain analysis
Firstly, the time domain acoustic signals for different gear conditions recorded using a microcontrollerbased data acquisition device and NI-cDAQ-9178 were compared. Three statistical parameters, viz. root mean square (RMS), kurtosis and crest factor were calculated for each gear condition. Annotations used for graphical representation are as follows: -microcontroller-based data acquisition device using general-purpose microphone: -GN45; and -NI-cDAQ-9178 using a G.R.A.S microphone: -46AE For demonstration purposes, 35 Hz rotational speed time domain data are highlighted. Figure 7 shows the comparison of variation in acoustic amplitude for a healthy gear measured by both 46AE and GN45 acoustic sensors at 35 Hz rotational speed. It has been observed that the developed system gives almost the same values like the standard data acquisition system. Following points were observed for healthy gear condition: -for 46AE sensor, the maximum and minimum values were 1.772 Pa and −2.303 Pa, respectively; whereas, for GN45 sensor, the maximum and minimum values were 1.937 Pa and −2.162 Pa, respectively; and -RMS value of acoustic data obtained from 46AE and GN45 were 0.4789 and 0.4791, respectively. whereas, for GN45 sensor, the maximum and minimum values were 3.664 Pa and −3.304 Pa, respectively; and -RMS value of acoustic data obtained from 46AE and GN45 were 1.028 and 1.021, respectively. Figure 9 depicts the comparison of variation in acoustic amplitude for RC50 measured by both 46AE and GN45 acoustic sensors. Following points were observed for the RC50 gear condition: -for 46AE sensor, the maximum and minimum values were 9.743 Pa and −10.33 Pa, respectively; whereas, for GN45 sensor, the maximum and minimum values were 8.942 Pa and −10.3 Pa, respectively; and -RMS value of acoustic data obtained from 46AE and GN45 were 2.09 and 2.001, respectively.
The variation of calculated statistical parameters for different gear conditions at various speeds is shown in figures 10-12. It was observed that the statistical parameters are well separated for different rotational speeds for all gear conditions. Also, the values of statistical parameters for both the acoustic sensors were found to be close to each other for each gear condition at all rotational speeds. It implies that the proposed system could acquire the acoustic data with the same accuracy as that of existing analysers. The per cent deviation of statistical parameters obtained using two acoustic sensors is listed in table 4. It was observed that the per cent deviation for RMS and kurtosis is under acceptable range (under 10%) for all running conditions, however, per cent deviation for crest factor was little high (above 10%).
Features evaluation and economics analysis
Features evaluation and economics analysis of the proposed SHM system was conducted by comparing the features with state-of-the-art benchtop equipment available in the market. Both hardware and software capabilities were compared as listed in table 6. The proposed system has a variety of software features incorporated using MATLAB algorithms, such as signal editing utility, FFT, waterfall and spectrogram, power spectral density, cepstrum and graphic display of data. With these features the proposed system can perform most of the functions available in complex, stand-alone analysis devices.
Conclusion
A microcontroller data acquisition device based on non-contact transducer has been developed for fault detection and preventive maintenance of rotating machinery. The present work has aimed to develop a general-purpose SHM device which is simple, economical and adaptable to individual problems. The system can perform most of the functions available in complex, stand-alone vibration analysers. The software is menu-driven and user-friendly. The device is capable of computing frequency spectra, cepstrum analysis, signal editing of vibration, acoustic and velocity. The proposed SHM device has been used for fault detection in gearboxes. The following conclusions have been obtained from the experimentation and analysis of results: (i) real-time non-contact condition monitoring of rotating machinery can be carried out by using the developed system; (ii) owing to non-contact data acquisition, the drawback of traditional sensors that have to be mounted on the machine has been diminished; (iii) the results were also compared and calibrated with the standard data acquisition device and acoustic sensor for the validation of results. It has been found that the developed system showed the same pattern of variation of acoustic amplitude as an standard system does; (iv) the FFT spectrum peak values for different gearbox running conditions were observed. It has been noticed that peak amplitude of FFT spectrum in all three cases, i.e. healthy, RC30 and RC50 for both the data acquisition system lie near the GMF; and (v) economic analysis revealed that the developed data acquisition system is economical, userfriendly and equipped with various signal processing tools required for health monitoring of rotating machinery.
Research trends
(i) Future research may focus on integrating artificial intelligence as an integral part of SHM systems with the use of emerging Information Communication Technologies to give real-time decision regarding status of the machine. | 3,369.8 | 2018-06-01T00:00:00.000 | [
"Engineering"
] |
Analysis of Web Spam for Non-English Content: Toward More Effective Language-Based Classifiers
Web spammers aim to obtain higher ranks for their web pages by including spam contents that deceive search engines in order to include their pages in search results even when they are not related to the search terms. Search engines continue to develop new web spam detection mechanisms, but spammers also aim to improve their tools to evade detection. In this study, we first explore the effect of the page language on spam detection features and we demonstrate how the best set of detection features varies according to the page language. We also study the performance of Google Penguin, a newly developed anti-web spamming technique for their search engine. Using spam pages in Arabic as a case study, we show that unlike similar English pages, Google anti-spamming techniques are ineffective against a high proportion of Arabic spam pages. We then explore multiple detection features for spam pages to identify an appropriate set of features that yields a high detection accuracy compared with the integrated Google Penguin technique. In order to build and evaluate our classifier, as well as to help researchers to conduct consistent measurement studies, we collected and manually labeled a corpus of Arabic web pages, including both benign and spam pages. Furthermore, we developed a browser plug-in that utilizes our classifier to warn users about spam pages after clicking on a URL and by filtering out search engine results. Using Google Penguin as a benchmark, we provide an illustrative example to show that language-based web spam classifiers are more effective for capturing spam contents.
Introduction
Web spamming (or spamdexing) is a process for illegitimately increasing the search rank of a web page with the aim of attracting more users to visit the target page by injecting synthetic content into the page [1,2]. Web spamming can degrade the accuracy of search engines greatly if this content is not detected and filtered out from the search results [3][4][5]. In general, spammers aim to illegally enhance the search engine ranks of their spam pages, which might lead to user frustration, information pollution, and distortion of the search results, thereby affecting the entire information search process.
Black hat search engine optimization (SEO) techniques are generally used to create web spam pages. For example, in content-based web spamming, spammers stuff spam keywords into the target page by listing them in the HTML tags (e.g., META tags) or by using an invisible font. In addition, scraper techniques are used where the spam content is simply a replica of another popular site [6][7][8]. These deception techniques are refused by search engines because they can lead to misleading search results [9].
Some web ranking algorithms give higher ranks to pages that can be reached from other web pages that are highly ranked, so the black hat SEO method exploits this feature to increase the ranks of spam pages [5,[10][11][12][13]. For example, in the cookie stuffing method, the user's browser receives a third-party cookie after visiting a spam page with an affiliate site so the cookie stuffer is credited with a commission after visiting the affiliate site and completing a particular qualifying transaction. Moreover, by utilizing a page cloaking mechanism, a search engine crawler can receive different content from the spam page compared with that displayed on the end-user's browser, where the aim is delivering advertisements or malicious content to the user, which is partially or completely irrelevant to that searched for by the user. Another link-based tactic is link farms where a set of pages are linked with each other.
Site mirroring is another black hat SEO method, which exploits the fact that many search engines grant higher ranks to pages that contain search keywords in the URL. Thus, spammers can create multiple sites with various URLs but similar content. Further, web spammers can create pages that redirect the user's browser to a different page that contains the spam content in order to evade detection by search engines [10].
Due to the success of email anti-spam tools based on machine learning, we consider that these techniques might also be effective for detecting web spamming. Typically, high detection accuracy and a low false positive rate are the main properties required for detection tools based on machine learning methods. This is particularly important for detecting spam pages and ensuring that benign web sites are not penalized.
Search engines enhance their anti-spamming techniques continuously. For example, Google developed their latest algorithm (called Penguin) in 2012 and they have continued updating it to lower the search engine ranks of web sites that use black hat SEO or that violate Google Webmaster Guidelines [21,22]. Google's latest web spam report urges publishers to verify the contents of their pages via the Search Console. In fact, Google sent over 4.3 million emails to webmasters during 2015 alone to warn them of identified spam-like content and to give them a chance of reconsideration [23].
The effectiveness of the Google Penguin algorithm is affected by the text language used in the page examined [24]. Several web spam detection features have been proposed but to the best of our knowledge, the effect of the language on these detection features has not been examined previously. In addition, to the best of our knowledge, the performance of the Google Penguin algorithm at detecting web spam pages that contain text in languages other than English has not been evaluated.
This study significantly extends our earlier conference paper [25,26], where the data set is expanded and updated, a new release of Google Penguin is explored, new spamming detection algorithms are introduced, and their results are presented. This study makes the following main contributions.
showing how and why the distribution of selected detection features differ according to a given page language. We used English and Arabic as languages in case studies.
2. COLLECTING AN ARABIC WEB SPAM DATA SET. We collected and manually labeled a corpus containing both benign and spam pages with Arabic content. We used this corpus to evaluate our proposed machine learning-based classifier and we have also made the corpus available for use by the research community in this domain.
3. ANALYSIS OF DETECTION FEATURES AND DEVELOPMENT OF A NOVEL CLASSIFIER. Using Arabic pages in a case study, we showed how to identify a set of web spam detection features with satisfactory detection accuracy. Employing supervised machine learning techniques, we then built a classifier for detecting web pages that contain spam content and showed that it yielded better accuracy compared with the Google Penguin algorithm.
4. CONSTRUCTION OF A BROWSER ANTI-WEB SPAM PLUG-IN. Using our proposed classifier, we developed a browser plug-in to warn the user before accessing web spam pages (i.e., after clicking on a link from the search results). The plug-in is also capable of filtering out spam pages from the search engine results.
The remainder of this paper is organized as follows. Section 2 presents our analysis of how the page language affects the detection rate for web spam using a set of classifiers. Section 3 describes the collection and labeling process for our data set. Section 4 illustrates our system architecture and design. Section 5 explains the feature extraction and selection process. Section 6 presents the proposed classifier and evaluations of its accuracy. Section 7 discusses the meaning and implications of our main findings, and Section 8 presents related research. Finally, we give our conclusions in Section 9.
Data Sets
Two web spam data sets were used in this study. First, we used UK-2011 [27], which is a subset of the WEBSPAM-UK2007 data set [28]. The UK-2011 data set was labeled by volunteers and each page is flagged as either "spam" or "non-spam." Second, we used an extended Arabic web spam data set [29], which included spam and non-spam Arabic pages (this data set was collected and labeled during the period from April 2011 to August 2011).
We used Wahsheh's web spam detection features [30] (see Table 1). We employed the J48 classifier, which is a Weka (version 3.7.6) implementation of the C4.5 decision tree classifier (decision trees are statistical machine learning algorithms that utilize a greedy top-down process to select attributes at selected nodes in the tree and divide the samples into subsets based on the values of these attributes). Cross-validation, a model evaluation method used to improve how a classifier generalizes to an independent data set, was used to ensure that each instance in the data set had an equal probability of appearing in either the training or testing sets. We performed a 10-fold cross-validation and we divided the data set into 10 chunks for training 10 times, where a different chunk was used as the testing set each time. For the decision tree classifier, the issue of overfitting was addressed by using a pruning technique, where the less significant tree nodes for classifying the data set instances were removed from the tree (we set the minimum number of instances to two).
Results and Analysis
We started our analysis by studying the selected detection features in both data sets. Fig 1 shows the probability density function (PDF) for different features in both data sets. A random sample of 1,500 web pages was used to determine the figure visibility (compared with 3,688 pages in data set (1) and 9,988 in data set (2)). According to the cumulative distribution function (CDF) for feature 2 in Fig 1A, almost 60% of the Arabic non-spam pages contained less than 270 words in their pages, whereas less than 15% of Arabic spam pages had less than 270 words. The figure shows that Arabic spam pages tended to have more words in their pages compared with Arabic non-spam pages. In addition, the CDFs for the number of words in Arabic non-spam pages and English pages were very similar. The same observation can be made based on Fig 1B and 1C, but there was more variation among them. In fact, most of the features exhibited greater variation between spam and non-spam pages in the Arabic data set compared with the UK data set. Furthermore, Fig 1B shows that Arabic spam pages tended to have shorter word lengths, where almost 80% of the Arabic spam pages had an average word length of six characters, whereas only 40% of the Arabic non-spam pages had an average word length of six characters. In terms of the number of characters per meta-element, as shown in Fig 1C, Arabic spam pages usually had more characters (80% had more than 400 characters) compared with Arabic non-spam pages (20% had more than 400 characters). Furthermore, Fig 1D shows that Arabic pages usually had more images in their pages compared with English pages, particularly in spam pages.
First, we used all 11 detection features to build the classifiers. Most of the Arabic web spam pages used more obvious spamming tactics compared with those in English, so the DR for English spam pages was lower than that for those in Arabic. We then selected different sets of features using the following feature selection algorithms implemented in Weka: CfsSubsetEval, PrincipalComponents, ConsistencySubsetEval, and FilteredSubsetEval. Brief descriptions of these algorithms and the results obtained from their execution are shown in Table 2. The CfsSubsetEval algorithm considers the individual predictive ability of every feature as well as the features' degree of redundancy in order to evaluate the value of a subset of features. Princi-palComponents performs principal components analysis and transforms the data. Based on the results obtained by these algorithms, we selected the following sets as training scenarios for the classifier: 1,5,7,11,1,5,8,9,1,5,7,10,11, and all 11 features. Tables 3 and 4 show the performance of each set of features using the classifiers described above, the performance measurement indices mentioned in Table 5, and the confusion matrix obtained by the classifier.
Limitations in Existing Data Sets
We found that the distributions of a selected set of features varied according to the underlying language used in the page examined. In addition, for both data sets, the results obtained by the classifiers showed that only a few common features yielded similar results. However, the significance of several of the remaining features varied according to the language used in the page
Measurement Indices Description
Detection rate (DR) Ratio of the number of correctly classified samples relative to the total number of samples.
Error rate (ER) Ratio of the number of incorrectly classified samples relative to the total number of samples. examined. The effect of language was due partly to the use of a similar set of web spamming techniques for a given language. It is important to note that these data sets are fairly old and they do not represent the current techniques of new spammers. In addition, given that the original contents of the web pages of the two data sets were not available, we could not examine other spam detection features (i.e., other than those of the 11 features provided within the two data sets). Furthermore, the method used to collect the web pages in these data sets did not consider specific search engines as the main goal of spammers in order to obtain higher ranks for their web pages in the search engine results and increase the number of hits. To overcome these limitations, we decided that a new data set must be collected carefully and made available.
Building an Arabic web spam corpus
In order to overcome the limitations described in the previous section, we followed a threestep process to collect a data set of Arabic pages, including both benign and spam web pages. First, we collected the top Arabic search keywords for the period from January 2004 to October 2012 on the Google Trends website. We then queried the Google search engine using the collected search keywords. The URLs of the top 50 result pages for each search keyword were then stored, thereby obtaining a total of 8,168 distinct domain names. Fig 2 shows the percentages of the URLs collected for each category in Google Trends. We note that the number of search keywords in a given category affected the corresponding percentage.
We identified multiple types of pages with malware and phishing content, where each URL was examined using six security scanners (these scanners were provided by selected antivirus vendors): 1) Sucuri SiteCheck scanner; 2) McAfee SiteAdvisor scanner; 3) Google Safe Browsing scanner; 4) Norton scanner; and 5) Sophos scanner (with Yandex ranking). The scanners examined every visible web page in the entire domain of a given URL. This scanning process was beneficial for studying the relationships between existing vulnerabilities, malicious content, and web spam [31]. The scanning results were then stored into a database (see Fig 3).
Finally, the URLs were labeled manually by several raters. Each link was classified into one of four categories: i) spam class; ii) borderline class; iii) benign class; and iv) unknown class. The raters were given a set of guidelines for labeling web spam pages (e.g., see [32]). A web application was utilized by the raters to view and rate the data set's links so every link was classified by at least one rater. Fig 4A shows the distribution of classes (i.e., non-spam, borderline, and spam) according to the raters. It should be noted that almost 26% of the Google search results were flagged as either the spam class (10%) or borderline class (16%), although the new update to the Penguin algorithm has been in place for several months.
Many spammers aim to compromise the machines of users and there was a clear correlation between spamming and the existence of web vulnerabilities, as shown in Fig 4B and 4C. We note that 15% of the positive URLs results obtained from the Sucuri scanner (i.e., containing malware and flagged as malicious) were manually labeled as spam, whereas 9% of the negative web pages were labeled as spam. Similarly, the percentage of URLs flagged as borderline represented (1) 13% of the Sucuri scanner-negative URLs and (2) 30% of the Sucuri scanner-positive URLs. However, the percentage of non-spam URLs represented more than 78% of the negative URLs and 55% of the positive URLs. Similar observations can be made for the sites scanned by the McAfee tool, as shown in Fig 4C, which indicates that spamming seems to be a preferred tool for attackers. Fig 5A, 5B and 5C illustrate the distributions of our three classes among Google Trends categories. The distribution is divided into two sets: malicious and benign, as found in the URL classification by the Sucuri scanner. The arts & entertainment, beauty & fitness, and online communities categories were most common for web spammers. Furthermore, we note that the numbers of positive and negative URLs according to the Sucuri scanner were proportional to those in the spam and borderline classes, unlike the non-spam category class.
System Architecture and Design
The system comprises two major components: (i) a back-end server and (ii) a browser plug-in. The plug-in represents the connection between the back-end server and the browser (see Fig 6). After the browser plug-in captures the URL (either clicked on or entered in the web browser address bar by the user), the URL is sent by the plug-in to the back-end server, which then extracts the values of the detection features from the URL and flags it as either benign or spam.
The page will be blocked by the plug-in if it is flagged as a spam page and it will display a pop-up dialog box to warn the user of spam content. The user has the option to proceed and browse the spam page. The plug-in maintains a cache with a blacklist and whitelist, so only Toward More Effective Language-Based Classifiers new URLs are examined by the back-end server. A database containing all the requests received from the plug-ins is also maintained by the back-end server, which serves as a local cache lookup mechanism to speed up the retrieval process.
The plug-in was implemented for the Chrome browser using standard web techniques, such as HTML, CSS, and JavaScript, and JavaScript Object Notation (JSON) is used for lightweight data interchange with the browser and the back-end server. The back-end server uses Apache tomcat as a web server, MySQL as a database server, and JavaServer Pages (JSP) as a server-side programming technology. In the back-end server, jsoup is used as a Java library to deal with HTML and xml document parsing and feature extraction. Most computations are performed on the server side, which maintains a cache containing both the blacklist and whitelist, so the waiting time tends to be very short compared with the loading time for the pages examined. Furthermore, the back-end server can easily be scaled up or down to serve the number of requests. The back-end server can also be used to collect crash reports from the plug-in, which may help to improve new releases.
Feature Selection and Extraction
Feature selection and extraction are crucial steps in the construction of a classifier. Several previous studies have proposed the detection of features that minimize the intra-class variability and maximize the inter-class variability (e.g., [33][34][35][36][37][38]). In general, the use of raw data for classification leads to classifiers with complex structures, thereby resulting in poor performance.
In addition to some known features from previous studies, we propose novel detection features that have not been used before to the best of our knowledge, as shown in Table 6. We calculated the CDF for the second feature in Fig 7A, the fifth feature in Fig 7B, the sixth feature in Fig 7C, and feature 7 in Fig 7D, thereby helping us to understand the nature of each feature, and thus the contribution of the features to the classifier's accuracy.
As shown in Fig 7A, 70% of the web spam and borderline pages had 18 links, whereas the benign pages had 10 links. Fig 7B shows that 90% of the benign web pages had 8 meta tags compared with 37 meta tags in the borderline and spam pages.
Similarly, Fig 7C and 7D show clearly that for features 6 and 7, the benign web pages were sufficiently easy to distinguish from both borderline and spam web pages. For instance, 90% of the benign pages had 12 unique words from Google Trends compared with 25-30 words in both the borderline and spam pages. Furthermore, 90% of the benign web pages had 70 Toward More Effective Language-Based Classifiers repeated words from Google Trends compared with 170-230 words in both the borderline and spam web pages. Features 6 and 7 were actually critical for distinguishing between spam and borderline URLs. In almost 50% of cases, the borderline and spam web pages differed from each other by 50-60 words (see Fig 7D). We also calculated the PDF for the same features, as shown in Fig 8. Fig 9A shows that 6% of the spam pages had one hidden iframe, whereas this was the case for only 2% of the borderline and benign pages. It should be noted that although some detection features might not prove useful in isolation, employing multiple features for detection could result in better detection performance when distinguishing between benign and spam pages because these features may complement each other (see Fig 9B and 9C). Fig 10A, we note that there is one obvious peak where the PDF for the non-spam pages was much greater than that for the spam pages (the x-axis represents feature F2 and the y-axis represents feature F5, as in Table 6; the non-spam class is shown in red and the spam class in green). Fig 10B shows the delta values (i.e., |P n − P s | (F2, F5)). Similarly, in Fig 10C, when the values of features F2 and F6 were relatively small, there was a clear peak where the PDF for the non-spam pages was greater than that for the spam pages. Fig 10D shows the delta values (i.e., |P n − P s | (F2, F6)). Similar observations can be made based on Fig 10E and 10F.
Classification and Evaluation
We tested four machine learning algorithms by using multiple variations to build our classifier, as follows. First, we tested decision trees (C4.5, logistic model tree, random forest, and Logit-Boost). Second, we tested Bayes Network, which is a probabilistic graphical model that represents the relationships and conditional dependencies between a set of random variables using a graphical model. Third, we tested a support vector machine (SVM), a statistical-based algorithm that separates classification classes using a set of hyperplanes. Fourth, we tested a multilayer neural network, which comprises a set of interconnected processing units (the weights of these interconnections are calibrated during the training phase to obtain the required knowledge).
Understanding the similarity between spam and borderline web pages is important for the prior training of classification models (see Section 5). To build our classifiers, we considered the following scenarios: (i) two-class classification with only two classes: class 1 for spam and borderline web pages, and class 2 for benign pages; and (ii) three-class classification where we had three classes: spam pages, borderline pages, and benign pages.
The classifiers were configured using Weka (version 3.7.6) for both scenarios [39]. The parameters settings for the three algorithms are shown in Table 7. We performed 10-fold cross-validations for each of the classifiers by using a subset of the observations to establish the classifier and to identify whether the classifier correctly flagged the eliminated observations. To address the overfitting problem for the decision tree classifier, we utilized a pruning technique to reduce the size of the tree by eliminating tree nodes with low significance for classifying instances. Pruning techniques are used for reducing the complexity of classifiers, which in turn helps to reduce the time required to execute the classifier in the browser plug-in. For the other classifiers, a validation threshold was used to stop the training process when the algorithm detected overfitting and misclassification increased in the validation set. In order to deal with an imbalanced data set, we used the Synthetic Minority Oversampling Technique (SMOTE), which is an oversampling technique for the minority in an imbalanced data set based on the use of "synthetic" examples. The letter "S" is used at the end of the abbreviations in the tables to indicate whether SMOTE was applied to the data set or not.
The results obtained after training the classifier in the three-class scenario are shown in Table 8, which demonstrate that decision trees performed the best, followed by the Bayesian network, multilayer neural network, and SVM classifiers. In particular, the random forest (RFT-S) decision tree scores were better than those produce by all of the other algorithms, with the highest precision (value of 84%), F-measure (value of 84%), and ROC (value of 95%) However, we note that the detection accuracy was relatively low due partly to two main causes: (1) the URLs in the spam and borderline classes (27% of the data set) were actually similar; and (2) the fact that spammers use clever tactics to evade detection by Google Penguin. For mitigation purposes, we only established the classification models for the two-class scenario. Table 9 shows that the performance of decision tree was better than that of the other classifiers (particularly the RFT-S algorithm where DR = 87% and ROC = 93%). Similarly, the BayesNet-S classifier was ranked second, where DR = 86% and ROC = 93%, followed by the multilayer neural network and SVM classifiers. Tables 10 and 11 show the confusion matrices (i.e., error matrix) for the three-class and two-class classifiers, respectively. In each confusion matrix, the first row represents the actual class and the second row represents the predicted class or that classified by a given classifier. Thus, for the RFT-S algorithm, the number of correctly detected spam instances (i.e., TPs) was 87, the number of spam instances mistakenly flagged as borderline was seven, and the number Table 7. Parameters used in the decision tree, Bayesian network, support vector machine (SVM), and multilayer neural network methods (see Part II of the WEKA Manual for descriptions of the various algorithms used in our study [40] Toward More Effective Language-Based Classifiers of spam instances mistakenly flagged as non-spam was five. Similarly, the number of correctly detected non-spam instances (i.e., true negatives) was 85, the number of non-spam instances mistakenly flagged as borderline was nine, and the number of non-spam instances mistakenly flagged as spam was five.
Further Discussion
In this study, we used two public data sets (see Section 2) to show that spammers who target different languages behave differently and develop their own new tactics to influence the results obtained by search engine ranking algorithms. In fact, this issue has been recognized by search engine companies and they are considering the development of ranking algorithms that are global and language-independent as far as possible in their new releases. In most web spam data sets, however, search engine ranking algorithms were not considered when the data sets were constructed. In this study, we constructed a new data set to address this issue (see Section 3). Our data set was carefully selected to contain highly ranked Web pages according to the Google Penguin ranking algorithm. However, this data set led to a concern about the effectiveness of the Google anti-spamming algorithm against spam pages containing Arabic content as well as other non-English languages. In particular, when the data set was examined using six security scanners, the results showed that a significant number of In a further study (see Sections 5 and 6), we explored the effectiveness of multiple detection features using our data set and we evaluated different classifiers. Despite that some of our classifiers obtained a detection rate of 87%, which might be lower than previous reported detection rates in other studies, we demonstrated that spammers employ clever techniques to avoid being detected by Google Penguin. We also confirmed the need to build more representative and realistic data sets that are suitable to the context of the outputs obtained by search engines.
Related Work
Numerous previous studies have investigated the prevalence of web spam and various detection techniques have been proposed using different approaches. Gyongyi and Garcia-Molina proposed a web spam taxonomy after the web spam problem emerged in the early 2000s [2]. Heymann et al. were the first to survey the detection, demotion, and prevention of web spam [41]. Recent surveys of existing spam detection techniques and mechanisms have analyzed their advantages and disadvantages (e.g., [42] and [43]). It should be noted that spam and automated accounts in social networks have also contributed to the prevalence of web spam (e.g., see [44][45][46][47][48]). The detection features used for web spam in previous studies belong to two categories: (1) those that exploit topology and network-related data; and (2) those that exploit the web page content.
Gyongyi et al. [1] proposed an algorithm for identifying pages that are likely to be spam and those that are likely to be reputable (also see [49] and [50] for improved versions of the algorithm). Fetterly et al. [51] utilized statistical analysis to show that there are outliers in the statistical distribution of the linkage structure, page content, and page evolution properties in spam pages compared with benign web pages. Wu et al. [52] proposed some alternative methods for propagating trust on the web and utilized distrust to demote web spam. In addition, Castillo et al. [53] built a machine learning classifier that utilizes both link-based and content-based detection features, which obtained TP = 88.4% and FP = 6.3%. Svore et al. [33] built a classifier to identify web spam pages by training a SVM classifier based on a selected set of page attributes.
Ntoulas et al. [15] proposed a C4.5 decision tree classifier, which could detect 86.2% of the spam pages examined. Becchetti et al. [37] explored the best combinations of spam detection features and selected classifiers that achieved high precision (DR = 80.4%) using a small set of features. Furthermore, Abernethy et al. [54] proposed a machine learning classifier that employs a variety of SVM for detecting web spam using both the page content and hyperlinks. Similarly, Becchetti et al. [55] proposed a link-based technique for detecting web spam pages by using a damping function for rank propagation and an approximate counting technique. By exploiting textual and extra-textual features in HTML source code, Urvoy et al. [56] investigated multiple HTML style similarity measures and proposed a flexible clustering algorithm for identifying web spam pages. In addition, Gan and Suel [57] proposed a classifier that uses the decision tree C4.5 algorithm and many detection features, including content-based and link-based, which obtained precision of around 88%. Webb et al. [58] identified a relationship between email and web spam, which they utilized to identify web spam. They also employed their method to collect a web spam corpus. Lee et al. [59] proposed a simplified swarm optimization method to solve the complexity problem that affects statistical classification and machine learning approaches, which increases when there are a large number of web spam detection features.
Previous studies also considered linguistic-based detection features and evaluated their effectiveness at web spam classification (e.g., [36,60]). However, to the best of our knowledge, no previous studies have investigated the advantages of using linguistic-based features to improve web spam detection in a particular language.
Conclusion and Future Work
Google continues to improve their Penguin algorithm, but web spammers are also developing creative evasion mechanisms to increase their web page ranks with the aim of attracting more users. In fact, we consider that web spam will remain a good method for both phishing attacks and malware spreading. In this study, we showed that Google anti-spamming methods are actually ineffective against web spam pages that contain non-English content, which raises a concern that the insufficient testing of pages with non-English content could potentially encourage spammers to target these pages.
As an illustrative example, we developed and tested a classifier in the form of a browser anti-spam plug-in for detecting Arabic spam pages, and we showed that our classifier captured most of the web spam pages not detected by the Penguin algorithm. We also created a labeled Arabic web spam data set to evaluate our classifier and to encourage other researchers to build upon our work.
In future work, we plan to extend our web spam data set, create similar data sets for other languages, and develop custom classifiers for these languages. Spammers and Google search engine developers are continually improving their techniques to defeat each other, so future experimental studies are important for understanding new trends and directions. In recent years, large-scale spamming campaigns using compromised Web sites have been performed to corrupt search engine results. These spamming campaigns are an emerging trend that needs to be investigated. Using Google Penguin as a benchmark, our illustrative example shows that language-based web spam classifiers are more effective at capturing spam content. We consider that the web spam problem requires a continuous effort from search engines as well as developers and webmasters based on appropriate vetting of their sites, and end-users should also report spam content. | 7,691.6 | 2016-11-17T00:00:00.000 | [
"Computer Science"
] |
Reversibility of Cardiac Involvement in Acromegaly Patients After Surgery: 12-Month Follow-up Using Cardiovascular Magnetic Resonance
Purpose Cardiac comorbidity is one of the leading causes of death among acromegaly patients. We aimed to investigate the reversibility of acromegalic cardiac involvement after surgical treatment using the gold standard method, cardiovascular magnetic resonance, and to explore the effects of endocrine remission and gender on reversibility. Methods In this single-center, prospective cohort study, fifty untreated acromegaly patients were enrolled. Comprehensive cardiac assessments were performed using a 3.0 T magnetic resonance scanner before and 3 and 12 months after transsphenoidal adenomectomy. Results Preoperatively, left ventricular (LV) enlargement (13.0%), LV systolic dysfunction (6.5%), right ventricular (RV) enlargement (4.3%), RV systolic dysfunction (2.2%) and myocardial fibrosis (12.0%) were identified. On average, the LV and RV ejection fractions of acromegaly patients were higher than the healthy reference values. Male patients had thicker LV myocardia, wider ventricular diameters and more dilated pulmonary artery roots than female patients. After surgery, LV myocardial hypertrophy was reversed, the left atrium was remodeled, and ventricular systolic dysfunction recovered to normal. Cardiac alterations were detected early in the 3rd postoperative month and persisted until the 12th month. The interventricular septum was initially thickened in the 3rd postoperative month and then recovered at the 12th month. Notable postoperative cardiac reversibility was observed in male patients but did not occur in all female patients. Patients achieving endocrine remission with normalized hormone levels had thinner LV myocardia than patients without normalized hormone levels. Conclusion Our findings demonstrated that some of the cardiac involvement in acromegaly patients is reversible after surgical treatment which lowers hormone levels. Endocrine remission and gender significantly impacted postoperative cardiac reversibility.
INTRODUCTION
Elevated levels of serum growth hormone (GH), combined with its product hormone insulin-like growth factor 1 (IGF-1), contribute to the systemic complications responsible for the increased mortality of acromegaly patients (1,2). Cardiovascular comorbidity is the most common comorbidity and represents one of the most important causes of death in acromegaly patients (3)(4)(5)(6). Myocardial hypertrophy, cardiac chamber enlargement, and diastolic and systolic dysfunction have been indicated to be common presentations of cardiac involvement in acromegaly patients (1,5). Transsphenoidal adenomectomy and medical treatment with somatostatin analogs are the first-line treatments for acromegaly patients (2,7,8). Based on echocardiography, cardiac abnormalities are reversible after treatment, especially in young patients with a short disease duration upon GH and IGF-1 remission (9)(10)(11)(12), while these abnormalities rarely improve in patients with uncontrolled GH and IGF-1 levels (13).
An accurate and comprehensive evaluation of cardiac performance is essential for the diagnosis and treatment of cardiac comorbidity (14). Although echocardiography is widely used and well studied in cardiac evaluations of acromegaly patients, it relies heavily on the experience of the operator and suitable echocardiographic windows (15). Evaluations of the heart with the gold standard cardiovascular magnetic resonance (CMR) offer objective and reproducible results, and CMR can precisely assess cardiac structure and function and even evaluate parameters that are difficult to precisely assess by echocardiography, such as myocardial fibrosis and ventricular volumes (16)(17)(18)(19). Several studies have used CMR to evaluate cardiac involvement in acromegaly patients (20)(21)(22)(23)(24)(25), and the results of these studies showed that the incidence rates of myocardial fibrosis, left ventricular (LV) systolic dysfunction (LVSD), and LV hypertrophy in acromegaly patients were 0% to 14.8%, 0% to 12.5%, and 5% to 72%, respectively. Some of these studies focused on the posttreatment reversibility of cardiac involvement, but the results are controversial. Bogazzi et al. (22) demonstrated that acromegaly patients with controlled disease after using somatostatin analogs had a greater reduction in the LV mass index than those without controlled disease. dos Santos Silva et al. (23) found no clinically relevant cardiac improvements after medical therapy with octreotide. Andreassen et al. (21), however, reported a decrease in cardiac function 3 months after treatment. Compared to medical treatment, transsphenoidal adenomectomy allows GH to decline drastically and immediately after tumor removal. The reversibility of cardiac involvement after surgery has not been systematically evaluated in acromegaly patients by CMR, and we propose that their postoperative cardiac reversibility may differ from that in patients who received medical treatment in the literature.
On the basis of our previous study on baseline cardiac involvement in a series of untreated Chinese acromegaly patients using CMR (25), the current study focused on the reversibility of cardiac involvement after transsphenoidal adenomectomy and the impacts of gender and endocrine remission (ER). Here, we applied CMR to 50 untreated acromegaly patients and analyzed the comprehensive qualitative and quantitative CMR data before and after surgery to test our hypothesis that cardiac involvement in acromegaly patients is reversible following successful surgical treatment and that gender and ER influence postoperative cardiac reversibility.
Patient Population
Untreated acromegaly patients were enrolled in this singlecenter, prospective, cohort study. This study was performed in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board at PUMCH (No. ZS-1293). Written informed consent was obtained from all patients.
The inclusion criteria were as follows: 1) an IGF-1 level exceeding the age-and gender-matched reference range and a GH level exceeding 1.0 ng/ml following an oral glucose tolerance test (OGTT) (3); 2) a pituitary tumor identified on sellar magnetic resonance imaging (MRI) (3,26); 3) adult age, with no restrictions on gender; 4) typical clinical manifestations of acromegaly; and 5) normal levels of other pituitary-related hormones, including in the patients on hormone replacement therapy, to eliminate the potential effects of hypopituitarism on the heart. The exclusion criteria were as follows: 1) hepatic disease, renal disease, implanted metal or claustrophobia; 2) recurrent pituitary adenomas; and 3) known primary cardiovascular diseases before the onset of acromegaly. Inappropriate use of the reference ranges might lead to improper data interpretation (23). Therefore, we referred to the article by Le et al. (18) as the control reference. The authors enrolled 180 healthy Chinese volunteers and grouped them according to gender and age in 10-year intervals. The mean values and upper and lower limits of the CMR parameters in each subgroup were given in that study.
Study Design and Follow-up
Acromegaly patients were clinically diagnosed, and CMR was performed before surgery. Clinical data (gender, age, body mass index (BMI) and disease duration) and hormone levels (random GH, GH nadir and fasting IGF-1) were recorded. The disease duration was defined as the duration from the onset of acromegaly symptoms until the clinical diagnosis. Microscopic transsphenoidal adenomectomy was performed for all patients, and the pathology of GH-secreting pituitary adenomas was confirmed.
The patients were re-evaluated 3 months and 12 months after transsphenoidal surgery. CMR scans were performed, and hormone levels were measured during the follow-up period. ER was defined as a random GH<1.0 ng/ml or a GH nadir<0.4 and age-gender normalized IGF-1 (27). The IGF-1 reference range used in this study referred to our published article (28). Repeated surgery, octreotide LAR treatment and observation were recommended for recurrent patients according to the radiological findings and hormone levels.
CMR Image Acquisition
CMR was performed on a 3.0T superconducting MR scanner (Siemens Healthineers, Germany). Cine images were acquired with an electrocardiogram-gated two-dimensional balanced steady-state free precession sequence during multiple breath holds. Two-, 3-and 4-chamber long-axis and short-axis slices were acquired. The key parameters were as follows: repetition time/echo time, 3.3/1.43 msec; flip angle, 55°-70°; voxel size, 1.6 × 1.6 × 6.0 mm; temporal resolution, 45.6 msec; and bandwidth, 962 Hz/pixel. A bolus of gadolinium (0.5 mmol/ml, Beijing BEILU Pharmaceutical Co., Ltd.) was injected at a dose of 0.05 mmol/kg and a flow rate of 4 ml/s for first-pass perfusion imaging. Another bolus of gadolinium at a dose of 0.1 mmol/kg and a flow rate of 1 ml/s was then given, followed by late gadolinium enhancement (LGE) imaging 10 to 15 min later.
LGE images were acquired with a 2D phase-sensitive inversionrecovery gradient-echo pulse sequence. Focal myocardial fibrosis was identified by the presence of LGE on a special focus shown on both short-and long-axis stacks.
CMR Parameter Extraction and Cardiac Abnormality Identification
CMR images were transmitted into Circle Cardiovascular Imaging software (version 5.3, Canada). Figure 1 shows the protocol for CMR analysis. For ventricular wall thickness, we evaluated LV anterior wall (LVAW) thickness, LV lateral wall thickness, LV posterior wall thickness, interventricular septum (IVS) thickness and right ventricular (RV) lateral wall thickness. For dimensions of the heart and arteries, we evaluated the LV longitudinal diameter (LVLD), LV transverse diameter (LVTD), RVLD, RVTD, left atrial LD (LALD), LATD, right atrial LD (RALD), and RATD. The pulmonary artery root diameter (PARD) and LV outflow tract diameter (LVOTD) were measured on the 3-chamber long-axis slice. For ventricular volume, we evaluated the indexed LV end-diastolic volume (LVEDV), indexed LV end-systolic volume (LVESV), and indexed RVEDV and RVESV. The LV ejection fraction (LVEF) and RVEF were calculated as the reduction in ventricular volume from the EDV to the ESV divided by the EDV.
The reference ranges for ventricular volume and systolic function in Le's study were used to identify qualitative cardiac abnormalities by transforming the quantitative values in this study into z scores (18). Therefore, the LVEDV, RVEDV, LVEF and RVEF values were transformed accordingly. A z score >2 standard deviations (SDs) or <−2 SDs was recognized as abnormal.
Hormone Assays and Sellar MRI
Blood samples were collected in the morning after an 8-h fasting period for hormone analysis. Chemiluminescence assays (Siemens Healthcare Diagnostics Products Ltd., UK) were used to measure fasting GH and IGF-1 levels. GH levels were also measured at 30, 60, 120, and 180 min after an OGTT. Contrast-enhanced sellar MRI (Discovery MR750, GE, USA) was performed. The typical radiological imaging features of a pituitary adenoma include a solid hypo/isointense mass on T1-weighted imaging, a hyper/isointense (13) and RA transverse diameter (RATD) (14) were measured on the 4chamber long-axis slice in the LV end-systolic phase (D). LV and RV volumes were contoured slice by slice on the short-axis stacks in diastole (E) and systole (F). Papillary muscles were excluded from the LV volume evaluation but included in the RV volume evaluation. Late gadolinium enhancement (LGE) was identified on both short-axis and long-axis stacks (G, H). mass on T2-weighted imaging and reduced enhancement after gadolinium administration.
Statistical Analysis
SPSS (IBM SPSS Statistics, version 23.0, USA) was used to analyze the data. GraphPad Prism (GraphPad Software, version 8.1, USA) was used to generate bar charts. Categorical variables are presented as numbers and percentages. Continuous values are presented as the means ± SDs. Student's t-test or the Mann-Whitney U test was used to compare continuous data according to the data distribution. The c 2 test was used to analyze correlations among categorical variables. Statistical significance was defined as p<0.05.
Characterization of the Study Population
The baseline clinical characteristics and hormone levels of the 50 untreated acromegaly patients, including 27 males and 23 females, are listed in Table 1. No differences in age or gender were found between the acromegaly patients and healthy controls. Seven patients had hypertension at diagnosis, and their arterial blood pressure was controlled normal using antihypertensive drugs. Diabetes mellitus (DM) was diagnosed in 6 patients. Insulin was subcutaneously injected during the perioperative period, and the fasting and postprandial blood glucose levels were controlled within the normal range.
Forty-five patients (23 males and 22 females) completed the 3-month follow-up, 24 of whom (53.3%) reached ER. Among the 21 patients without ER, 12 continued observation, 5 underwent repeated surgery, and 4 received octreotide LAR treatment. Thirty-five patients (18 males and 17 females) completed the 12-month follow-up, 25 (71.4%) of whom reached ER. We compared the postoperative clinical characteristics, including age, gender, hypertension, DM, BMI and hormone levels, of the patients with the preoperative characteristics. Apart from significantly reduced postoperative GH and IGF-1 levels, no other clinical differences were found.
Quantitative CMR parameters and LGE were evaluated in 50, 45 and 35 patients before surgery, 3 months after surgery and 12 months after surgery, respectively. However, the z scores of 4 patients before surgery, 4 patients at 3 months after surgery, and 2 patients at 12 months after surgery were not calculated because of age restrictions in Le's study (20-69 years). Therefore, cardiac abnormalities, including ventricular systolic dysfunction (<−2 SDs) and ventricular enlargement (>2 SDs), were evaluated in 46, 41 and 33 patients before surgery, 3 months after surgery and 12 months after surgery, respectively.
The quantitative CMR parameters of the acromegaly patients are listed in Table 1. Acromegaly patients had a larger LV mass and indexed LV mass than the healthy controls. The average LVEF and RVEF of the acromegaly patients were both significantly higher. The LVEDV was markedly elevated, and the RVESV was notably reduced. In the acromegaly patients, the thicknesses of the LVAW, LV lateral wall and LV posterior wall were all thinner than the IVS and thicker than the RV lateral wall, the LALD was shorter than the LATD, the RALD was longer than the RATD, and the LVOT was significantly wider than the PAR. Patients with a history of hypertension had significantly shorter LVLD and RVLD, longer LALD and RALD and larger LVOTD (all p<0.05). Patients with a history of DM had significantly thicker LVAW and LV lateral walls, lower LVEF and RVEF and larger PARD (all p<0.05). Age, BMI, and disease duration, but not smoking, GH and IGF-1, also affected CMR parameters (Supplementary Table 1). Six acromegaly patients (12.0%) had LGE on CMR. Areas of LGE in all these patients were focal and located in the midmyocardium, with 3 in the IVS, 1 in both the IVS and LVAW, 1 in both the IVS and LV lateral wall, and 1 in the LV posterior wall. Patients with LGE had a thicker IVS than patients without LGE, while the thicknesses of the other ventricular walls were similar between the two groups.
Reversibility in the Cardiac Structure and Function After Surgery
We gathered the preoperative and postoperative CMR data of acromegaly patients who completed all 3 CMR scans (Figure 2). The results showed that at both 3 months and 12 months after transsphenoidal surgery, the LVAW was thinned, and the LVOT was significantly narrowed. The structure of the left atrium changed dramatically in the acromegaly patients, with marked LALD shortening and LATD extension, which is defined here as "LA remodeling," while the structures of the other heart chambers remained unchanged. The IVS was initially thickened at 3 months and then recovered to the preoperative level at 12 months after surgery.
Effects of Gender on Postoperative Cardiac Reversibility
In male acromegaly patients, at 3 months after surgery, the LVAW was significantly thinner, the IVS was significantly Table 2). In female patients, however, the LVAW thickness change was statistically nonsignificant at this time point. At the 12 th postoperative month, in male acromegaly patients, in addition to the cardiac alterations detected at the 3 rd postoperative month, the RVLD was newly found to be elongated, and the PAR was narrowed; the IVS thickness decreased and recovered to the preoperative level. However, only LA remodeling and PAR narrowing were detected in female patients at the 12 th postoperative month.
Effects of ER on Postoperative Cardiac Reversibility
Although postoperative hormone levels were not normalized in patients without ER, the reductions in these hormone indices were all significant ( Table 3). LA remodeling and LVOT narrowing were found in both groups of patients with ER and without ER at both the 3 rd postoperative month and the 12 th postoperative month. The LVAW was notably thinned in patients with ER but not in patients without ER. The absolute reduction in the LVOTD in patients with ER was significantly greater than that in patients without ER at the 3 r d postoperative month.
DISCUSSION
In this study, we systematically assessed the dimensions of the atria, ventricles and artery roots, ventricular myocardium thicknesses, ventricular volumes, cardiac systolic function, and presentations of cardiac fibrosis in acromegaly patients using CMR and longitudinally analyzed the reversibility of cardiac involvement at 3 and 12 months after transsphenoidal adenomectomy. We found that some of the cardiac involvement in acromegaly patients was reversible after surgical treatment and that gender and ER had significant impacts on postoperative cardiac reversibility. Ventricular systolic dysfunction, which is defined as a reduced ventricular ejection fraction, has been studied in acromegaly patients using CMR (20)(21)(22)(23)(24). In this study, the rates of LVSD and RVSD were 6.5% and 2.2%, respectively, which are lower than previous findings. Apart from LVSD and RVSD, our results showed elevated average ejection fractions for both ventricles in acromegaly patients compared to those in healthy controls. Since the disease duration was shorter on average in this cohort than in the abovementioned CMR studies, this parameter might provide a clue regarding the entire alteration process of acromegalic cardiac function from elevated systolic function at the beginning of the disease course to reduced systolic function and even heart failure at the end of the disease course. Notably, the etiologies of elevated LVEFs and RVEFs differed. The LVEF was elevated because of an increased LVEDV and unchanged LVESV, whereas the RVEF was elevated due to a decreased RVESV and an unchanged RVEDV.
Myocardial fibrosis, which appears as LGE on CMR, occurs in ischemic coronary disease or nonischemic hypertrophic cardiomyopathy (29). The most common location of LGE in this cohort was the IVS, and the thickness of the IVS in patients with LGE was significantly greater than that in patients without LGE. LGE in acromegaly patients in this study was present only in the midmyocardium, showing different radiological features from those in patients with coronary heart diseases (30). LGE was found in 12% of the patients in this study, which is similar to other CMR studies (23,24). However, in a study using biopsy (31), myocardial fibrosis was found in 53.7% of the cohort. Possible reasons for this discrepancy include a long disease duration of 10 years, the inclusion of patients with coronary heart disease, and limited diagnosis and treatment of acromegaly during the study period (31).
Our results showed that male acromegaly patients had thicker ventricular walls, larger ventricular chambers and wider PARs than female patients before surgical treatment. Lei et al. (32) demonstrated that healthy males had a longer LVLD than healthy females (53 mm vs. 50 mm). We found that the LVLD in male acromegaly patients was also longer than that in female patients (97 mm vs. 88 mm). Thus, the increase in LVLD in male patients was 83%, which is larger than the 76% increase among female patients. The study by Le et al. (18) showed that the average LVEF and RVEF of healthy men were lower than those of healthy women by 4% and 7%, respectively. However, in male and female acromegaly patients, these gaps were reduced to 2% and 3%, respectively. Therefore, although hormone levels and disease durations were similar between males and females in this study, cardiac chamber enlargement and ventricular systolic function increases were more obvious in male patients than in female patients.
Cardiac abnormalities in acromegaly patients were reported to improve after treatment according to echocardiography (1,4,5) in both humans and cats (33,34). However, no studies have systematically focused on the reversibility of acromegalic cardiac involvement after adenomectomy using the gold standard of CMR. In this cohort, the ejection fraction of all acromegaly patients with ventricular systolic dysfunction recovered to normal during the follow-up, showing that the functional cardiac abnormalities of acromegaly patients could be reversed by postoperative hormone reduction. Additionally, the average ventricular thickness changed markedly after surgery. These results using CMR confirmed the reversibility of some of the structural and functional cardiac involvement in acromegaly patients after adenomectomy.
LV remodeling and LA enlargement are typical presentations of hypertensive heart diseases (35,36). Similarly, LV hypertrophy in acromegaly patients, which is caused by both excessive hormones and secondary hypertension, was identified and reported to be reversible after treatment (1,5,37). Our results revealed LA remodeling after surgery in acromegaly patients regardless of gender and ER for the first time. Preoperatively, a short LALD and long LATD with an average ratio of 3.2:4.8 were noted in the acromegaly patients, as evidenced by a flat atrium in the anteroposterior view that p1 indicates the differences in quantitative CMR parameter changes between the patients with and without ER at 3 months after surgery, and p2 indicates the differences at 12 months after surgery. *(p < 0.05) and **(p < 0.01) indicate statistically significant differences between preoperative parameters and postoperative parameters.
differed from the right atrium. However, after surgery, an increased LALD and a reduced LATD resulted in a reversed LALD/LATD ratio of 4.8:4.3 at 3 months and 5.0:4.4 at 12 months, manifesting as a tall atrium in the anteroposterior view. We hypothesized that this change might be a specialized response of an acromegalic left atrium to a sudden and drastic postoperative hormone reduction. The distribution of hormone receptors in the heart and their roles in postoperative LA remodeling require further verification.
Little is known about the effects of tumor resection on cardiac involvement in male acromegaly patients compared to female patients. In this study, although the female patients experienced several postoperative cardiac alterations similar to those in the male patients, including LA remodeling, transient IVS thickening and LVOT narrowing, the reversibility of LVAW hypertrophy after surgery was a unique finding in male patients. Before surgery, the LVAW of the male patients was thicker than that of the female patients, while the reduction in LVAW thickness after adenomectomy in male patients was much greater. We hypothesize that estrogen and estrogen receptors may impose potential effects on the differences in cardiac reversibility between female and male patients. Further studies are needed to explore the underlying mechanisms. Additionally, we found that the current criteria for ER (27) served as a good predictor of CMR-based postoperative cardiac reversibility. LVAW reduction was marked in patients with postoperative ER but not in patients without ER. This correlation between hormone normalization and myocardial thinning verified conclusions from previous studies determined with echocardiography (13,38) and reiterated the importance of hormone normalization to postoperative cardiac improvements.
Our study has some limitations. First, given the low incidence of acromegaly of 1.1/100,000 (2) and the strict inclusion criteria, we enrolled only 50 patients in this study, which might contribute to false-negative results. However, this study is currently among the largest of its kind, and the highly objective and reproducible assessments of CMR were able to yield reliable results with limited samples (19). Second, three articles were candidates for healthy Chinese CMR reference values for this study (18,38,39). Dong and Lei's studies had smaller sample sizes than Le's study, and their reference ranges were not given according to age, precluding our acquisition of qualitative data by converting absolute CMR values into z scores. Thus, we finally chose the reference ranges from Le's study. Third, 5 patients and 15 patients were lost to follow-up at the 3 rd and 12 th postoperative months, respectively. We analyzed the possible reasons. Our institute is one of the largest centers for pituitary surgery in China, and patients always come from different regions of China. Some who live far away or have difficulty affording travel expenses were more likely to complete the postoperative follow-up at local medical centers instead of returning to our institute. However, we compared the general data of the acromegaly patients before and after surgery, and the results showed consistency. Thus, the loss of patients in this study had little impact on the acquisition and interpretation of our results. Fourth, detailed metabolic parameters were not acquired, cardiac comorbidities, e.g., atherosclerosis, coronaropathy and valvular insufficiency, were not estimated, and postoperative gonadal hormones were not recorded in this study, which might have potential influence on CMR findings.
In conclusion, we applied the gold standard technique, CMR, to acromegaly patients and systematically assessed the postoperative reversibility of acromegalic cardiac involvement. Using CMR, we obtained and analyzed comprehensive cardiac parameters that were difficult to precisely and objectively evaluate by echocardiography and compared them with the corresponding parameters in healthy subjects. We detected postoperative LA remodeling and improvement of cardiac systolic function and myocardium thickness and found that gender and ER significantly impacted postoperative cardiac reversibility. The application of CMR furthered our understanding of the nature and postoperative reversibility of cardiac comorbidity in acromegaly patients.
DATA AVAILABILITY STATEMENT
All datasets presented in this study are included in the article/ Supplementary Material.
ETHICS STATEMENT
The studies involving human participants were reviewed and approved by the Institutional Review Board at Peking Union Medical College Hospital. The patients/participants provided their written informed consent to participate in this study.
AUTHOR CONTRIBUTIONS
XG, YW, and BX designed the study. XG, ZW, LG, and XB enrolled the patients and completed the arrangements for CMR for all patients before, 3 months after, and 12 months after surgery. YC, JC, XL, and PL acquired the scans and analyzed and interpreted the data. XG and YC analyzed the data and created the figures. XG drafted the manuscript. YW and BX critically revised the manuscript. All authors contributed to the article and approved the submitted version. | 5,805.8 | 2020-10-21T00:00:00.000 | [
"Medicine",
"Biology"
] |
A Parsimonious Approach to Estimate Soil Organic Carbon Applying Unmanned Aerial System (UAS) Multispectral Imagery and the Topographic Position Index in a Heterogeneous Soil Landscape
Remote sensing plays an increasingly key role in the determination of soil organic carbon (SOC) stored in agriculturally managed topsoils at the regional and field scales. Contemporary Unmanned Aerial Systems (UAS) carrying low-cost and lightweight multispectral sensors provide high spatial resolution imagery (<10 cm). These capabilities allow integrate of UAS-derived soil data and maps into digitalized workflows for sustainable agriculture. However, the common situation of scarce soil data at field scale might be an obstacle for accurate digital soil mapping. In our case study we tested a fixed-wing UAS equipped with visible and near infrared (VIS-NIR) sensors to estimate topsoil SOC distribution at two fields under the constraint of limited sampling points, which were selected by pedological knowledge. They represent all releva nt soil types along an erosiondeposition gradient; hence, the full feature space in terms of topsoils’ SOC status. We included the Topographic Position Index (TPI) as a co-variate for SOC prediction. Our study was performed in a soil landscape of hummocky ground moraines, which represent a significant of global arable land. Herein, small scale soil variability is mainly driven by tillage erosion which, in turn, is strongly dependent on topography. Relationships between SOC, TPI and spectral information were tested by Multiple Linear Regression (MLR) using: (i) single field data (local approach) and (ii) data from both fields (pooled approach). The highest prediction performance determined by a leaveone-out-cross-validation (LOOCV) was obtained for the models using the reflectance at 570 nm in conjunction with the TPI as explanatory variables for the local approach (coefficient of determination (R2) = 0.91; root mean square error (RMSE) = 0.11% and R2 = 0.48; RMSE = 0.33, respectively). The local MLR models developed with both reflectance and TPI using values from all points showed high correlations and low prediction errors for SOC content (R2 = 0.88, RMSE = 0.07%; R2 = 0.79, RMSE = 0.06%, respectively). The comparison with an enlarged dataset consisting of all points from both fields (pooled approach) showed no improvement of the prediction accuracy but yielded decreased prediction errors. Lastly, the local MLR models were applied to the data of the respective other field to evaluate the cross-field prediction ability. The spatial SOC pattern generally remains unaffected on both fields; differences, however, occur concerning the predicted SOC level. Our results indicate a high potential of the combination of UAS-based remote sensing and environmental covariates, such as terrain attributes, for the prediction of topsoil SOC content at the field scale. The temporal flexibility of UAS offer the opportunity to optimize flight conditions including weather and soil surface status (plant cover or residuals, moisture and roughness) which, otherwise, might obscure the relationship between spectral data and SOC content. Pedologically targeted selection of soil samples for model development appears to be the key for an efficient and effective prediction even with a small dataset.
Introduction
Future sustainable and digitalized agriculture requires precise soil information from farm to field scale. This especially holds true for soil organic matter (SOM) distribution, as it represents one of the most important soil properties for crop production. It strongly influences soil structure as well as water, air and nutrient supply for plant growth. Unfortunately, SOM data are commonly scarce at field and farm scale. The sampling density in agricultural practice is often too low to cover the spatial heterogeneity of SOM in soil landscapes.
Remote sensing offers an option for a spatially explicit imagery of SOM. The spectral reflectance of SOM, and soil organic carbon (SOC) in particular, in the visible (VIS), the near-infrared (NIR) and the short-wave infrared (SWIR) between 350 nm and 2500 nm wavelength have been investigated in numerous studies with regard to either the qualitative composition of SOM or the quantification of SOC content [1][2][3][4][5][6]. The VNIR region between 350 nm and around 1100 nm SOC did not show any specific spectral features. Reflectance in general decreases with increasing SOC content caused by wide absorption bands related to chromophores affecting the soil color [7]. Unlike laboratory conditions, SOC related spectral features will be modified or even superposed by combinations of various factors under field conditions such as soil moisture [8], particle size, chemical constituents [9,10], surface roughness [10][11][12] and crop residues (e.g., stubbles and straw cover) [13,14].
In awareness of these uncertainties, several studies have been conducted using the wide range of available multi-and hyperspectral sensors mounted on different spaceand airborne platforms aiming for the retrieval of SOC content at different scales. The review of [15] showed a wide range of prediction accuracies in terms of coefficients of determination (R 2 s between 0.23 and 0.95 considering all platforms). Relationships were mainly affected by the spectral-and spatial resolution, signal-to-noise-ratios, atmospheric conditions, the abovementioned surface conditions and finally scale [16,17]. The authors concluded the high potential of most sensors and different retrieval models for a reliable description of regional and in-field SOC variability but emphasized the need for ultrahigh spatial resolution RS techniques that address the agricultural monitoring scale.
The rapidly advancing technology of Unmanned Aerial Systems (UAS) offers the opportunity to meet the requirements of this specific issue and has the potential to circumvent or minimize some of the disadvantages related to space-and airborne sensors, in particular atmospheric and surface conditions. The usage is limited according to legal restrictions, payload capacity and flight duration, but advanced fixed-wing UAS have the ability to carry lightweight multispectral sensors [18] and to realize a spatial coverage up to a few hundred hectares in ultrahigh spatial resolution (<1 m). The temporal flexibility enables users to conduct missions under clear sky conditions and has the potential to reduce the effects of unfavorable surface conditions (e.g., soil moisture, crop residues). According to technical constraints (weight and/or dimension) and costs, most sensors available for UAS are currently restricted to the VNIR region between 400 nm and around 1000 nm, which is not necessarily a disadvantage. [19] compared the performance of two VNIR multispectral cameras arrays (Tetracam Mini-MCA6 and Parrot Sequoia) with hyperspectral data sampled by a spectroradiometer (ASD Fieldspec 3 FR) for a set of prepared topsoil samples (laboratory and outdoor illumination conditions, but no field conditions) and reported similar SOC prediction accuracies. In a comprehensive field study, [20] recently compared prediction abilities of multispectral Sentinel-2, Landsat-8, PlantScope satellite and UAS imagery with airborne hyperspectral data using different advanced multivariate regression techniques to identify the best sensor specific predictor variables (spectral bands) for SOC content. The results confirmed the general applicability of all sensors, but the lowest prediction accuracies were achieved for the UAS data. However, a previous experiment of [21] demonstrated the high performance of a Tetracam Mini-MCA6 for SOC predictions in a range between 0.5% and 3.8%. Despite SOC content for calibration and validation Remote Sens. 2021, 13, 3557 3 of 20 being sampled from different soil depths and principal components being used instead of soil reflectance, high prediction accuracies could be achieved.
Despite the high potential of remote sensing techniques, the involvement of environmental covariates have been proposed in numerous studies. Beside present and historic land use/land cover and climatic data, which are the most commonly used co-variables, first-order (e.g., aspect, slope) and second-order (e.g., topographic indices) derivatives of digital elevation models (DEM) contribute significantly to the reduction of SOC prediction errors at different scales [22]. So far, to the best of our knowledge, only few studies used remote sensing data in conjunction with terrain attributes for estimation of SOC. At the regional scale, [23] used a set of first-and second-order derivatives of a DEM and Landsat satellite imagery to create a spatial SOC distribution map using regression kriging. [24] reported correlations between the spatial variability of SOC derived from hyperspectral airborne imagery across and within different terrain classes delineated by the topographic wetness index (TWI). Terrain attributes strongly reduced the uncertainty of estimating SOC using satellite imagery compared to other predictors in a study of [25].
Most of the studies required a large dataset of soil samples to provide sufficient input for calibration and validation, which is unfavourable under practical considerations (field work, laboratory work and costs) [16,26]. Hence, the need of an efficient sampling strategy to avoid pseudo-replications and selecting truly representative locations for an accurate prediction of SOC have been emphasized [17].
Generally, the spatial distribution of soil types and soil properties in hummocky ground moraines mainly results from lateral topsoil translocation (including SOC) by tillage erosion, as has been demonstrated across continents [27][28][29][30][31]. These soil landscapes have been intensively studied under aspects of erosion feedbacks on crop biomasses and yields [31][32][33][34], as well as on carbon dynamics [35] and greenhouse gas fluxes [27,36,37]. The landscape is characterized by a large number of closed depressions (kettle holes) which act as ultimate sediment traps of transported topsoil material; hence SOC. Furthermore, capillary rise of shallow groundwater into the rooting zone of soils in closed depressions enhances plant growth, which increases the C input into the soil-plant system. At the same time highly reactive iron oxide phases are precipitated along the capillary fringe which enforces carbon stabilization mechanisms. Because of all these interacting processes and mechanisms soils of closed depressions show highest SOC content and stocks in this landscape type. At the convex landscape elements strong tillage erosion lead to a recurrent SOC decrease due to the admixture of subsoil horizons (low SOC) into plough layers by each tillage operation [35]. In essence, the terrain-related, interacting geomorphic and pedogenic processes and mechanisms control the strong contrast in SOC observed in these landscapes. Therefore, the inclusion of terrain parameters (TPI, TWI) into a model offers a promising pathway to improve the spatial prediction of SOC content and stocks.
In this study we investigate the potential of UAS-based multispectral, high spatial resolution imagery in combination with a second-order terrain attribute for the prediction of SOC content at field scale. In terms of sampling strategy, we performed a parsimonious approach; i.e., we used a small dataset of ground-based measurements at two representative arable fields for model development. We compare local models with a model build from an enlarged dataset consisting of samples from both fields (pooled approach). In addition, by applying local models to the respective other field, the ability for cross-field prediction will be evaluated.
Study Area
This study was performed at the "AgroScapeLab Quillow" (ASLQ)-ZALF's landscape laboratory covering 260 km 2 of the Uckermark region in NE Germany. The ASLQ is part of the Weichselian-age ground moraine landscape with a pronounced hummocky relief. This landscape type is quite important for global crop production as it covers~10% of the arable land at the global temperate climatic zone [27,38].
Convex knolls and closed, concave depressions take turns over a short distance (50-100 m), which leads to strong local soil-moisture gradients (dry to wet in <100 m distance). The area is characterized by a subcontinental climate with a mean annual air temperature of 8.7 • C. Mean annual precipitation decreases from 550 mm in the western to 450 mm in the eastern part of the ASLQ (1990-2019, ZALF Research Station Dedelow; 53 • 21 55.46 N/13 • 48 17.78 E). Soils mainly developed from illitic-calcareous glacial till (sandy loam) and-to a minor portion-from carbonate-free, sandy sediments (glacial, glaciofluvial, periglacial). The actual soil pattern of the area is strongly affected by soil erosion over the past centuries [29,[39][40][41]. Recent studies confirm the dominant role of tillage erosion compared to water erosion [40]. Only 20% of the arable land shows soils unaffected by soil erosion (Calcic Luvisols), mainly at lower midslopes. Extremely eroded soils (Calcaric Regosols) occur at convex landscape positions, especially hilltops (knolls), as well as on steep slopes. Strongly eroded soils (Nudiargic Luvisols) stretch from upslopes to upper midslopes. Footslopes and closed depressions comprise 20% of the landscape. Here, groundwater-influenced colluvial soils have developed (Gleyic-Colluvic Regosols, partly overlying peat). The study area is under intensive arable land use. Typical crops grown in the "ALSQ" are winter wheat (Triticum aestivum L.), winter canola (Brassica napus L.), maize (Zea mays L.), and winter barley (Hordeum vulgare L.). Soil erosion together with a strong wetness gradient leads to a very high spatial variability of growth conditions, hence crop biomasses [18].
The study was carried out at two contrasting arable fields of the ASLQ, designated as "hd02" ( This study was performed at the "AgroScapeLab Quillow" (ASLQ)-ZALF's landscape laboratory covering 260 km 2 of the Uckermark region in NE Germany. The ASLQ is part of the Weichselian-age ground moraine landscape with a pronounced hummocky relief. This landscape type is quite important for global crop production as it covers ~10% of the arable land at the global temperate climatic zone [27,38].
Convex knolls and closed, concave depressions take turns over a short distance (50-100 m), which leads to strong local soil-moisture gradients (dry to wet in <100 m distance). The area is characterized by a subcontinental climate with a mean annual air temperature of 8.7 °C. Mean annual precipitation decreases from 550 mm in the western to 450 mm in the eastern part of the ASLQ (1990-2019, ZALF Research Station Dedelow; 53°21′55.46″ N/13°48′17.78″ E). Soils mainly developed from illitic-calcareous glacial till (sandy loam) and-to a minor portion-from carbonate-free, sandy sediments (glacial, glaciofluvial, periglacial). The actual soil pattern of the area is strongly affected by soil erosion over the past centuries [29,[39][40][41]. Recent studies confirm the dominant role of tillage erosion compared to water erosion [40]. Only 20% of the arable land shows soils unaffected by soil erosion (Calcic Luvisols), mainly at lower midslopes. Extremely eroded soils (Calcaric Regosols) occur at convex landscape positions, especially hilltops (knolls), as well as on steep slopes. Strongly eroded soils (Nudiargic Luvisols) stretch from upslopes to upper midslopes. Footslopes and closed depressions comprise 20% of the landscape. Here, groundwater-influenced colluvial soils have developed (Gleyic-Colluvic Regosols, partly overlying peat). The study area is under intensive arable land use. Typical crops grown in the "ALSQ" are winter wheat (Triticum aestivum L.), winter canola (Brassica napus L.), maize (Zea mays L.), and winter barley (Hordeum vulgare L.). Soil erosion together with a strong wetness gradient leads to a very high spatial variability of growth conditions, hence crop biomasses [18].
Soil Sampling and Soil Analysis
Sampling locations at each field were selected according to pedological knowledge along an erosion-deposition gradient, which covers the full feature space in terms of soils' SOC status. Topsoil samples (0-15 cm) in both fields (hd02: 11, kr01: 8) were taken at points for which auxiliary soil data were available from previous sampling campaigns. The points were formerly selected by a stratified-random-sampling scheme, which included the TPI, the Normalized Difference Vegetation Index (NDVI) from satellites (Quickbird, RapidEye) as well as the apparent electrical conductivity (ECa) from geophysical mappings (EMI 38) as strata. All strata are involved in landscape scale carbon cycling, hence influence SOC content in topsoils. Crop growth conditions, related C inputs and resulting SOC content and stocks are determined by erosion states of soils in hummocky ground moraine landscapes. Erosion state itself is reflected by terrain attributes, especially TPI in the case of tillage erosion, which is by far the most important mode of erosion in the study area. Spatial distribution of the ECa served as a proxy for the clay content as the most important subsoil property (water, nutrient supply). Finally, crop biomass derived from satellite imagery is directly related to C-inputs into soils and reflect local conditions in an integrated manner.
We carried out a composite sampling to capture the local, small-scale variability. Three topsoils in 1 m distance to the central coring point were sampled along a triangle and bulked into one sample (=physical averaging). The achieved average sampling density of one soil sample per two hectares depicts a low, but common sampling density of agricultural practice in NE Germany.
Bulk soil samples were air dried, gently crushed and sieved at 2 mm to separate the fine earth fraction (<2 mm) from the gravel (>2 mm). The particle size distribution of the fine earth was determined by a combined wet sieving (>63 µm) and pipette (<20 µm) method (DINISO 11277 1998). Pretreatment for particle size analysis was performed by wet oxidation of the soil organic matter using H 2 O 2 (10 vol. %) at 80 • C and dispersion by shaking the sample end over end for 16 h with a 0.01 M Na 4 P 2 O 7 solution [33]. The total C content was determined by dry combustion at 1050 • C using an elemental analyzer (Vario EL, Elementar Analysensysteme). Soil inorganic carbon content (SIC, mainly CO 3 -C) was determined conductometrically using a Scheibler apparatus [42]. The SOC content was calculated as the difference between total C and SIC. Pedogenic iron (Fe) oxides were characterized by dithionite (Fe d ) and acid-oxalate extraction (Fe o ) [42]. The Fe concentrations in the solutions were determined by ICP-OES. All basic soil analyses were performed in two laboratory replicates.
Topographic Position Index (TPI)
The TPI is the difference between the elevation of each grid cell in a DTM and the averaged elevation of a neighbourhood defined by circles of arbitrary radius. According to [43], the TPI for a single grid cell is calculated by: where h is the elevation of a grid cell in meter above sea level, x h is the mean elevation of grid cells in the neighbourhood with radius r. Values of the TPI range between −1 and +1 where negative values represent DTM grid cells lower than their surroundings and positive values higher ones, respectively ( Figure 2). At the time of UAS missions no GNSS (Global Navigation Satellite System) reference station was available that provides the required positioning correction data for the generation of a high precision DTM from simultaneously captured RGB images. The DTM used here was generated in accordance with the latest quality standards for the entire ASQL from a LiDAR (Light Detection and Ranging) airborne mission conducted in 2010 by order of the "Amt für Geoinformation, Vermessung-und Katasterwesen" (AfGVK, Federal state of Mecklenburg-Vorpommern). The original resolution of 1 m showed unwanted details and structures related to actual farming practices (e.g., tillage operations and tractor lanes). The DTM was resampled to a resolution of 5 m to smooth out these effects.
of Mecklenburg-Vorpommern). The original resolution of 1 m showed unwanted details and structures related to actual farming practices (e.g., tillage operations and tractor lanes). The DTM was resampled to a resolution of 5 m to smooth out these effects.
The ArcView extension "TPI" (Jenness Enterprises, Flagstaff, AZ, USA) was used to calculate the TPI for the entire ASLQ. A radius of 5 grid cells (25 m) was found to be appropriate for this landscape [44] to identify small relief elements (knolls, closed depressions) and related soil types within both fields ( Figure 3). The ArcView extension "TPI" (Jenness Enterprises, Flagstaff, AZ, USA) was used to calculate the TPI for the entire ASLQ. A radius of 5 grid cells (25 m) was found to be appropriate for this landscape [44] to identify small relief elements (knolls, closed depressions) and related soil types within both fields ( Figure 3). ation of a high precision DTM from simultaneously captured RGB images. The DTM used here was generated in accordance with the latest quality standards for the entire ASQL from a LiDAR (Light Detection and Ranging) airborne mission conducted in 2010 by order of the "Amt für Geoinformation, Vermessung-und Katasterwesen" (AfGVK, Federal state of Mecklenburg-Vorpommern). The original resolution of 1 m showed unwanted details and structures related to actual farming practices (e.g., tillage operations and tractor lanes). The DTM was resampled to a resolution of 5 m to smooth out these effects.
The ArcView extension "TPI" (Jenness Enterprises, Flagstaff, AZ, USA) was used to calculate the TPI for the entire ASLQ. A radius of 5 grid cells (25 m) was found to be appropriate for this landscape [44] to identify small relief elements (knolls, closed depressions) and related soil types within both fields ( Figure 3).
UAS Remote Sensing
The fixed wing UAS Tron (Quantum Systems GmbH, Gilching, Germany) is a vertical take-off and landing (VTOL) system (Figure 4a). With a wingspan of 3.5 m and the four electric tiltrotors the Tron combines the benefits of both long-range aerodynamic flight capabilities up to 2 h and independency from airstrips in close distance to the flight area. The total take-off weight is about 14.5 kg including the camera payload of approximately 2 kg. The core of the sensor equipment is a miniature multi-camera array Mini-MCA 6 (MCA hereafter) (Tetracam Inc., Chatsworth, CA, USA) supplemented by a FLIR Tau 640 thermal camera incorporated in the same rugged chassis. The MCA consists of an array Remote Sens. 2021, 13, 3557 7 of 20 of six individual CMOS sensors (1280 × 1024 pixels; pixel size 5.2 µm), lenses (focal length 9.6 mm) and mountings for user defined, interchangeable band-pass filters. The six mounted narrow-band filters, manufactured by Andover (Andover Corp., Salem, MA, USA), cover the spectral range from visible to near infrared wavelengths (550 nm, 570 nm, 610 nm, 656 nm, 760 nm and 900 nm). The bandwidths range between 10 nm (the first five bands) to 40 nm (band 6). A skyward looking electronic incident light sensor (eILS) collects downwelling solar irradiation using an identical band filter set as mounted on the MCA. The eILS images are captured simultaneously to each MCA image and are used in the postprocessing chain for radiometric calibration purposes. Camera lenses are partly covered by view caps, which narrow the field of view to minimize vignetting effects. An off-the-shelf 24 MP Sony alpha 5100 camera (Figure 4b) complements the sensor package. Both cameras are triggered by the Tron autopilot in a 2 s interval.
The fixed wing UAS Tron (Quantum Systems GmbH, Gilching, Germany) is a vertical take-off and landing (VTOL) system (Figure 4a). With a wingspan of 3.5 m and the four electric tiltrotors the Tron combines the benefits of both long-range aerodynamic flight capabilities up to 2 h and independency from airstrips in close distance to the flight area. The total take-off weight is about 14.5 kg including the camera payload of approximately 2 kg. The core of the sensor equipment is a miniature multi-camera array Mini-MCA 6 (MCA hereafter) (Tetracam Inc., Chatsworth, CA, USA) supplemented by a FLIR Tau 640 thermal camera incorporated in the same rugged chassis. The MCA consists of an array of six individual CMOS sensors (1280 × 1024 pixels; pixel size 5.2 μm), lenses (focal length 9.6 mm) and mountings for user defined, interchangeable band-pass filters. The six mounted narrow-band filters, manufactured by Andover (Andover Corp., Salem, MA, USA), cover the spectral range from visible to near infrared wavelengths (550 nm, 570 nm, 610 nm, 656 nm, 760 nm and 900 nm). The bandwidths range between 10 nm (the first five bands) to 40 nm (band 6). A skyward looking electronic incident light sensor (eILS) collects downwelling solar irradiation using an identical band filter set as mounted on the MCA. The eILS images are captured simultaneously to each MCA image and are used in the post-processing chain for radiometric calibration purposes. Camera lenses are partly covered by view caps, which narrow the field of view to minimize vignetting effects. An off-the-shelf 24 MP Sony alpha 5100 camera (Figure 4b) complements the sensor package. Both cameras are triggered by the Tron autopilot in a 2 s interval. The missions were carried out at bare fields on 16 August 2018 (hd02) and 30 April 2019 (kr01) close to the respective maximum possible solar zenith angle. We chose a flight altitude of 200 m above ground to achieve a sufficient ground sampling distance of approximately 0.11 m. Sky conditions were cloud free and winds speeds were moderate during both missions. The images of field hd02 were partly affected by cloud shadows in the eastern part of the field and were excluded from further image processing. Both missions were carried out during a longer rain-free period a couple of days after the preparation of the sowing bed of subsequent crops.
After milling and rolling, soil surfaces showed gentle roughness and little variation between the two fields but larger differences across different soil types. Soils with higher clay content formed larger soil aggregates (Figure 5a) than the sandier soils ( Figure 5b) and showed respectively higher shading effects. Soil moisture in the upper 2 cm was low in almost all topographic positions with the exception of some small areas in the center of depressions. The missions were carried out at bare fields on 16 August 2018 (hd02) and 30 April 2019 (kr01) close to the respective maximum possible solar zenith angle. We chose a flight altitude of 200 m above ground to achieve a sufficient ground sampling distance of approximately 0.11 m. Sky conditions were cloud free and winds speeds were moderate during both missions. The images of field hd02 were partly affected by cloud shadows in the eastern part of the field and were excluded from further image processing. Both missions were carried out during a longer rain-free period a couple of days after the preparation of the sowing bed of subsequent crops.
After milling and rolling, soil surfaces showed gentle roughness and little variation between the two fields but larger differences across different soil types. Soils with higher clay content formed larger soil aggregates (Figure 5a) than the sandier soils ( Figure 5b) and showed respectively higher shading effects. Soil moisture in the upper 2 cm was low in almost all topographic positions with the exception of some small areas in the center of depressions.
Additionally, two of the depressions in field hd02 were rummaged by boars. The main difference concerning surface conditions on the two fields was caused by management effects. While field kr01 was almost free from crop residues, the surface of field hd02 was sparsely covered by straw residues of winter wheat. However, the coverage was higher in depressions, were proper seedbed preparation failed due to technical limitations of the cultivator or harrow. Remote Sens. 2021, 13, 3557 8 of 21 Additionally, two of the depressions in field hd02 were rummaged by boars. The main difference concerning surface conditions on the two fields was caused by management effects. While field kr01 was almost free from crop residues, the surface of field hd02 was sparsely covered by straw residues of winter wheat. However, the coverage was higher in depressions, were proper seedbed preparation failed due to technical limitations of the cultivator or harrow.
Image Pre-Procesing
The retrieval of SOC from a bare soils surface requires an absolute calibration of the collected imagery because a recorded digital number (DN) is not only a function of the spectral characteristics of a bare soil surface but also of environmental conditions [45]. These include in particular the atmospheric conditions during the flight and the respective illumination geometry (solar zenith and sensor viewing angles). A straightforward way to convert DN to at-surface reflectance is the use of eILS measurements. An in-flight calibration of images using each corresponding eILS measurement showed unsatisfactory results clearly apparent by means of slight banding artefacts in the final mosaic of all calibrated images. Without having a reasonable explanation, the patterns disappeared with the use of a single eILS measurement selected by two criteria: (i) flight direction and (ii) flight attitude. We obtained best results by using an eILS measurement captured in flight direction towards the sun and almost no UAS movement along the pitch-and roll-axis (correspond to nadir-view conditions). We used the PhotoScan-Pro (V. 1.7.) software (Agisoft LLC, St. Petersburg, Russia) for image mosaicking. The workflow involves common photogrammetric procedures including the search for conjugate points by feature detection algorithms used in the bundle adjustment procedure, approximation of camera positions and orientation, geometric image correction, point cloud and mesh creation, automatic georeferencing and finally the creation of an orthorectified mosaic [46]. This workflow (lens distortion correction included) was applied to create an orthorectified mosaic from RGB images collected by the RGB camera (GSD ≈ 4.8 cm) and flight attitude data. The involvement of flight attitude data in the workflow produced discernable misalignments in the resulting multispectral MCA orthomosaic. Therefore, eILS-calibrated images captured by each MCA band were mosaicked individually. The GPS-measured positions of white markers (15 cm × 15 cm) were used to improve the spatial accuracy of the automatically generated RGB orthomosaic applying the ArcMap (V. 10.4.1) georeferencing tool. The orthomosaic was transformed to the local coordinate system ETRS 89 UTM 33, which finally served as a spatial reference for the co-registration of the six MCA orthomosaics. On average, 15 ground control points were sufficient to yield a co-registration accuracy between 1-2 pixels using a second order polynomial function. We used an ASD Field-Spec 4 Wide-Resolution spectrometer (ASD Inc., Boulder, CO, USA) to sample bare soil
Image Pre-Procesing
The retrieval of SOC from a bare soils surface requires an absolute calibration of the collected imagery because a recorded digital number (DN) is not only a function of the spectral characteristics of a bare soil surface but also of environmental conditions [45]. These include in particular the atmospheric conditions during the flight and the respective illumination geometry (solar zenith and sensor viewing angles). A straightforward way to convert DN to at-surface reflectance is the use of eILS measurements. An in-flight calibration of images using each corresponding eILS measurement showed unsatisfactory results clearly apparent by means of slight banding artefacts in the final mosaic of all calibrated images. Without having a reasonable explanation, the patterns disappeared with the use of a single eILS measurement selected by two criteria: (i) flight direction and (ii) flight attitude. We obtained best results by using an eILS measurement captured in flight direction towards the sun and almost no UAS movement along the pitch-and rollaxis (correspond to nadir-view conditions). We used the PhotoScan-Pro (V. 1.7.) software (Agisoft LLC, St. Petersburg, Russia) for image mosaicking. The workflow involves common photogrammetric procedures including the search for conjugate points by feature detection algorithms used in the bundle adjustment procedure, approximation of camera positions and orientation, geometric image correction, point cloud and mesh creation, automatic georeferencing and finally the creation of an orthorectified mosaic [46]. This workflow (lens distortion correction included) was applied to create an orthorectified mosaic from RGB images collected by the RGB camera (GSD ≈ 4.8 cm) and flight attitude data. The involvement of flight attitude data in the workflow produced discernable misalignments in the resulting multispectral MCA orthomosaic. Therefore, eILS-calibrated images captured by each MCA band were mosaicked individually. The GPS-measured positions of white markers (15 cm × 15 cm) were used to improve the spatial accuracy of the automatically generated RGB orthomosaic applying the ArcMap (V. 10.4.1) georeferencing tool. The orthomosaic was transformed to the local coordinate system ETRS 89 UTM 33, which finally served as a spatial reference for the co-registration of the six MCA orthomosaics. On average, 15 ground control points were sufficient to yield a co-registration accuracy between 1-2 pixels using a second order polynomial function. We used an ASD FieldSpec 4 Wide-Resolution spectrometer (ASD Inc., Boulder, CO, USA) to sample bare soil spectra directly after image acquisition. White reference spectra were collected before soil measurements using a Spectralon ® reference panel. Since the distance between fibre optics and soil was approximately 1 m, the collected spectra represented an average reflectance from a circle of 0.44 m in diameter. Measurements from all locations were acquired to check agreement with eILS calibrated MCA spectra. We then compared the averaged MCA reflectance at ASD sample locations (5 × 5 pixels) with mean ASD reflectance sampled in respective MCA bands according to bandwidth.
Spectral Band Selection
Due to the limited numbers of samples, a simple Multiple Linear Regression (MLR) model was used to predict SOC content. The TPI as a stand-alone predictor was not considered to be an appropriate explanatory variable for SOC content, nor for the pooled approach or the cross-field predictions since the same TPI values may correspond with different SOC content in other fields caused by slightly different soil properties (e.g., clay content) or management (e.g., fertilization). To avoid overfitting, only one spectral band, together with the TPI, was selected by a leave-one-out-cross-validation (LOOCV) using the hydroGOF package in R software (R development Core Team, Vienna, Austria). To assess accuracies and predictive power of the models built from the respective covariates, the coefficient of determination (R 2 ), the Root Mean Square Error (RMSE; Equation (2)) and the Ratio of Performance to Interquartile Range (RPIQ; Equation (3)) were computed for both the pooled approach and the two individual fields. We evaluated the contribution of the TPI to the model performance by repeating the LOOCV for the three single bands without the TPI as a covariate. Analysis of variance (ANOVA) was applied to assess the significance of model improvement by the TPI.
where y o and y p are observed and predicted values and n is the number of data pairs. IQ is the difference between Q3 (values below 75%) and Q1 (values below 25%) of the sample distribution. Due to the small dataset, no upper and lower limits for IQ were set, which usually detects outliers. Finally, the variance inflation factor (VIF; Equation (4)) was calculated to examine multicollinearity among predictor variables for the best performing variable combination.
where R j 2 is the R 2 value obtained by regressing the j th predictor on the remaining predictor. A VIF value of 1 indicates no correlation among predictor variables whereas values larger than 5 are suggested for detecting multicollinearity. However, there is no general agreement regarding a threshold value clearly indicating serious multicollinearity [47].
Site Properties
Descriptive statistics of relevant site properties are summarized in Table 1. The SOC content in the upper 15 cm ranged from 0.8% at strongly eroded hilltops (field hd02) to 3.0% at the center of a depression (field kr01). Within field variation was higher at field kr01 due to the high SOC content in the central depression, which is a periodically flooded (former) kettle hole.
Image Processing
The mosaicking procedure failed for the eILS-calibrated images captured in bands b610 and b656 of the hd02 mission exactly for the relevant part of the field containing most of the sampling points. Although MCA reflectance values agreed well with ASD values in the single images (data not shown), we excluded both bands from further evaluation. ASD spectra of the remaining bands (b 550 , b 570 , b 760 and b 900 ) revealed inconclusive results. Bands b 550 and b 570 were in good accordance with ASD values for both fields. Mean differences ranged between −0.007 and 0.008 (field kr01) and 0.027 and 0.014 (field hd02) for b 550 and b 570 , respectively. On average, band b 760 matched perfectly for field kr01 but showed a considerable difference (0.055) at field hd02 ( Figure 6). The worst agreement existed for the NIR band b 900 . MCA values exceeded ASD values to more than 100% suggesting a systematic error in the calibration file shipped with the camera and was therefore rejected. The scatterplot of the remaining three bands indicated slightly higher MCA reflectance for the hd02 mission in the VIS bands and a generally lower dynamic range in all MCA bands compared to ground measured ASD values.
Predictor Selection and Statistical Analysis
Regarding the statistical measures produced by the LOOCV, b570 was found the best predictor for SOC in all evaluated approaches regardless of using a single band or in conjunction with the TPI (Table 2). Among the control variants without the TPI, the local model for field kr01 produced the lowest prediction error and showed the highest predic-
Predictor Selection and Statistical Analysis
Regarding the statistical measures produced by the LOOCV, b 570 was found the best predictor for SOC in all evaluated approaches regardless of using a single band or in conjunction with the TPI (Table 2). Among the control variants without the TPI, the local model for field kr01 produced the lowest prediction error and showed the highest prediction ability (RMSE = 0.24%; R 2 = 0.84; RPIQ = 2.72). Despite the use of all 19 samples (pooled approach), the prediction performance decreased significantly (RMSE = 0.34%; R 2 = 0.60; RPIQ = 1.42), but outperformed the local hd02 model (RMSE = 0.30%; R 2 = 0.52; RPIQ = 0.82). By adding the TPI as a second explanatory variable, all measures computed for the local model kr01 indicated an increased prediction accuracy (RMSE = 0.11%; R 2 = 0.91 and RPIQ = 4.27). Regarding the prediction performance, the effect was less pronounced for the field hd02 model, where the RPIQ was slightly higher (RPIQ = 0.84 vs. 0.82), but the RMSE increased from 0.30% to 0.33%. In contrast, the TPI improved the prediction performance and accuracy for the pooled data (RMSE = 0.28%; R 2 = 0.73; RPIQ = 1.73). Correlations among the most successful predictor variables b 570 and TPI were moderate for the local models (field kr01: R 2 = 0.61, field hd02: R 2 = 0.61) and low for the pooled dataset (R 2 = 0.14). Calculated VIF values of 2.54 (field kr01), 2.56 (field hd02) and 1.16 (pooled dataset) indicated no serious multicollinearity.
The ANOVA showed a minor but significant (p < 0.01) deterioration of the model performance when the TPI was included. Nonetheless, regarding the small number of samples and the respective distribution especially on field kr01, the results of the LOOCV were highly sensitive to the prediction of samples marking the upper and lower tail of the SOC range.
Spatial Distribution of SOC
According to the better prediction performance using the TPI, the spatial distribution of SOC content was computed using the MLR with the two predictor variables b 570 and TPI. We compared the mapped SOC content (local versus pooled) to assess quantitative differences and the impact on the spatial distribution apart from performance metrics when using a small rather than a large number of samples. Furthermore, we evaluated the ability of the local models for cross-field predictions.
Comparison of Local Models with the Pooled Approach
R 2 s and RMSE for each approach are given in Table 3 and corresponding model equations are given in Table 4. The respective relationships between predicted and observed SOC are shown in Figure 7a,b. The R 2 s ranged between 0.79 (field hd02) and 0.88 (field kr01). The low RMSE values between 0.05 and 0.07 indicated high prediction accuracy. Figure 8a,b depicts the spatial distribution of SOC computed from the local models for field hd02 and kr01, respectively. Regardless of the higher level of SOC content on field kr01, the spatial distribution clearly showed a general accordance with terrain. However, in contrast to field hd02, higher SOC content occurred in up-and midslope positions on field kr01 (e.g., points 11 and 13), which hardly fell below the content in the central depressions. This might be related to the dominance of stagnant water soils, which are known to reveal higher SOC in topsoils due to: (i) O2 deficiency during wetter periods (inter-/intraannual) and (ii) higher C sorption capacity of newly formed, amorphous Fe oxides (redox processes). Figure 8a,b depicts the spatial distribution of SOC computed from the local models for field hd02 and kr01, respectively. Regardless of the higher level of SOC content on field kr01, the spatial distribution clearly showed a general accordance with terrain. However, in contrast to field hd02, higher SOC content occurred in up-and midslope positions on field kr01 (e.g., points 11 and 13), which hardly fell below the content in the central depressions. This might be related to the dominance of stagnant water soils, which are known to reveal higher SOC in topsoils due to: (i) O 2 deficiency during wetter periods (inter-/intraannual) and (ii) higher C sorption capacity of newly formed, amorphous Fe oxides (redox processes).
Compared with observed values, highest prediction errors occurred at the lower and upper tail of the SOC range. At both tails, SOC was underestimated by the local models in extreme terrain positions (e.g., point 25 on field hd02; hilltop, point 5 on field kr01; hilltop, point 0 on field kr01; depression). Overestimations occurred at points in transitional terrain positions with medium SOC content (e.g., point 33 on field hd02; footslope, point 14 on field kr01; midslope). The MLR developed from all 19 points from both fields (R 2 = 0.86; RMSE = 0.05%) revealed no improvement of the prediction accuracy compared to the local models regarding SOC levels and the spatial patterns. The relative accuracy, calculated as the difference between the models, was negligible for field hd02 and showed deviations lower than ±0.1% SOC (Figure 9a). Less than 2% of the area showed larger deviations again occurring in extreme terrain positions (central depression and one of the hilltops) of field kr01 (Figure 9b). However, the comparison with observed SOC values at point level indicated an outperformance of the local models at the upper and lower tail of the SOC range. With the exception of point 5 on field kr01, predicted content clearly exceeded the errors produced by the local models unless a larger dataset was used. Figure 8a,b depicts the spatial distribution of SOC computed from the local models for field hd02 and kr01, respectively. Regardless of the higher level of SOC content on field kr01, the spatial distribution clearly showed a general accordance with terrain. However, in contrast to field hd02, higher SOC content occurred in up-and midslope positions on field kr01 (e.g., points 11 and 13), which hardly fell below the content in the central depressions. This might be related to the dominance of stagnant water soils, which are known to reveal higher SOC in topsoils due to: (i) O2 deficiency during wetter periods (inter-/intraannual) and (ii) higher C sorption capacity of newly formed, amorphous Fe oxides (redox processes). to the local models regarding SOC levels and the spatial patterns. The relative accuracy, calculated as the difference between the models, was negligible for field hd02 and showed deviations lower than ±0.1% SOC (Figure 9a). Less than 2% of the area showed larger deviations again occurring in extreme terrain positions (central depression and one of the hilltops) of field kr01 (Figure 9b). However, the comparison with observed SOC values at point level indicated an outperformance of the local models at the upper and lower tail of the SOC range. With the exception of point 5 on field kr01, predicted content clearly exceeded the errors produced by the local models unless a larger dataset was used. Figure 9. Spatial distribution of deviations between SOC predicted from the pooled approach and the local models: (a) field hd02; (b) field kr01; numbers (white background) indicate selected deviations between SOC predicted by the pooled approach and measured values.
Cross-Field Prediction Ability
The local model developed for each field was used to predict SOC content in the topsoil of the respective other field. Despite the smaller influence of the TPI on field hd02, the application of Equation (4) (field kr01) to field hd02 data showed a high correlation between cross-field predicted and observed SOC content ( Figure 10) but tended to underestimate lower values.
Cross-Field Prediction Ability
The local model developed for each field was used to predict SOC content in the topsoil of the respective other field. Despite the smaller influence of the TPI on field hd02, the application of Equation (4) (field kr01) to field hd02 data showed a high correlation between cross-field predicted and observed SOC content ( Figure 10) but tended to underestimate lower values. Regarding the difference between the local model and the cross-field prediction, the underestimation is larger than 0.1% for 60% of the total field area (Figure 11a). Nevertheless, SOC content was underestimated by only 9.6% (absolute 0.11%) on average. As observed for the local model, the largest difference between predicted and observed values occurred at point 25 (0.27% and 0.83%, respectively), compared to the local model the prediction error which increased significantly from −0.29% to −0.56% SOC. In contrast, the differences for higher SOC content were negligible (e.g., point 33; local 0.24%, across-field 0.22%). The cross-field prediction of SOC content on field kr01 using Equation (5) showed minor differences regarding the spatial distribution and the SOC level. The deviation, depicted in Figure 11b, ranged between −0.1% and 0.1% SOC for 96% of the field area compared to the local model. Regarding the difference between the local model and the cross-field prediction, the underestimation is larger than 0.1% for 60% of the total field area ( Figure 11a). Nevertheless, SOC content was underestimated by only 9.6% (absolute 0.11%) on average. As observed for the local model, the largest difference between predicted and observed values occurred at point 25 (0.27% and 0.83%, respectively), compared to the local model the prediction error which increased significantly from −0.29% to −0.56% SOC. In contrast, the differences for higher SOC content were negligible (e.g., point 33; local 0.24%, across-field 0.22%). The cross-field prediction of SOC content on field kr01 using Equation (5) showed minor differences regarding the spatial distribution and the SOC level. The deviation, depicted in Figure 11b, ranged between −0.1% and 0.1% SOC for 96% of the field area compared to the local model. SOC content was underestimated in the central depression, where the prediction error, represented by point 0, increased from −0.29% to −0.45%. The larger dark colored area in the surrounding of point 5 indicated an overestimation towards the local model. As already observed for the pooled approach, the prediction error produced by the local model was larger than the error produced by the across-field prediction. The difference between predicted and observed SOC content at point 5 decreased from −0.28% to −0.07%. Since the error for transitional terrain positions (e.g., point 14) equaled more or less the error produced by the local and the pooled approach, the cross-field prediction performed equally well for the major part of the field area and even outperformed the local model on eroded hilltops with low SOC content (point 7). occurred at point 25 (0.27% and 0.83%, respectively), compared to the local model the prediction error which increased significantly from −0.29% to −0.56% SOC. In contrast, the differences for higher SOC content were negligible (e.g., point 33; local 0.24%, across-field 0.22%). The cross-field prediction of SOC content on field kr01 using Equation (5) showed minor differences regarding the spatial distribution and the SOC level. The deviation, depicted in Figure 11b, ranged between −0.1% and 0.1% SOC for 96% of the field area compared to the local model. Figure 11. Spatial distribution of deviations between SOC from cross-field prediction and local prediction: (a) field hd02 (b) field kr01; numbers (white background) indicate selected deviations between SOC from cross-field prediction and measured values. Figure 11. Spatial distribution of deviations between SOC from cross-field prediction and local prediction: (a) field hd02 (b) field kr01; numbers (white background) indicate selected deviations between SOC from cross-field prediction and measured values.
Discussion
We presented a simple multivariate approach for the prediction of spatially distributed SOC content in topsoils at the agricultural field scale incorporating soil reflectance data by UAS remote sensing and topography, based on a small dataset of ground-based measurements. Unfortunately, among the six MCA bands, only three bands were available for SOC prediction in topsoils of both fields showing sufficient calibration quality and full spatial coverage. However, we found the highest prediction ability according to the LOOCV statistics for the two local models and the pooled approach for the reflectance at 570 nm, irrespective of using the reflectance as a single predictor variable or together with the TPI as an independent covariate. Regardless of remote sensing platforms, sensors, scales and data evaluation techniques used in previous studies, our results generally confirmed the suitability of spectral bands located in the VIS-NIR wavelength domain, provided that sensors are not equipped with spectral bands in the SWIR region. Since SOC exhibits no specific absorption features in this spectral region the reported results reflected the simple relationship between soil colour and reflectance-the higher the SOC content, the lower the reflectance.
However, a bundle of factor combinations is known to modify surface reflectance (e. g., soil type, surface and atmospheric conditions) and the variety of available sensors (including non-imaging spectroradiometers) provides reflectance data with different spectral and spatial resolutions. Consequently, a range of spectral bands between 400 and 700 nm were used successfully for SOC prediction. A significant influence of SOC content on reflectance of different soils using spectroradiometers was reported for the VIS wavelength domain (550 to 700 nm, SOC range 0.8% to 9.0%) [10]; 500 to 700 nm, SOC range 0.1% to 4.0% [48]; around 600 nm, SOC range 0.2% to 40.2% [3]). Several statistical measures (Variance Importance in Projection (VIP) and factor loadings from Partial Least Square Regression (PLSR) technique or Relative Importance (RI)) were calculated to identify the most important spectral regions in multi-and hyperspectral data. Similar results were found for different spaceborne sensors [49], (SOC range 0.1% to 16%) and airborne HyMap spectra [50]; (SOC range 1.1% to 3.9%), [51] (SOC range 1.1% to 3.0%). Recently, a comprehensive study was conducted to compare current multi-and hyperspectral sensors (including a UAS-mounted multispectral camera) with respect to their SOC prediction ability for soil samples of the same field (SOC range 0.8% to 2.6%) and found the most important spectral bands, less precise, in the VIS-NIR region [20].
Using only the reflectance at 570 nm, our prediction performance showed satisfying results in terms of statistical metrics of the LOOCV especially for field kr01 (RMSE = 0.24%, RPIQ = 2.72). Remarkably worse values were found for field hd02 (RMSE = 0.30%, RPIQ = 0.82). If reported, aforementioned authors found similar prediction accuracies in terms of RMSE between 0.16 and 0.50 but RPIQs ranged between 1.53 and 3.34 (only reported by [20,48]. The performance of the UAS-mounted multispectral camera (Parrot Sequoia, 4 spectral bands) was slightly lower (RMSE = 0.31%, RPIQ = 1.77) compared to all other sensors. Better results were achieved by an earlier UAS mission conducted over bare soils of a field trail (SOC range 0.5% to 4.7%) at Rothamsted Research (Harpenden, UK) using a Tetracam Mini-MCA6 (RMSE = 0.21%, RPIQ = 2.3) [21]. However, results of most studies based on a larger number of observations compared to our approach, which in turn increased the model performance. This was partly confirmed by our pooled approach. Here, the RPIQ increased to 1.42, which is higher than the RPIQ obtained for field hd02, but lower than the value for field kr01. Nonetheless, the prediction accuracy decreased (RMSE = 0.34%) compared to the local approach. As proposed by different authors [17,20], the prediction accuracy might be improved by incorporating other environmental covariates into prediction models. By using the TPI in our study, the prediction error clearly decreased, and the model performance clearly increased in case of field kr01 (RMSE = 0.11%; RPIQ = 4.27). The RPIQ increased only slightly for field hd02 (RPIQ = 0.84 vs. 0.82 without TPI). For the pooled approach, the prediction error decreased (RMSE = 0.28% vs. 0.34% without TPI) and the RPIQ indicated a clearly better performance when using the TPI as an explanatory covariate (RPIQ = 1.73 vs. 1.42 without TPI). In conclusion, our results are in the range of worst and best model performances achieved by aforementioned studies.
The generally positive impact of the TPI was clearly related to the pronounced hummocky terrain of our study area. Convex and concave terrain features take turns over short distances and selected sample locations almost covered the full range of possible TPIs (−1 to +0.8). The close relationship between terrain position and SOC was founded in soil redistributions by tillage erosion [24,52,53]. Nevertheless, the TPI should be used carefully because values represent a spatially demarcated terrain attribute calculated from the immediate vicinity of a single grid cell. Therefore, limited conclusions can be drawn on the spatial dimension of, for example, a depression, which in turn determines the amount of locally sequestrated or relocated (by erosion) SOC. Hence, higher SOC content may occur in the center of wide, flat depressions as observed for point 0 on field kr01 (TPI = −0.4; SOC = 3%) than in small ones as observed for point 29 on field hd02 (TPI = −1.0; SOC = 2%). Therefore, the TPI is not suited for SOC prediction purposes without spectral covariates. Our MLR for spatial prediction of SOC content achieved high correlations and low prediction errors (field kr01: R 2 = 0.88, RMSE = 0.07%; field hd02: R 2 = 0.79, RMSE = 0.06%) showing similar intercepts and coefficients for the 570 nm band and the TPI.
As expected, the resulting maps preferentially showed a pattern of SOC content, which was consistent to topography. The highest errors regarding individual sample locations occurred in depressions (high SOC content) and hilltop positions (low SOC content), where observed SOC content was underestimated in both fields. The first case was also observed in two studies using high spatial resolution UAS imagery: for Colluvic Regosols in depressions and weakly eroded Chernozems [20] and for field trials with long-term applications of organic manure [21].
For hummocky soil landscapes, we assume decreasing SOC content with depth in topsoil. While observed SOC represent the average content in the sample depth (topsoil), the spectral reflectance complies with the SOC content at the surface (~2 cm). Assuming a homogeneous uniform SOC distribution in topsoil of Colluvic Regosols and a continuous decrease in other soils, the surface reflectance will yield SOC values closer to the average for Colluvic Regosols. Consequently, the surface reflectance tends to overestimate average SOC content of other soil profiles rather than underestimate those of Colluvic Regosols. So far, the assumption of [21] was likely, since most SOC values in this study were low or medium. The underestimation of SOC content in strongly eroded hilltop positions at points 25 (field hd02) and 5 (field kr01) appeared to be related to soil mineralogy. The soil analysis showed that both points had remarkable high SIC content (12.2% and 9.9% respectively) compared to the average for all other points (1.8% on field hd02 and 0.4% on field kr01), which in turn increased reflectance in the VIS region [9,10]. Compared with soils having low SIC but similar SOC content, a disproportionate high reflectance probably decreased the predicted SOC content at these points. In our study there was no evidence for a significant effect of other surface properties such as roughness and straw residues. We consider these effects to be negligible compared with uncertainties related to spectral data, the TPI and the laboratory SOC determination. Finally, our sample design was not dedicated to the evaluation of these effects and thus most of them remained unaffected.
One of the most interesting issues is the ability of local models for reliable cross-field predictions of SOC content without having any information than spectral data and, if available, environmental covariates of explanatory value. Our results are promising, since deviations between local models and the respective cross-field prediction were negligible concerning the spatial distribution, the overall level and finally the SOC prediction of individual samples. From this, we conclude that the selected samples represent well the overall SOC range of both fields. Although a larger number of samples would probably increase the prediction accuracy, there was no evidence for a lack of samples representing other constellations of soil and topography than the selected.
Although soils showed a wide range of properties in terms of texture and mineralogy on both fields, the spectral response to SOC content was similar. Despite of higher calcium carbonate and lower iron oxide content on field hd02, soils developed under the same climatic conditions from the same parent material and were managed over a comparable time period. Insofar, our results are not surprising and support the findings and subsequently recommendations of other authors to stratify soils, sampled at regional or larger scales, according to their soil forming factors and management regimes yielding higher prediction accuracies compared to global calibration approaches [16,17,54].
Our approach is easily transferable and applicable in soil landscapes showing a similar geomorphological setting and tillage erosion as a major driver of soil distribution. As already stated in the introduction, this includes all landscapes of hummocky ground moraines, which cover approximately 1 Mio. km 2 arable land world-wide (Sommer et al., 2004). It will also work out in landscapes, where a strong contrast in SOC is explainable by a terrain-related interaction of geomorphic and pedogenic processes influencing carbon dynamics. The inclusion of the TPI as explanatory variable should be tested in further areas dominated by tillage erosion. However, the TPI will not improve the prediction of SOC content by remote sensing techniques on flat terrain, where other environmental conditions are considered as driving factors for spatial differences (e.g., parent material, hydrological conditions). For soil erosion by water, the Topographic Wetness Index (TWI) might outperform the TPI as the TWI will reproduce spatial domains of overland flow more robust. For wind erosion, a 3D DEM was used to capture luv-lee effects on wind speed and related sink-source domains of dust [55,56].
Conclusions
Our approach offers a high potential for spatially distributed SOC predictions from UAS remote sensing and generally available environmental covariates. This especially holds true for areas of limited soil data. A successful application of our approach requires a targeted selection of sampling locations. Coring points need to represent the full range of soils and related drivers (soil forming factors and processes), i.e., they need to cover the full feature space. Furthermore, we emphasize the great importance of optimized conditions during image acquisition, mainly soil surface characteristics. We strongly recommend the avoidance of high roughness (e.g., after ploughing) causing strong contrasts between illuminated and shaded areas. Reflectance of dead plant material such as straw and stubbles can hardly be distinguished from soil reflectance and thus increases uncertainties in SOC predictions. This is more important than a huge number of soil samples, predictor variables or sensor specifications. In this landscape type, which is characterized by predominately conventional agriculture, the chance for optimal conditions is best fulfilled during a time period (approximately 10 days) after seed bed preparation. Conducting flight and field campaigns by minimizing effects obscuring the relationship between spectral data and SOC content even enable reliable cross-field prediction accuracy.
Beside the basic issues concerning the natural (physical and biological) processes controlling SOC content, this demand is directly linked to the question of how farmers can react in a site-specific manner on small-scale varying and increasing SOC losses. Regardless the high spatial resolution, the spatial coverage is a general limitation even for fixed-wing UASs due to actual legal and technical restrictions, but sufficient for addressing an agricultural monitoring for at least two reasons.
First, considering the importance of framework conditions, agricultural landscapes consist of many fields but only few of them offer ideal conditions at the same time due to management practice (e.g., crop rotation, seedbed preparation, crop residues), soil moisture and surface roughness. Hence, simultaneous SOC predictions for numerous agricultural fields (e.g., from satellite imagery) require detailed information of surface conditions and local corrections. This, in turn, relativizes the advantage of the large spatial coverage provided by space-borne sensors.
Secondly, the application of UAS remote sensing is feasible even for the monitoring of changes in SOC content since loss and accumulation are long-term processes showing minor changes from year to year even on intensively managed agricultural fields. The detection of significant changes beyond the level of laboratory accuracy requires low repetition rates in the order of magnitude of years. Most disadvantages associated with the use of UASs, and mounted sensors are of technical nature. However, rapid technical progress provides more and more enhanced: (i) UASs with extended flight endurance and high accurate positioning capabilities by using real-time or post-processed kinematic and (ii) sensors equipped with in-flight calibration techniques and lower signal-to-noise ratios. Decreasing acquisition costs and good prospects for a further liberalization of the current legal framework (e.g., flying beyond visual line of sight) for research and governmental organizations make UASs an efficient alternative to space-and airborne sensors at the field and landscape scale. | 14,039.2 | 2021-09-07T00:00:00.000 | [
"Environmental Science",
"Mathematics",
"Agricultural And Food Sciences"
] |
Shear modulus loss as damage indicator of structural integrity in bonded joints
Damage in single lap joints under shear load was assessed. Specimens were tested in accordance with ASTM D1002-10(2019). Controls were taken on adhesive and surface preparation, adhesive thickness, and load-specimen alignment. Digital Image Correlation was used to measure stresses over the adhesive, and shear modulus variation and percent of damage were assessed. Results show shear modulus losses against shear stress. Finally, early damage evolution was determined up to 10% of the shear load rate and later, increasing up to 70%-96%.
INTRODUCTION
Fracture mechanics is an important part and logic result of the field of strength of materials. Early work on the strength of materials focused first on basic properties and later, as the theories of elasticity and plasticity evolved to include structures with flaws were developed, on the strength of structures components containing stress concentrators. Fracture mechanics deals with the strengths of materials and structures which contain flaws in the form of detectable or visible sharp cracks.
The fundamental concept of damage has several meanings according to the context and application; however, Worden et al. 1 described the concepts of failure, damage, and defect in a hierarchical form as described below.
Failure. It is when a structure cannot operate satisfactorily. Damage. It is when the structure no longer operates in ideal conditions, but can operate in satisfactory conditions, but in a sub-optimal way. Defect. It is inherent in the material and statistically all materials will contain a known number of defects. This means that the structure will operate optimally even if the constituent materials contain defects.
Therefore, the main objective of a damage theory is to allow prediction of lifetime of the structure. Then the remaining life concept is a natural way to define damage. Damage is the center of the study of the Continuum Damage Mechanics (CDM) addresses among others the study of the growth of micro-cracks and micro-voids (and other defects) and their effect on the mechanical behavior of the material. Further consideration is the direct effect of damage on the mechanical properties of materials which result in a significant influence on the safety aspect of engineering structures. Due to the presence of damage at the microscopic level that generates an effect on the macroscopic properties of materials or structures, several groups of researchers have proposed that a damage index can be associated with changes observed in some properties such as the change of the area of the cross section in a standardized specimen or the change in the rigidity of the structure compared to a pristine or undamaged condition. Proposing 2 several scalar damage variables. The first scalar damage variable is defined in terms of the reduction in cross sectional area, while the second scalar damage variable is defined in terms of the reduction in the elastic modulus or elastic stiffness. Based in Voyiadijis et al., 2 Balieu et al., 3 Lamaitre, 4 and C. Basaran et al. 5 the Damage Index ( ) can be determined with the scalar variable of stiffness change damage : Where is the elastic modulus of the body or structure with the presence of damage and it is the effective elastic module, in pristine condition or without failure.
ADHESIVE JOINTS DAMAGE
One of the most important challenges of structural adhesive bonds is related to the difficult to detect the damage in advance or until the defect evolves and a "symptom" reveals its presence inside the material or structure. The study of the mechanics of damage in adhesive joints is carried out in a multidisciplinary context; however, it is an open and extensive field to research and with an impact on the strength, durability, and application of adhesive joints in high-tech structures and primary structures with high demands of mechanical performance and safety like aeronautics, space, automotive, wind energy, shipping, defense, etc. Wahab et al. 6 published the effect of many parameters for optimization and maximizing the fatigue resistance and durability of adhesive joints. These parameters were identified as geometric parameters, material parameters, surface loading and treatment conditions, and curing conditions. In 2009 Banea et al. 7 and in 2013 Krishnaraj et al. 8 worked on predicting strength on joints with both adhesive and rivets. In theory, adding a rivet to the adhesive bond allows to have a predictable fault path first in the rivet acting as a mechanical fuse before the adhesive bond fails; however, this premise is not always the safest, acceptable, or certifiable. For instance, in the case of composite materials, the adhesive bond can fail before the rivet due to damage caused to the substrate and the adhesive itself generating concentrations of stresses that significantly impact the structural integrity and can significantly reduce the mechanical strength of the complete joint. In 2015 Antunes 9 published a study about the flaws in adhesive bonds, the aim of the research was to analyze the strength and performance of the commercial epoxy adhesive 3M 9323 B/A with dissimilar substrates using Carbon-fiber-reinforced plastic (CFRP). As a preliminary approach the interest was to create a simple and realistic procedure able to produce practical results that could be used for standard bonded model calculations. The literature review and the state of the art aimed to get perfect results, on the other hand, the objective was to find results to be implemented on the daily basis, adjusted to the imperfections made for specimen manufacture. Antunes comments that there are factors during the elasticity test methods that affects the quality of the results. The slip from grips and the air entrapped on the adhesive during mixing and spreading affects the elastic parameters and important deviation for literature was found. Since the research was focused on the adhesive´s parameters assessment, a peeling test was carried out. It was found that specimen manufacturing errors, misalignment between bonded areas and the thickness control may affect the results and the failure evolution. The dissimilar substrates study there was not conclusive.
In 2016 de Sousa et al. 10 reported numerical and analytical models using the cohesive zone along with the extended finite element method to assess the damage in simple overlap adhesive joints. The analytical models were based on the Volkersen method, 11 the Goland-Reissner model 12 and the Hart and Smith elastic model. 13
SHEAR STRENGTH AND DAMAGE INDEX ON ADHESIVE JOINTS
Bonding by adhesives is usually the most appropriate solution when it comes to joining physically dissimilar or metallurgically incompatible materials, thermoset polymers, ceramics, elastomers, very thin materials, or substrates of very small size. Adhesive bonding is generally also suitable when it comes to joining large areas or when the use of adhesives means great improvements in terms of productivity. The central component of structural adhesive joints are high-performance adhesives such as epoxy polymer resins with thermo-mechanical and viscoelastic properties. An adhesive is a substance capable of containing at least two surfaces to bind them together in a strong and permanent way. Adhesives are chosen for their retention and bonding power. In general, they are materials that have high shear and tensile strength. The most common type of structural adhesive bonding is single lap joint (SLJ) in which the load is transferred from one substrate to another by shedding stresses in the adhesive (figure 2). Because the applied loads are off-center in a simple overlap joint, the bending action caused by the applied load creates normal stresses (splitting stress) in the direction of the adhesive thickness. The combination of cutting stress and normal stress reduces the strength of the joint at the ends of the overlap. For the purposes of selection and characterization of adhesives, it is very common to test SLJ joints because they are easy to manufacture.
As shown in the Figure 3 there are four types of failures that may occur during an adhesive joint test: Type 1: Adhesive failure is a failure at the interface between the adherent and the adhesive. Type 3: The combined failure is the one that presents partial zones with adhesive and cohesive fault in the overlap area. Substrate failure occurs when the bond between the adhesive and the substrate material is stronger than the internal strength of the substrate material. The latter type of failure is common in the adhesion of composite materials. Type 4: In substrate failure, the original materials fail away from the joint or near the junction area when the original materials detach. Another expected failure mode is adhesive cohesion failure, in which the adhesive splits into the bonding area but remains firmly attached to both substrates. Adhesive failure, where adhesions are released from substrate materials, is considered a weak bond and is generally unacceptable. An initial and simplified approach is supposed that the axial or tensile load V is exerted longitudinally to one of substrates (base material to be joined) and that is converted into a uniform shear stress ( = ⁄ ) through the contact area of the adhesive and subsequently again this shear stress is converted into another longitudinal force on the opposite substrate as schematized in the Figure 4.a.
In fact, the substrates are not rigid and deforms closer to their final load, as shown in the Figure 4.b. Thus the shear stress generated is not uniform but grows at the ends of the overlap, as illustrated in figure 4.c. Finally, the shear stress present in the adhesive by the off-center or non-collinear load causes a bending in a substrate. In addition to the non-uniform distribution of the shear stress, this flexion increases the transverse stresses, sometimes the well-known peeling stress (peeling) is formed, or at other times called cleavage or cleavage stress, figure 4.d.
Because of the chemical, mechanical, thermal, and viscoelastic processes involved in adhesive joints, the strength evaluated not only corresponds to the body of the adhesive, but to the effects of the substrates and bonding interfaces, which is why adhesive bonding is considered a complex system.
Equation 3
Where is the shear stress on the adhesive body, is the force applied to the inner adherend, the reciprocal of ω is the shear-lag distance as a measure of how quickly the load is transferred from one adherent to the other, is the distance measured from the middle of the overlap length, is the adhesive shear module, is the elastic modulus of the substrate, is the adhesive thickness, is the thickness of above substrate, is the thickness of below substrate, is the Poisson´s ratio. Shear modulus G is directly proportional to the shear stress at the adhesive body. That is the reason which changes in the shear modulus can be a measure of the damage or integrity of the joint system.
% = |1 − | × 100 Equation 4
Where is the shear module (quasi static) after increased uniaxial load and is the initial shear module in pristine condition.
COUPONS MANUFACTURING
The strength of adhesive bond specimens is sensitive to the repeatability in the manufacturing process and variability of some of the properties such as adhesive thickness, overlap length, surface preparation, adhesive mixing properties and proportions, adhesive application and curing. Due to the last, a specimen manufacturing process was established based on ASTM D1002-10 Standard Test Method for Apparent Shear Strength of Single-Lap-Joint Adhesively Bonded Metal Specimens by Tension Loading (Metal-to-Metal), adhesive manufacturer recommendations. 15, 16, 17 A special steel mold was manufactured to have repeatability and accuracy at the specimen geometry, reliability and parameter deviation control like adhesive preparation and application, wetting, uniform pressure, adhesive thickness, overlapping area, geometry, uniform heat treatment and easy demolding and cleaning previously to the test. Figure 5 shows the special mold, and the final SLJ specimens. There was manufactured 20 SLJ specimens and tested at quasistatic conditions.
SETUP FOR SHEAR STRESS MEASUREMENTS IN ADHESIVE JOINTS
To perform an assessment of damage in an adhesive joint the experimental setup of Quasi-static unidirectional shear condition or QSSC was established.
Quasi-static unidirectional shear. A Shimadzu AGS-X universal machine was used to carried out the traction cutting test of the single overlap joint specimens with reference to ASTM D1002-10 Standard Test Method for Apparent Shear Strength of Single-Lap-Joint Adhesively Bonded Metal Specimens by Tension Loading (Metal-to-Metal) ( figure 6). A quasi-static load was applied with a speed of 1,270 mm/min for a maximum load range of 10,000 N. Simultaneously, the Digital Image Correlation Technique was applied mapping the shear strain near the border between the adhesion zone to capture the deformations between the adhesive and adherent and getting in the post-processing stage the shear stress-strain curve (t−) and the shear modulus (G, MPa) in "without damage" or "pristine" condition. Digital Image Correlation. A high-resolution camera Allied Vision, Model Mako U503-B with 5 megapixels was used along with a 50 mm lens and extension tubes. Due to the high magnification and achieving a high accuracy approach, a support mounted on remote-controlled linear guides was used, with small linear displacements for each servomotor revolution. The configuration achieved a total of 940 pixels over the width of the joint, i.e. a resolution of 0.0072 mm/pixel, enough to accurately measure both deformations and displacements.
The measurement area of the adhesive joint with a width of 6.8 mm, with the dark contrast pattern and bright random spotting which allows to obtain the deformation and displacement vectors in the subspace that is selected for analysis. In addition, the resolution measured between the two reference points indicated as point 1 and 2, which is 940.93 pixels/6.8 mm, i.e. 138 pixels/mm, is displayed. The cameras were connected to a Dell Precision 3510 Workstation for sequential image capture and synchronized with the data acquisition rate of the 10 kN Shimadzu AGX-10 universal machine. Frame rate image acquisition was programmed to 1 image per second for static testing. In a complete case, the camera was synchronized at a speed of 5 images per second to capture a complete static test sequence.
EXPERIMENTAL RESULTS
There is a huge amount of information derived from the experimentation, that's because we present the results for one specimen only. The average results of the L3M57140N specimen are shown in figure 7 where the average maximum shear stress is tmax prom=9.32±2.86 MPa and the average maximum shear strain is max prom=0.07±0.02 rad.
The average stress-strain shear curve does not have a linear or proportional zone as in other materials, so the literature suggests that the yield shear stress and therefore the shear modulus can be obtained by the offset yield method 18,19 or by the method of the derivatives of the stress-strain shear curve. 20 tav=4.12 MPa for and av= 0.01 rad, so the shear modulus is Gav prom G=343.17 MPa. Figure 9 shows the percentage of damage rate based on the change of the quasi-static shear modulus %ID G , obtained from the digital correlation measurements of images and the derivative of the curve L5M4205KN. The figure shows that the percentage of damage rate %ID G is increased to 75.2% for a slight increase or load ratio rV of 0.01 in quasi-static condition. Subsequently, the %ID G decreases to 29.2% for load ratio rV of 0.08. After that, the %ID G is greatly increased up to 81.8%, to remain at values around 97% until the final failure.
DISCUSSION
As mentioned by Wahab, 6 "There are difficulties associated with modeling the nucleation of a crack and the ability to control and detect the initiation phase". However, in this work, it was possible with unexpensive tools and precise technical references have repeatability and accuracy at the specimen geometry, reliability and parameter deviation control like adhesive preparation and application, wetting, uniform pressure, adhesive thickness, overlapping area, geometry, and uniform heat treatment for the adhesive curing.
In the other hand, we can probe that shear modulus evolves against time according with the application of quasi-static load rate. This effect is identified as the Loss Stiffens or Shear Modulus Loss and its derivative respecting time is a possible signal of damage evolve and it can be assessed in function of time, with the shear load rate or even the shear strain rate. This behavior has been stablished at the literature even for numerical methods as the elasticity bi-lineal model, however there is not many results of its variations with adhesives and shear loading conditions. It was found that the quasi-static shear modulus G has a nonlinear variation, reaching a maximum level within the first 10% of the shear stress range and progressively decreases as the stress and strain level increases. This behavior was observed for all the points analyzed on the adhesion line. Last may be due to the viscoelastic behavior of the epoxy adhesive as a polymeric resin the shear modulus behavior can be interpreted as a degradation of the quasi-static elastic properties of the adhesive because internal molecular chains acquire a forced resorting in the direction of the shear load. These polymeric chains alignment tends to remain in time because degrees of freedom of adjacent molecules have been modified and therefore exhibit non-elastic behavior. The degradation of the quasi-static shear modulus is directly related to the viscoelastic creep and relaxation phenomena and these in turn the thermo-elastic properties of the adhesive. Regarding the percentage of damage rate (%ID G ) based on the variation of the shear modulus shows a nonlinear behavior close to a logarithmic function of the form = + ( ) where and can be calculated through the method of least squares. This percent index of damage (% = + ( )) is proposed as a predictive model of damage because can provide early information about damage evolution under quasi-static load for the adhesive at the single lap joint.
CONCLUSION
It is possible to assess with high accuracy and reliability the evolve of damage in SLJ under shear load, but more measures should be performed to reduce uncertainties derived from materials properties, preparation process, human manipulation vias, postprocessing procedures, etc. A different joint configuration should be tested with another physical conditions like variations at the load speed, types of loads (cyclic, impact, random) and different environmental conditions. When all these conditions have been analyzed the following steps should be testing real parts and components like aerospace or automotive components and after that testing with structures at scale and with real dimensions. The experimentation using large scale components using design of experiments could be expensive, that is why it is very important the continuous research in this field using coupon level samples. Techniques like Digital Image Correlation can provide a lot of information that can be used to validate theoretical models and computational algorithms about damage progress. DIC have its enough research field, however, is a powerful help to visualize the strain-stress distribution in a body.
Finally theoretical models can be fitted as a possible average behavior of the damage inside the adhesive body during the shear load application. This is a little step focused to know the early damage stages evolution and understand the hidden behaviors or phenomena that can primarily affect the adhesive joint both strength and safety, after affect components and finally a compromise a safety application. | 4,436.6 | 2022-01-30T00:00:00.000 | [
"Engineering",
"Materials Science"
] |
Distinguishing Effect of Buffing Operation on Surface Residual Stress Distribution and Susceptibility of 304 L SS and 321 SS Welds to Chloride Induced SCC
Stress corrosion cracking (SCC) of austenitic stainless steels (ASS) and its weldments in presence of chloride ions is a key concern in its successful application. AISI 304L SS in surface milled, turned, ground conditions have high tensile residual stresses on the surface which lead to early cracking in an aggressive environment. Spot welds of AISI 321 SS have shown multiple failures due to chloride induced SCC as a result of high magnitude of tensile residual stresses and improper post weld heat treatment. The present study proposes a simple surface engineering method to prevent the initiation of stress corrosion cracking in austenitic stainless steel and its welds in presence of chloride ions. 304L SS in milled, turned and ground conditions and 321 SS in spot welded condition was subjected to surface buffing operation. Surface roughness was calculated using a surface profilometer and residual stresses were determined. Residual stress distribution, and phase transformation were calculated using X-ray diffraction measurements. The detailed microstructural characterization was performed using field emission scanning electron microscopy (FESEM). 304L plates and 321 SS welds in buffed and un-buffed conditions were tested for SCC susceptibility by exposing these to boiling MgCl2 as per ASTM G36. Results showed that 304L SS and 321 SS were resistant to SCC in buffed conditions as no cracking occured even after prolnged exposure to boiling MgCl2. Buffing being a very simple, economic and portable operation can be easily applied on large components of AISI 304L SS after the conventional surface finishing operations and AISI 321 SS weld also can be extended to components in service in an aggressive environment. Introduction Austenitic stainless steels have good corrosion resistance and mechanical properties. However, these become highly susceptible to SCC when subjected to different surface finishing and welding operations [1-3]. Laser peening and shot peening are generally used for enhancing SCC resistance, but these processes induce plastic strain in the material and are not economic. Hence a simple and economic route is essential in preventing SCC. We have shown in our previous report that simple surface buffing can be used to enhance the SCC resistance of austenitic stainless steels [4]. The present study substantiates the effectiveness of buffing in preventing the initiation of Cl induced SCC of austenitic stainless steel in presence of chloride ions. 304L SS in milled, turned and ground conditions and 321 SS in spot welded condition when subjected to surface buffing operation. Residual Stresses 2018 – ECRS-10 Materials Research Forum LLC Materials Research Proceedings 6 (2018) 139-144 doi: http://dx.doi.org/10.21741/9781945291890-22 140 Experimental Materials and methods In the present study two steels namely, AISI 304L SS (C 0.03, Cr 18, Ni 8, Mn 1.6, P 0.04, Si 0.4,S 0.03, balance Fe) wt %, and AISI 321 SS welds (C 0.024, Cr 17.41, Ni 9.14, Mn 1.64, Ti 0.23, Mo 0.37, P 0.04, Si 0.4, S 0.03) wt% were used. AISI 304L SS samples were cut into a dimension of 100 mm X 25 mm X 5 mm plates and they were solution annealed in order to remove the internal stresses present in the material. AISI 304 L SS was subjected to three different surface working conditions namely a) milling b) turning and c) grinding operations at a feed rate of 0.1mm/rev to remove 0.5 mm from the surface [7]. Subsequently, the samples were buffed at 3600 rpm using and 50μm was removed from the surface. In another study, AISI 321 SS tubes having 72 mm diameter, 3 mm thickness and length of 3.6 m having spot welds on the surface were subjected to buffing and 50μm was removed. Both the samples were tested for SCC susceptibility in buffed and unbuffed condition. During buffing compressive forces are applied on the sample substrate as shown in the schematic given below (Figure 1). The buffing wheel is rotated at a set speed and is rastered on the surface of the workpiece imparting compressive stresses to the entire surface of the workpiece. No Ti2N particles were present in the near surface region of buffed samples in each case. Fig 1: a) Schematic of the buffing operation b) cross sectional micrograph of 321 SS in buffed condition and c) cross sectional micrograph of 321 SS in un-buffed condition. Surface roughness measurements Surface roughness measurements were done on all samples using a surface profilometer (contact mode) with a scan length of 1μm at a scan speed of 0.01 mm/sec with a minimum resolution step of 1 nm. 321 SS ring samples were measured using an optical surface profilometer (non contact mode) with a scan length of 1mm at a scan speed of 47μm/sec with a resolution of 0.2 nm. The average roughness (Ra) was determined in each case. XRD phase analysis XRD studies were conducted on all the samples to confirm the phases present in the material using BRUKER G8 powder XRD, Cu-Kα source, λ= 1.54 Å, Bragg angle, 2θ from (30-100o), step size 0.01 and step/scan 0.5 at an accelerating voltage of 40kV and 30 A. Residual stress measurements Residual stresses measurements were conducted on Xstress G2R (High-resolution XRD) to find the surface residual stresses present in the material. By using CrKα source, applied voltage 27 kV, a current of 70 mA, λ= 2.28, Bragg angle 147.6o was kept constant for all the samples, (hkl 311) peak was considered for diffraction with a step size of 0.1o [5]. Collimeter diameter 4 mm and exposure time of 20 s was used. The multiexposure side inclination and fixed (χ) chi method was adopted, at 0o and 90o for each measurement. 2D stress state was assumed sinχ technique was applied for residual stress analysis. Residual Stresses 2018 – ECRS-10 Materials Research Forum LLC Materials Research Proceedings 6 (2018) 139-144 doi: http://dx.doi.org/10.21741/9781945291890-22 141 Determination of SCC resistance The SCC susceptibility of AISI 304L SS and AISI 321 SS in buffed and un-buffed condition was determined using ASTM G 36 [6] in surface worked and welded conditions respectively. SCC test was conducted for 3 h and 9 h in AISI 304L SS samples and 5 h and 10 h in AISI 321 SS samples. As per ASTM G 36, 600 g of magnesium chloride hexahydrate (MgCl2.6H2O) was melted and test temperature was maintained at 155± 1o C throughout the test. Care has been taken to prevent vapor losses Results and Discussion Surface roughness measurements: Surface roughness values (Ra) for different surface working conditions have been tabulated in Table1. In both 304L SS and 321 SS surface roughness in buffed conditions was much less as compared to un-buffed conditions. Higher the surface roughness, greater the tendency to form localized pockets of high chloride concentration and early initiation of SCC. Table 1: Average surface roughness values of 304L SS and 321 SS in un-buffed and buffed conditions Material Material conditions Surface roughness (Ra) in μm 304L SS Milled 2.1 ± 0.15 Turned 4.3 ± 0.30 Ground 0.6 ± 0.04 Milled + Buffed 0.13 ± 0.06 Turned + Buffed 0.11± 0.04 Ground + Buffed 0.08 ± 0.02 321 SS Spot welded 1.74±0.24 Spot welded +Buffed 0.94±0.12 XRD studies Figure 2 shows the XRD spectra of 304L SS under different surface working conditions. Solution annealed sample showed austenitic (γ) phase and other surface worked samples showed strain-induced martensite (αʹ) phase due to the metastable nature of austenitic stainless steel. Fig. 2: Shows the XRD spectra of 304L SS in different surface working conditions. Figure 3 shows the XRD spectra of the spot welded region of 321 SS in both buffed and unbuffed condition. Characteristic austenitic peaks were observed in both the cases together with the presence of stress induced martensite. The number of peaks for stress induced martensite was Residual Stresses 2018 – ECRS-10 Materials Research Forum LLC Materials Research Proceedings 6 (2018) 139-144 doi: http://dx.doi.org/10.21741/9781945291890-22 142 much higher for buffed condition. The Ti2N present in the surface layers of 321SS gets removed on buffing. Optical microstructures support the observation. Fig. 3: Shows the XRD spectra of 321 SS in spot welded and spot welded followed by a buffed condition Residual stress measurements Table 2 gives the residual stresses of 304L SS and 321 SS under different conditions. The measurements were taken in longitudinal and transverse direction. The result shows that for 304L SS in milled, turned and ground conditions have high magnitude of tensile residual stresses. A similar result has been reported in earlier studies performed by some of the authors [7-8]. However, compressive residual stresses were found to be present on the surfaces in buffed condition. Similarly, the HAZ and the fusion zone of spot welds of 321 SS were found to have tensile residual stresses in un-buffed condition and compressive residual stresses in buffed condition. Table 2: residual stress values of 304L SS and 321 SS in different conditions 304L SS conditions 0o (MPa) 90o (MPa) Milled 740±86 639±71 Turned 397±82 69±85 Ground 192±40 15±39 Milled + Buffed -386±21 -378±16 Turned + Buffed -523±17 -504±26 Ground + Buffed -481±22 -409±16 321 SS conditions 0o (MPa) 90o (MPa) Base material -239±23 -306±57 Spot + buffed -351±17 -433±24 Determination of susceptibility to stress corrosion cracking (SCC) Figure 4 and Figure 5 shows the FESEM surface micrographs of 304L SS in a) milled b) turned c) ground before and after buffing after exposure to 3 h and 9 h of SCC test respectively. The results showed that the samples in milled ground and turned condition were highly susceptible to SCC, whereas the samples in buffed condition after 3 h and 9 h of exposure. Pit initiation was observed in some cases after 9 h of exposure due to the preferential dissolution of martensite on the surface. Residual Stresses 2018 – ECRS-10 Materials Research Forum LLC Materials Research Proceedings 6 (2018) 139-144 doi: http://dx.doi.org/10.21741/9781945291890-22
Introduction
Austenitic stainless steels have good corrosion resistance and mechanical properties.However, these become highly susceptible to SCC when subjected to different surface finishing and welding operations [1][2][3].Laser peening and shot peening are generally used for enhancing SCC resistance, but these processes induce plastic strain in the material and are not economic.Hence a simple and economic route is essential in preventing SCC.We have shown in our previous report that simple surface buffing can be used to enhance the SCC resistance of austenitic stainless steels [4].The present study substantiates the effectiveness of buffing in preventing the initiation of Cl -induced SCC of austenitic stainless steel in presence of chloride ions.304L SS in milled, turned and ground conditions and 321 SS in spot welded condition when subjected to surface buffing operation.
Materials and methods
In the present study two steels namely, AISI 304L SS (C 0.03, Cr 18, Ni 8, Mn 1.6, P 0.04, Si 0.4,S 0.03, balance Fe) wt %, and AISI 321 SS welds (C 0.024, Cr 17.41, Ni 9.14, Mn 1.64, Ti 0.23, Mo 0.37, P 0.04, Si 0.4, S 0.03) wt% were used.AISI 304L SS samples were cut into a dimension of 100 mm X 25 mm X 5 mm plates and they were solution annealed in order to remove the internal stresses present in the material.AISI 304 L SS was subjected to three different surface working conditions namely a) milling b) turning and c) grinding operations at a feed rate of 0.1mm/rev to remove 0.5 mm from the surface [7].Subsequently, the samples were buffed at 3600 rpm using and 50µm was removed from the surface.In another study, AISI 321 SS tubes having 72 mm diameter, 3 mm thickness and length of 3.6 m having spot welds on the surface were subjected to buffing and 50µm was removed.Both the samples were tested for SCC susceptibility in buffed and unbuffed condition.During buffing compressive forces are applied on the sample substrate as shown in the schematic given below (Figure 1).The buffing wheel is rotated at a set speed and is rastered on the surface of the workpiece imparting compressive stresses to the entire surface of the workpiece.No Ti 2 N particles were present in the near surface region of buffed samples in each case.
Surface roughness measurements
Surface roughness measurements were done on all samples using a surface profilometer (contact mode) with a scan length of 1µm at a scan speed of 0.01 mm/sec with a minimum resolution step of 1 nm.321 SS ring samples were measured using an optical surface profilometer (non contact mode) with a scan length of 1mm at a scan speed of 47µm/sec with a resolution of 0.2 nm.The average roughness (Ra) was determined in each case.
XRD phase analysis
XRD studies were conducted on all the samples to confirm the phases present in the material using BRUKER G8 powder XRD, Cu-Kα source, λ= 1.54 Å, Bragg angle, 2θ from (30-100º), step size 0.01 and step/scan 0.5 at an accelerating voltage of 40kV and 30 A.
Residual stress measurements
Residual stresses measurements were conducted on X-stress G2R (High-resolution XRD) to find the surface residual stresses present in the material.By using Cr-Kα source, applied voltage 27 kV, a current of 70 mA, λ= 2.28, Bragg angle 147.6º was kept constant for all the samples, (hkl 311) peak was considered for diffraction with a step size of 0.1º [5].Collimeter diameter 4 mm and exposure time of 20 s was used.The multiexposure side inclination and fixed (χ) chi method was adopted, at 0º and 90º for each measurement.2D stress state was assumed sin 2 χ technique was applied for residual stress analysis.
Determination of SCC resistance
The SCC susceptibility of AISI 304L SS and AISI 321 SS in buffed and un-buffed condition was determined using ASTM G 36 [6] in surface worked and welded conditions respectively.SCC test was conducted for 3 h and 9 h in AISI 304L SS samples and 5 h and 10 h in AISI 321 SS samples.As per ASTM G 36, 600 g of magnesium chloride hexahydrate (MgCl 2 .6H 2 O) was melted and test temperature was maintained at 155± 1º C throughout the test.Care has been taken to prevent vapor losses
Surface roughness measurements:
Surface roughness values (Ra) for different surface working conditions have been tabulated in Table1.In both 304L SS and 321 SS surface roughness in buffed conditions was much less as compared to un-buffed conditions.Higher the surface roughness, greater the tendency to form localized pockets of high chloride concentration and early initiation of SCC.
XRD studies
Figure 2 shows the XRD spectra of 304L SS under different surface working conditions.Solution annealed sample showed austenitic (γ) phase and other surface worked samples showed strain-induced martensite (αʹ) phase due to the metastable nature of austenitic stainless steel.
Fig. 2: Shows the XRD spectra of 304L SS in different surface working conditions.
Figure 3 shows the XRD spectra of the spot welded region of 321 SS in both buffed and unbuffed condition.Characteristic austenitic peaks were observed in both the cases together with the presence of stress induced martensite.The number of peaks for stress induced martensite was much higher for buffed condition.The Ti 2 N present in the surface layers of 321SS gets removed on buffing.Optical microstructures support the observation.
Fig. 3: Shows the XRD spectra of 321 SS in spot welded and spot welded followed by a buffed condition
Residual stress measurements
Table 2 gives the residual stresses of 304L SS and 321 SS under different conditions.The measurements were taken in longitudinal and transverse direction.The result shows that for 304L SS in milled, turned and ground conditions have high magnitude of tensile residual stresses.A similar result has been reported in earlier studies performed by some of the authors [7][8].However, compressive residual stresses were found to be present on the surfaces in buffed condition.Similarly, the HAZ and the fusion zone of spot welds of 321 SS were found to have tensile residual stresses in un-buffed condition and compressive residual stresses in buffed condition.
Determination of susceptibility to stress corrosion cracking (SCC)
Figure 4 and Figure 5 shows the FESEM surface micrographs of 304L SS in a) milled b) turned c) ground before and after buffing after exposure to 3 h and 9 h of SCC test respectively.The results showed that the samples in milled ground and turned condition were highly susceptible to SCC, whereas the samples in buffed condition after 3 h and 9 h of exposure.Pit initiation was observed in some cases after 9 h of exposure due to the preferential dissolution of martensite on the surface.Figure 6 shows the FESEM micrographs of 321 SS spot welded region after exposure to boiling MgCl 2 for a) 5 h and b) 10 h. Figure 6(c-d) shows spot welded and buffed samples after 5 h and 10 h test respectively.High densities of cracks were observed in un-buffed condition whereas no cracking was observed in buffed samples.
Fig 1 :
Fig 1: a) Schematic of the buffing operation b) cross sectional micrograph of 321 SS in buffed condition and c) cross sectional micrograph of 321 SS in un-buffed condition.
Fig. 4 :
Fig. 4: Shows surface micrographs after 3 h SCC test in 304L SS a) milled b) turned c) ground d) milled and buffed e) turned and buffed f) ground and buffed.
Fig. 5 :
Fig. 5: Shows surface micrographs after 9 h SCC test in 304L SS a) milled b) turned c) ground d) milled and buffed e) turned and buffed f) ground and buffed conditions.
Fig. 6 :
Fig. 6: Shows surface micrographs of 321 SS in spot weld condition after 5 h and 10 h SCC test respectively (a-b) before buffing and (c-d) after buffed condition.
Table 1 :
Average surface roughness values of 304L SS and 321 SS in un-buffed and buffed conditions
Table 2 :
residual stress values of 304L SS and 321 SS in different conditions | 4,134.2 | 2018-09-11T00:00:00.000 | [
"Materials Science",
"Engineering"
] |
Large-Eddy Simulation of a Hydrocyclone with an Air Core Using Two-Fluid and Volume-of-Fluid Models
: Large-eddy simulations have been conducted for two-phase flow (water and air) in a hydrocyclone using Two-Fluid (Euler–Euler) and Volume-of-Fluid (VOF) models. Subgrid stresses are modeled using a dynamic eddy–viscosity model, and results are compared to those using the Smagorinsky model. The effects of grid resolutions on the mean flow and turbulence statistics have been thoroughly investigated. Five block-structured grids of 0.72, 1.47, 2.4, 3.81, and 7.38 million elements have been used for the simulations of Hsieh’s 75 mm hydrocyclone Mean velocity profiles and normal Reynolds stresses have been compared with experimental data. Results of the two-fluid model are in good agreement with those of the VOF model. A fine mesh in the axial and radial directions is necessary for capturing the turbulent vortical structure. Turbulence structures in the hydrocyclone are dominated by helical vortices around the air core. Energy spectra are analyzed at different points in the hydrocyclone, and regions of low turbulent kinetic energy are identified and attributed to stabilizing effects of the swirling velocity component.
Introduction
Hydrocyclones are used in many industries for separation of different phases. In the mineral-processing industry, classifying hydrocyclones are used to classify ores after grinding based on their particle sizes ( [1][2][3]). A typical geometry of classifying hydrocyclone consists of a cylindrical section followed by a conical section with two outlets that are open to the atmosphere. The top outlet is called the overflow opening and the bottom one is called the underflow opening. Slurry of water and solid particles is forced tangentially into the hydrocyclone top, and it swirls as it goes downward. The swirling flow field generates a column vortex with pressure decreasing radially towards the center. The strength of the vortex depends on the rate of input angular momentum about hydrocyclone centerline. The radial pressure gradient affects the motion of particles based on their specific gravity and/or sizes. Because of the low pressure created by the swirling layer of slurry, air enters through the underflow opening and forms an air core around the hydrocyclone centerline. The structure and motion of the air core depends on the operating conditions of the hydrocyclone ( [4]). Hararah et al. [5] found that the distribution of tangential and axial velocity of air inside the core depends on the pressure drop and geometry of hydrocyclone. Devulapalli [6] experiments show that the inlet flow rate has a direct effect on the air core diameter.
Laboratory experiments and CFD techniques can be used for performance evaluation and improvements of hydrocyclones. Laboratory experiments are important to determine mean velocity profiles and turbulence statistics. However, such measurements are expensive and time consuming. The high cost of experiments limits their effectiveness to optimize the geometry of hydrocyclones. CFD techniques are viable tools to predict the flow structure and turbulence in hydrocyclones. Current computing capabilities enable conducting high fidelity numerical simulations such as Large-Eddy Simulation (LES) of two-phase flow in hydrocyclones at an affordable cost ( [7][8][9][10][11][12]). However, validation of numerical models is an important step to assess their accuracy. In 1988, Hsieh [13] conducted both laboratory experiments and axisymmetric numerical simulations of flow in a hydrocyclone. His pioneering work shows good comparison between experimental velocity profiles and those from numerical simulations. However, the role of turbulence is essential in the flow behavior inside the swirling layer and air core. The experimental velocity profiles in Hsieh [13] show flow asymmetry. Therefore, 3D time-accurate simulations are essential to capture the spatial variations of the mean flow and instantaneous turbulence structures as those affect particles dispersion.
Structural stability of flow in hydrocyclones is achieved by feeding appropriate rate of angular momentum that drives the tangential and axial flow. The wall shear layers in the feeding duct become free shear layers upon entering the hydrocyclone volute. Instabilities on those layers produce new turbulence structures as the swirling flow go downward. The axial velocity distribution on the cross-section of hydrocyclone has a complex structure. Downward flow prevails near the outer cylindrical/conical walls whereas upward flow prevails near the interface with the air core. Local instabilities of such a complex shear flow produce turbulent vortical structures that continuously interact with a swirling flow. Because of the high tangential velocity of the swirling flow, the turbulent fluctuations in the tangential direction are not the same as those in the radial and axial directions indicating anisotropic large-scale turbulence. The nonlinear interaction between the mean flow and turbulent fluctuations affects the distribution of mean velocity components and hence the performance of a hydrocyclone. Therefore, it is important to use accurate turbulence models in numerical simulations. The first numerical approach to simulate turbulent flows is Unsteady Reynolds-Averaged Navier-Stokes (U-RANS) where ensemble-averaged equations are solved for mean flow and Reynolds stresses are modeled. Turbulence models of varying fidelity have been applied. Eddy-viscosity models such as 2-equation models k − and k − ω are well known. These models assume isotropic relations between local Reynolds stresses and the mean rate-of-strain tensor. Because of the flow structure in hydrocyclone, this isotropic condition in turbulent fluctuations is invalid. A more sophisticated turbulence model such as Reynolds Stress Transport Model (RSM) can be used to overcome the shortcomings in 2-equations turbulence models ( [14][15][16][17][18]). In RSM, 6 transport equations are solved for Reynolds stress components and this shall increase computations time tremendously. The second approach is Large-Eddy simulation (LES) where the Navier-Stokes equations are space-filtered and solved on a fine grid to resolve large scales of turbulence [19]. The filter width is directly determined by the size of grid element. The effects of small scales of turbulence on the evolution and motion of large eddies are modeled by an eddy-viscosity subgrid stress model. Direct Numerical Simulations (DNS) approach is more accurate than LES where all the turbulent eddies down to Kolmogorov scale are resolved and turbulence modeling is not required. However, for industrial applications it is not feasible because of very large computational resources and time. Delgadillo and Rajamani ([14,15]), and Brennan ( [16,17]) conducted two-phase LES and U-RANS simulations for Hsieh's hydrocyclone [13]. They modeled the two-phase flow of air and water using Volume-Of-Fluid (VOF) and mixture models. Vakamalla and Mangadoddy [20] conducted a comprehensive study of the role of turbulence on mean velocity, turbulence statistics, and classification performance of different sizes of hydrocyclones. They tested different models in conjunction with VOF and mixture models. They demonstrated the superiority of LES approach over RSM for high turbulence hydrocyclones.
The mixture and VOF models are considered to be One-Fluid models where one set of momentum and continuity equations are solved in space and time. The air/water interface is reconstructed and tracked in VOF method by solving continuity equation for volume fraction of air phase. The experiments conducted by Hararah et al. [5] show that the air core destabilizes in the overflow pipe because of decrease in centrifugal forces and produces a mixture of air and water droplet (Aero-Suspension). The use of VOF approach is expensive in terms of computational resources and time to resolve this breakup and coalescence of water droplets and/or air bubbles in overflow pipe. The mixture model is computationally less demanding and simple where one set of momentum and continuity equations are solved for a homogeneous mixture of fluids phases. The slip velocity between dispersed phase and continuous phase is estimated by balancing drag and body force ( [21]). The mixture model can be applied for a wide variety of turbulent multiphase flows where the response time is very small (i.e., very small inertia) relative to time scale of corresponding turbulent scales. In many applications of particulate and/or bubbly flows, the spatial distribution of dispersed phases cannot be accurately captured by a mixture model [22]. The interfacial momentum transfer in terms of drag, added mass and history forces between dispersed and continuous phases is the driving mechanism to suspend solid particles and/or distribute air bubbles in the liquid phase. Therefore, accurate modeling of interfacial phenomena is important to account for the effect of turbulent fluctuations on the motion of dispersed phases. The mixture multiphase model does not accurately model these interfacial forces. The Two-Fluid model is highly recommended as discussed by Ishii and Mishima [23]. In the Two-Fluid model we solve a set of averaged momentum and continuity equations for each of the phases. The motions of the phases are coupled by modeling interfacial forces such as drag, added mass, lift and history force. This approach of multiphase flows is more expensive than mixture model, but it is more accurate for multiphase problems that have large spatial variations in void fractions of dispersed phases. Moreover, the two-fluid model is the natural choice for air-sparged hydrocyclones ( [24][25][26]) where there are a huge number of bubbles in addition to an air core or froth layer and resolving the gas-liquid interfaces is not computationally feasible by VOF.
This paper aims at validating the two-fluid model (Euler-Euler) for a hydrocyclone with an air core using LES by comparison with experimental data and simulation results using VOF model. Furthermore, a fine mesh of about 7.4 million elements for the 75 mm hydrocyclone designed by Hsieh [13] enables in-depth analysis of large-scale turbulence structures, statistics, and spectra. The influence of the subgrid-scale stresses model on the simulations results is evaluated. An eddy-viscosity model is used where the coefficient is determined by the dynamic model and the results are compared with predictions using Smagorinsky model. Validation of the LES in conjunction with the two-fluid model is important for future LES modeling of three-phase (gas-liquid-solid) in hydrocyclones and fluidized bed reactors [27].
Two-Fluid Model
The motion of the two continuous phases is described by the unsteady filtered Navier-Stokes equations: Continuity equation: Momentum equation: where V i is the filtered velocity of phase i, P is the filtered pressure field which is shared by all phases, ρ i and α i are density and local volume fraction of phase i, respectively. g is the gravitational acceleration. The index (i = 1) denotes water phase and (i = 2) denotes air phase. The turbulent eddy viscosity ν t in the above equation is calculated as, The coefficient C m is known as the subgrid model constant. This coefficient is determined locally in the computational domain using Germano dynamic model [28]. For Smagorinsky model, the coefficient C m is replaced by a constant C 2 s . M i is the interfacial force that acts on phase (i) due to the presence of other phases. In the present simulations, momentum exchanges between the two phases due to drag and buoyancy on bubbles are the only mechanisms that couple the motion of the two phases. For the two-fluid model (Euler-Euler), a bubble diameter of 0.1 mm is assumed, while no such an assumption is needed with the VOF. Schiller-Naumann drag model [29] has been used to estimate drag coefficient of air bubbles.
Volume-of-Fluid Model
In the volume-of-fluid model (VOF) of two-phase flow, a single momentum equation is solved at every cell where density and viscosity are volume-weighted average of the density and viscosity of the two phases. Water is declared as the primary phase, and the air-water interface is tracked by solving the volume fraction equation of the air phase using a high-resolution scheme. Volume-weighted is used for smoothing the air-water interface. Brackbill et al. [30] Continuum Surface Force (CSF) model is used for the effects of surface tension. A constant surface tension of 0.07 N/m is assumed.
Computational Grid and Boundary Conditions
We used ICEMCFD software to construct the geometry of the hydrocyclone and generate Block-Structured O-type grid. The dimensions of the hydrocyclone are given by Hsieh [13] as shown in Figure 1a. The x and z axes of a right-handed coordinate system xyz are depicted in Figure 1b, the y axis is vertically upward and coincides with the hydrocyclone centerline. Axial locations are measured relative to the roof of the hydrocyclone such that a horizontal plane that is 60 mm below the roof is called y = 60 mm. Figure 1b also shows the definitions of the radial v r and tangential v t velocity components. The axial velocity v a is in the y−direction. The inlet pipe is circular with diameter of 25 mm. We modified the cross-section of inlet pipe from circular at inlet to a square of equal area at the intersection with the cylindrical part. This modification of inlet pipe has been done to facilitate the construction of the block-structured mesh. The cylindrical part of hydrocyclone has inner diameter of 75 mm and its height equals 75 mm. The conical part of hydrocyclone extends vertically below the cylindrical part a distance of 186 mm. The diameter of this conical part decreases from 75 mm at intersection with cylindrical part and reaches 12.5 mm at intersection with spigot pipe. The cone angle is approximately 20 • . The spigot pipe extends downward a distance of 25 mm. The overflow pipe (Vortex finder) has a diameter of 25 mm and extends a distance of 50 mm inside the cylindrical part of the hydrocyclone. This overflow pipe takes the shape of an elbow as shown in Figure 1a and maintains constant diameter. The exit boundary of the overflow pipe (on the right) is located at the same level of the roof of the hydrocyclone. The interior of hydrocyclone is divided into 5 subdomains to generate appropriate grid-block in each subdomain and optimize the size of each hexagonal grid. As shown in Figure 2, the grid in the core of hydrocyclone and overflow pipe has been designed to resolve the high gradient of tangential and axial velocity components near the expected air core. We constructed five different grid sizes for the same set of blocks to study the effects of grid size on the quality of the results. Table 1 provides a summary of the grid sizes in our simulations. Resolving small scales vortical structures in the boundary layer requires a very fine mesh near the wall such that the first grid point off the wall is at y + ≈ 1. To alleviate this stringent requirement, a near-wall model is employed, which is the default set up in ANSYS CFX. A wall function is used if the grid is too coarse to resolve the viscous sub-layer (say y + > 5). However, if the near-wall grid is fine (y + < 2), the near-wall treatment provides a smooth transition from wall function modeling to grid-resolved viscous flow. For the three coarse meshes, where the surface average y + is around 18, the near-wall flow is modeled. For the fine mesh f2, the normal distance from the first grid point to the wall is 0.01 mm, and the surface average y + is 2.2. We expect the viscous sub-layer to be resolved for the fine mesh. The step sizes in the directions parallel to the wall are also estimated. In the axial direction, the step size in wall-units is 110, and in the circumferential direction it is 220. These are adequate mesh steps since the vortical structures in hydrocyclones are elongated in the circumferential direction.
Boundary Conditions and Numerical Method
The LES Euler-Euler model in CFX-ANSYS (V17.2) has been used to simulate the flow of air and water where the two phases are treated as interpenetrating continua. However, for the purpose of momentum exchange between the two phases, the water is declared as the continuous phase and air as dispersed phase. The boundary conditions are setup at inlet as velocity-inlet with a uniform value of 2.28 m/s that achieves mass flow rate of 1.117 kg/s of water. The water volume fraction at inlet is 1. The opening boundary condition has been selected for the exit boundaries of the overflow elbow and underflow pipe. The pressure at these boundaries has been specified as zero gauge pressure (i.e., atmospheric pressure) and the back flow has been selected to be 100% air. No-slip condition with zero velocities is applied on hydrocyclone walls.
We use bounded central difference to discretize the continuity and momentum equations of air and water phases. The bounded central-difference schemes are second-order accurate and can damp dispersive errors in central-difference schemes. A second-order implicit Euler method is used for time integration. The value of time step is important to stabilize calculations of LES models. Most of the results presented here are obtained for a time step ∆t = 10 −4 s. Some of the results as indicated are obtained with a smaller time step ∆t = 5 × 10 −5 s. The residence time in Hsieh's hydrocyclone is about 1 second. The first simulation uses coarse grid c1 (2.4 million elements) and is initialized with 100% water. It is executed for at least 3 s until the air core is established and the flow reached statistically steady state. The flow on this grid is then interpolated to initialize simulations on other grids c2, c3 and f1. Converged solution on f1 is interpolated to initialize solution on the finest grid f2. The simulation on each grid after interpolation is continued for at least one second before mean flow and turbulence statistics are computed over an additional one second.
Inlet Pressure, Split Ratio, and Air Core
The water volume flow rate at the inlet is prescribed and is steady in time. However, the inlet pressure and water volume flow rates to the overflow and underflow are predicted in the course of simulations and they fluctuate in time. Their time-averaged values are given in Table 1. %Overflow is the water volume flow rate exiting from the top opening as a percentage of water volume flow rate at the inlet of hydrocyclone. Different mesh resolutions show good comparison with experimental data by Hsieh [13] for these overall measures of the hydrocyclone performance. Although the coarsest mesh shows lower percentage errors than finer meshes, the mean velocity profiles and turbulence statistics for such a coarse mesh are much less accurate. Air is drawn from the underflow opening as a result of the low pressure on the axis of hydrocyclone. The air core is nearly a cylindrical surface that extends axially from the underflow opening to the overflow exit. The instantaneous air-core surface fluctuates radially and twists by its interaction with the upward flowing swirling turbulent water flow. The time-averaged of air volume fraction is used as an indicator of the air core. Figure 3a,b) show mean air volume fraction profiles on a horizontal line at 60 and 120 mm from the roof of hydrocyclone. For the finest mesh f 2, the air volume fraction is nearly 1 over a diameter of 10 mm and drops to 0 beyond a diameter of 15 mm. Hsieh [13] reported an air-core diameter of 10 mm. Because of the fluctuating position of the air-water interface in addition to numerical diffusion, a sharp interface is difficult to predict. The figures show that as the mesh is refined the thickness of air-water mixture decreases, it is the largest for the coarsest mesh c3, but the profiles are nearly identical for the fine meshes f 1 and f 2.
Mean Water Superficial Axial Velocity
The axial flow in a hydrocyclone has a complex structure. The near-wall flow is downward, and it appears as a swirling wall jet, while on the inside the flow is a swirling upward flow bounded by a free surface (air core). Comparisons between experimental data by Hsieh [13] [(0, 180 • ) data] and LES results for the mean superficial water axial velocity are shown in Figure 4a,b at y = 60 and y = 120 mm, respectively. The maximum upward axial velocity increases as the grid is refined, and convergence is established for the fine meshes f 2 and f 1 where the profiles are almost identical near the air core. The experimental data indicates a smaller core diameter with maximum axial velocity inward. The data also show a plateau in the axial velocity at y = 120 mm which cannot be predicted by grid refinements. Further comparisons with experimental data [Hsieh [13] (90 • , 270 • ) data] are shown in Figure 5a,b. The overall agreement between LES and experimental data is fair.
Mean Water Superficial Tangential Velocity
Comparisons between experimental data [Hsieh [13] (0, 180 • ) data] and LES results for the mean superficial water tangential velocity are shown in Figure 6a,b at y = 60 and y = 120 mm, respectively. The maximum tangential velocity increases with grid refinement, but it is consistently higher than the experimental data. It is a coincidence that the maximum tangential velocity of the coarsest grid c3 agrees with the experimental data, and as a results the predicted inlet pressure shown in Table 1 is also in agreement with the measured pressure. Except for the coarsest grid c3, the tangential velocity profile is less sensitive to grid resolution. All grids approach the same profile. The persistent difference between the experimental data and LES cannot be attributed to grid effects. No attempt is made here to adjust the exit pressures or other operating conditions to reduce that shift in the tangential velocity. The effects of the turbulence model on the tangential velocity profile is presented in Section 6.
Mean Water Superficial Radial Velocity
In hydrocyclones, the radial velocity is of a much smaller magnitude than the axial and tangential components, but it may have important effects on particles transport. The present simulations show a complex distribution of the radial velocity component. Figure 7a,b show mean water superficial velocity contours in planes y = 60 and 120 mm, respectively. The maximum magnitude is less than 0.4 m/s. On a diagonal, the velocity is radially outward (positive) on one side of the air core, and radially inward (negative) on the opposite side of the air core. Such a distribution also rotates with the swirling flow. Isosurfaces of mean water superficial radial velocity, v r = ±0.2 m/s are shown in Figure 8a, the spiral structure causes bending of the air core as shown in Figure 8b which shows velocity vectors colored by the radial velocity in the xy−plane. These figures also show a streamwise vortex initiated at the inlet to the cylindrical part of hydrocyclone.
Resolved Normal Reynolds Stresses and Energy Spectra
Investigation of the radial distribution of circulation on circles centered on the axis shows that the circulation increases with radius except in a very thin layer near the solid wall where the circulation drops to zero. Such a distribution is centrifugally stable [31]. Sreedhar and Ragab [32] conducted LES for an Oseen vortex whose circulation increases with radius and showed that initial random disturbances are damped by the centrifugal field. Thus, the swirl in the boundary-layer flow and the shear of the axial flow must be sources of turbulence production in a hydrocyclone. After the flow reached statistically steady state, time-averaging over at least one second is performed for computing turbulence statistics such as Reynolds stresses and turbulent kinetic energy. Normal Reynolds stresses of the resolved large-scale turbulence are presented as the root-mean-square (RMS) of the water superficial velocity fluctuations. Hsieh's [13] experimental data are provided on a radial line. For the sake of comparison, LES data are averaged on four radial lines along the x and z− axes. Figure 9a,b show the axial component v a at elevations y = 60 and 120 mm, respectively. The LES data show consistent convergence to the experimental data with mesh refinement. The two fine meshes f 1 and f 2 are in satisfactory agreement with experimental data, but the coarser meshes over-predict the axial fluctuations. A fine mesh in the axial direction is necessary to capture the axial turbulence length scales. In the cylindrical region of hydrocyclone at y = 60 mm, the near-wall region is well predicted by all grids, but the significant difference happens near the air core. In the conical region at y = 120 mm, the finest mesh f 1 shows the best agreement with experimental data; however, further refinement might be of little effects. Figures 10a,b show the tangential component v t at elevations y = 60 and 120 mm, respectively. LES data for meshes f 1 and f 2 are in very good agreement with Hsieh's [13] data. The results for the coarse mesh c1 are also in fair agreement with experimental data. The tangential fluctuations show less sensitivity to grid resolution than the axial fluctuations. It will be shown later that the turbulent eddies are elongated in the circumferential direction relative to the axial and radial directions which requires much finer mesh size in these directions than the circumferential direction. Turbulent kinetic energy of resolved scales in a vertical plane (x = 0), which is normal to the inlet stream, are shown in Figure 12a. High turbulent kinetic energy is generated just below the volute and continues in the cylindrical section. Contours in a horizontal plane y = 24 mm, which is very close to the bottom of inlet volute, depicted in Figure 13a show the high levels of turbulent kinetic energy generated as the flow enters the hydrocyclone.
However, as turbulence becomes subjected to the stabilizing swirling flow, turbulent kinetic energy is reduced at the end of the cylindrical section and beginning of the conical section as shown by the annular region of low turbulent kinetic energy in Figure 12a. The reduced turbulent kinetic energy is also shown in Figure 13b by contours in a horizontal plane y = 82 mm below the roof of hydrocyclone. Localized regions of low turbulent kinetic energy can be seen also in Cui et al. [33]; however, their locations are affected by the vortex finder diameter. Figure 14a and the corresponding turbulent energy spectra (computed using the Welch method [34]) are shown in Figure 14b. The spectra for the tangential and radial velocity components show a peak at a frequency of approximately 100 Hz. Such a peak is absent in the spectrum of the axial velocity. Thus, the air core is subjected to larger fluctuations in horizontal planes. The stabilizing effect of the swirling flow is further demonstrated by comparing the tangential velocity fluctuations at points 1 and 2 as shown in Figure 14a and their energy spectra shown in Figure 14b. The flow is almost locally re-laminarized in the low turbulent kinetic annular region. Using LES, Ragab and Sreedhar [35] found that a centrifugally stable vortex flow can quench the large-scale turbulent structures generated by instabilities of an axial velocity with inflection point and renders the flow to a laminar state. Except near the wall, the mean tangential velocity profile in the current hydrocyclone is centrifugally stable. In the conical region, the turbulent kinetic energy is maximum near the air-water interface and near the wall. The energy spectra at two points in the high kinetic energy regions are shown in Figure 15a for point 3 and in Figure 15b To visualize turbulent large-scale structures in the hydrocyclone, we define the instantaneous fluctuating component of tangential vorticity ω θ Isosurfaces of ω θ = ±500 s −1 are depicted in Figure 12b, the high values near the walls are filtered for clarity of the interior region. The length scales in the axial and radial directions are much smaller than the circumferential length scale. In the annular region of low turbulent kinetic energy those high levels of vorticity are suppressed, but they persist around the air core. The generation of such vortical structures must be attributed to the shear associated with complex form of the axial velocity profile.
Resolved Shear Reynolds Stresses
The three components of resolved shear Reynolds stresses in cylindrical coordinates (< v r v a >, < v r v t >, < v a v t >) are shown in Figures 16-18 at the two elevations y = 60 and 120 mm. Results for different grids demonstrate the importance of mesh refinement for capturing these Reynolds stresses. Satisfactory grid convergence is established for the fine meshes f 1 and f 2. No experimental data are available for validation. The near-wall and the air-core regions are regions of intense shear stresses. All three components acquire stronger magnitudes as the flow go down the conical section.
Comparison of Volume-of-Fluid and Two-Fluid Models
It is important to validate the two-fluid (Euler-Euler) model used in the present paper by comparing the results with VOF model, the later model is commonly used to resolve air core-water interface in hydrocyclones of the present type. However, the two-fluid model is the natural choice for air-sparged hydrocyclones where there are a huge number of bubbles and resolving the gas-liquid interfaces is not computationally feasible by VOF, [26].
The finest mesh f2 is used for both models. Comparisons of mean water superficial velocity profiles are shown in Figure 19a,b for the axial velocity and in Figure 20a,b for the tangential velocity. Velocity profiles predicted by the two models are in good agreement. The maximum axial velocity near the air-water interface is somewhat lower in VOF model than in the two-fluid model. The persistent difference between experimental data and simulation in the tangential velocity could not be reduced by the VOF model. Large-eddy simulations by Durango et al. [36] also show the shift in tangential velocity. Axial and tangential turbulence velocity fluctuations (RMS) profiles are also compared in Figures 21a,b and 22a,b, respectively. The VOF results show lower axial velocity fluctuations in the conical section at section y = 120 mm from the hydrocyclone top. Overall, excellent agreement between the two models is observed.
Effects of Subgrid-Scale Turbulence Model
It is unsettling to see the persistent shift in the mean tangential velocity between experimental data and LES results. This shift is predicted on all five meshes and for the VOF and two-fluid models. In the present study, an eddy viscosity is used, and the dynamic model is used to compute the model coefficient. The dynamic subgrid stress turbulence model is deemed more adequate than Smagorinsky model for turbulent flows that may be subjected to re-laminarizing mechanisms. Using Smagorinsky model with model constant C s = 0.12, simulations are repeated for the two-fluid model on fine grid f1. Figure 23a,b show comparison between the dynamic model and Smagorinsky model results for the mean axial velocity. The two models are in good agreement. The results for the mean tangential profiles are shown in Figures 24a,b. Mean tangential velocity profiles predicted by Smagorinsky model, with adjusted model coefficient of 0.12, are in good agreement with the experimental data. However, the maximum velocity is under-predicted and hence the pressure drop between inlet and outlet is also under-predicted. Simulations using a smaller value for the model coefficient brought tangential velocity profiles of Smagorinsky model closer to the dynamic model profiles. Feiz et ai. [37] conducted LES for turbulent flow in a rotating pipe and compared their results with direct numerical simulations by Orlandi and Fatica [38]. They concluded that the dynamic model is more accurate than Smagorinsky model. The centrifugal field in the rotating pipe problem can have stabilizing effects as the swirling flow of the hydrocyclone.
Conclusions
Turbulent two-phase flow in a classifying hydrocyclone has been investigated using LES. The two-phase aspect of the flow is modeled by the two-fluid (Euler-Euler) and volume-of-fluid (VOF) models. Mean velocity and turbulence statistics show good comparison between the two-phase models. Thus, the present study validates the two-fluid model against the commonly used VOF for this type of hydrocyclone. The two-fluid model can be used for simulations of classifying as well as air-sparged hydrocyclones for which VOF is impracticable. The present study encourages application of the multi-fluid model to three-phase flow (liquid, solid particles, gas) in hydrocyclones. Simulations on fine meshes 5.86 and 7.4 million elements enable capturing helical vortices as the dominant large-scale turbulent structures in the hydrocyclone. The turbulence structure in hydrocyclones is complex because of the competing effects of a stabilizing swirling flow and turbulence producing shear associated with the axial flow. Except near the wall, the free-vortex tangential velocity decays slower than the inverse of the radial distance measured from the axis. Such a distribution is centrifugally stable, and it can quench turbulence generated by other instabilities in the flow. Significant reduction in the turbulent kinetic energy is predicted at the end of the cylindrical section and beginning of the conical section. The flow is nearly re-laminarized in an annular region at the beginning of the conical section. High turbulent kinetic energy prevails near the air-water interface. The energy spectra at points near the interface show a short inertial subrange where energy decays as f −5/3 , followed by damping where energy drops as f −7 , where f is frequency. For points in the boundary layer where high turbulent kinetic energy is found, the energy spectra exhibit f −4 decay.
Mean velocity profiles and normal Reynolds stresses are in fair agreement with experimental data by Hsieh [13]. However, LES results with the dynamic subgrid model over-predict the experimental data for tangential mean velocity. Using Smagorinsky model, one may obtain better agreement with the experimental data by adjusting the model constant. However, such an approach is not adequate for turbulent flows that involve regions of intense turbulence that are subjected to stabilizing or re-laminarizing mechanism as is the case of hydrocyclones. | 7,753.2 | 2021-10-14T00:00:00.000 | [
"Physics",
"Environmental Science"
] |
Complete Chloroplast Genome Sequence of Omani Lime (Citrus aurantiifolia) and Comparative Analysis within the Rosids
The genus Citrus contains many economically important fruits that are grown worldwide for their high nutritional and medicinal value. Due to frequent hybridizations among species and cultivars, the exact number of natural species and the taxonomic relationships within this genus are unclear. To compare the differences between the Citrus chloroplast genomes and to develop useful genetic markers, we used a reference-assisted approach to assemble the complete chloroplast genome of Omani lime (C. aurantiifolia). The complete C. aurantiifolia chloroplast genome is 159,893 bp in length; the organization and gene content are similar to most of the rosids lineages characterized to date. Through comparison with the sweet orange (C. sinensis) chloroplast genome, we identified three intergenic regions and 94 simple sequence repeats (SSRs) that are potentially informative markers with resolution for interspecific relationships. These markers can be utilized to better understand the origin of cultivated Citrus. A comparison among 72 species belonging to 10 families of representative rosids lineages also provides new insights into their chloroplast genome evolution.
Introduction
Citrus is in the family of Rutaceae, which is one of the largest families in order Sapindales. Flowers and leaves of Citrus are usually strong scented, the extracts from which contain many useful flavonoids and other compounds that are effective insecticides, fungicides and medicinal agents [1][2][3]. Citrus is of great economic importance and contains many fruit crops such as oranges, grapefruit, lemons, limes, and tangerines. However, due to a long cultivation history, wide dispersion, somatic bud mutation, and sexual compatibility among Citrus species and related genera, the taxonomy of Citrus remains controversial [4,5] and the origination of many Citrus species and hybrids is still unresolved [6,7].
The chloroplast (cp) genome sequence contains useful information in plant systematics because of its maternal inheritance in most angiosperms [8,9] and its highly conserved structures for developing promising genetic markers. The only complete cp genome available in Citrus is sweet orange (Citrus sinensis) [10], which has provided valuable information to the position of Sapindales in rosids. Although a genome sequencing project is in progress for C. clementine, its complete chloroplast genome sequence is not available yet. To identify the cp genome regions that are polymorphic and may be used as molecular markers for resolving the evolutionary relationships among Citrus species, a second cp genome within the genus is necessary for comparative analysis. For this purpose, the major aim of this study is to determine the complete cp genome sequence of C. aurantiifolia.
C. aurantiifolia, which is commonly known as Key lime, Mexican lime, Omani lime, Indian lime, or acid lime, is native to Southeast Asia and widely cultivated in tropics and subtropics. Oman is known to be a transit country for lime, from which lime spread to Africa and the New World [11]. In Oman, Omani lime is considered the fourth most important fruit crop in terms of cultivated area and production. The products of Omani lime can be used for beverage, food additives and cosmetic industries [12]. Omani lime is sensitive to several biotic agents, the most serious of which is 'Candidatus Phytoplasma aurantifolia', the cause of witches' broom disease of lime (WBDL). Recent studies on WBDL focused on effect of genetic diversity of Omani limes on the disease [13], transcriptome and proteomic analysis of lime response to infection by phytoplasma [14][15][16] and effect of phytoplasma on seed germination, growth and metabolite content in lime [17,18].
Here, we present the complete chloroplast genome sequence of Omani lime (C. aurantiifolia). To identify loci of potential utility for the molecular identification and phylogenetic analyses of Citrus cultivars and species, we compared the intergenic regions and SSRs in the cp genomes of C. aurantiifolia and C. sinensis. Furthermore, we performed phylogenetic analyses to infer the history of gene losses in the cp genome evolution among representative rosids lineages.
Sample Preparation and Sequencing
The Omani lime leaves were collected from a 5-year-old lime tree at a private farm located in the Omani territory of Madha (GPS coordinates: 25.276318, 56.318909). This farm is owned by one of the co-authors of this work, Dr. Abdullah M. Al-Sadi, whom should be contacted for future permissions. This study does not involve endangered or protected species and does not require specific permission from regulatory authority concerned with protection of wildlife. The sample was stored in a cool box and transported to the Plant Pathology Research Laboratory at Sultan Qaboos University (Al Khoud, Oman) for DNA extraction following a protocol of Maixner et al. [19]. The leaves were washed with clear water before the isolation procedure. 1 g of leaves were used and crushed in 3 ml CTAB extraction buffer (2% CTAB, 1.4 M NaCl, 500 mM EDTA pH8, 1 M Tris-HCl pH8 and 0.2% beta-mercaptol). 1.5 ml of the leave extract was transferred to a 2 ml tube and incubated in a water-bath at 65uC for 15 min. The tube was turned up and down twice during incubation, centrifuged at 960 g for 5 min, and the supernatant was subsequently transferred to a clean eppendorf tube. An equal volume of chloroform-isoamyl alcohol mix (24:1) was added and the tube was centrifuged at 21000 g for 20 min. The supernatant was transferred to a new tube and then 0.6 volume of isopropanol was added to the supernatant and incubated at 220uC for 30 min. The DNA pellet was collected by centrifugation at 21000 g for 20 min and then washed with 1 ml of 70% ethanol. The final DNA was resuspended in 100 ml TE (Tris 10 mM, EDTA 1 mM pH8) and was stored at 280uC until used.
The library construction and sequencing were done at the Genome Analysis Centre (Norwich, UK). The Illumina TruSeq DNA Sample Preparation v2 Kit was used to prepare an indexed library. The DNA sample was sheared to a fragment size of 500-600 bp using a sonicator, followed by end-repair and the addition of a single A base for binding of the indexed adapter. The appropriate sized library (500 bp) was selected by gel electrophoresis, followed by PCR enrichment. The 251 bp paired-end sequencing run was performed on an Illumina MiSeq instrument using the SBS chemistry and Illumina software MCS v2.3.0.3 and RTA v1.18.42. The raw reads were deposited at the NCBI Sequence Read Archive under the accession number SRR1611615.
Genome Assembly and Analyses
The procedures for genome assembly and annotation were based on our previous studies of cp genomes [20,21]. In addition to the standard de novo assembly approach by using Velvet v1.2.10 [22] with the k-mer size set to 243, a reference-based approach for assembly as described below was used in parallel. All of the raw reads were initially mapped onto the published cp genome of C. sinensis [10] using BWA v0.6.2 [23]. The sequence variations were identified with SAMtools v0.1.19 [24] and visually inspected using IGV v2.3.25 [25]. The variants were corrected with the raw reads and the regions without sufficient coverage were converted into gaps. This corrected sequence was then used as the new draft reference for the next iteration of verification. Gaps were filled using the reads overhang at margins and the process was repeated until the reference was fully supported by all mapped raw reads. The final assembly, which was supported by our de novo and reference-based approaches, resulted in an average of 1,441-fold coverage of paired-end reads with a mapping quality of 60 and the region with the lowest coverage is 506-fold.
The preliminary annotations of the C. aurantiifolia cp genome were performed online using the automatic annotator DOGMA [26] and verified using BLASTN [27,28] searches (e-value cutoff = 1e-10) against other land plant cp genomes. Each annotated gene was manually compared with C. sinensis cp genome for start and stop codons or intron junctions to ensure accurate annotation. The codon usage was analyzed by using the seqinr R-cran package [29]. A circular map of genome was produced using OGDRAW [30].
To identify the differences between C. aurantiifolia and C. sinensis, the two sequences were aligned using Mauve v2.3.1 [31] and the result was analyzed using custom Perl scripts. Intergenic gene regions were parsed out from the two Citrus cp genomes and aligned using MUSCLE v3.8.31 [32] with the default settings. The pairwise distances were calculated using the DNADIST program in the PHYLIP package v3.695 [33].
The positions and types of simple sequence repeats (SSRs) in the two Citrus cp genomes were detected using MISA (http://pgrc. ipk-gatersleben.de/misa/). The minimum number of repeats were set to 10, 5, 4, 3, 3, and 3 for mono-, di-, tri-, tetra-, penta-, and hexanucleotides, respectively. For long repeats, the program REPuter [34] was used to identify the number and location of direct and inverted (i.e., palindromic) repeats. A minimum repeat size of 30 bp and sequence identity greater than 90% setting were used according to the study of C. sinensis cp genome [10]. The redundant or overlapping repeats were identified and filtered manually.
Phylogenetic Inference
Phylogenetic analysis of the representative rosids lineages with complete cp genomes available was performed using PhyML v20120412 [35] with the GTR+I+G model. A total of 72 rosids species were chosen as the ingroups and Vitis venifera was included as the outgroup, the accession numbers were provided in Table S1. The protein-coding and rRNA genes were parsed from the selected cp genomes and clustered into ortholog groups using OrthoMCL [36]. The presence/absence of orthologous genes in each genome was examined and further verified using TBLASTN [27,28] searches (e-value cutoff = 1e-10). The nucleotide sequences of the conserved genes were aligned individually by using MUSCLE with the default settings. The concatenated alignment was used to infer a maximum likelihood phylogeny as described above. The bootstrap supports were estimated from 1,000 resampled alignments generated by the SEQBOOT program in the PHYLIP package.
Investigations of orf56 and ycf68
To investigate the presence/absence of orf56 and ycf68 in the selected cp genomes, the gene sequences from C. aurantiifolia was used as the queries to perform BLASTN [27,28] searches (e-value cutoff = 1e-10). The significant hits were examined to investigate the presence of intact open reading frames (ORFs). Phylogenetic analysis of the cp orf56 genes and the homologous mitochondrial sequences was performed as described above. The final alignment contains 190 aligned nucleotide sites and a total of 70 sequences, including two sequences of Amborella as the outgroup.
General Features of the Omani Lime Chloroplast Genome
The complete cp genome of C. aurantiifolia (Christm.) Swingle (GenBank accession number KJ865401.1) is 159,893 bp in length, including a large single copy (LSC) region of 87,148 bp, a small single copy (SSC) region of 18,763 bp, and a pair of inverted repeats (IRa and IRb) of 26,991 bp each ( Figure 1 and Table 1). A total of 137 different genes, including 93 protein-coding genes, 30 tRNA genes, and four rRNA genes, were annotated (Table S2). Among these, 12 protein-coding genes and 7 tRNA genes are duplicated in the IR regions. Most of the protein-coding genes are Table 2. Differences between the C. aurantiifolia and C. sinensis cp genomes. composed of a single exon, while 14 contain one intron and three contain two introns. The gene rps12 was predicted to undergo trans-splicing, with the 59 exon located in the LSC region and the other two exons located in the IR regions. The protein-coding regions contain a total of 27,159 codons (Table S3). Isoleucine and cysteine are the most and least frequent amino acids and have 2,892 (10.7%) and 359 (1.2%) codons, respectively. The codon usage is biased towards a high ratio of A/ T at the third position, which is also observed in many land plant cp genomes [37].
Sequence Comparisons with Sweet Orange
The general characteristics of the two Citrus cp genomes are summarized in Table 1, overall the compositions are quite similar. The GC content of these Citrus cp genomes is approximately 38.5%, which is slightly higher than the average of the 72 representative rosids lineages (36.7%). In these two Citrus cp genomes, the genic regions, introns, and intergenic regions account for ca. 58%, 11%, and 31%, respectively ( Table 1).
The pairwise sequence alignment between the two Citrus cp genomes revealed approximately 1.3% sequence divergence (Table 2), including 1,780 indels (1.11%) and 330 substitutions (0.21%). The LSC region contains more sequence polymorphisms than expected by its size, including 1,360 (76.4%) indels and 235 (71.2%) substitutions. In contrast, the two IR regions account for ca. 34% of the cp genome yet contain only 16 (0.9%) indels and 12 (3.6%) substitutions. The size differences in the LSC and SSC regions between these two cp genomes are mostly explained by one large indel in each region. The LSC sizes differ by 596 bp and a 523-bp indel was found in the spacer between rps16 and trnQ- UUG. The SSC sizes differ by 370 bp and a 354-bp indel was found in the spacer between rpl32 and trnL-UAA.
To identify the intergenic regions that may be useful for phylogenic analysis or molecular identification, we searched for the spacers that are .400 bp in length and exhibit above-average sequence divergence between the two Citrus species (i.e., .1.3%). A total of three regions satisfied these criteria, including the spacer between trnH-GUG and psbA (449 bp, 1.6% divergence), the spacer between rpl32 and trnL-UAG (1141 bp, 1.5% divergence), and the spacer between trnD-GUC and trnY-GUA (469 bp, 1.3% divergence).
The junctions between the IR, LSC, and SSC regions in C. aurantiifolia are similar to that of C. sinensis except for the LSC-IRb boundary. A total of 23 indels and five substitutions were found at this region, resulting in one copy of rpl22 spanning across the LSC-IRb junction in C. aurantiifolia. Comparing the IR junctions of Citrus with Theobroma and Gossypium in Malvaceae [38], it was found that the IRs in Citrus have expanded to include rps19 and 252 nt of rpl22, whereas in Malvaceae, rps19 is located in LSC and rpl22 was missing [38][39][40].
Analyses of Repetitive Sequences
A total of 109 SSR loci were found in the cp genome of C. aurantiifoliaa, accounting for 1,352 bp of the total sequence (ca. 0.8%). Among these, 94 were also found in C. sinensis and 42 exhibit length polymorphism (Table 3). Most SSRs are located in intergenic regions, but some were found in coding genes such as matK and ycf1. Concerning the controversial status of Citrus taxonomy, the SSRs identified in this study may provide new perspective to refine the phylogeny and elucidate the origin of the cultivars. Furthermore, these SSRs may be used as molecular markers for population studies.
In addition, 62 large repeats that are longer than 30 bp were found in the C. aurantiifolia cp genome ( Table 4). Most of these repeats are located in intergenic spacers, except for three that are located in the coding regions of rps4, psaA and psaB. Twelve of these long repeats were also found in C. sinensis, indicating that these repeats might be widespread in the genus.
Gene Content Analyses within the Rosids
A maximum likelihood phylogenetic analysis of 72 representative rosids lineages was conducted based on a concatenated alignment of four rRNA and 58 protein-coding genes with 54,689 sites ( Figure 2). Citrus represents Sapindales and is sister to the clade containing Malvales and Brassicales. These relationships are congruent with the previous reports [10,[41][42][43]. Based on this phylogeny and the gene content, we inferred the gene loss events during the cp genome evolution in rosids.
The translation initiation factor gene infA in cp has been lost independently at least 24 times in angiosperms and evidence provided from some cases suggested functional replacement by a nucleus copy [44]. Although the majority of infA in our selected cp genomes were found to be pseudogenized or completely lost, an intact infA was found in Quercus, Francoa, and two Cuscumis species.
The rpl22 were found to be lost in Fabaceae [45] and Castanea of Fagaceae [46] following independent transfers to nucleus. Furthermore, another putative loss of rpl22 was detected in Passiflora [46]. The rpl22 in Malvaceae, including Theobroma and three Gossypium species, were found to be pseudogenized in our analysis. In Citrus, the ORF of rpl22 was shortened to 252-264 nt compared to the typical length of 399-489 nt in other rosids [10,46]. However, compared with the pseudogenized rpl22 found in Malvalvace, the rpl22 homologs in Citrus still show high sequence conservation. Additionally, the rpl22 transcripts can be identified in the EST database for various Citrus species (data not shown). Taking account into the above consideration, we did not annotate rpl22 as a pseudogene in Citrus.
The parallel losses of rps16 were found in several rosids lineages (Figure 2), including one time in Salicaceae, two times in Fabaceae and another two times in Brassicaceae. The loss of rps16 in Medicago and Populus was found to be substituted by a nuclearencoded copy that transferred from the mitochondrion (mt) [47]. Because the nuclear-encoded RPS16 was found to target both mt and cp in Arabidopsis, Lycopersicon, and Oryza [47], it is possible that the cp genome-encoded rps16 would not be maintained by selection and will eventually become lost in these lineages.
There are only a few gene loss events of photosynthetic genes found in rosids. In addition to the loss of psaI in Lathyrus sativus [48], the accD seems to be lost independently in Trifolium subterraneum and several Gerantiaceae species except for Geranium palmatum. In Trifolium, a nuclear-encoded accD copy has been reported [48], which presented another example of horizontal gene transfer from cp to nucleus. Successful gene transfers from cp to the nucleus in angiosperms are rare and have been only documented for four genes in rosids. Other than the three genes described above (i.e., infA, rpl22, and accD), the rpl32 in Populus (Salicaceae) is the fourth example [49][50][51].
The IR has been reported to be independently lost at least five times among seed plants, two of which are within rosids [51]. In addition to the inverted repeat lacking clade (IRLC) of papilionoid Fabaceae [52] and Erodium of Gerantiaceae [53,54], the IR was found to be lost in two lineages of Fragaria (Rosaceae), which are F. vesca ssp. bracteatea and F. mandschurica (accession: NC_018767, not shown in Figure 2). Based on the Fragaria phylogeny shown in a previous study [55], it seems that IR loss was not a single event in Fragaria.
Molecular Evolution of orf56 and ycf68 within the Rosids
In the comparison of gene content between the two Citrus cp genomes, C. aurantiifolia was found to contain two additional protein-coding genes. The first gene, orf56, is located in the trnA-UGC intron that contains one sequence homologous to previously recognized mitochondrial ACRS (ACR-toxin sensitivity gene) in Citrus [56]. In addition to the 171-bp identical sequences between cp orf56 and the ORF sequences of ACRS in mt, the full length of 355-bp region of ACRS that conferred sensitivity to ACR-toxin in E. coil are also identical. Furthermore, the whole trnA-UGC among two Citrus cp regions and C. jambhiri mitochondrial ACRS shared more than 96% identity ( Figure S1), which highlight the conservation of this region between cp and mt.
The gene orf56 has also been included in the annotation of complete cp genomes of Calycanthus [57] and Pelargonium [58]. Our BLAST search against the rosids genome database revealed that in addition to Citrus and Pelargonium, all of the species examined in Cucurbitaceae and Myrtales also contain an intact orf56 ( Figure 3). Moreover, an intact ACRS ORF is also present in the mt genomes of Liriodendron [59] and Silene [60] and the ORF sequences between cp and mt are identical. Goremykin et al. [57] suggested that the ACRS gene was relative recently transferred from cp to mt. Based on the phylogeny containing the cp orf56 and the mt ACRS ( Figure S2), it appears that orf56 has been independently transferred from cp to mt in different lineages.
The second gene, ycf68, is located in the trnI-GAU intron. A nearly identical sequence was found in C. sinensis but an additional T insertion near the C-terminus abolished the stop codon at the corresponding position. The intact ycf68 can be detected in several monocots and Nymphaeaceae [61,62]. However, in the majority of other rosids ( Figure 3) and the rest of the eudicots [61], the ycf68 homologs all contain premature stop codons. Although Raubeson et al. [61] argued that ycf68 is not a protein-coding gene based on the lack of intron-folding pattern, the high levels of sequence conservation among the ORFs of identified homologs suggest that the true identity and functionality of this putative gene remains to be further investigated.
Conclusions
We reported the complete cp genome sequence of Citrus aurantiifolia (Rutaceae) in this study. The genome organization and gene content is typical of most angiosperms and highly similar to that of C. sinensis (i.e., 98.7% identical at the nucleotide level). The only difference in the gene content between the two Citrus cp genomes is the C. aurantiifolia-specific presence of a proteincoding gene (ycf68) in the trnI-GAU intron. Notably, three long intergenic spacers with high sequence divergence and 94 shared SSR regions were identified in the C. aurantiifolia-C. sinensis comparison. These regions may provide phylogenetic utility at low taxonomic levels and could be applied to the molecular identification of Citrus cultivars. Finally, our comparative analysis of gene content among 72 representative rosids lineages highlighted multiple events of gene losses within this group. Figure S1 Alignment of the orf56-containing sequences of two Citrus cp genomes and C. jambhiri mitochondrial ACRS sequences. (TIF) Figure S2 The maximum likelihood phylogeny of the cp orf56 and mt ACRS ORF sequences.
(TIF)
Table S1 List of the complete chloroplast genome sequences included in the phylogenetic analysis. (XLSX) | 5,055.4 | 2014-11-14T00:00:00.000 | [
"Biology",
"Environmental Science"
] |
On hazard ratio estimators by proportional hazards models in matched-pair cohort studies
Background In matched-pair cohort studies with censored events, the hazard ratio (HR) may be of main interest. However, it is lesser known in epidemiologic literature that the partial maximum likelihood estimator of a common HR conditional on matched pairs is written in a simple form, namely, the ratio of the numbers of two pair-types. Moreover, because HR is a noncollapsible measure and its constancy across matched pairs is a restrictive assumption, marginal HR as “average” HR may be targeted more than conditional HR in analysis. Methods Based on its simple expression, we provided an alternative interpretation of the common HR estimator as the odds of the matched-pair analog of C-statistic for censored time-to-event data. Through simulations assuming proportional hazards within matched pairs, the influence of various censoring patterns on the marginal and common HR estimators of unstratified and stratified proportional hazards models, respectively, was evaluated. The methods were applied to a real propensity-score matched dataset from the Rotterdam tumor bank of primary breast cancer. Results We showed that stratified models unbiasedly estimated a common HR under the proportional hazards within matched pairs. However, the marginal HR estimator with robust variance estimator lacks interpretation as an “average” marginal HR even if censoring is unconditionally independent to event, unless no censoring occurs or no exposure effect is present. Furthermore, the exposure-dependent censoring biased the marginal HR estimator away from both conditional HR and an “average” marginal HR irrespective of whether exposure effect is present. From the matched Rotterdam dataset, we estimated HR for relapse-free survival of absence versus presence of chemotherapy; estimates (95% confidence interval) were 1.47 (1.18–1.83) for common HR and 1.33 (1.13–1.57) for marginal HR. Conclusion The simple expression of the common HR estimator would be a useful summary of exposure effect, which is less sensitive to censoring patterns than the marginal HR estimator. The common and the marginal HR estimators, both relying on distinct assumptions and interpretations, are complementary alternatives for each other. Electronic supplementary material The online version of this article (doi:10.1186/s12982-017-0060-8) contains supplementary material, which is available to authorized users.
Background
Matching is a useful sampling method employed in cohort studies, in which the control of confounders is indispensable [1]. The simplest matching design is a 1:1 matched (matched-pair) cohort study, in which each matched pair comprising an exposed and an unexposed member is followed up through the study period. The standard choices of effect measures are common odds ratio (OR) and risk ratio (RR) conditional on matched pairs. As the number of pairs increases, asymptotically unbiased estimate of common OR across matched pairs is the ratio of the number of "discordant" pairs [2,3]; using the numbers of pairs shown in Table 1, the conditional maximum likelihood estimator (CMLE) of common OR is B/C [2]. This estimator coincides with the Mantel-Haenszel OR estimator [4] and the unconditional maximum likelihood estimator using multinomial distribution of (A, B, C, D) parameterized under common OR [5]. Common RR may be estimated by the Mantel-Haenszel RR estimator, which simplifies to the crude RR (A + B)/(A + C) [3,6], by estimating equations for parameters in conditional multiplicative risk models [7], or by conditional Poisson regression models, which are mimicked by stratified Cox model-fitting with Breslow or Efron-type tie breaking [8,9].
One of the concerns in cohort studies is censoring owing to unequal follow-up period or loss to follow-up before the end of the study. In the presence of censoring, common hazard ratio (HR) is a viable alternative. Common HR can be estimated by the Mantel-Haenszel rate ratio [6] or partial maximum likelihood estimators (PMLE) of Cox proportional hazards models stratified on matched pairs [10][11][12]. However, perhaps because of the ease of Cox model-fitting by modern computer software, it is lesser known that the PMLE of common HR can also be transcribed in a simple formula as in the case of CMLE of common OR [10]. The formula motivates us to focus on the relationship of the PMLE to the matchedpair analog of C-statistics for time-to-event, which has been recently discussed in the literature for evaluating discriminatory ability in prediction [13][14][15].
This representation would be useful because HR (like OR) is known to be a noncollapsible measure: even under homogeneity of conditional HR across strata and in the absence of confounding, marginal (unadjusted) HR is not necessarily equal to the conditional one [16,17]. Moreover, the assumption of homogenous (common) HR across strata may be too restrictive. To circumvent interpretational difficulties, marginal HR estimated by unstratified Cox models with robust variance estimator is often of primary interest than common HR [18]. Even when HR is not constant over time, it may be interpreted as the "average" HR of time-varying HR [19]. However, we argue that the uncritical "average" view of marginal HR may have limited value because the estimator depends on censoring distribution that is nuisance to inference for exposure effect on outcome [20,21].
In this paper, we showed the simple expression of the common HR estimator and its alternative interpretation as the odds of the matched-pair analog of C-statistic for censored time-to-event data. Through simulation studies, assuming proportional hazards within the matched pairs, we evaluated the influence of various censoring patterns on the marginal and common HR estimators of unstratified and stratified proportional hazards models, respectively. For illustration, several estimators were compared in a propensity score-matched dataset of primary breast cancer from the Rotterdam tumor bank.
Methods
In this section, we provide the simple formula for the common HR estimator under a stratified proportional hazards model in matched-pair cohort studies. The common HR is linked to overall C-index with matched-pair analog to improve its interpretation. By simulation studies under the stratified proportional hazards models, we compare the performance in competing estimators as well as statistical tests used in matched-pair cohort studies in various censoring distributions. Finally, we illustrate the methods in a real dataset.
Stratified PMLE of common HR in matched-pair cohorts
Consider matched-pair cohort studies comparing timeto-event outcome T, in which each pair k (k = 1,…, n) is comprised of an exposed (e = 1) and an unexposed member (e = 0). Because outcome T ke of member e in pair k may be censored by drop-out time U ke , or the end of follow-up τ, we observe follow-up time as X ke = min(T ke , U ke , τ ). Define Y ke as an indicator of event (Y ke = 1 if X ke = T ke , 0 otherwise). Suppose all risk factors have the same distribution within each pair.
If we are interested in common HR across all matched pairs throughout the follow-up period, an appropriate model is the Cox proportional hazards model stratified on the matched-pair k: where λ ke (t) and β are a hazard function of T ke and logarithm of common HR, respectively.
Partial likelihood of (1) is given by the product of the contribution at each event time from each stratum k, expressed as follows [10][11][12]: To express the contribution L k (β) from each stratum, we classify each pair observable in matched-pair data into eight types (Table 2). For clarity, the only tie we additionally consider is caused by the end of follow-up, i.e., X k1 = X k0 = τ (type 9). Let n 1 ,…, n 9 denote the number of pairs of types 1-9. Partial likelihood in the presence of other types of ties are shown in "Appendix 1". Pairs of types 1 and 3 contribute to partial likelihood by exp(β) 1+exp(β) , pairs 2 and 4 contribute by 1 1+exp(β) , and pairs of types 5-9 do not contribute to it. Eventually, the only contributors for the PMLE are those who are in the pairs in which the pair-member with shorter observed time experienced an event; this is the necessary and sufficient condition for "comparable" pairs in C-statistic for timeto-event, which we revisit later [13][14][15]. The resulting partial likelihood (2) is where G = n 1 + n 3 denotes the number of pairs where the exposed member has shorter observed time and experienced an event (types 1 and 3) and H = n 2 + n 4 denotes the number of pairs where the unexposed member has shorter observed time and experienced an event (types 2 and 4). Therefore, all information regarding matched-pair data for common HR is concentrated on the number of only two types of pairs.
By maximizing partial likelihood, we can write the PMLE of common HR as G/H. Substituting G/H into the observed Fisher information [11], the approximate variance estimator of log(G/H) can be obtained by 1/G + 1/H. These are of the same form as the logarithm of the CMLE log(B/C) and its variance estimator 1/B + 1/C [2,3].
Tests of null association
To test the null hypothesis of common OR in matchedpair data, McNemar's test is often recommended [22][23][24]. The test statistic is (B−C) 2 B+C , which is also a function of B and C. Similarly, by using the definitions of G and H from above Klein and Moeschberger [12] have developed a stratified log-rank test statistic (G−H ) 2 G+H as a weighted e β 1 + e β G 1 1 + e β H , rank statistic. As the number of pairs grows, (G−H ) 2 G+H has an asymptotic Chi-squared distribution with one degree of freedom under β = 0. Similar to McNemar's test that can be considered as the score test of OR = 1 in a conditional logistic model [3], the stratified log-rank test can be considered as the score test for β = 0 in a stratified Cox model (1). Note that Wald and score tests for the hypothesis of conditional HR (or OR) = 1 can be shown to be asymptotically equivalent to test statistics for marginal HR (or OR) = 1 [25]. Therefore, tests for both conditional and marginal null hypotheses in different models may be used interchangeably, although OR and HR are both noncollapsible measures.
Stratified PMLE as overall C-statistic for matched pairs
For binary exposure E (1 if exposed, 0 if unexposed) and time-to-event outcome T, overall C-index (C-index for time-to-event) is defined as where τ is the time of the end of follow-up or an arbitrary time interval set by analysts [13,14]. Assuming the absence of censoring except at the end of followup, Pencina and D' Agostino proposed to estimate C τ by restricting all possible pairs in the sample to "comparable" pairs, in which the member with a shorter observed time experienced an event, i.e., "X i < X j , Y i = 1, X i < τ" [14].
Due to censoring by U ke , event times T k1 and T k0 are not always observable. As shown in the "Appendix 2", G/(G + H) converges in probability (as pair number n grows) to Note that (3) is not equal to Pr(T k1 < T k0 |T k1 < τ or T k0 < τ ) in general. Thus, C τ,pair cannot be apparently calculated by the observed data if censoring before τ occurs. "Appendix 2" shows, however, that under the model (1), the odds of (3) equal exp(β) even if T ke is censored by U ke that is independent to T ke conditional on matched pairs and exposure. Thus, we can estimate C τ,pair as well as β based on only "comparable" matched pairs introduced by design even if censoring depends on both matched pairs and exposure.
Simulation studies
To examine the performance of the stratified PMLE under the assumption (1) compared to competitive PMLEs used in matched-pair cohort studies, we simulated 2000 cohorts with size 2n = 100, 500 (n = 50, 250 exposed-unexposed pairs). SAS code for generating data will be provided in the Additional file 1.
We also employed Weibull time-to-event variables in additional scenarios to emulate the situations in which (1) baseline hazards increase or decrease instead of the time-constant hazard λ 0 , or (2) the shape parameter varied between the strata while keeping stratum-specific HR fixed as a constant across strata. As the results from these additional scenarios were similar to those from the above exponential-normal frailty model, the parameter settings and the results are provided in the Additional file 1.
We fitted pair-stratified Cox models, unstratified Cox models with or without robust sandwich variance estimator [26], as well as true frailty Cox models as a reference. Note that stratified and frailty models assume that the conditional parameter is constant across matched pairs, while unstratified models only model a marginal parameter and do not assume such constancy across pairs.
With the frailty Cox models used in the data generation, the marginal distributions of time do not follow proportional hazards except for the positive-stable distributed frailty [12]. Thus, the unstratified Cox model is known to be misspecified. One way around this problem is to define the model parameters as the asymptotic means of the maximum-likelihood estimators that are free from censoring, which is always well-defined and interpretable (even if the models are not correct) [20,27,28]. Therefore, the targeted marginal HR in this study is defined as a mean of the estimate of unstratified Cox models calculated in a large (n = 5,000,000) dataset where no member is censored.
The performance of the above estimators was evaluated by mean bias (the average of 2000 log-HR estimatestrue log HR), empirical standard error (standard deviation of 2000 estimates), mean estimated standard error, root mean squared error (RMSE; the square root of the sum of the squared bias and the empirical variance of the estimator), and coverage proportion of 95% confidence intervals. The empirical power (or type I error when HR = 1) tested by Wald statistics (log-HR estimates divided by their estimated standard errors) of the above was also compared, accompanied by a stratified log-rank test statistic. We used PHREG procedure in SAS version 9.4 (Cary, NC).
Application: propensity-matched cohort data from the Rotterdam tumor bank
To illustrate the methods in a real dataset, we used the records from the Rotterdam tumor bank, which includes follow-up data on 2982 women with primary breast cancer. The dataset is available in the R package AF developed by Dahlqwist and Sjölander [29] and the details of the dataset have been described elsewhere [30]. The outcome T is relapse-free survival, which is defined as time to developing relapse of breast cancer or death from any cause before the end of the follow-up period. Women remained in the dataset until they experienced relapse or death, were lost to follow-up or were at the end of the follow-up period, whichever came first. The exposure of interest is the absence of chemotherapy (1 if treated without chemotherapy, n = 2402; 0 if treated with chemotherapy, n = 580).
One notable feature in this dataset is that the amount of confounding is very strong-adjusting for the possible confounders reverses the sign of the association [30]. Thus, we turned the dataset into a matched cohort based on propensity score (the conditional probability of exposure given possible confounders). Propensity score is conditional on the following prognostic variables: age at surgery (years), menopausal status (0 if premenopausal, 1 if postmenopausal), tumor size (≤20 mm, >20-50 mm, and >50 mm), tumor grade (2 or 3), progesterone receptors, (fmol/l), oestrogen receptors (fmol/l), and the number of positive lymph nodes [ranging between 0 and 34; transformed into exp(-0.12 * the number of nodes)]. We estimated the propensity score for each woman by fitting a logistic model, and then matched women on the estimated propensity scores by caliper-based pair-matching algorithm without replacement (allowable caliper width was 20% of the standard deviation of estimated propensity scores in a chemotherapy group). The resulting matched cohort is comprised of n = 446 exposed-unexposed pairs. The SAS code for forming the propensitymatched cohort from the Rotterdam dataset is provided in the Additional file 1. Table 3 shows the results in independently censored data (similar results were obtained for n = 50, provided in online supplementary material). The marginal HR defined in the unstratified models is towards null from conditional HR, similar to the well-known result that the marginal OR is closer to null than common stratum-specific OR [16]. In fact, marginal HR always lies between the conditional HR and 1 under the exponential survival model [17]. Null exposure effect in conditional HR implies that marginal HR is also null. In this case, no estimator has a bias. Coverage of confidence intervals maintains almost nominal level except for the unstratified model without robust variance that overestimates the true variability.
Simulation results
For non-null HR (β ≠ 0), PMLE for unstratified models have "bias" from conditional HR that partly reflects the noncollapsible property of HR [17] and the dependency on censoring distribution. The latter also impedes its interpretation as "average" marginal HR that is independent of censoring. Frailty to disease structurally changes hazard among the remaining risk-set over the follow-up period [19,31]: under our simulation model, HR constancy during the follow-up period only holds conditionally on frailty but does not hold marginally with non-null exposure effect. Estimates for unstratified models are indeed valid as marginal effect-measures in pair-matched data if there are no other covariates that need to be controlled and if no censoring occurs [20,32]. If no observation is censored, estimates from unstratified models are unbiased for the marginal HR parameter (data not shown). As censoring increases, the bias in unstratified PMLE from the marginal HR parameter becomes larger and the coverage probability decreases. Table 4 shows the results for censorship dependent on matched pair and exposure. The pair effect on censoring alone (from the rows "Censoring rate ratio = 1") does not invalidate any estimate for null exposure effect but biases unstratified PMLE from both conditional and marginal HRs under non-null exposure effect, as expected from Table 3. Exposure effect on censoring also affects the distribution of unstratified PMLE for both null and non-null exposure effects. This censoring mechanism also makes bias in PMLE for frailty Cox models despite that it models true hazard. Only stratified PMLE, G/H, has no bias in this censoring pattern, which is guaranteed with the assumption (1), as shown in "Appendix 2".
The shortcomings of stratified and frailty PMLEs are their variability. Even if conditional HR is of primary interest, their RMSE can be greater than that from unstratified models. However, in the moderate sized samples (e.g., n > 250), the variability around conditional HR by stratified and frailty PMLEs can be outweighed by the bias in unstratified models: the "bias" from conditional and marginal HRs. Frailty models also failed to converge a few times in 2000 repetitions.
Matched analysis of the Rotterdam cohort
The Kaplan-Meier estimates of relapse-free survival from the original (2982 women) and the propensity-matched (2n = 892 women) Rotterdam cohorts were depicted in Fig. 1. While the unadjusted curves in the original cohorts favored the absence of chemotherapy, the propensity-matched curves adjusting possible confounders reversed the association of exposure and outcome.
Censoring before the end of the follow-up period was not negligible in the matched cohort, but censoring distribution is similar across exposure groups (Fig. 2): the situation would resemble the simulation pattern 1 (independent censoring). Among 446 pairs in the matched cohort, the number of pairs where the exposed member has shorter observed time and experienced an event (G) was 198 and the number of pairs where the unexposed member has shorter observed time and experienced an event (H) was 135. The PMLE of common HR is G/H = 1.47 with 95% confidence limits exp(log(1.47) ± 1 .96 × √(1/198 + 1/135)) = 1.18 and 1.83; these estimates coincide with the result from the stratified Cox model fitting program in SAS/PHREG procedure. On the contrary, marginal HR estimated from the unstratified Cox model with robust sandwich variance estimator in the same matched cohort is 1.33 (95% CI 1.13-1.57). As seen from simulation, the common HR estimate was further from null than marginal HR estimate in this dataset.
These results were compared to other marginal or conditional HR estimates adjusting for the seven prognostic variables: Table 5 shows the estimates from inverse-probability weighted Cox model with robust variance estimator, multivariable-adjusted Cox model, and multivariable-adjusted Cox model with inverse-probability weighting and robust variance estimator. Although the target populations were not the same between propensity-matched and inverse-probability weighted analyses [18,33], the PMLE of stratified and unstratified Cox models from the matched cohort approximate the conditional and marginal estimates from the original cohort, respectively.
Discussion
The PMLE for common HR in matched-pair cohort studies can be expressed by a simple formula based on only two numbers: the number of pairs in which the exposed has a shorter observed time and experienced an event (G) and the number of pairs in which the unexposed has a shorter observed time and experienced an event (H). Such a simple form of HR estimators is unique to PMLE. Corresponding Poisson rate regression may be a stratified Poisson model with common HR, with the likelihood conditional on the total number of events in each stratum (0, 1, or 2), which reduces to binomial likelihood [34]. The CMLE for common HR is the solution of k Y k1 X k0 −HR·Y k0 X k1 X k0 +HR·X k1 = 0, which is dependent on observed time X. It is slightly different from the Mantel-Haenszel rate ratio estimator, k Y k1 X k0/ (X k0 +X k1 ) k Y k0 X k1/ (X k0 +X k1 ) , which approximates the Poisson CMLE around HR = 1.
The current simple expression of stratified PMLE and its relationship with C τ,pair is also unique to a binary exposure. We could not find any simple expression of estimators of multiple effect-parameters for more than 2 exposure levels, or even a single parameter in the stratified Cox model (i.e., linear effect on log-hazard) for continuous exposure, say, V. In the latter case, an adequate definition of matched-pair overall C-index may be C τ,pair = Pr(V k1 > V k2 | T k1 < T k2 , T k1 < τ) (if log-HR β > 0). This may be estimated by redefining a "binary" exposure E, such that E k1 = I(V k1 > V k2 ) and E k2 = I(V k1 < V k2 ), and calculating G/(G + H) as if the dataset comes from a pair-matched cohort. However, the limiting value of this statistic is now dependent on the censoring distribution irrespective of the underlying model form. Instead, the stratified PMLE obtained by iterative maximization (e.g., by Newton-Raphson algorithm) of partial likelihood may be used to estimate overall-C; following the similar discussion of Gönen and Heller [35], the average of exp(β·|V k1 −V k2 |) across n matched pairs converges to Pr(V k1 > V k2 | T k1 < T k2 , T k1 < τ) if the assumed linear-effect Cox model is correct and if no tie-event occurs [36]. Although the simple expression of PMLE is not applicable to continuous/ multiple-level exposures, the stratified PMLE is still relevant to interpretation of a matched-pair C-index for time-to-event outcomes.
It is well recognized that whenever matching variables in case-control studies are associated with either exposure or disease in an original cohort, unless exposure effect on disease is absent, they must be adjusted in analysis irrespective of whether they are confounders [1,37]. Unlike case-control studies, ignoring matching in cohort studies generally produce valid point estimates when the matching ratio is constant across strata and no censoring occurs [32,37]. This phenomenon is due to the design that completely balances the matched variables between exposed and unexposed groups. In the theory of causal diagrams, design unfaithfulness occurs, i.e., exposure and matching variable are independent in the matched subpopulation despite being connected in the causal diagram [37][38][39]. However, when additional confounders are adjusted in the analyses, such cancellation breaks down and ignoring matching variables results in biased estimates [32]. Moreover, as shown in our simulation, if the proportionality of hazards holds given matching variables and if censoring is present, the estimated HR under marginal models would be biased away from both conditional HR and an "average" of time-varying HR [19]. The bias depends on the censoring rate even if events are independently censored. Similar to this phenomenon, if the matching variable is a risk factor for outcome and competing risks or losses to follow-up are associated with both exposure and the matching variable, estimates not adjusting for the matching variables are no longer unbiased even in the null exposure effect [1,37]. Our simulation also showed that censorship dependent on the exposure and the matched pair invalidates the marginal estimate and statistical test.
Matching is often conducted in analysis stage, especially with estimated propensity scores in order to reduce confounding, as in our real data example. Contrary to actually matched data, analytical subtleties in propensity-matching have been discussed in recent literature. First, whether propensity-matching should adjust for as sampling variation remains controversial [18,40]. Second, conditional effect parameters are usually not targeted in propensity-matching because conditioning on propensity-matched pairs has little interpretability. Third, at the cost of balancing between exposure groups, propensity-matching discards some proportion of available data. If one is interested in the effect on the exposed (the marginal effect-measure if 1 unexposed is matched on 1 exposed), one can expect more efficient estimates are obtained by differential propensity-weighting for exposed and unexposed groups [41] than marginal modeling with robust variance estimator after propensity-matching. While weighting directly uses the estimated propensity scores from fitted models [42,43], matching only uses the ranks of estimated propensity scores within an allowable caliper width. As ranks are less sensitive to misspecification of the model form, one may argue the propensity-matching analyses are more robust than estimates using propensityweighting or outcome-regression, or both [44]. Detailed investigation of this bias-variance trade-off between propensity-matching and weighting for marginal estimates is interesting future work. From these viewpoints, however, a PML estimator of common HR may be of little use along with propensity-matching.
Conclusion
Although common HR itself may have limited value in public health literature because of its noncollapsibility and built-in selection bias [19], the simple and intuitive representation of its estimator would be a useful summary of the exposure effect. The common HR estimator may be a good alternative to the marginal HR estimators if loss-to-follow-up is not negligible and/or depends on exposure and matching variables. Otherwise, survival time or risk comparisons should be used to overcome the problems with causal interpretation of HR [17,19]. Table 6 shows the additional types of pairs when tied events occur in the same pair. By convention, when censoring and event simultaneously occur in the data, that pair is treated as if censoring was measured after event: types 10 and 11 (Table 6) are treated as types 3 and 4 ( Table 2), respectively. Then, we only have to consider how to treat tie-event (type 12) by each method: the exact, Breslow's, or Efron's methods are commonly used [11,12].
Additional file
Additional file 1. Additional simulation results and SAS codes for simulation and for matching women based on estimated propensity scores in the Rotterdam tumor database.
If tie-events are treated by the exact method, the contribution of that type of pair is exp(β)/{1 + exp(β)} + 1/ {1 + exp(β)} = 1, which means these pairs do not contribute the partial-likelihood. Thus, we only have to modify the matched-pair common HR estimator as G/H in which type 12-pairs are counted in neither G nor H: G = n 1 + n 3 + n 10 and H = n 2 + n 4 + n 11 . On the contrary, if tie-events are treated by Breslow's methods, the contribution of that pair type is exp(β)/{1 + exp(β)} 2 , or by Efron's methods, 2 × exp(β)/{1 + exp(β)} 2 . This means that the partial likelihood counts type 12-pairs in both G and H: a modified matched-pair common HR estimator corresponding Breslow's or Efron's methods is G/H in which type 12-pairs are counted in both G and H (i.e., G = n 1 + n 3 + n 10 + n 12 and H = n 2 + n 4 + n 11 + n 12 ).
We provide in Table 7 a typical dataset (50 pairs) generated with true HR = 2 with independent censoring (rate = 1) in the main text. To view the impact of tied event data, we rounded the observed time to one decimal place. Among the comparable pairs, n 1 + n 3 + n 10 = 18, n 2 + n 4 + n 11 = 7 and n 12 = 1 (pair 30). Using exact tiebreaking method, G = 18 and H = 7; PMLE of common HR is G/H = 2.57 with 95% confidence limits exp{log (2.57) ± 1.96 × √(1/18 + 1/7)} = 1.07 and 6.16. Using Breslow's and Efron's tie-breaking method, G = 19 and H = 8 and the estimate (95% confidence limits) is 2.38 (1.04, 5.43). These data are in perfect accordance with the results obtained by fitting a Cox model stratified on matched-pairs via the PHREG procedure with options "ties = exact" and "ties = Breslow" (default in current version of SAS; "ties = Efron" provide the same result), respectively.
Appendix 2: Equivalence between the limiting value of G/H and C τ with censoring under stratified proportional hazards model
Following the main text, the Cτ estimator G/(G + H) is expressed as n k=1 Y k1 I(X k1 < X k0 ) n k=1 Y k1 I(X k1 < X k0 ) + Y k0 I(X k1 > X k0 ) , Table 6 Pair types with tied data with at least one event
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. | 6,965 | 2017-06-05T00:00:00.000 | [
"Mathematics"
] |
DeStripe: frequency-based algorithm for removing stripe noises from AFM images
Background Atomic force microscopy (AFM) is a relatively recently developed technique that shows a promising impact in the field of structural biology and biophysics. It has been used to image the molecular surface of membrane proteins at a lateral resolution of one nanometer or less. An immediate obstacle of characterizing surface features in AFM images is stripe noise. To better interpret structures at a sub-domain level, pre-processing of AFM images for removing stripe noises is necessary. Noise removal can be performed in either spatial or frequency domain. However, denoising processing in the frequency domain is a better solution for preserving edge sharpness. Results We have developed a denoising protocol, called DeStripe, for AFM bio-molecular images that are contaminated with heavy and fine stripes. This program adopts a divide-and-conquer approach by dividing the Fourier spectrum of the image into central and off-center regions for noisy pixels detection and intensity restoration; it is also applicable to other images interfered with high-density stripes such as those acquired by the scanning electron microscope. The denoising effect brought by DeStripe provides better visualization for image objects without introducing additional artifacts into the restored image. Conclusions The DeStripe denoising effect on AFM images is illustrated in the present work. It allows extracting extended information from the topographic measurements and implicitly enhances the molecular features in the image. All the presented images were processed by DeStripe with the raw image as the only input without any requirement for other prior information. A web service, http://biodev.cea.fr/destripe, is available for running DeStripe.
Background
Unlike other optical-based microscopes, atomic force microscope (AFM) is a sensing instrument [1]. In brief, a nano-sized tip located beneath a micro-cantilever scans across a field of deposited molecules. The cantilever deflection can be detected by a laser beam that reflects off the back of the cantilever. With a set of piezoelectric ceramics connected to the cantilever, the so-called height image of molecules can be made at a constant applied force. Because of its exceptional high signal/noise ratio, AFM is able to measure the topography of a single isolated molecule with a lateral resolution of a few nanometers and a vertical resolution of a few Angstroms [2].
Acquisition of high-resolution images of macromolecules in aqueous solution using AFM does not require sample staining [3]. Development of AFM imaging techniques in life sciences is progressing [4] including imaging single isolated molecules at high speed [5]. To date very high-resolution imaging by AFM has been obtained on membrane proteins 2 D crystals [6,7] as well as on densely packed proteins in native membranes [8][9][10]. Although AFM is primarily an imaging tool, it also allows measuring inter-and intra-molecular interactions on the pico-Newton scale [11][12][13][14] and force-probing the surfaces of living cells at the single molecule level [15,16] in order to map protein receptors for example [17].
Image quality is highly involved in feature interpretation and extraction for sampled objects in all kinds of imaging systems. Noise is a critical artifact that influences image quality and is mainly produced during image acquisition. It needs to be removed or reduced for further data processing to acquire desirable information or feature interpretation. There is no exception for AFM images. Among all types of noises, stripe noise is the notorious one that profoundly degrades image quality acquired by AFM, and is a consequence of the scanning pattern. During the scanning along one line of image, the sample surface height is acquired by an oscillating motion of the cantilever tip in the perpendicular direction to the substrate plate. Stripe noise may occur, for example using the tapping mode of AFM [18], from a loss or inadequate acquisition of height information. Any abrupt increase of the force from samples exerted on the tip would make a dramatic change in the tip vibration such that the noise cannot stay constant during the scanning; moreover these noise errors cannot be averaged off. A critical factor causing this change is the interaction between the sample and the tip. Thereby, these stripes vary in intensity, length, local density and frequency range, subjective to a variety of factors, such as sample preparation conditions and the constituents therein on the substrate plate. In particular, sharp and irregular boundaries or protuberances in the object distribution usually produce serious stripes across the image.
Noise is usually modeled as an additive term in the intensity distribution with a Gaussian or Poisson form of zero mean and constant variance [19]. This assumption becomes inapplicable if the image intensity is nonrandomly distributed. From the noise origin described above, the high-density and fine stripes observed in AFM images do not occur randomly (neither in time nor in space). An image frequency spectrum can depict the noise characteristics [20]. In traditional strategies of Fourier transform, a filter with high or low pass is often set up to gate the frequencies composed of stripe noises. However, this method sometimes removes some image details when stripe noises are mixed with components from object textures. Recently, a combination of wavelet and Fourier transform has been developed to remove both stripe and ring artifacts from images [21]. Alternatively, a method based on the heterogeneity of image frequency spectrum has been proposed to remove periodic and quasi-periodic stripes [22]. However, as shown in this paper, the stripe noises observed in AFM images are not in a unique or specific form. Although we adopted a similar approach for developing the denoising protocol, some new designs are highlighted for the removal of non-uniform and high-density stripe noises from AFM images.
Implementation
To be visualized by the human eye, the denoising procedure was not performed straightforwardly on the spectrum amplitude but on its logarithmic scale, the corresponding image of which is called LogF. We proposed a divide-and-conquer strategy to proceed the denoising in two separate regions of the frequency domain, i.e., the central and the off-center regions, as most of the high intensities are concentrated in the neighborhood of the origin. In order to best preserve the original data and reduce the computational task, we employed a decision-based algorithm [23] to select pixels for the variance test [24] and intensity reconstruction. Heterogeneity measurements were performed for detecting a spectral pixel potentially responsible for stripe noise in the raw AFM image.
Regarding the image restoration, the intensity replacement was also performed in the image spectrum, and the processed image was obtained by the inverse Fourier transformation. In general, a smoothing procedure applied in the image spatial domain can be exactly used in its frequency domain, e.g., a median-like filter [22]. Yet, processing in the frequency domain is a better approach than in the spatial domain for that the edge sharpness can be better preserved. Moreover, we attempted choosing a size for the local window to restore the image intensity as small as possible. Consequently, the accuracy of image restoration is only jeopardized at a minimal degree in the case where pixels are falsely picked up as noisy. The design of our denoising protocol is outlined in Figure 1 and details are described below.
Heterogeneity function
The heterogeneity of LogF determines whether a pixel is noisy or not. The heterogeneity measurement ranges from 0 to 1; the larger the value, the more heterogeneous is the intensity. There are two components in the heterogeneous function, i.e., abrupt change in intensity and intensity itself. The former is represented by the Laplacian of LogF.
where L(i, j) and I(i, j) represent the Laplacian and intensity values at pixel (i, j), respectively, L min , L max , I min and I max correspond to their maxima and minima. As a result, the pixels with higher H values possess higher intensity and experience more dramatic change in intensity.
Global sampling of pixels
A preliminary sampling of noisy pixels was done by thresholding the H value based on an internally determined value (H ref ) extracted from the heterogeneity histogram which is analog to the homogeneity level in removal of impulse noises [25]. For convenience, we denoted this set of sampled pixels as P n 1 and the algorithm is described below: 1. Calculate H values of LogF. 2. Form the histogram of H with 20 bins such that the variation of H in each bin is within 0.05.
3. Find the most spreading group of consecutive bins with non-zero populations in the histogram. 4. From the most spreading group, find the threshold bin in the direction of increasing heterogeneity. The threshold bin is defined as the first bin encountered such that the ratio of its population relative to that of the most populated bin is ≤ 0.5. Figure 1 The flow chart of DeStripe, see details in Implementation. The dash-lined box describes the steps of reducing the number of potentially noisy pixels. The denotation of pixel sets at various steps is stated in Implementation. Table 1 The discrete Laplacian operator used in DeStripe
Divide-and-Conquer Strategy
Due to dramatic variations in intensity in the central region of the frequency domain, we divided P n 1 into two groups; one was referred to as the central region and the other as the off-center region.
Formation of the central region
The central region was considered as a circular disk. The initial radius of the disk was derived from the moment of inertia tensor of P n 1, where the mass magnitude was replaced by the intensity value in the tensor array [26]. We calculated the eigenvalues (s x , s y ) and eigenvectors (ê x , ê y ) of the tensor array. The initial radius equals the square root of s x + s y . Starting from the center of intensity distribution, (i 0 , j 0 ), we expanded stepwise the region outwardly with an increment of 1/10 of the initial radius value. At each expansion step, we counted the P n 1 pixels and the total within the newly expanded region; if the ratio of the two numbers was ≤ 0.85, then the expansion was stopped and all the visited pixels P n 1 were included as members in the central region; we denoted it as C 0 and P n 1 -C 0 as P n 2.
Sampling of noisy pixels in the central region
In order to avoid vain data treatments and reduce the number of false noisy pixels recruitment, we modeled the intensity distribution of C 0 by an anisotropic Gaussian function. We used the Levenberg-Marquadt algorithm [27] to fit the intensity nonlinearly into the model function. In other words, we minimized the objective where I(i 0 , j 0 ), c 1 and c 2 are fitting parameters; I(i 0 , j 0 ) is the restored intensity at (i 0 , j 0 ), and c 1 and c 2 shape the anisotropic breadth of the Gaussian fitting; I(i, j) is the intensity at (i, j). The relationship between (s x ', s y ') and (s x , s y ) can be obtained through (ê x , ê y ) as [28] where θ is the angle that rotates counterclockwise the Cartesian x, y-coordinate axes aligned with (ê x , ê y ). Note that (s x , s y ) and (ê x , ê y ) were computed based on the right-handed rule while (s x ', s y ') were defined according to the convention of image presentation, i.e., the origin is located at the left-top corner of the image, the image j row and i column represent the x-and y-axes, respectively. For a pixel (i, j) C 0 , if f (i, j) ≤ 0, then the pixel was not considered as corrupted with noise, otherwise it was included in a set denoted by C n 1.
We selected data points in C n 1 with a 10-bin histogram using the same thresholding method as described previously for global sampling. The collected pixels were further screened and clustered based on two criteria: (1) if the pixel distribution in the form of horizontal or vertical line was > 2/3 of the region of interest; (2) if the length of consecutive pixels was ≥4. The qualified pixels with either condition were considered as noisy and formed the set, C n 2. The same procedure was also applied to the off-center region, P n 2; the resulting pixel set is denoted by P n 3. Maps of these sets are presented in additional file 1 Figure S1, for each study image. The imposed constraint is that only non-noisy pixels were counted for N var within the (2N S +1) × (2N S +1) local window, i.e., (i+i c , j+j c ) neither C n 2 nor P n 2.
Constrained variance (CVAR) test and intensity restoration
In the present work, N S = 1. Define std = var , if I(i c , j c ) -ave >s std , then I(i c , j c ) was replaced by ave, otherwise the original I(i c , j c ) was reserved. For P n 2, the CVAR test was not performed point by point; instead, we clustered the connected pixels and the test was performed starting at the boundary pixels of each cluster.
Filter function
The major component of DeStripe is the filter function that turns a noisy image into a clean one. Consider S(i, j) and S(i, j)' as the measured and restored intensity values at pixel (i, j) in the LogF image, respectively, then the filter function was calculated as Φ( , ) value range is (0, 1], i.e. 0 < Φ(i, j) ≤ 1. Accordingly, the restored and the associated noise images were obtained by the inverse FFT of the products of exp[S(i, j)] with Φ (i, j) and 1-Φ(i, j), respectively. The image formed by the Φ(i, j) values is henceforth called Φ-image.
Results and Discussion
We present two biological systems of which the topographic images were measured by AFM. One contains biomembranes and the other is constituted of proteins belonging to a large family of hydrolase enzymes, GTPase. We applied DeStripe to these raw AFM images for stripe noise reduction. The main purpose of the denoising is to better reveal molecular features in images that are distinguishable to human vision. Lastly, the applicability of DeStripe was also tested for stripe removal from images acquired by a scanning electron microscope, SEM [29].
Denoising vs. high-resolution AFM imaging
In high-resolution AFM imaging, one goal is to measure molecular topographies down to a sub-nanometer scale.
In most occasions, one may seek an AFM tip as small as possible for probing the sample surface; the usual radius for AFM tips is 5-10 nm [30]. However, we found that the AFM image quality also profoundly depended on the study system. In other words, one set of instrumentation parameters may not guarantee to gain similar quality for images with different biological constituents, as the sample-related stripe noise is always a major hindrance. The denoising results of AFM imaging on biomembrane surfaces are shown in Figure 2. Two sets of images are arranged in two columns to indicate that the topographic measurements were independently performed with different experimental preparations and AFM setups. The representation of restored and noise images sharing the same intensity range with raw images is referred to additional file 1 Figure S2. We present here no details in AFM instrumentation or sample preparation, implying no need to know how this image was obtained for denoising. The top row presents the AFM raw images while the second and third rows illustrate the corresponding restored and noise images. The last row shows the Φ-image. Similar presentation for GTPase enzymes is presented in Figure 3.
We found that the Gaussian model was appropriate for fitting the intensity distribution of the central region, and for assigning an appropriate value to the center pixel, (i 0 , j 0 ). A strong noise occurred at the center pixel usually leads to the appearance of heavy or notorious stripes in the image. The Gaussian model greatly reduced the number of pixels processed; one may notice much less pixels restored in the central region than in the off-center region. Naturally, one may speculate that there is more chance to recruit false noisy pixels in the off-center region. Recall that the FFT amplitude in the off-center region is much smaller and more homogeneous than that in the central region. The replacement of intensity values at false noisy pixels may not dramatically affect the image quality; that is the underlined rationale for dividing the candidate pixels into two regions.
In structural biology, the distribution of molecules is one feature of concern for AFM imaging. In general, observable individuality is a prerequisite prior to interpretation of molecular features for imaged objects. One may overlook important details simply because the entire image looks so dim due to the presence of stripe noises with very high intensity. The consequence of stripe removal can be evidenced by comparison of raw and processed images. First, the visibility of fine structures in the image was enhanced by the protocole. The left system in Figure 2 clearly demonstrates this effect. The restored image revealed that the dark areas in between bright segmental regions in the raw image were in fact distributed with granular particles. Second, the particle shape or cluster form became better observed. After noise reduction, stripes were eliminated or separated into short segments, and some brightly fused regions or islands were resolved into assemblies of individual granular particles; even the individual shape of background particles was noticeable.
Recently, enhancing image contrast by subtraction of a smoothed image from the raw image has been used for better visualizing the image objects [31]. The subtraction method is equivalent to extraction of edge features. In the present study, the purpose of denoising is to remove the noise prior to any other image processing and enhanced visibility on existent features in the image is a natural result.
Noise vs. the image quality of AFM
For the same biological molecule but prepared in different conditions, the occurring pattern of noise is dissimilar as illustrated in Figures 2 and 3. Note that stripes can be due to horizontal or vertical noises [32]. In AFM surface measurements, vertical stripe is the major pattern of noise that intensely affects image quality. To a worse degree, their intensity distribution is not uniform, as seen in the raw image on the left of Figure 2. This complicates noise characterization and estimation. Likewise, these inhomogeneous stripes are also observable form the right noise image of Figure 3. In the right image of Figure 3 two observed dark bands are however due to scan lines misalignments along the height ordinate. DeStripe does not aim to remove such type of noise, yet it can be eliminated or reduced by flattening the image [31].
The noise images in Figures 2 and 3 present the noise components peeled off from the raw image. Comparing AFM images on different systems, the degree of noise removal by DeStripe was found to vary. This reflects in the distribution of potentially noisy pixels and the values in the Φ-image. From the intensity distribution of Φ-image, one may perceive the noisy pixels identified and the degree of intensity modification. The smaller intensity value in the Φ-image, the greater proportion of the spectrum amplitude is removed. The performance of DeStripe in denoising can be evaluated by visual inspection of the noise image. Our results show that the stripe noise pattern is very different from one system to the other. It implies that DeStripe is able to automatically tune the denoising performance. It is noteworthy that here all the images processed by DeStripe use the same set of parameter values; these values were chosen by trial and error such that the denoising can be effective for various images. Consequently, the only userprovided input is the raw image.
Taken together with restored images, we found that DeStripe under-denoises somewhere but over-denoises elsewhere in the raw image. On the one hand, it is trivial to diagnose a case of under-denoising if there is any stripe noise visible in the restored image. By lowering the heterogeneity threshold to include more candidate pixels in C n 2, we found that the visible stripe noises can be further removed from the restored image (data not shown). We have also attempted to use lower values for restored intensities and some stripes diminished. On the other hand, if there is any non-stripe structure pattern in the noise image correlated with object features in the restored image, it may imply that DeStripe cut off too much the image intensity. In order to obtain the true surface measured by AFM, the structure pattern of noise image can be used as a guide for judging whether the denoising is appropriate. All the noise images presented here exhibit almost purely stripe noises, mainly in linear form. Indeed, a compromise between noise reduction and preservation of structural features remains challenging.
By comparing intensity range between restored and raw images, our results yield a smaller size for restored images than for the raw ones. Nevertheless, except for the left system in Figure 2 all other systems show comparable sizes for both images. From the results of the left system in Figure 2 the range of restored intensities is almost half that of the raw ones; it implies that near 50% of the intensity magnitude measured at some pixels by AFM is attributed to the noise. Certainly, those topographic measurements cannot represent a true value for the molecular surface; one can no longer consider these topographic measurements as true surface heights if heavy and bright stripe noises contaminate the image at such level. Consequently, the denoising is inevitable for better characterizing the surface feature of AFM images for a realistic biological system.
Comparisons of denoising effects with Gwyddion
Gwyddion is open-source software for AFM image processing [33]. We applied the "correct lines" tool for comparing the denoising effects on high-density stripe removal. The results from Gwyddion are presented in Figure 4. By visual inspection, Gwyddion scarcely removed any noise from the image presented in row 4; this is observable either by comparing the restored and the raw images or from the intensity range of the noise image. As a matter of fact, the intensity values are essentially zero in the noise image on the right of Figure 3. The comparison reveals that DeStripe is more effective than Gwyddion on stripe reduction for AFM images. As a result, DeStripe manifests better the individuality of objects observed from the AFM image.
For topographic measurements, we assume that the noise is a positive quantity overlaid over the molecular surface; therefore the denoising effect should not yield restored images with greater intensity values than the raw measurements. In the present work, the processed images by DeStripe satisfy this important feature, revealing no other additive component imposed on the original image during the denoising procedure. This also reflects in the positive range of intensity in the noise images. In contrast, some noise intensities fall within the negative range in the denoising results from Gwyddion, see Figure 4. This indicates that Gwyddion may yield some restored intensities greater than the original data, thus creating new artifacts in the raw AFM image. Consequently, the two advantageous attributes, i.e., effective denoising and no extra artifact, make DeStripe superior to Gwyddion in improving the AFM image quality.
Application to SEM image
DeStripe performs denoising using an anisotropic Gaussian function to fit the spectral amplitude distribution close to the origin of the frequency domain. This modeling has been shown to provide an appropriate estimate of intensity for the origin of the frequency domain that is corrupted with noise. To further explore this aspect, we ran the program for a SEM micrograph. The results are presented in Figure 5. The originally measured image is seriously interfered with stripe noises, and we found that these severe stripes were mainly attributed to the contribution of the amplitude value at the origin. This ascertains that DeStripe strategy design is useful for images corrupted with heavy stripe noises. This image was processed using the same set of parameter values as for the AFM images.
Conclusions
AFM emerges as a new nanotechnology for investigating the structure of biological systems with mesoscopic sizes, ranging from cellular morphology to single protein topography. Although this technique can probe biomolecular surface down to a sub-nanometer scale, noise artifacts produced along with the measurements are also visible. We aimed to unveil the layer of noise from the image in order to witness the true topography beneath, and thus developed the automatic program DeStripe. By ripping off the noise contribution from the measured intensity, one obtains the surface heights closer to the true value and is able to inspect the surface features more easily; implicitly, the denoising processing enhances the molecular features of concern.DeStripe involves several methods for reducing the number of pixels for intensity restoration, including global sampling, non-local variance test, Gaussian fitting, and local CVAR test. Gaussian modeling revealed its superiority over conventional variance test in identifying the noisy pixels in the central region where intensity drastically changes. Moreover, the criterion of non-negative noise was insistently used for the reason that AFM imaging was directed for probing surfaces unexplored before; denoising simply for pleasing human eye is a false pursuit. | 5,913.6 | 2011-02-01T00:00:00.000 | [
"Physics"
] |
TNF, but not hyperinsulinemia or hyperglycemia, is a key driver of obesity‐induced monocytosis revealing that inflammatory monocytes correlate with insulin in obese male mice
Abstract Inflammation contributes to obesity‐related hyperinsulinemia and insulin resistance, which often precede type 2 diabetes. Inflammation is one way that obesity can promote insulin resistance. It is not clear if the extent of obesity, hyperinsulinemia, or hyperglycemia, underpins changes in cellular immunity during diet‐induced obesity. In particular, the requirement for obesity or directionality in the relationship between insulin resistance and monocyte characteristics is poorly defined. Inflammatory cytokines such as tumor necrosis factor (TNF) can contribute to insulin resistance. It is unclear if TNF alters monocytosis or specific markers of cellular immunity in the context of obesity. We measured bone marrow and blood monocyte characteristics in WT and TNF −/− mice that were fed obesogenic, high fat (HF) diets. We also used hyperglycemic Akita mice and mice implanted with insulin pellets in order to determine if glucose or insulin were sufficient to alter monocyte characteristics. We found that diet‐induced obesity in male mice increased the total number of monocytes in blood, but not in bone marrow. Immature, inflammatory (Ly6Chigh) monocytes decreased within the bone marrow and increased within peripheral blood of HF‐fed mice. We found that neither hyperinsulinemia nor hyperglycemia was sufficient to induce the observed changes in circulating monocytes in the absence of diet‐induced obesity. In obese HF‐fed mice, antibiotic treatment lowered insulin and insulin resistance, but did not alter circulating monocyte characteristics. Fewer Ly6Chigh monocytes were present within the blood of HF‐fed TNF −/− mice in comparison to HF‐fed wild‐type (WT) mice. The prevalence of immature Ly6Chigh monocytes in the blood correlated with serum insulin and insulin resistance irrespective of the magnitude of adipocyte or adipose tissue hypertrophy in obese mice. These data suggest that diet‐induced obesity instigates a TNF‐dependent increase in circulating inflammatory monocytes, which predicts increased blood insulin and insulin resistance independently from markers of adiposity or adipose tissue expansion.
Abstract
Inflammation contributes to obesity-related hyperinsulinemia and insulin resistance, which often precede type 2 diabetes. Inflammation is one way that obesity can promote insulin resistance. It is not clear if the extent of obesity, hyperinsulinemia, or hyperglycemia, underpins changes in cellular immunity during diet-induced obesity. In particular, the requirement for obesity or directionality in the relationship between insulin resistance and monocyte characteristics is poorly defined. Inflammatory cytokines such as tumor necrosis factor (TNF) can contribute to insulin resistance. It is unclear if TNF alters monocytosis or specific markers of cellular immunity in the context of obesity. We measured bone marrow and blood monocyte characteristics in WT and TNF À/À mice that were fed obesogenic, high fat (HF) diets. We also used hyperglycemic Akita mice and mice implanted with insulin pellets in order to determine if glucose or insulin were sufficient to alter monocyte characteristics. We found that diet-induced obesity in male mice increased the total number of monocytes in blood, but not in bone marrow. Immature, inflammatory (Ly6C high ) monocytes decreased within the bone marrow and increased within peripheral blood of HF-fed mice. We found that neither hyperinsulinemia nor hyperglycemia was sufficient to induce the observed changes in circulating monocytes in the absence of diet-induced obesity. In obese HF-fed mice, antibiotic treatment lowered insulin and insulin resistance, but did not alter circulating monocyte characteristics. Fewer Ly6C high monocytes were present within the blood of HF-fed TNF À/À mice in comparison to HF-fed wild-type (WT) mice. The prevalence of immature Ly6C high monocytes in the blood correlated with serum insulin and insulin resistance irrespective of the magnitude of adipocyte or adipose tissue hypertrophy in obese mice. These data suggest that diet-induced obesity instigates a TNF-dependent increase in circulating inflammatory monocytes, which predicts increased blood insulin and insulin resistance independently from markers of adiposity or adipose tissue expansion.
Introduction
Obesity is associated with chronic, low-grade systemic inflammation that also impacts endocrine control of metabolism (McPhee and Schertzer 2015). This metabolic inflammation (or metaflammation) can impair insulin action in tissues that participate in blood glucose control. The consequent insulin resistance is often matched by hyperinsulinemia, which independently promotes obesity and insulin resistance that generally precedes hyperglycemia and type 2 diabetes (T2D) (Mehran et al. 2012). In combination with other environmental and genetic factors, prolonged hyperinsulinemia, insulin resistance and metaflammation during obesity, coupled with insufficient insulin secretion, can lead to elevated blood glucose and the development of T2D (Ashcroft and Rorsman 2012).
Obesity alters endocrine and immunological communication between cells and tissues. For example, increased inflammatory mediators in adipose tissue contribute to reduced insulin sensitivity within adipocytes and other tissues that help control blood glucose. Endocrine and paracrine actions of cytokines, chemokines and adipokines produced by expanding adipose tissue and tissue-localized immune cells during obesity contribute to local and systemic inflammation and the development of insulin resistance (Weisberg et al. 2006;Lumeng et al. 2007a,b). The molecular mechanisms that connect insulin resistance and inflammatory cytokines/chemokines are widely studied. Comparatively little is known about how cellular immunity relates to compartmentalized, tissue-specific inflammatory cues and insulin resistance during obesity.
Accumulation of pro-inflammatory polarized macrophages within metabolic tissues is a key component of metabolic inflammation, as ablation or inhibition of macrophage inflammatory signaling can attenuate obesity-induced insulin resistance in adipose and liver tissue (Arkan et al. 2005;Cai et al. 2005;Han et al. 2013;Revelo et al. 2016;Desai et al. 2017). Adipose tissue macrophages in particular are derived from local proliferation, resident hematopoietic stem cells, and infiltration of peripheral blood monocytes (macrophage precursors) (Weisberg et al. 2003;Xu et al. 2003;Oh et al. 2012;Amano et al. 2014;Luche et al. 2017). Murine monocytes can be divided into subsets by their surface expression of the glycoprotein Ly6C: Ly6C À (also referred to as Ly6C low ) and Ly6C + (Ly6C int and Ly6C high ) (Geissmann et al. 2003;Gordon and Taylor 2005). Ly6C high monocytes mature into Ly6C À patrolling monocytes in the steady state but migrate to sites of acute inflammation during infection via CCL2/CCR2-dependent chemotaxis, extravasate into tissues, and differentiate into macrophages (Geissmann et al. 2003;Gordon and Taylor 2005;Serbina and Pamer 2006;Tsou et al. 2007). The proinflammatory cytokine TNF contributes to monocyte migration during infection by promoting CCR2 expression (Boekhoudt et al. 2003). CCR2 À/À mice are protected against obesity-induced macrophage accumulation, inflammation, hyperinsulinemia, and insulin resistance (Chen et al. 2005;Weisberg et al. 2006;Ito et al. 2008). TNF is produced by tissue-localized macrophages in obese mice, and its neutralization improves insulin sensitivity (Hotamisligil et al. 1995;Uysal et al. 1997). Thus, we hypothesized that during diet-induced obesity there is a TNF-dependent egress of Ly6C high monocytes from bone marrow into circulation, these monocytes contribute to tissue-associated macrophage accumulation, and they are associated with metainflammation and insulin resistance.
We initially assessed the effects of diet-induced obesity on monocyte prevalence in bone marrow and blood and their expression of inflammatory and maturity markers (Ly6C, F4/80) in wild-type (WT) male mice. We found that monocyte reprogramming during obesity begins in the bone marrow and is marked by increased circulating Ly6C À , Ly6C int , and Ly6C high monocytes in HF-fed mice. Using TNF À/À male mice as a model of reduced systemic inflammation, we demonstrated that despite diet-induced obesity in TNF À/À mice there was a reduction in circulating Ly6C high inflammatory monocytes and macrophage accumulation in adipose tissue in comparison to WT mice. More importantly, this discovery allowed for an assessment of insulin resistance characteristics in obese mice that had different levels of circulating Ly6C high monocytes. We determined that during diet-induced obesity the TNFdependent prevalence of blood monocytes, and inflammatory Ly6C high monocytes in particular, were better predictors of indices of insulin resistance than body weight or parameters of adiposity, such as adipocyte size. We found no evidence that this relationship is bidirectional since manipulation of insulin levels, insulin resistance, or blood glucose did not alter monocyte characteristics. Importantly, obesity-induced increases in circulating Ly6C high monocytes directly correlated with blood insulin levels irrespective of adipose tissue/cell expansion in obese mice.
Materials and Methods
Animals WT C57BL/6J male mice and TNF À/À male mice on a C57BL/6J background were originally purchased from The Jackson Laboratory (WT no. 00064; TNF À/À no. 003008; Bar Harbor, ME) and bred at the McMaster Central Animal Facility (Hamilton, ON, Canada), as described (Zganiacz et al. 2004;Puchta et al. 2016). Heterozygous C57BL/6-Ins2 Akita /J (Akita +/À ) male mice were purchased from The Jackson Laboratory (no. 003548). All animals were housed under specific pathogen-free conditions with a 12-hour light/dark cycle. Diets were manipulated at 8 weeks of age. WT and TNF À/À male mice (not littermates) were allocated to either ad libitum standard chow diet (18% kcal from fat; Envigo Teklad Diets 7913, Madison, WI) or an obesogenic, low fiber, high fat (HF) diet (60% kcal from fat; Research Diets Inc. D12492, New Brunswick, NJ) for 18-24 weeks. Akita +/À male mice and WT male mice were used for hyperglycemia experiments at 8 weeks of age. WT male mice used for hyperinsulinemia experiments were maintained on an ad libitum standard chow diet. To induce hyperinsulinemia, subcutaneous implantation of insulin pellets, or sham surgery was performed on WT male mice at 10-12 weeks of age as recommended by the manufacturer (~0.1U/day release; LinBit for Mice Pr-1-B, LinShin Canada Inc., Toronto, ON, Canada). WT male mice used for antibiotic microbiota depletion experiments were maintained on HF diet for 20 weeks prior to antibiotic treatment, which consisted of 1 g/L ampicillin (Sigma A6140) and 0.5 g/L neomycin (Sigma N1878) in drinking water. Antibiotics were changed every 2 days over 4 weeks. All experiments were performed in accordance with Institutional Animal Utilization Protocols approved by McMaster University's Animal Research Ethics Board following the recommendations of the Canadian Council for Animal Care. Data from each individual mouse are indicated by a single symbol in all figures.
Metabolic assessments
Mice were fasted for 6 h with ad libitum water. Blood glucose and insulin were measured to calculate HOMA-IR as we have done previously (Cavallari et al. 2017). Fasting blood glucose was measured via tail vein using the Accu-Chek Inform II system glucometer and test strips (Roche Diagnostics, Mississauga, ON, Canada). Blood (50 lL) was collected via tail vein, incubated at room temperature for 20 min, and spun at 7500g for 5 min at 4°C. Insulin was quantified in serum by ELISA according to the manufacturer's instructions (
Immunohistochemistry
Adipose tissue macrophages were analyzed as published (Denou et al. 2015). Epididymal fat pads were fixed in 10% formalin at room temperature and embedded in paraffin. 5lm sections cut at 50-lm intervals were mounted on positively charged glass slides, deparaffinized in xylene, treated with proteinase K, and stained with an anti-F4/80 (1:500) monoclonal antibody (Bio-Rad Antibodies MCA-497) on the Leica Bond RX automated staining system. Adipocyte size and total number of F4/80-expressing cells (macrophages) were assessed using ImageJ (Schneider et al. 2012).
Statistical analyses
Monocyte population prevalence and phenotype differences (for individual monocyte populations) were evaluated with the Mann-Whitney U test or Student's t test according to normality by the D'Agostino and Pearson omnibus test. Differences in HF-fed TNF À/À and WT adipose tissue macrophages, adipocyte cross-sectional area, and fasting blood glucose were evaluated with the Mann-Whitney U test or Student's t test according to normality. Body mass and metabolic parameter comparisons by genotype and diet were performed by two-way ANOVA with Tukey's post hoc test. Correlation analyses between adipose tissue characteristics, monocytes, and metabolic parameters, were performed using Spearman's rank correlation rho or Pearson correlation according to normality. Data were analyzed with GraphPad Prism version 6 (GraphPad Software, La Jolla, CA). A P < 0.05 was considered statistically significant.
Diet-induced obesity alters monocyte populations in bone marrow and blood
Leukocytosis, an increase in white blood cells, occurs in obese and/or diabetic humans (Kullo et al. 2002;Ohshita et al. 2004;Tong et al. 2004). More recently, an elevation in circulating monocytes (monocytosis) has been associated with poorly controlled T2D, cardiovascular disease, and increased adiposity (Poitou et al. 2011;Barrett et al. 2017). Similar to Nagareddy and colleagues (Nagareddy et al. 2014), we found that WT male mice fed a HF diet for 18 weeks had a significant increase in circulating leukocytes in peripheral whole blood (Fig. 1A). HF-fed mice had an elevated ratio of circulating monocytes to lymphocytes in comparison to chow-fed control mice (Fig. 1B), due to a disproportionate increase in the prevalence of circulating monocytes (Fig. 1C) and a small decrease in the prevalence of circulating neutrophils (Fig. 1D). To further investigate the effect of diet-induced obesity on monocyte characteristics, we assessed their prevalence and expression of surface markers of inflammation (Ly6C) and maturity (F4/80). When we assessed the circulating monocyte subsets according to their expression of Ly6C it was apparent that male mice on a HF diet had elevated circulating populations of Ly6C À , Ly6C int , and Ly6C high monocytes (Fig. 1E). All monocyte subsets were increased more than twofold (as a proportion of total leukocytes) in the circulation of HF-fed mice. Absolute cell counts were also elevated in HF-fed mice (data not shown). Ly6C high monocytes were significantly less mature (decreased F4/80 expression) in HF-fed mice ( Fig. 1F) and Ly6C high had higher surface expression of CCR2 in comparison to the other monocyte subsets (data not shown) (Chen et al. 2005;Tsou et al. 2007). Intracellular staining also indicated that circulating Ly6C high monocytes from HF-fed mice produced higher levels of the pro-inflammatory cytokine IL-6 in response to LPS stimulation (Fig. 1G). Thus, diet-induced obesity in male mice is accompanied by an increase in circulating, immature inflammatory Ly6C high monocytes. It was previously demonstrated that HF-fed mice have increased proliferation of the hematopoietic common myeloid progenitors within bone marrow (Nagareddy et al. 2014). We Each data point indicates a single mouse. Statistical significance was determined by Mann-Whitney tests. Data are presented as box and whiskers plots, minimum to maximum, where the center line represents the median. *P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001. MFI -Geometric Mean Fluorescence Intensity. (M) flow cytometry gating strategy for the identification of leukocytes, neutrophils, and monocytes (bone marrow-resident and circulating). Representative images from a blood sample are shown. A width gate was created to exclude cell aggregates. CD45 + cells (leukocytes) were first gated. Subsequently, CD11b +/À and AF700 +/À population gating allowed separation of CD11b + Ly6G À (monocytes), CD11b + Ly6G + (neutrophils), and CD11b mid/ À CD3 + CD19 + NK1.1 + (lymphocytes: T cells, B cells, NK cells) cell populations. The CD45 + CD11b + Ly6G + Ly6C + SSC high cells were identified as neutrophils and the CD45 + CD11b + Ly6G À Ly6C + SSC low monocyte cells were divided into subsets by their expression of Ly6C: Ly6C À , Ly6C int , and Ly6C high . determined that leukocyte numbers and monocyte prevalence, as well as the prevalence of neutrophils, were unchanged within bone marrow of HF-fed mice ( Fig. 1H-J). However, Ly6C À and Ly6C int monocytes were elevated whereas Ly6C high monocytes were decreased in the bone marrow of HF-fed mice (Fig. 1K). In addition, bone marrow-resident Ly6C high monocytes in HFfed mice had lower expression of the F4/80 maturity marker (Fig. 1L). These data are consistent with the concept that diet-induced obesity promotes egress of immature, inflammatory Ly6C high monocytes from the bone marrow into circulation. The flow cytometry gating strategy is illustrated in Figure 1M, and representative plots of monocyte populations in chow-fed and HF-fed mice are shown in Figure 1N. We subsequently assessed whether specific metabolic factors (such as glucose, insulin, insulin resistance or extent of obesity/adiposity) were associated with elevated circulating Ly6C high monocytes.
Elevated blood glucose or insulin alone is insufficient to alter Ly6C high monocytes
The HF diet model used leads to increased adiposity, blood glucose, insulin, and insulin resistance (Cavallari et al. 2017). We next tested if elevation of blood glucose or insulin alone could increase circulating inflammatory monocytes. Chow-fed Akita +/À mice (C57BL/6-Ins2 Akita /J) were used to model diabetes to test the effect of hyperglycemia on monocyte characteristics, in the absence of obesity (Yoshioka et al. 1997;Gurley et al. 2006). Random fed blood glucose was significantly elevated in chow-fed Akita +/À mice in comparison to chow-fed WT mice ( Fig. 2A). No differences were identified in the percentage of circulating total monocytes or monocyte subsets ( Fig. 2B and C). Ly6C high monocyte surface expression of F4/80 was unchanged in hyperglycemic Akita +/À mice (Fig. 2D). CCR2 expression was also unchanged in hyperglycemic Akita +/À mice (Fig. 2E). However, hyperglycemic Akita +/À mice had significantly elevated neutrophils in circulation (Fig. 2F). We next tested the effects of short-term hyperinsulinemia that was sustained for 2 weeks after implanting slow release insulin pellets in WT chow-fed mice. The increased insulin load reduced random fed blood glucose (Fig. 2G), but monocyte percentages were similar between sham and insulin pellet implanted mice ( Fig. 2H and I). Expression of F4/80 and CCR2 on Ly6C high monocytes was also unchanged in mice with insulin pellets (Fig. 2J and K).
To contrast, we observed that mice implanted with insulin pellets had decreased circulating neutrophils (Fig. 2L).
These results suggest that overt hyperglycemia, and (short-term) chronic hyperinsulinemia are not sufficient to account for the increase in circulating immature, inflammatory Ly6C high monocytes that we observed in HF diet-fed mice.
Antibiotic-mediated lowering of glucose and insulin during obesity are insufficient to alter circulating Ly6C high monocytes We also considered whether circulating Ly6C high monocytes can be modulated by glucose and insulin in the context of diet-induced obesity. We used our previously established antibiotic treatment protocol in mice that modestly reduces peripheral glucose and insulin within 4 weeks (Denou et al. 2015). Mice with established dietinduced obesity after 20 weeks of HF diet feeding were allocated to receive water as usual or drinking water supplemented with antibiotics for 4 weeks. Antibiotic treatment did not alter body weight (Fig. 3A) or adiposity (Fig. 3B). Fasting blood glucose was lower after 4 weeks of antibiotic treatment (Fig. 3C). Further, fasting blood insulin decreased at 2 weeks and 4 weeks of antibiotic treatment (Fig. 3D). Insulin resistance, measured by the homeostatic model assessment of insulin resistance (HOMA-IR), was lower at 2 weeks and 4 weeks of antibiotic treatment (Fig. 3E). Total circulating leukocyte numbers were unaffected by antibiotic treatment (data not shown). The prevalence of circulating Ly6C high monocytes was not altered by antibiotics at 2 and 4 weeks (Fig. 3F). These results were consistent with our previous data showing that manipulation of glucose and insulin is not sufficient to alter the prevalence of circulating Ly6C high monocytes.
TNF contributes to inflammatory Ly6C high monocyte prevalence during diet-induced obesity We used TNF À/À mice to examine the role of TNF in mediating obesity-associated changes to inflammatory Ly6C high monocytes. Consistent with previous studies (Uysal et al. 1997;Ventre et al. 1997;Koulmanda et al. 2012), we initially confirmed that TNF À/À mice are partially protected from a HF diet-induced increase in body mass, as well as insulin resistance assessed by HOMA-IR ( Fig. 4A and B). Lower insulin resistance was due to lower fasting serum insulin rather than changes in blood glucose in HF-fed TNF À/À mice ( Fig. 4C and D). We next showed that HF diet-fed TNF À/À mice had significantly fewer adipose tissue-resident macrophages compared to HF diet-fed WT mice (Fig. 4E). When we examined peripheral monocyte populations, we found there was a twofold reduction in the proportion of circulating monocytes relative to lymphocytes in HF-fed TNF À/À mice (Fig. 5A). This difference in the ratio of 7) and Akita +/À (n = 5) mice. (A) random fed blood glucose in WT and Akita +/À mice. (B and C) total monocytes and Ly6C À , Ly6C int , and Ly6C high monocyte subsets as a percentage of total leukocytes (CD45 + cells) in WT and Akita +/À mice. (D and E) Ly6C high monocyte F4/80 and CCR2 surface expression in WT and Akita +/À mice. (F) total neutrophils as a percentage of total leukocytes in WT and Akita +/À mice. Effects of hyperinsulinemia on circulating monocytes were assessed in peripheral blood of sham (n = 8) and insulin pellet implanted (n = 7) chow-fed, WT male mice. (G) random fed blood glucose preimplantation and 2-weeks post-insulin pellet implantation.
(H and I) total monocytes and Ly6C-expressing monocyte subsets as a percentage of total leukocytes (CD45 + cells) in mice 2 weeks after sham and post-insulin pellet implantation. (J and K) Ly6C high monocyte F4/80 and CCR2 surface expression. (L) total neutrophils as a proportion of total leukocytes in mice 2 weeks after sham and post-insulin pellet implantation. Each data point indicates a single mouse. Two-tailed Mann-Whitney U tests were used to assess statistical significance between diet groups. Data are presented as box and whiskers plots, minimum to maximum, where the center line represents the median. **P ≤ 0.01, ***P ≤ 0.001. MFI: Geometric Mean Fluorescence Intensity. monocytes/lymphocytes in HF-fed WT mice versus HFfed TNF À/À mice is analogous to the ratio of monocytes/ lymphocytes in HF-fed WT mice compared to chow-fed WT mice (Fig. 1B). Accordingly, total and Ly6C high monocytes were lower in the circulation of HF-fed TNF À/À mice compared to HF-fed WT mice ( Fig. 5B and C), though Ly6C high monocyte prevalence was similar in HF-fed WT and TNF À/À mouse bone marrow (Fig. 5D).
No differences were observed in circulating Ly6C high monocyte maturity (Fig. 5E) or CCR2 expression between HF-fed WT and TNF À/À mice (Fig. 5F). These data indicate that TNF contributes to obesity-associated changes in circulating Ly6C high monocyte prevalence in addition to hyperinsulinemia and macrophage accumulation in metabolic tissues. Indeed, we observed a strong positive correlation in these HF-fed mice between circulating monocytes and body weight, as well as circulating monocytes and macrophage accumulation in adipose tissue ( Fig. 5G and H).
Ly6C high monocytes correlate with insulin during obesity
The relationship between adiposity and insulin sensitivity is complicated by many factors, including the lipid storage location (Wagenknecht et al. 2003;Muller et al. 2012). In mice, the mass of specific fat pads and adipocyte cross-sectional area have previously been shown to mark specific aspects of diet-induced adiposity (Weisberg et al. 2003;Chusyd et al. 2016). We assessed if body weight, gonadal adipose tissue mass, or adipocyte hypertrophy (i.e., cross-sectional area of adipocytes), correlated with serum insulin levels in HF-fed WT and TNF À/À mice ( Fig. 6A-C). We combined HF-fed WT and TNF À/À data and observed that there was a positive correlation between total body weight and insulin. In contrast, neither epididymal adipose tissue mass nor adipocyte crosssectional area, as markers of adiposity, correlated with fasting blood insulin. In fact, adipocyte cross-sectional area was not different in the gonadal fat pads of HF-fed WT and TNF À/À mice (Fig. 6D). Given the importance of adipose tissue macrophages in driving obesity and metabolic dysfunction (Olefsky and Glass 2010;Osborn and Olefsky 2012), and our observations of the effects of HF diet on the prevalence of circulating monocytes, we next examined whether macrophage and monocyte characteristics related to insulin. The number of adipose tissue macrophages did not correlate with fasting blood insulin (Fig. 6E). However, the prevalence of blood Ly6C high monocytes positively correlated with fasting blood insulin ( Fig. 6F and Table 1). Given the direct correlation of insulin and blood Ly6C high monocytes it was logical that we also found an association between Ly6C high monocytes and HOMA-IR (Table 1). Correlations were also identified between total, Ly6C À and Ly6C int monocytes and fasting blood insulin. We did not observe an association between Ly6C high monocytes and fasting blood glucose ( Fig. 6G and Table 1). These data indicate that there is a direct association between serum insulin and the prevalence of circulating immature inflammatory Ly6C high monocytes rather than markers of adiposity such as fat pad mass or adipocyte size in obese male mice.
Discussion
In this study we examined the effects of diet-induced obesity on monocyte prevalence and phenotype. We found that our model of long-term diet-induced obesity elevates circulating leukocytes and results in an expansion of peripheral monocytes. We identified that alterations to Ly6C high monocyte phenotype during obesity are TNFdependent, begin in the bone marrow, and lead to a higher prevalence of circulating immature and migrationprimed inflammatory Ly6C high monocytes that positively correlate with serum insulin, linking peripheral immune function with endocrine dysregulation. Monocytosis, an increase in peripheral blood monocytes, has been recognized for many years to coincide with inflammation (Dutta and Nahrendorf 2014), and it has been more recently recognized to contribute to chronic . TNF contributes to insulin resistance and adipose inflammation during obesity. WT and TNF À/À male mice were allocated to a chow or HF diet for 24 weeks. Body weight (A) and HOMA-IR (B) determined at endpoint for chow-fed and HF-fed WT and TNF À/À mice. Fasting blood insulin (C) and glucose (D) in HF-fed mice WT and TNF À/À mice. (E) quantification of adipose tissue-resident macrophages and representative images of immunohistochemistry staining for macrophages (F4/80-positive cells) in gonadal adipose tissue of HF-fed WT and TNF À/À mice. Data were derived from two independent cohorts of mice, where n = 5-9 mice per group. Each data point indicates a single mouse. Two-way ANOVA with Tukey's post test was performed for A and B, where inset P values represent main effects and bars with different letters indicate significant differences. Significance was assessed by Mann-Whitney test for C to E. Data in C to E are shown as bar graphs with mean +/À standard deviation. **P ≤ 0.01, ***P ≤ 0.001. low-grade inflammation during obesity (Schmidt et al. 1999;Ford 2002;Kullo et al. 2002). We observed a decrease in Ly6C high monocytes in the bone marrow and an increase in this population in circulation in HF diet-fed obese male mice that are well known to have mild hyperglycemia and hyperinsulinemia. Circulating monocytosis, and a particular elevation of Ly6C high monocytes, has also been reported in Ob/Ob mice (Nagareddy et al. 2014). Ly6C high mouse monocytes are comparable to human classical (CD14 + CD16 À ) and intermediate (CD14 + CD16 + ) monocytes (Gordon and Taylor 2005). Classical monocytes both in circulation and adipose tissue increase with obesity (G) correlation of fasting blood glucose and the prevalence of circulating Ly6C high monocytes in all HF-fed mice. Data in A and B are from two independent cohorts of HF-fed mice (n = 7-9/genotype). Data in C-E are from a subset of HF-fed mice from those cohorts (n = 6/genotype). Data in F and G are from one cohort of HF-fed mice (n = 7/genotype). Each dot is a mouse. Data in D are shown as a bar graph with mean AE standard deviation. Correlations for A-C and E-G were determined by Spearman or Pearson's tests and Mann-Whitney U test was used to assess D. (Wouters et al. 2017), and their activation has also been reported in diabetics (Cipolletta et al. 2005). An increase in intermediate monocytes in obese humans has been linked to an increased risk of subclinical atherosclerosis and cardiovascular events (Seidler et al. 2010;Poitou et al. 2011;Rogacev et al. 2012;Ziegler-Heitbrock 2015), as well as insulin resistance (Krinninger et al. 2014). Expansion of the circulating intermediate monocyte population has also been reported in T1D and T2D patients (Mysliwska et al. 2012;Terasawa et al. 2015;Ren et al. 2017).
In addition to assessing circulating monocyte prevalence, we assessed how obesity affected their progression toward a mature tissue macrophage phenotype by examining their expression of F4/80, a surface marker related to maturity. Our observations of decreased F4/80 expression on the surface of Ly6C high monocytes in bone marrow and blood of HF-fed mice suggest that these monocytes are less mature. These immature monocytes are known to migrate to tissues in the context of infection via the CCR2/CCL2 chemotactic axis and thus may be primed for migration into metabolic tissues (Weisberg et al. 2006;Tsou et al. 2007;Ito et al. 2008). There were lower Ly6C high monocytes in the bone marrow and higher Ly6C high monocytes in blood, indicating that diet-induced obesity may alter egress of inflammatory monocytes from the bone marrow into circulation.
We have previously demonstrated a role for TNF in mediating dysfunction of monocyte development and function with aging, as well as increasing susceptibility to infection (Puchta et al. 2016). In addition to their association with increased circulating inflammatory CD14 + CD16 + intermediate monocytes, high levels of circulating TNF are characteristic of obesity as well as chronic inflammatory conditions. Anti-TNF therapies used to treat rheumatoid arthritis can decrease risk of developing T2D (Klaasen et al. 2011;Solomon et al. 2011;Gremese et al. 2013). We expanded on the pleiotropic role of this pro-inflammatory cytokine by demonstrating its involvement in inflammatory Ly6C high monocyte egress from the bone marrow into circulation and accumulation of macrophages in metabolic tissues during obesity. TNF may induce NLRP3 inflammasome priming and activation (Alvarez and Munoz-Fernandez 2013;McGeough et al. 2017). Our data illustrating the importance of TNF in obesity-associated monocytosis agree with that of Nagareddy and colleagues, who found that obesity-associated changes in adiposity and monocytosis are dependent on the NLRP3 inflammasome, and in particular NLRP3-dependent production of IL-1b (Alvarez and Munoz-Fernandez 2013;Nagareddy et al. 2014;McGeough et al. 2017). The changes we observed to monocyte phenotype in TNF À/À mice were coincident with attenuation of hyperinsulinemia and insulin resistance and a slight reduction in body mass. Lower insulin levels underpinned lower insulin resistance (measured by HOMA-IR) in the HF-fed TNF À/À mice. This was expected as TNF is known to promote insulin resistance during obesity (Uysal et al. 1997;Ventre et al. 1997;Hivert et al. 2010;Koulmanda et al. 2012), but this mouse model also provided a unique model to correlate insulin and inflammatory monocytes during diet-induced obesity. Deletion of TNF reduced accumulation of macrophages in adipose tissue despite similar levels of adipocyte hypertrophy during HFD-fed feeding. Therefore, our data show that TNF is a key regulator of dietinduced changes to monocyte/macrophage-driven inflammation in obesity, and that TNF can contribute to metaflammation in a manner that is independent of adipocyte hypotrophy and regulation of glucose. These data also reinforced the concept that obesity-induced changes in inflammatory monocytes correlated with insulin rather than markers of adiposity such as adipocyte size.
Our data indicated that tissue-specific effects of the Ly6C high monocytes and/or metabolic tissue changes predicted the development of some aspects of insulin resistance, which led us to examine the association of monocyte phenotype and insulin levels. Comparisons of blood monocyte characteristics and circulating insulin levels in HF-fed WT and TNF À/À mice during dietinduced obesity demonstrated that circulating immature inflammatory Ly6C high monocytes were more strongly associated with serum insulin levels compared to indices of adiposity, total body weight, and other monocyte populations in male mice. Circulating Ly6C high monocytes appear positioned to predict or propagate hyperinsulinemia or insulin resistance during diet-induced obesity. A recent publication examining the role of CX3CR1 and Gr1 low monocytes (equivalent to our Ly6C À monocyte population) in diet-induced obesity, by using female mice deficient for CX3CR1 in blood or hematopoietic compartments, reported a negative association between Gr1 low monocytes and HOMA-IR (B eliard et al. 2017). Our data in male mice show an association between Ly6C À monocytes and insulin levels rather than HOMA-IR. Our data were determined through modulation of TNF-related inflammation in obesity rather than directly modifying monocytes, which may explain this discrepancy. In addition, there are likely sex-specific differences that alter monocyte characteristics in obesity that warrant further study.
Hyperglycemia and hyperinsulinemia have both previously been shown to alter monocyte epigenetic programming and function in conditions of infection and stress (Xiu et al. 2014), and it has recently been reported that insulin signaling in obesity has a significant role in mediating adaptive T cell inflammatory responses (Tsai et al. 2018). Consequently, the positive correlation that we observed of an increasing prevalence of inflammatory Ly6C high monocytes with increasing insulin resistance in male mice suggested insulin may mediate phenotypic changes to Ly6C high monocytes in obesity. We found that raising insulin was not sufficient to alter circulating Ly6C high monocytes in the absence of obesity. In surprising contrast to a previous study (Nagareddy et al. 2013), we found that hyperglycemic Akita +/À mice did not have changes to circulating Ly6C high monocytes. Our Akita +/À mice were assessed at 8 weeks of age whereas the previous study performed measures on mice at 12-16 weeks of age (Nagareddy et al. 2013). Longer exposure to elevated peripheral glucose may explain this discrepancy. We did confirm that circulating neutrophils were significantly higher in hyperglycemic Akita +/À mice (Nagareddy et al. 2013), and further showed that ongoing low-dose insulin exposure (i.e., from a slow-release implant) can reduce the proportion of circulating neutrophils, likely due to the accompanying reduction in blood glucose. We further demonstrated that reduction of peripheral glucose and insulin via antibiotic treatment in obese wild-type mice does not alter Ly6C high monocyte populations in circulation. We acknowledge in vivo manipulation of glucose and insulin is interdependent. Nevertheless, our data suggest that elevated blood glucose or insulin in the absence of cellular mediators, hormones, hyperlipidemia, and other factors that accompany obesity, are insufficient to alter monocyte prevalence or measurements of maturity (F4/80) or chemotactic potential (CCR2). We cannot exclude the possibility that dyslipidemia or hyperleptinemia may also contribute to the changes in monocyte maturity and phenotype that we observed with HF diet (Desai et al. 2017;Rahman et al. 2017;Short et al. 2017). A significant limitation of our data is that our WT and TNF À/À experiments were not conducted in littermate mice. Although metabolic data were consistent with previous data, our study design cannot rule out the possibility that differences in genetic background, microbiota composition, or husbandry may have influenced the results (Hotamisligil et al. 1995;Uysal et al. 1997;Ventre et al. 1997). The use of non-littermate mice is a weakness of this study to provide a direct role of TNF, but the association of inflammatory monocytes with insulin may span a diverse genetic background and should be further investigated in humans. Our current data support a model where obesity-related cellular mediators alter monocyte characteristics, which contribute to cellular inflammation and hormone regulation within metabolic tissues. Our data suggest directionality in the relationship between insulin and monocyte changes during obesity, where dietinduced changes in Ly6C high monocytes predict insulin and insulin resistance, but neither glucose, insulin, nor insulin resistance appears to alter these inflammatory monocytes during obesity.
Monocyte prevalence and phenotype have been proposed as biomarkers in cardiovascular disease and chronic inflammatory disorders (Yang et al. 2014;Chara et al. 2015;Meeuwsen et al. 2017;Loukov et al. 2018). Our data links monocyte/macrophage phenotype to changes in whole body insulin regulation due to diet-induced obesity and suggests that monocyte characteristics in obese individuals could serve as predictive biomarkers of diabetes risk and may represent a mechanism linking inflammation and insulin regulation. Modulation of low-grade inflammation through selectively targeting monocyte populations in obese individuals may therefore improve insulin sensitivity. Examining the role of all monocyte subsets in driving peripheral and tissue-specific metainflammation will improve our understanding of the development of hyperinsulinemia, insulin resistance, and type 2 diabetes in obese individuals. | 8,206.8 | 2018-12-01T00:00:00.000 | [
"Medicine",
"Biology",
"Environmental Science"
] |
Using Artificial Neural Network Condensation to Facilitate Adaptation of Machine Learning in Medical Settings by Reducing Computational Burden: Model Design and Evaluation Study
Background Machine learning applications in the health care domain can have a great impact on people’s lives. At the same time, medical data is usually big, requiring a significant number of computational resources. Although this might not be a problem for the wide adoption of machine learning tools in high-income countries, the availability of computational resources can be limited in low-income countries and on mobile devices. This can limit many people from benefiting from the advancement in machine learning applications in the field of health care. Objective In this study, we explore three methods to increase the computational efficiency and reduce model sizes of either recurrent neural networks (RNNs) or feedforward deep neural networks (DNNs) without compromising their accuracy. Methods We used inpatient mortality prediction as our case analysis upon review of an intensive care unit dataset. We reduced the size of RNN and DNN by applying pruning of “unused” neurons. Additionally, we modified the RNN structure by adding a hidden layer to the RNN cell but reducing the total number of recurrent layers to accomplish a reduction of the total parameters used in the network. Finally, we implemented quantization on DNN by forcing the weights to be 8 bits instead of 32 bits. Results We found that all methods increased implementation efficiency, including training speed, memory size, and inference speed, without reducing the accuracy of mortality prediction. Conclusions Our findings suggest that neural network condensation allows for the implementation of sophisticated neural network algorithms on devices with lower computational resources.
Introduction
Machine learning applications for health care can have a great impact on people's lives. Currently, the possibilities for machine learning in health care include diagnostic systems, biochemical analysis, image analysis, and drug development. One of the most significant challenges in using machine learning for health care applications is that data is usually huge and sparse, requiring important computational resources, especially for overparameterized deep neural networks (DNNs). Consequently, the availability of computational resources to use such tools can limit their widespread use, such as for people who live in low-income countries and for those who want to run diagnostic apps on their own mobile devices.
In this study, we set in-hospital mortality prediction as a case study to explore the various ways of improving efficiency (ie, training speed, memory size, and inference speed) of neural network-based algorithms. Mortality prediction is a well-tried medical machine learning application wherein the mortality of a patient after being transferred to the intensive care unit (ICU) can be predicted based on their vital signs, laboratory tests, demographics, and other factors. Mortality prediction is important in clinical settings because such a prediction can help determine the declining state and need for intervention. We built baseline models with either recurrent neural network (RNN) or dense neural network architectures, based on which we explored efficiency improvements via neural network condensation without sacrificing the prediction accuracy. An RNN is a class of artificial neural networks wherein connections between nodes form a directed graph along a temporal sequence that consider a sequence of input in a recurrent manner. RNNs are widely used in clinical informatics in tasks such as temporal data analysis and clinical natural language processing.
Reduction of complexity and improvement of efficiency of artificial neural networks is an active field of research, wherein a wide range of methods have been explored. One representative example is neural network pruning, wherein a fraction of weights is removed from the trained model and the "lottery ticket" is found when the remaining weight can still be quickly trained with competitive loss and accuracy [1][2][3]. There are more fancy pruning approaches where the authors use another neural network to learn and conduct the best pruning decisions considering the network to be pruned (ie, the backbone neural network). For example, Lin et al [4] developed a method called runtime neural pruning to model their pruning process as a Markov decision process and use reinforcement learning for training via an additional RNN. Zhong et al [5], on the other hand, used long short-term memory (LSTM) to guide an end-to-end pruning of the backbone neural network. Some other previous works have converted the neural network condensation into an optimization problem where parameters are penalized under some norm [6][7][8][9].One RNN-specific condensation method is that instead of embedding information into multiple recurrent layers, we only use one recurrent layer but extend the capacity of the RNN unit (cell) by incorporating more hidden layers within the cell. Dai et al [10] showed that DNNs were inserted between the recurrent layer and the input (masking) layer for each gate in the LSTM to form an LSTM embedded with hidden layers (ie, hLSTM). Such an architecture can, in principle, be more efficient (ie, fewer number of parameters and higher training speed). There is another posttraining condensation method called quantization, wherein parameters originally stored in a 32-bit floating point format are forcibly converted to 8 fixed bits [11]. Other methods used for neural network condensation include, but are not limited to, binarization of neural networks [12], knowledge distillation [13], and Huffman coding [14]. In this paper, we describe the use of hLSTM, neural network pruning, and quantization to condense the size of neural networks and increase speed while maintaining their prediction accuracy.
Intensive Care Unit Data
We used the Medical Information Mart for Intensive Care-III (MIMIC-III) critical care database for the implementation of our models [15]. In all, 53,423 distinct hospital adult patients admitted to critical care units between 2001 and 2012 are included in this database. We excluded all neonatal and pediatric patients (aged 18 years or younger at the time of ICU stay) because the physiology of pediatric critical care patients differs significantly from that of adults [16]. We also excluded any hospital admissions with multiple ICU stays or transfers between different ICU units. The final cohort comprised 33,798 unique patients, with a total of 42,276 hospital admissions and ICU stays. Of these 33,798 patients, we defined a test set of 5070 (15%) patient stays. In-hospital mortality was determined by comparing patient date of death with hospital admission and discharge times. The mortality rate within the cohort was 10.9%. The median age of adult patients was 65.8 (SD 11.3) years, and 55.9% (18,893/33,798) patients were male. A mean of 4579 (SD 721.7) charted observations and 380 (SD 215.8) laboratory measurements, as well as other static information, are available for each hospital admission.
Data Prepossessing
Data were collected from the MIMIC-III database. Only data from the first 48 hours were used as inputs in our analysis. For the purpose of this study, 76 features were selected for analysis (see examples listed in Textbox 1). Some features may appear multiple times (in different means or conditions) and are thus regarded as independent features. We resampled the time series into regularly spaced intervals. If there were multiple measurements of the same variable in the same interval, we used the value of the last measurement. We imputed the missing values using the previous value, if it exists, or a prespecified "normal" [16] value, otherwise. In addition, we added a binary mask input for each variable, indicating the time steps that contain a true (vs imputed) measurement [17]. Categorical variables were encoded using a one-hot vector at each time step. Then, the inputs were normalized by subtracting the mean and dividing it by the SD value. Statistics were calculated per variable after imputation of missing values.
RNN Model
Our RNN baseline model is designed as an RNN consisting of a masking layer, two LSTM layers, a dropout layer, and a dense output layer, as shown in Figure 1. We chose two layers of LSTM because, based on a literature review, we identified this structure to be the one with the best performance in the MIMIC-III mortality prediction work [16]. The masking layer masks (skips) the time step for all downstream layers if the values of input tensor at the time step are all equal to zero, which represents missing data for that time step. The first layer of LSTM takes in the original 76 features and generates a 16-feature hidden state based on the hidden state of the previous step and the new incoming observation. Then, such a hidden state is forward to the entrance of the second LSTM layer, which produces another 16-feature hidden state at each step. A dropout layer is followed by the last-step hidden state of the second LSTM layer to prevent complex coadaptations of the neurons. Finally, a dense layer is used to generate a soft 0/1 mortality prediction. The training was conducted using Adam algorithm with a dropout rate of 0.3 between layers and a learn rate of 0.001. In this study, hyperparameters were chosen by grid searching based on performance on the validation set.
hLSTM Model
Besides pruning upon RNN, we also tried another way by inserting an additional hidden dense layer into the inner gates of LSTM, which we called hLSTM, to improve the "power" of the LSTM. For a traditional LSTM, the inner structure is as follows: where * is the matrix product; ⊗ is the element-wise product; W represents recurrent kernel matrices of the gates; and b represents corresponding bias terms. Moreover, f, i, o, c, x, h and c represent the forget gate, input gate, output gate, vector for cell updates, input, hidden state, and cell state, respectively. Subscript t indicates the time step. For hLSTM, the recurrent layer in equation 1 is modified as follows:
Feedforward DNN Model
Our baseline feed forward artificial neural network-commonly called DNN-used in this project consists of three fully connected layers, a dropout layer, and an output layer. The fully connected layers have 256, 128, and 64 neurons, respectively, and they use rectified linear unit (ReLU) as the activation function. The dropout layer has a probabilistic dropout rate of 0.5. Sigmoid function was used as activation at the output layer.
Neural Network Pruning
All neural network prunings were conducted at the channel level, which means a neuron and all its inputs and outputs were removed from the model if the neuron is pruned. Keras surgeon library in python was used for pruning. In each layer, neurons were pruned if their mean weight across all inputs from the previous layer were below the set quantiles (ie, 25% and 50% in this study). The original model was trained for 1 epoch before pruning and was trained for another 19 epochs after pruning.
Neural Network Quantization
Quantization was applied on the DNN model post training. Parameters, including weights and activation, originally stored in a 32-bit floating point format were converted to 8 bits using TensorFlow Lite. A uniform quantization strategy was used, as previously described [11]. Considering the range of float point values in the model to be (F min ; F max ), all the floating-point values were quantized into the range (0; 255) as 8 bits in a uniform manner, where F min corresponds to 0 and F max corresponds to 255.
The quantization process is where x is the floating-point variable, x q is the quantized variable, and
Recurrent Artificial Neural Network Condensation: hLSTM and Pruned LSTM
Recurrent artificial neural networks (or simply, RNNs) are a group of machine learning models widely used in clinical settings that take sequential or time series information as the input. However, training of RNNs and running inference from RNNs are relatively computationally intensive. In order to enable the machine learning algorithms to be used on devices with limited computational power, such as those in high-income countries and on mobile devices, we used three strategies to reduce the storage size of the model and to increase the speed of training and inference ( Figure 2).
We built a baseline RNN using two layers of LSTM neurons to predict ICU mortality rates using MIMIC-III dataset [15]. After training, the baseline RNN model achieved a decent performance of AUROC of 0.85 (Table 2). The first strategy was to modify the LSTM cell to increase the representation power of each layer. We modified the original neural network structure and added an additional hidden layer into the original LSTM class, wherein one additional layer called "hidden kernel" was inserted between the input kernel and the recurrent kernel (see equation 4). By using this strategy, we replaced the old 2-layer LSTM with only one layer of hLSTM, such that we simplified the overall structure by trying to embed the same quantity of information in this single "condensate" layer. Both the baseline model and the hLSTM model with only one layer of hLSTM are trained under the same settings. The comparison of AUROC and accuracy is shown in Figure 3. The number of parameters for these two models are listed in Table 2. This simplified model with a single layer of hLSTM beats the baseline model 2-fold in training speed, achieving a 32% reduction in parameter numbers while simultaneously maintaining a higher AUROC at the same time.
Feedforward Neural Network Condensation: Pruning and Quantization
Feedforward neural network, or commonly called DNN if it has multiple hidden layers, is another widely used form of machine learning in clinical settings. We trained DNN with 3 hidden layers, consisting of 256, 128, and 64 neurons in each layer, to enable ICU mortality prediction. The baseline DNN achieved an AUROC of 0.82, using patient data collected within the first 48 hours after admission. We explored two methods to condense the size of the DNN. The first method, called pruning, used the pruning strategy as in RNN; for this purpose, 50% of the channels were pruned after the first epoch of training, the prediction accuracy of the pDNN maintained at the same level as the original DNN, and the inference speed doubled ( Table 3). The second strategy involved quantization, which refers to the process of reducing the number of bits that represent a number. In the context of this project, the predominant numerical format used was a 32-bit floating point. We used an after-training-quantization strategy to represent the parameters of the DNN model using 8-bit integers (ie, quantized DNN or qDNN). This method reduced storage size of the DNN model by 5 times without incurring significant loss in accuracy (Table 3). We also compared the overall performances of DNN condensation with those of RNN, as shown in Figure 3.
Discussion
In this study, we were able to use data from the MIMIC-III database [15] to train in-hospital mortality neural network models with high accuracy and conduct model condensation with different methods to gain efficiency (eg, memory size reduction and increased speed) without compromising accuracy.
We implemented different neural network architectures for both RNNs and dense neural networks; thus, our methods can add value in both settings. We pioneered RNN pruning with clinical implementation and our condensation treatments aiming at higher efficiency can be extended to other medical applications using similar data, and probably to nonmedical applications as well. In addition, in medical settings, model calibration is conducted after initial model training. Calibration can be conducted using various training schemes and early stopping strategies. The model condensation method proposed in this study significantly reduces the number of parameters and will help make model calibration easier. The major limitation of the neural network condensation method is that although our proposed method significantly reduces the sizes of different models and their computational costs in training, the final model sizes after condensation are still proportional to the original model sizes. Therefore, if further model size reduction is warranted, a combination of better model design and neural network condensation will be required. | 3,651.2 | 2020-05-28T00:00:00.000 | [
"Medicine",
"Computer Science"
] |
The SIC Question: History and State of Play
Recent years have seen significant advances in the study of symmetric informationally complete (SIC) quantum measurements, also known as maximal sets of complex equiangular lines. Previously, the published record contained solutions up to dimension 67, and was with high confidence complete up through dimension 50. Computer calculations have now furnished solutions in all dimensions up to 151, and in several cases beyond that, as large as dimension 844. These new solutions exhibit an additional type of symmetry beyond the basic definition of a SIC, and so verify a conjecture of Zauner in many new cases. The solutions in dimensions 68 through 121 were obtained by Andrew Scott, and his catalogue of distinct solutions is, with high confidence, complete up to dimension 90. Additional results in dimensions 122 through 151 were calculated by the authors using Scott's code. We recap the history of the problem, outline how the numerical searches were done, and pose some conjectures on how the search technique could be improved. In order to facilitate communication across disciplinary boundaries, we also present a comprehensive bibliography of SIC research.
I. INTRODUCTION
The problem of symmetric, informationally complete quantum measurements [1][2][3][4] stands at the confluence of multiple areas of physics and mathematics. SICs, as they are known for short, tie into algebraic number theory [5][6][7][8], higher-dimensional sphere packing [9], Lie and Jordan algebras [10,11], finite groups [12,13] and quantum information theory [14][15][16][17][18][19][20][21][22][23]. Without the study of SICs, one might think that the intersection of all these subjects would have to be the empty set. And yet, for all that, a SIC is a remarkably simple mathematical structure, as structures go. Consider the complex vector space C d . To a physicist, this is the Hilbert space associated with a d-level quantum system. Let {|ψ j } be a set of exactly d 2 unit vectors in C d such that | ψ j |ψ k | 2 = 1 d + 1 (1) whenever j = k. The set {|ψ j }, which can be associated with a set of pairwise equiangular lines through the origin, is a SIC. One can prove that no more than d 2 vectors in a d-dimensional Hilbert space can be equiangular. That is, if {|ψ j } is a set of vectors, and | ψ j |ψ k | 2 = α for every j = k, then that set can have at most d 2 elements. In addition, for a maximal set the value of α is fixed by the dimension; it must be 1/(d + 1). So, a SIC is a maximal equiangular set in C d ; the question is whether they can be constructed for all values of the dimension. Despite a substantial number of exact solutions, as well as a longer list of high-precision numerical solutions [4,8,24], the problem remains open. Exact solutions, found by hand in a few cases and by computer algebra software in the others, are known in the following dimensions: 28, 30, 31, 35, 37, 39, 43, 48, 124. ( The historical record of exact solutions has been spread over several publications [4,8,25,26]. For several years, the most extensive published set of numerical results went as high as dimension d = 67 [4]. Now, numerical solutions are known in all dimensions up to and including d = 151, as well as a handful of other dimensions up to d = 844. These numerical solutions were found using code designed and written by Scott, who extended the results of [4] through d = 121 using his personal computer over several years of dedicated effort. In addition, Scott found solutions in a set of dimensions (d = 124,143,147,168,172,195,199,228,259, 323) by taking advantage of particular simplifying assumptions that are applicable in those dimensions [24]. Further close study of these properties led to a solution for d = 844 [26]. Because dimension d = 121 was pushing the limits of what was computationally feasible without those simplifying assumptions, the authors calculated solutions in dimensions 122 through 151 by running Scott's code on the Chimera supercomputer at UMass Boston. In turn, Scott was able to employ another algorithm (outlined below) to refine the numerical precision of these results. The solutions from all of these search efforts are available together at the following website: An intriguing feature of the SIC problem is that some numerical solutions, if extracted to sufficiently high precision, can be converted to exact ones [8,25]. Most recently, this technique was used to derive an exact solution in dimension d = 48. Another interesting aspect is that the number of distinct SIC constructions varies from one dimension to another. (The sense in which two SICs can be equivalent will be discussed in detail below.) One reason computational research is valuable, beyond extending the list of dimensions in which SICs are known, is that it provides what is likely a complete picture for many values of the dimension. This is important for understanding the subtle connection between SICs and algebraic number theory [5], a connection that brings a new angle of illumination to Hilbert's twelfth problem [6,7].
SICs are so called because, thanks to the rules of quantum theory, a SIC in C d specifies a measurement procedure that can, in principle, be applied to a d-level quantum system. For example, a SIC in C 2 is a set of four equiangular lines, and it is a mathematical model of a measurement that a physicist can perform on a single qubit. The term "informationally complete"-the "IC" in "SIC"-means that if one has a probability distribution for the possible outcomes of a SIC experiment, one can compute the probabilities for the possible outcomes of any other experiment carried out on the target system [17]. So, while one can pose the question of their existence using pure geometry, SICs are relevant to applied physics. Indeed, SIC measurements have recently been performed or approximated in the laboratory [27][28][29][30][31][32][33][34], and they are known to be optimal measurements for quantum-state tomography [35].
A SIC provides a frame-more specifically, an equiangular tight frame-for the vector space C d . Given a finitedimensional Hilbert space H with an inner product ·, · , a frame for H is a set of vectors {v j } ⊂ H such that for any vector u ∈ H, for some positive constants A and B. The frame is equal-norm if all the vectors {v j } have the same norm, and the frame is tight if the "frame bounds" A and B are equal. The ratio of the number of vectors to the dimension of the space is known as the redundancy of the frame [36]. For more on this terminology and its history, we refer to Kovačević and Chebira [37,38]. In our experience, the language of frames is more common among those who come to SICs from pure mathematics or from signal processing than among those motivated by quantum physics.
Any vector in C d can be represented by its inner products with all the SIC vectors. In quantum physics, one also considers the set of Hermitian operators on C d . This set in fact forms a Hilbert space itself, with a dimension of d 2 , and the inner product given by the Hilbert-Schmidt formula A, B = tr(AB).
Rewriting the SIC vectors {|ψ j } as rank-1 projection operators, we construct a nonorthogonal basis for the Hilbert space of Hermitian operators. Because the inner products of these projectors are uniform, given by then it is straightforward to find a shifting and rescaling that orthogonalizes the basis {Π j }, at the cost of making the operators non-positive-semidefinite. In fact, there are two choices: The bases {Q ± j } have interesting properties with regard to Lie algebra theory [11] and the study of quantum probability [20,39].
II. GENERATING SICS WITH GROUPS
All known SICs have an additional kind of symmetry, above and beyond their definition: They are group covariant. Each SIC can be constructed by starting with a single vector, known as a fiducial vector, and acting upon it with the elements of some group. It is not known in general whether a SIC must be group covariant. Because such an assumption greatly reduces the search space [4,5], it has been the only method used so far: The fact that we only know of group-covariant SICs could potentially be an artifact of this. (However, we do have a proof that all SICs in d = 2 and d = 3 are group covariant [40].) In all cases but one, the group that generates a SIC from a fiducial is an instance of a Weyl-Heisenberg group. We can define this group as follows. First, fix a value of d, and let ω = e 2πi/d . Let {|0 , |1 , . . . , |d − 1 } be an orthonormal basis for the Hilbert space H d = C d . Then, construct the shift and phase operators where the shift is modulo d. These operators satisfy the Weyl commutation relation, In a sense, the operators X and Z come as close as possible to commuting, without actually doing so: The only cost to exchanging their order is a phase factor determined by the dimension. The Weyl-Heisenberg displacement operators in dimension d are defined by The product of two displacement operators is, up to a phase factor, a third: Therefore, by allowing the generators to be multiplied by phase factors, we can define a group, known as the Weyl-Heisenberg group in dimension d. This group dates back to the early days of quantum physics. Weyl introduced the generators X and Z as long ago as 1925 in order to define what one might mean by the quantum theory of discrete degrees of freedom [41][42][43] (see also [44, pp. 2055-56]). This group, and structures derived from it, are critically important in quantum information and computation; for example, this is the basic prerequisite for the Gottesman-Knill theorem, which indicates when a quantum computation can be efficiently simulated classically [45]. The close relationship between SICs and the Weyl-Heisenberg group suggests that SICs are a kind of structure that quantum physics should have been studying all along. Zhu has proved that in prime dimensions, group covariance implies Weyl-Heisenberg covariance [46]. The one known exception to the rule of Weyl-Heisenberg covariance is the Hoggar SIC [47,48], which lives in a prime-power dimension, d = 8. As in all other dimensions, there is a Weyl-Heisenberg SIC, but there is also the Hoggar SIC. Like many other exceptions to mathematical classifications, it is related to the octonions [9,13].
One example of a Weyl-Heisenberg SIC can be constructed by taking the orbit of the following two-dimensional vector under the Weyl-Heisenberg displacement operators: This orbit is a set of four vectors. In the Bloch sphere representation, they form the vertices of a regular tetrahedron inscribed within the sphere. An example in dimension d = 3, one which is remarkable for the further subtle symmetries it possesses beyond even group covariance, is the orbit of under the Weyl-Heisenberg displacements. This set of vectors is known as the Hesse SIC [40,49,50], thanks to its relation with the Hesse configuration familiar from design theory and the study of cubic curves, specifically nonsingular cubic curves in complex projective two-space [12,16,51,52].
III. HISTORICAL OVERVIEW
In order to understand the current state of SIC research, one must grasp how people came to the SIC question, what other structures they think are related, what tools they suspect are applicable, and so forth. A physicist, motivated by quantum information theory, is apt to have a different mental context than a pure mathematician driven by the abstract appeal of geometry. To attempt to foster an interdisciplinary discussion, we provide in this section a brief historical overview.
The Hesse SIC can be extracted from a 1940 article by H. S. M. Coxeter [53], which discusses what he later called the "Hessian polyhedron" [51]. This polyhedron lives in three-dimensional complex vector space and has 27 vertices (which correspond to the 27 lines on a cubic surface [54]). The vectors in the Hesse SIC are a subset of those vertices. Mathematicians were studying SICs in d = 2 and d = 3 more explicitly as early as the 1970s, in the terminology of complex equiangular lines [55]. Stuart Hoggar found a SIC in dimension d = 8 a few years later [47], by considering the diagonals of a quaternionic polytope and converting their coordinates to complex numbers. The SICs in d = 2 and d = 3, together with the Hoggar lines in d = 8, still stand out among the known SICs; various unusual attributes they possess have led them to be designated the sporadic SICs [7,9].
In a 1987 article, Richard Feynman used a construction that is in retrospect a d = 2 SIC to study the probability theory of a qubit. SICs entered quantum theory more generally starting with the work of Gerhard Zauner, who began to consider the problem in the 1990s [44, p. 1941]. By 1999, Zauner had found the connection with the Weyl-Heisenberg group and proven SIC existence up to d = 5 [1]. He also posed a conjecture that the search for Weyl-Heisenberg SICs could be simplified by considering a particular unitary operator [1], a conjecture we will describe in detail below.
Independently of Zauner, Carlton Caves developed the idea of a SIC as representing a quantum measurement [2], motivated originally by attempts to prove the "quantum de Finetti theorem" [56]. SICs turned out to have more symmetry than was necessary for that proof, but they soon took on a life of their own in quantum information theory [23]. The term "SIC" (first pronounced as "sick," and later like "seek") dates to this period. A 2004 paper, "Symmetric Informationally Complete Quantum Measurements" [3], introduced them to the mainstream of the quantum information community, and reported numerical solutions for dimensions up to d = 45. In a later survey, Andrew Scott and Markus Grassl extended these numerical results up to dimension d = 67 [4]. Further work by Scott established Weyl-Heisenberg SICs in all cases up to d = 121 without exception. Moreover, these results together are probably a complete catalogue of all distinct Weyl-Heisenberg SICs up to dimension d = 90 [24]. In later sections, we will give an overview of how these searches were done.
Exact solutions were found more slowly. Having exact expressions for SIC fiducial vectors allowed Appleby et al. to discover a connection with Galois theory [57]. In turn, this led to further relations with algebraic number theory, a frankly mysterious development that is still under active investigation [5][6][7][8].
Before moving on, we note that the real analogue of the SIC problem, i.e., finding maximal sets of equiangular lines in real vector spaces, has also been of considerable interest to mathematicians [79][80][81]. The maximal number of equiangular lines in a d-dimensional vector space is not d 2 , but only d(d + 1)/2. That is, if we have a set of N unit vectors {v i } in a d-dimensional vector space, such that then the size N of the set cannot exceed d(d + 1)/2. Moreover, while the complex bound of d 2 has been saturated in every dimension that we have been able to check, it is known that the real bound of d(d + 1)/2 is not even attained for all values of d. For example, in d = 7, one can construct a set of 7 · 8/2 = 28 equiangular lines, but this is also the best that can be done in d = 8. In fact, the only known instances where the bound of d(d + 1)/2 can be attained are dimensions 2, 3, 7 and 23 [81].
There is a sign freedom in this definition of the angle, since Eq. (14) is satisfied if the inner product v i ,v j is either +α or −α. The presence of this discrete choice means that investigations of real equiangular lines often have a rather combinatorial flavor. In contrast, when we take the magnitude of a complex inner product, we discard a continuous quantity, a phase that in principle can be anywhere from 0 to 2π. Generally speaking, the "feel" of the real and complex problems differ, as is evidenced by the different areas of mathematical expertise brought to bear upon them. However, subtle and unanticipated points of contact between the real and complex cases do exist [13].
IV. HOW TO SEARCH FOR SICS NUMERICALLY
As before, let {|j } be an orthonormal basis for the Hilbert space H d = C d . In this basis, the fiducial vector can be written for some set of coefficients {a j }.
Acting with the Weyl-Heisenberg operator D lβ on the fiducial vector |ψ 0 produces a new vector, whose squared innner product with the fiducial vector is The right-hand side has the form of the magnitude squared of a Fourier coefficient, i.e., of a power spectrum. Specifically, the set of squared inner products between ψ (l,β) and |ψ 0 for any given value of l is the power spectrum of the sequence By the Wiener-Khinchin theorem, we know that the power spectrum of a sequence is the Fourier transform of the autocorrelation of that sequence [82]. Therefore, the autocorrelation of the sequence f (l) j , when put through the Fourier transform, will yield the sequence [F] βl . The set of autocorrelation sequences for all values of l forms a matrix. Using to denote the correlation of two sequences, we can write the elements of this matrix as The matrix G is in many situations more convenient to work with than the original matrix F, because G lacks phase factors and treats both of its indices on equal footing. For example, it is apparent from the definition of G that This is equivalent to a property of the matrix F that is less obvious to the eye: If we take the Fourier transform of the columns of the matrix G, we recover the squared inner products between the candidate SIC vectors and the fiducial. This means that if the vectors { ψ (l,β) } really do comprise a SIC, then the matrix G must take a very specific form. Every entry in F ({[G] kl }) β must equal 1/(d + 1), except for the element at l = β = 0, which equals 1. Recalling that a constant sequence is the discrete Fourier transform of a Kronecker delta function, we can deduce the desired values of [G] kl . The result is that if |ψ is a Weyl-Heisenberg fiducial vector, then This implication also works in reverse, thanks to the transitivity of the group action.
The basic idea of finding SICs numerically is to use standard optimization methods to find a fiducial vector that makes [G] kl as close to the desired form as possible. Note that, when G is constructed from a SIC fiducial, and the inequality is saturated if and only if the input vector is truly a SIC fiducial. This naturally suggests a way to find SIC fiducial vectors: Minimize the LHS of the inequality in Eq. (24), aiming for the lower bound given on the RHS. During our time investigating the SIC question, we have at various points implemented this idea in Mathematica, in Python and in C++ using the GNU Scientific Library. We find in general that numerical optimization finds a local minimum quickly, but a local minimum might only imply inner products between the vectors that are correct to a few decimal digits. A way around this problem is to repeat the optimization many times, starting from different points in the search space. Since these trials can run concurrently, the problem is amenable to parallelization. This is the approach we followed when using the Chimera supercomputer to obtain solutions in dimensions 122 through 151. Scott's implementation, which we employed on Chimera, uses a C++ code for a limited-memory quasi-Newton optimization algorithm, L-BFGS, due to Liu [84].
As is evident from Figure 1, the time required to obtain solutions did not increase steadily with the dimension. Once we have a numerical result in hand, we can refine its precision. This requires a code that uses multi-precision arithmetic, which will run more slowly than the optimization in the first step [24]. The fiducial vectors available at [24] and at the website referenced above were obtained in this way and are accurate to 150 digits.
Before moving on, we note a conjecture, based on numerical evidence, that hints at additional hidden structure in the SIC problem. Note that the definition of G implies Here, indices are to be interpreted modulo d. Because any autocorrelation attains its maximum at zero offset, we also know immediately that the elements of G cannot be larger in the middle of the matrix than they are on the edges: The Fourier transform of [G] kl over the index k is, by the Wiener-Khinchin theorem, the power spectrum of f (l) . Because power spectra are nonnegative, we can say that Are there additional symmetries or redundancies, not so apparent from the definition? By happenstance, one of the authors (CAF) observed that imposing a subset of the constraints in Eq. (22) was sufficient to find a SIC fiducial vector [44, pp. 1252-59]. Specifically, by finding a solution to the simultaneous equations one finds a solution to all the equations in (22). The redundancies in Eq. (25) are sufficient to imply that this holds up to d = 5. We call the idea that it remains true in all dimensions the "3d conjecture." It has been verified numerically up to dimension d = 28 [83]. If the 3d conjecture is indeed true, it would reduce the complexity of the problem, as measured by the number of simultaneous equations to solve, from quadratic in the dimension to linear.
V. ZAUNER SYMMETRY
Is there any way to narrow the search space for SIC fiducial vectors? To see how to answer this in the affirmative, we must elaborate upon the group theory we discussed in the previous sections. The Clifford group for dimension d is the "normalizer" of the Weyl-Heisenberg group: It is the set of all unitary operators that, acting by conjugation, map the set of Weyl-Heisenberg operators in dimension d to itself. We saw earlier how the orbit of a vector under the Weyl-Heisenberg group can be a SIC; likewise, we can study the orbit of a vector under the entire Clifford group.
For our purposes in this note, the important point is that if we conjugate a Weyl-Heisenberg operator D kl by a Clifford unitary U , we obtain a Weyl-Heisenberg operator D k l , possibly with an additional phase. Details on the construction and representation of the Clifford group in any finite dimension d can be found in Appleby [85].
It was conjectured by Zauner [1], and independently by Appleby, that in every dimension d > 2, a Weyl-Heisenberg SIC fiducial exists that is an eigenvector of a certain order-3 Clifford unitary, which is now known as the Zauner unitary. Acting on the Weyl-Heisenberg generators, the Zauner unitary effects the change up to an overall phase factor. (See equation (3.10b) of Zauner [1], or equation (127) of Appleby [85].) Applying this again yields and a third iteration gives confirming that this operation has order 3. How might assuming the Zauner conjecture simplify the search for SICs? First, we will make some remarks on this from an algebraic perspective, and then we will address the point in a way suited to numerical optimization. Let |ψ be a candidate fiducial vector, and suppose that it is an eigenvector of the Zauner unitary U with unit eigenvalue: Consequently, As U is a Clifford unitary, requiring that |ψ is an eigenvector of U implies degeneracies among the elements of the matrix F. Because the Zauner unitary sends the left edge of F to the top edge and then to the main diagonal. More generally, specifying a column of F (which is equivalent to fixing a column of G) and imposing the Zauner condition means that the same constraint also simultaneously fixes a row and a diagonal. Earlier, in Eq. (20), we saw that a symmetry of G implied a Fourier-type relation among the elements of F. We have expressed the Zauner condition as a degeneracy within F, but what does it imply for G? The result can be found straightforwardly, and it closely resembles Eq. (20): This the expression of Zauner symmetry in the G matrix. A special case of note: If we set k = l = 0, then We note that the assumption that the fiducial is a Zauner eigenvector is enough to prove some additional cases of the 3d conjecture, up to dimension d = 9, in a straightforward way. In dimension d = 5, the basic symmetries of G imply that the 3d constraints automatically specify the entire matrix G, and thus also fix F to have the desired form for a SIC. By imposing the condition that our initial vector is a Zauner eigenvector, we can extend this up to dimension d = 8. This can be seen directly by drawing an 8 × 8 grid and shading in the appropriate squares.
In fact, we can carry this argument a little further. We obtain G by Fourier transforming the columns of F.
Recalling that we therefore find that In other words, the Zauner condition implies that if we add up the entries in a column, leaving out the entry on the top row, they must all cancel out and leave zero. We knew already, thanks to the Wiener-Khinchin theorem and Eq. (27), that the imaginary parts will sum to zero. Now we can establish this for the real parts as well.
The left-most column of G is also its top row, which tells us the averages of each column of F. So, if the Zauner orbits leave only one element in a column unspecified, then we can fill in that element, because we know the average over the whole column vector. This proves the 3d conjecture in dimension d = 9.
We now turn to the simplification that the Zauner conjecture provides for numerical search efforts. By postulating that the SIC fiducial we are looking for is a Zauner eigenvector, we can significantly reduce the effective size of the search space. First, suppose that U is a unitary of order n, so that and the eigenvalues of U can all be written with m an integer. The projector onto the eigenspace with this eigenvalue is We can restrict our numerical search to this subspace by projecting our vectors into it, at each iteration of the optimization algorithm.
Most of the known solutions were found by postulating Zauner symmetry. Scott has also found several solutions by assuming that the fiducial was an eigenvector of another Clifford unitary. For an in-depth exposition of these variations, see [24].
VI. EXHAUSTIVE SEARCHES
Suppose that |ψ 0 ∈ H d is a vector in a Weyl-Heisenberg SIC, and let U be a Clifford unitary. Applying U to the vector |ψ 0 will yield some vector, The Weyl-Heisenberg orbit of |ψ 0 is a SIC, so what about the orbit of |χ 0 under the same group? We define and we consider the squared magnitudes of the inner products Because U is a Clifford unitary, it maps D kl to some Weyl-Heisenberg operator D k l , with any phase factor dropping out when we take the magnitude of the inner product. So, meaning that the image of our original SIC under the mapping U is also a SIC. One way in which the Hesse SIC is remarkable is that it is invariant under the entire Clifford group. For contrast, we can take the vector which differs from the Hesse SIC fiducial in Eq. (13) by a sign. The orbit of this vector under the Clifford group is a set of four separate SICs, comprising 36 vectors in all-the so-called Norrell states, which are significant in the theory of quantum computation [52,86]. We consider two SICs equivalent if they can be mapped into each other by a Clifford unitary. In fact, it is convenient to extend the Clifford group by including the anti-unitary operation of complex conjugation. The extended Clifford group for dimension d, EC(d), is the set of all unitary and anti-unitary operators that send the Weyl-Heisenberg group to itself. For (extensive) details, we again refer to Appleby [85,87].
In order to search the space as exhaustively as possible and create a catalogue of all essentially unique SICs, Scott's code chooses initial vectors at random under the unitarily invariant Haar measure on the complex projective space CP d−1 . Once enough solutions are found-generally, this means hundreds of them-the code then refines their precision, as described above. Then, we must identify unique orbits under the extended Clifford group. This last step is computationally demanding, because we must translate each solution vector |ψ by each element in the extended Clifford group EC(d). However, in the process, Scott's algorithm also finds the stabilizer group of each fiducial, which is important information. The task of determining when two SICs are equivalent up to a unitary or anti-unitary transformation has been discussed in depth by Zhu [18], and we expect that additional theoretical insights may lead to an improved algorithm for this step.
Following this procedure, Scott has carried out exhaustive searches in dimensions up to d = 90. We strongly expect his catalogue of solutions to be complete up to that point: All Weyl-Heisenberg SICs in those dimensions are equivalent to the ones tabulated, up to equivalence under the extended Clifford group.
VII. DISCUSSION
In the preceding sections, we have described the process of finding SIC fiducial vectors numerically. However, some patterns among SICs have only become apparent when exact solutions were studied carefully. Suppose we refrain from taking the magnitude-squared in our definition of a SIC, Eq. (1). Then for some set of phases {e iθ jk }. (In fact, one can reconstruct the SIC from knowing the phases [10].) It was recently discovered that when d > 3, for all the known Weyl-Heisenberg SICs, these phases have a remarkable meaning in algebraic number theory: They are units in ray class fields and extensions thereof [5]. This is a topic to which we can hardly do justice here, and indeed, treatments accessible to anyone who is not already an algebraic number theorist have only recently been attempted [6,7]. For now, we content ourselves with the observation that this area of number theory is the territory of Hilbert's twelfth problem, one of the still outstanding questions on history's most influential list of mathematical challenges [88]. (Specialists may recall that according to the Kronecker-Weber theorem, any abelian extension of the rationals is contained in a cyclotomic field. When we instead consider abelian extensions of real quadratic fields, the analogue of the cyclotomic fields are the ray class fields. The phases of Weyl-Heisenberg SICs appear to be playing a role regarding ray class fields much like the role that roots of unity play with cyclotomic fields. Moreover, recalling Eq. (14), it is intriguing that in the real-vector-space version of equiangular lines, we discard a phase factor that is a unit among the ordinary integers, while in the complex Weyl-Heisenberg case, the phases turn out to be units among algebraic integers.) From Hilbert space to Hilbert's twelfth problem! What physicist would ever have anticipated that? And who could turn down the opportunity to intermingle two subjects that had seemed so widely separated? SICs have found relevance, not just in quantum computation, but in signal-processing tasks like high-precision radar [89] and speech recognition [90]. In February 2016, our colleague Marcus Appleby attended a conference in Bonn, Germany on uses of the Weyl-Heisenberg group. Many participants were engineers, including representatives from the automotive and cell-phone industries. Appleby was told that if he managed to construct a SIC in dimension 2048, he should patent it [91]. At the moment, dimension 2048 is beyond our abilities for algebraic or numerical solutions, but this may not always be the case.
VIII. ACKNOWLEDGMENTS
We are deeply indebted to Andrew J. Scott, the guru of SIC numerical solutions, for code and for discussions. We also thank Marcus Appleby for many conversations, and Gary McConnell for email feedback on the original arXiv version of this article. This research was supported in part by MCH's Oracle Undergraduate Research Fellowship at UMass Boston.
IX. AUTHOR CONTRIBUTIONS
MCH ran the calculations on Chimera to find SICs in dimensions 122 through 151. BCS wrote the paper. CAF directed the research, contributed to the bibliography and worked with BCS in revising the paper. | 7,670.6 | 2017-03-23T00:00:00.000 | [
"Computer Science"
] |
Understanding LTP in pain pathways
Long-term potentiation (LTP) at synapses of nociceptive nerve fibres is a proposed cellular mechanism underlying some forms of hyperalgesia. In this review fundamental properties of LTP in nociceptive pathways are described. The following topics are specifically addressed: A concise definition of LTP is given and a differentiation is made between LTP and "central sensitisation". How to (and how not to) measure and how to induce LTP in pain pathways is specified. The signal transduction pathways leading to LTP at C-fibre synapses are highlighted and means of how to pre-empt and how to reverse LTP are delineated. The potential functional roles of LTP are evaluated at the cellular level and at the behavioural level in experimental animals. Finally, the impact of LTP on the perception of pain in human subjects is discussed.
Background
Long-term potentiation (LTP) is a much studied cellular model of synaptic plasticity. It is generally defined as the long-lasting but not necessarily irreversible increase in synaptic strength [1,2]. At least two different stages of LTP can be distinguished depending upon its duration and the signal transduction pathways involved. Early phase LTP is independent of de-novo protein synthesis and lasts for up to three hours. Late phase LTP involves protein synthesis and lasts longer than three hours, up to the life span of an animal and may involve structural changes at synapses [3]. Short-term potentiation of synaptic strength lasts less than half an hour. Synaptic strength is the magnitude of the post-synaptic response (i.e. the post-synaptic potential or the post-synaptic current, but not action potential firing, see below) in response to a pre-synaptic action potential. LTP can be expressed pre-and/or postsynaptically, i.e. synaptic strength can increase if the release of neurotransmitter(s) is enhanced and/or if the postsynaptic effects of the neurotransmitter(s) become stronger [4]. LTP at synapses in hippocampus is the prime model for learning and memory formation [1]. Recent studies have shown that LTP can also be induced in pain pathways and may contribute to hyperalgesia caused by inflammation, trauma or neuropathy. This review deals with the latter form of LTP.
What is "central sensitisation"?
"Central sensitisation" is used in the literature in at least two mutually exclusive definitions. Some of the authors use "central sensitisation" as an umbrella term for all forms of changes within the central nervous system which ultimately lead to enhanced pain perception. If using this definition of "central sensitisation" one should keep in mind that none of the presently known phenomena in the central nervous system (CNS) which can be observed in experimental or clinical models of hyperalgesia or allodynia has a proven, causative role for the perception of pain. Thus, all presently proposed mechanisms of "central sensitisation" in this definition would have the status of hypotheses. This includes but is not be limited to the topic of the present review, the LTP at C-fibre synapses in the superficial spinal dorsal horn.
The task force for taxonomy of the International Association for the Study of Pain (IASP) and many other authors define "central sensitisation" as "an enhanced responsiveness of nociceptive neurons in the CNS to their normal afferent input". This is a very clear-cut definition with little error-proneness. The mechanisms underlying "central sensitisation" in this definition can be well studied experimentally. Nociceptive neurons in the CNS may, however, serve very distinct and also antagonistic functions not all of which are related to the perception of pain. Some nociceptive neurons are excitatory, others are inhibitory. Spinal nociceptive neurons may project to different areas in the brain, with a wide spectrum of functions. Other nociceptive spinal neurons may project segmentally to motoneurons and still other may be interneurons with no known function at all. A remarkable exception are those nociceptive neurons which are located in lamina I of the spinal dorsal horn and which express the NK1 receptor for substance P, many of which project supraspinally. If these neurons are destroyed selectively by a toxin conjugated to substance P, then neither inflammation nor neuropathy leads to full expression of hyperalgesia [5,6]. Interestingly, animal responses to acute noxious stimuli are not altered [5,6]. Thus, these neurons are indispensable for the development of hyperalgesia but not essential for acute pain. Enhanced responsiveness of different nociceptive neurons in the CNS may thus have distinct and perhaps opposing consequences on pain. Thus, "central sensitisation" defined in this way is fundamentally different from the former definition. Some forms of "central sensitisation" as defined by the IASP task force could contribute to hyperalgesia and/or allodynia while other forms may rather lead to the contrary, i.e. stronger feedback inhibition and endogenous pain control. Many of the presently known forms of "central sensitisation" defined in the latter way would be assigned a yet unknown (if any) role for pain perception.
An utter confusion is created if the term "central sensitisation" is used but not defined or when switching from one definition to the other within the same publication, i.e. if both definitions are not clearly distinguished. This would incorrectly imply that any form of enhanced responsiveness of nociceptive neurons in the CNS to their normal afferent input would lead to an enhanced perception of pain. Before we evaluate the evidence suggesting that a prolonged increase in synaptic strength at C-fibres contributes to hyperalgesia we will first review the fundamental properties of LTP in pain pathways.
How to (and how not to) measure LTP
LTP is measured as an increase in mono-synapticallyevoked post-synaptic currents or potentials in response to a single pre-synaptic action potential. LTP is often studied in in vitro preparations which allow reliable recordings of synaptic strength. Whole-cell patch-clamp recording is now the most often used technique. It enables some control over the composition of the intracellular fluid of the post-synaptic neurons which may be advantageous to study post-synaptic mechanisms of LTP. If, however a diffusible mediator is involved and dialysis of the post-synaptic neuron has to be avoided, perforated patch-clamp recordings or intracellular recordings with sharp electrodes can be used. To evaluate LTP at the first synapses in nociceptive pathways, transverse slices with long dorsal roots attached can be prepared from lumbar spinal cord of rats or mice to study mono-synaptic, Aδ-fibre or C-fibre evoked excitatory postsynaptic potentials or currents in identified dorsal horn neurons [7,8].
Some aspects of LTP can only be studied in the entire animal with primary afferent nerve fibres and descending pathways from the brain intact. In vivo C-fibre-evoked field potentials can be measured in superficial spinal dorsal horn e.g. in response to high intensity electrical stimulation of the sciatic nerve for up to 24 h [9]. These extracellulary recorded field potentials reflect summation of post-synaptic, mainly mono-synaptically-evoked currents but not action potential firing [9,10].
Monitoring presynaptic activity at synapses of primary afferent nerve fibres is technically quite demanding. In an attempt to monitor pre-synaptic activity in primary afferents optical recordings techniques have been utilised. Some voltage-sensitive dyes can be anterogradely transported in primary afferents to the central terminals mainly in lamina I [11] and may serve as an indicator for pre-synaptic electrical activity but not for transmitter release.
LTP can not be directly investigated by recording action potential discharges of post-synaptic neurons, as action potential firing not only depends upon synaptic strength but also on membrane excitability and the balance between excitatory and inhibitory input to the neuron. For the same reasons poly-synaptically-evoked responses can generally not be used to study synaptic strength and changes thereof.
How can LTP be induced in pain pathways? LTP induction by high frequency electrical nerve stimulation
The most frequently used form of conditioning stimulation to induce LTP at synapses in the brain consists of high frequency electrical stimulation (HFS, around 100 Hz) of an input pathway. Likewise, LTP can be induced at spinal synapses of small diameter primary afferents by conditioning high intensity, high frequency burst-like stimulation (typically 100 Hz bursts given several times for 1 s at C-fibre strength) both, in vitro and in vivo. In spinal cord slice preparations, both, Aδ-fibre [7] and C-fibre [8,12]evoked responses are potentiated by HFS when post-syn-aptic neurons are mildly depolarized to -70 to -50 mV. The same HFS induces, however, long-term depression (LTD) of Aδ-fibre-evoked responses if cells are hyperpolarized to -85 mV suggesting that the polarity of synaptic plasticity is voltage-dependent [7].
Neurons in spinal cord lamina I which express the NK1 receptor play a pivotal role for hyperalgesia in behaving animals [5,6]. Most of these neurons send a projection to supraspinal areas. Interestingly, HFS induces LTP selectively at C-fibre synapses with lamina I neurons which express the NK1 receptor and send a projection to the parabrachial area (Fig. 1). In contrast, HFS fails to induce LTP at synapses with neurons which express the NK1 receptor and send a projection to the periaqueductal grey or at synapses with neurons which do not express the NK1 receptor and which have no identified supraspinal projection [8,12].
HFS at C-fibre intensity of sciatic nerve fibre afferents induces LTP of C-fibre, but not Aβ-fibre-evoked field potentials in superficial spinal dorsal horn of adult, deeply anaesthetized rats [9,13,14]. In contrast, conditioning HFS at A-fibre intensity fails to induce LTP of either A-or C-fibre-evoked field potentials in intact animals. In spinalised animals, conditioning HFS at Aδ-fibre intensity induces, however, LTP of C-fibre-evoked field potentials [15].
LTP induced by low frequency electrical nerve stimulation
For most of the C-fibre afferents it is not typical to discharge at rates as high as 100 imp. s -1 . Some C-fibres may, however, discharge at these high rates but only for short periods of time, e.g. at the beginning of a noxious mechanical stimulus [16]. Many C-fibres discharge at considerably lower rates, around 1-10 imp. s -1 , e.g. in response to an inflammation or an injury [17]. Conditioning stimulation within this lower frequency band is successfully used to induce LTP at C-fibre synapses. In a spinal cord-dorsal root slice preparation conditioning electrical low frequency stimulation (LFS 2 Hz for 2-3 min, C-fibre strength) of dorsal root afferents induces LTP selectively at C-fibre synapses with lamina I neurons that express the NK1 receptor and project to the periaqueductal grey (Fig. 1) [12]. C-fibre synapses with lamina I neurons which express the NK1 receptor and project to the parabrachial area or with no identified supraspinal projection are, in contrast, not potentiated by LFS [12]. Thus, the pattern and the frequency of discharges in C-fibres determine which synapses at the origin of different ascending pain pathways are potentiated.
Pairing strong post-synaptic depolarization (to +30 mV) with low frequency (2 Hz) stimulation of afferent nerve fibres also leads to a robust LTP in superficial spinal dorsal horn neurons in vitro [18]. In spinal cord slices from neonatal rats field potentials evoked by electrical stimulation in the tract of Lissauer are potentiated by repetitive burstlike stimulation at 10 Hz [19].
In deeply anaesthetized adult rats with their spinal cords left intact LFS (at 2 Hz for 2-3 min) of sciatic nerve fibres at C-fibre intensity but not at Aδ-fibre intensity also triggers LTP of C-fibre-evoked potentials (Fig. 1) [12].
In conclusion, HFS and LFS may have fundamentally different effects on LTP induction at different C-fibre synapses. This finding is in line with previous reports also illustrating that the frequency of afferent barrage in Cfibres may have qualitatively different effects in spinal cord. For example, brain-derived neurotrophic factor is released from primary afferents in spinal cord slices in an activity-dependent manner by HFS at 100 Hz but not by 1 Hz LFS of primary afferent nerve fibres [20]. Furthermore, in spinal cord slices from mice HFS (100 Hz) of primary afferent nerve fibres at C-fibre intensity, but not LFS (900 pulses at 1 Hz) selectively induces phosphorylation of extracellular receptor-activated MAP Kinases (ERK1/2) in spinal dorsal horn lamina I [21].
Natural noxious stimulation induces LTP in pain pathways
At synapses in the brain LTP induction requires synchronous, high-frequency pre-synaptic activity or pairing of low-level pre-synaptic activity with strong post-synaptic depolarization. At least some of the C-fibre synapses are apparently unique in that LTP can be induced by LFS and by natural, low or high frequency, asynchronous and irregular discharge patterns in sensory nerve fibres. In animals with spinal cord and descending pathways intact, intraplantar, s.c. injections of capsaicin (100 μl, 1%) or formalin (100 μl, 5%) induce slowly rising LTP (Fig. 1) [12]. Some forms of low level afferent input can induce LTP only if descending, presumably inhibitory pathways are interrupted or weakened. Noxious radiant heating of the skin at a hind paw induces LTP in spinalised animals but not in animals with spinal cord intact [22]. Likewise, repetitive, noxious squeezing of the skin or the sciatic nerve induces LTP of C-fibre-evoked field potentials in spinalised rats only [22]. These findings indicate that endogenous antinociceptive systems not only raise thresholds for nociception but also those for the induction of LTP.
LTP of A-fibre-evoked responses
A-fibre-evoked spinal field potentials are depressed by conditioning 50 Hz stimulation of sciatic nerve fibres. After GABA A receptor antagonist bicuculline (1 mg/kg i.p.) the same conditioning stimulus now produces LTP rather than LTD [25]. Similarly, 50 Hz conditioning stimulation produces short lasting potentiation followed by LTD in control animals but LTP in animals with a CCI of sciatic nerve [26]. Topical application of muscimol (10 μg), a GABA A receptor agonist to spinal cord prevents tetanus-induced LTP of A-fibre-evoked field potentials in animals with a CCI [27]. This again suggests that the polarity of synaptic plasticity is context-sensitive and not solely dominated by the type of afferent input.
Signal transduction pathways leading to LTP at C-fibre synapses
In principle, LTP can be induced and/or expressed by presynaptic [28] or by post-synaptic [29,30] mechanisms or by any combination thereof. At present, there is clear evidence for a post-synaptic, Ca 2+ -dependent form of LTP induction in spinal cord lamina I neurons. Induction of LTP at C-fibre-synapses requires co-activation of NK1 and NK2 receptors [9], opening of ionotropic glutamate receptors of the NMDA type [8,12,13], opening of T-type voltage-gated calcium channels [8,12], and activation of group I but not group II or III metabotropic glutamate receptors [31]. Activation of NK1 receptors by substance P may directly enhance single NMDA channel opening [32] and NMDA receptor mediated currents in lamina I neurons [8] and all this may lead to substantial rise in postsynaptic [Ca 2+ ] i (Fig. 2). It is presently unknown if Ca 2+ influx through Ca 2+ -permeable a-amino-3-hydroxy-5methyl-4-isoxazolepropionic acid (AMPA) receptors is required for LTP induction in pain pathways. Some indirect evidence suggests, however, that this might be the case [21,33].
In any case, a rise in post-synaptic [Ca 2+ ] i is essential for LTP induction and the magnitude in [Ca 2+ ] i rise is linearly correlated with the magnitude of LTP in vitro [8]. Recent data demonstrate that LTP-inducing stimuli cause substantial rise in [Ca 2+ ] i in lamina I neurons not only in slice preparations, but also in intact animals [12]. Not surprisingly therefore signal transduction involves Ca 2+ -dependent pathways including activation of protein kinase C, calcium-calmodulin-dependent protein kinase II (CaM-KII), protein kinase A (PKA) phospholipase C (PLC), inositoltriphosphate-3 (IP 3 ) receptors, nitric oxide synthase (NOS) and members of the mitogen-activated protein kinase family (MAPK), including extracellular signalregulated kinase (ERK) (Fig. 2) [8,12,18,[34][35][36].
When assessed with voltage-sensitive dyes the pre-synaptic facilitation of electrical activity in primary afferents after LTP-inducing stimuli is partially sensitive to iNOS inhibitor (AMT), a blocker of glial cell metabolisms (MFA), and an mGluR group I antagonist (LY367385) [11].
Inhibition of protein synthesis in spinal cord by either cycloheximide or anisomycin selectively inhibits the maintenance of the late-phase of spinal LTP but does not affect either LTP induction or baseline responses of Cfibre evoked field potentials [37].
Importantly, the very same signal transduction pathways are required for full expression of hyperalgesia in animal models of inflammatory and neuropathic pain e.g. [38][39][40][41].
How to pre-empt LTP induction in pain pathways
LTP induction can be prevented by blockade of any of the above mentioned essential elements of signal transduction for LTP. In mature rats deep (surgical) level of anaesthesia with either urethane, isoflurane or sevoflurane is, however, insufficient to pre-empt LTP induction of Cfibre-evoked field potentials [42]. LTP is prevented by low dose intravenous infusion of μ-opioid receptor agonist fentanyl [42]. Similarly, LTP of spinal field potentials elicited by stimulation in the tract of Lissauer in spinal cord slices is blocked by DAMGO, a more specific agonist at these receptors [19]. Activation of spinal a 2 -adrenoreceptors by clonidine [43] or spinal application of the benzodiazepine diazepam [44] also prevents LTP induction in vivo.
Functional blockade of glial cells by i.t. administration of fluorocitrate changes the polarity of HFS induced synaptic plasticity. When HFS is given 1 h, but not 3 h after fluorocitrate LTD but no LTP of C-fibre-evoked field potentials is induced [14].
LTP can be reversed
LTP of C-fibre-evoked field potentials can be reversed by brief, high frequency conditioning electrical stimulation of sciatic nerve fibres at Aδ-fibre intensity [15]. Reversal of LTP by Aδ-fibre stimulation is time-dependent and effec-tive only when applied 15 or 60 min but not 3 h after LTP induction [45].
Spinal application of either, NK1 or NK2 receptor antagonists one to three hours after HFS, i.e., after LTP is established, does not affect maintenance of LTP [9], suggesting that activation of these receptors, which are required for the induction of LTP are not essential for its maintenance.
What is the functional role of LTP in pain pathways?
Modulation of synaptic strength is a powerful mechanism to control signal flow in selected pathways. A typical con-sequence of LTP at excitatory synapses would be an increase in action potential firing of the same and perhaps also of downstream neurons in response to a given stimulus. And indeed, LTP-inducing conditioning stimuli have been found to facilitate action potential firing of multireceptive neurons in deep dorsal horn e.g. [46][47][48]. This is likely due to LTP at the first synapse in the nociceptive pathway but other mechanisms of facilitation should not be excluded. Action potential firing would also be enhanced if membrane excitability is increased, i.e. the thresholds for action potential firing are lowered, if inhibition is less effective or if inhibition is even reversed and Potential mechanisms of potentiation, prevention and de-potentiation at synapses between C-fibres and spinal cord projection neurons Figure 2 Potential mechanisms of potentiation, prevention and de-potentiation at synapses between C-fibres and spinal cord projection neurons. Conditioning electrical nerve stimulation or natural noxious stimulation triggers release of glutamate and substance P which causes opening of NMDA receptor channels and T-type voltage-gated Ca 2+ channel and Ca 2+ release from intracellular stores. This activates Ca 2+ -dependent signal transduction pathways including protein kinases and transcription factors. Synaptic strength is probably increased by phosphorylation of synaptic proteins including AMPA receptor channels [55], altered trafficking of synaptic proteins, e.g. increased insertion of AMPA receptors into the sub-synaptic membrane [56] and de-novo protein synthesis. According to this model, LTP can be prevented if release of glutamate and/or substance P is inhibited, for example by activation of pre-synaptic, G-protein-coupled μ-opioid receptors, or if opening of voltage sensitive and Ca2+ permeable ion channels is blocked, e.g. via postsynaptic inhibition by an opioid. Depotentiation could result from de-phosphorylation of synaptic proteins, changes in receptors trafficking and degradation of synaptic proteins.
becomes excitatory e.g. due to a reversal of the anion gradient in the post-synaptic neuron [49,50].
HFS of sciatic nerve fibres which induces LTP at synapses of C-fibres in spinal cord has behavioural consequences in rats and causes thermal hyperalgesia at the ipsilateral hind paw for six days [34]. This suggests that LTP at C-fibre synapses has an impact on nociceptive behaviour.
Perceptual correlates of LTP in pain pathways in human subjects
An indispensable proof for any proposed mechanism of hyperalgesia is an appropriate correlate in the human. And indeed, conditioning HFS of cutaneous peptidergic afferents in humans cause increased pain perception in response to electrical test stimuli applied through the same stimulation electrode [51]. Noxious stimulation with punctate mechanical probes in skin adjacent to the HFS conditioning skin site uncovers a marked (2-3 fold) increase in pain sensitivity, i.e. secondary hyperalgesia [51]. Touching the skin around the conditioning stimulation electrode with a soft cotton wisp evokes pain only after HFS. Thus, HFS also induces secondary mechanical allodynia. Hyperalgesia at the conditioned site but not secondary hyperalgesia or allodynia at adjacent skin areas is prevented by pre-treatment with ketamine [52], a clinically used substance which, among other effects, also blocks NMDA receptors.
Interestingly, all thermal modalities comprising cold and warm detection thresholds, cold and heat pain thresholds as well as pain summation (perceptual "wind-up") remain unaltered after conditioning HFS of peptidergic skin nerve fibres [53].
When verbal pain descriptors are used to evaluate pain in addition to its perceived intensity after HFS, a significant long-term increase in scores for sensory but not for affective descriptors of pain is detected [54]. Within the sensory descriptors, those describing superficial pain, those for heat pain and those for sharp mechanical pain are all potentiated. The authors conclude that brief painful stimuli rarely have a strong affective component and that perceived pain after HFS exhibits predominantly a potentiation of the C-fibre-mediated percepts hot and burning [54].
In humans subjects conditioning LFS causes also an increased pain sensitivity in the area around the LFS conditioned skin site but a depression of pain evoked by stimulation through the same electrode [51].
Conclusion
LTP at synapses between primary afferent C-fibres and a group of nociceptive neurons in spinal cord lamina I which express the NK1 receptor for substance P is a potential mechanism underlying some forms of pain amplification in behaving animals and perhaps human subjects. Both, LTP and hyperalgesia involve the same essential elements, i.e. primary afferent C-fibres and lamina I neurons which express the NK1 receptor. Further, induction protocols, pharmacological profile and signal transduction pathways are virtually identical.
Competing interests
The author declares that he has no competing interests. | 5,132.4 | 2007-01-01T00:00:00.000 | [
"Medicine",
"Biology"
] |
Three Metric-Based Method for Data Compatibility Calculation
This article analyzes ways of calculating characteristics of data and most common data structure types that allow comparison between them or on a time axis. To achieve this, it studies the key aspects of relational databases, XML, JSON and RDF structure types. These data structure types are compared to multiple isolated approaches to data quality and other data characteristics measurements. The goals of the article are the calculation method itself and a storage structure for calculated values. The article presents a method of characterization of data and data structure types based on the calculation of three metrics: the amount of structuredness, the amount of hierarchicallity and the amount of information. This triad of metrics allows comparison between various data sets (objects), for example evaluating the complexity of the transformation of data from one data object to another, as well as with data structure types (as mentioned above). Based on the vector of three metrics, the calculation method of the compatibility between data and data structure type is proposed. This method can help select the most compatible data format for existing data. The calculated values of metrics can also detect non-optimal storage design and classify data transformations. The method was evaluated on an example case study, which showed its usability on an example demonstration data set. It can be used in the process of data modelling to help select optimal data structure type, to design a data transformation process and to optimize existing data storages.
Introduction
The motivation for this article is to design a method for characterization of various data (independently on its structure type, format, or data model) by three calculable metrics. This calculation allows comparison between data sets (data with another data) and the calculation of their compatibility with any data structure type (data with relational database, XML etc.). Compatibility calculation helps to design the transformation of any data set to the most compatible data structure type. The calculated compatibility can be compared with the complexity of data transformation represented by its transformation length (vector of metrics values changes).
To demonstrate the problem addressed by our approach, we shall consider complex enterprise data stored in a relational database. Multiple questions can be asked: 1. Can the data be more easily converted and stored by XML or JSON (or any other format or data model)? 2. Which parts of the data can most easily be converted and stored by XML? 3. Which parts of the data are unsuitable for storing in a relational database? 4. Which part of the data changes the most in time? 5. Which parts of the data most increase its size and what is the probable reason? 6. Which parts of the data are the most redundant?
This article builds on the idea that to answer such questions, a single three metric-based method can be applied. These metrics are the amount of structuredness, hierarchicallity and information.
Background
There have been multiple isolated researches comparing single or two data structure types, their usage advantages and disadvantages and specifics for their most compatible data. Typically, relational database and XML Ma et al., 2020;Shanmugasundaram et al., 2001;Song & Haw, 2020) or XML and JSON (Helland, 2017). But current research lacks a unified systematic approach that allows comparing various pairs of data structure types and their corresponding data, which is the perspective of this article.
The problem of the amount of data structuredness is usually limited to focusing on data at a specific structuredness level. Paper (Florescu, 2005) represents a general consideration of the problem of less structured information and data as one of the main limitations of relational databases. Article (Šperková, 2014) mentions that unstructured data is generated mostly by social networks nowadays. This paper extends this structuredness approach into a continuous calculable metric. Article (Wellenzohn et al., 2020) propose an indexation method for semi-structured (and also hierarchical) data. This paper also addresses the problem of measuring how hierarchical the data is (the structure of a tree). Despite multiple definitions of a tree, no work approaching a measurable continuous value has been found. For example, (Musca et al., 2011) only say that a structure is hierarchical if "lower level units are part of higher level units". This article uses this definition to identify non-hierarchical parts of data and to calculate the amount of hierarchicallity based on their share.
The amount of information has been well known since the approach of Shannon's entropy. However, its main philosophy of the amount of information as a "non-redundancy" expressing the "unpredictability" of data is extended by Algorithmic information theory (AIT) based on Kolmogorov complexity, which deals with "the relationship between the computational method, information and randomness" and thus measures in particular the complexity of the relevant data (Hutter, 2007). This paper further applies AIT to measure the amount of information.
There are two main objectives of this article. The first objective is the calculation method, that takes a data object (any bounded set of data) and computes a triad of values -its amount of structuredness, hierarchicallity and information, belonging to the interval 〈0; 1〉. The second objective is the storage method that allows saving computed values referenced to specific data and time into a structured database called "metadata warehouse". Both objectives are strongly interconnected -each method requires using the other (the calculation method uses previously stored values). Their usability is evaluated on experimental data examples.
The calculated values that characterize various data can help to demonstrate their characteristics in simple tables or graphically and to facilitate communication between data experts, software engineers as well as statisticians and content managers.
Many universal or specifically targeted metrics of data have been designed and used. A large number of them refers to the concepts of "Big Data" and "data quality". Article (Morton, 2014, p. 4) presents the following characteristics of Big Data -volume, variety, velocity, value and validity. Regarding data quality, (Floridi, 2013) presents "accessibility, accuracy, availability, completeness, currency, integrity, redundancy, reliability, timeliness, trustworthiness and usability". Article (Närman et al., 2011) present the main dimensions of data quality, which they study in the context of Enterprise architecture: completeness, consistency, timeliness, relevance and accuracy, which is, however, a concept of metric rather than a dimension. This approach is also related to the "data governance" topic, which (Begg & Caira, 2012) define as a framework -a logical structure for classifying, organizing and communicating complex activities involving making decisions about them and taking action on business data.
Multiple papers have studied topics related to the specific three metrics used in this article. The structuredness of data is widely understood as a discrete rather than continuous variable, having three levels -structured, semi-structured and unstructured, described e.g. by (Bartmann et al., 2011). Similarly, (Pokorný, 2010) extends the concept of classic and traditional database, because the goal is to manage a rich collection of structured, semi-structured, and unstructured data, spread in various enterprise repositories and on the Web. No relevant publications about data hierarchicallity have been found, just general definitions of hierarchies and the author's previous poster (Vodňanský & Zamazal, 2016) measuring hierarchicallity on OWL ontologies.
The amount of information is a popular topic since Shannon's concept of information entropy (Shannon, 2001), which is limited to a linear sequence of messages and their probability in an information channel. This concept is usable on single-dimension structured data only, but it is not flexible enough for usage on the multiple data structure types in this paper. According to (Hutter, 2007), the algorithmic amount of information is the "length of the shortest description" of a string or a data set which can be created by a compression (e.g. by Lempel-Ziv, bzip2 and PPMZ algorithms). Authors in article (Grünwald & Vitányi, 2008) mention a "minimum number of bits from which a particular message can effectively be reconstructed". Paper (Meinsma, n.d.) discusses the relation to lossy compression and proves that "there is no lossless compression that strictly reduces every file" and (Grünwald & Vitányi, 2008) recognize the extension of the AIT concept to lossy algorithms, which "allows formalizing more sophisticated concepts such as 'meaningful information' and 'useful information'".
Applied methods
This article uses a synthesis of sorted information sources of current theoretical knowledge (summarized in chapter 2). Data structure types are gradually qualitatively analyzed in part 4.1 and the result of their analysis is formalized by the three metrics described in part 4.2. By summarizing the possible combinations of metrics value changes, the data transformation types are derived in part 4.3.
The evaluation method is an experimental verification of the applicability of the metrics calculation method on an example case study (chapter 5), which uses three sets of demonstration data and also analyses their transformations. The output of this paper can be used as a quantitative analytical method based on deduction, which is practically verified on demonstration and real data.
Designed calculation method
This paper designs a method which allows describing various data by a triad of metrics.
Considered data structure types
This paper studies the following common data structure types.
Relational databases (RDBs) are undoubtedly one of the dominant data storage technologies, popular on a theoretical basis from (Codd, 1990), strongly popular in the enterprise sector (Ramel, 2015). RDBs are oriented primarily on structured data, despite the support of XML and JSON storing typical for most RDBS systems and the NoSQL concept. An important aspect of RDBS is normalization -their main intention is to save structured data with a minimum redundancy.
XML is a format complementary with RDBs especially in data interchange and semi-structured data storage. Compared with RDBs, it is self-describing and more flexible (Helland, 2017). It can store data from unstructured to fully structured level. The storage is hierarchical (if we ignore the non-uniform mechanism of elements referencing each other). XML documents can contain any level of redundancydata normalization can be impossible in this format when the information represented by the XML document is not hierarchical.
JSON represents a brackets-based notation of objects originally from the JavaScript language. However, the following description corresponds to multiple formats with similar structures including Redis and other "key-value" storages. JSON has almost same characteristics as XML, but it is focused primarily on structured data and probably the record can be shorter due to using brackets instead of elements, which decreases data redundancy.
RDF is a data format used for publishing Linked Data together with schema defined by RDBs and OWL. Its specific form depends on a serialization of RDF, but the main aspect is its graph orientation (even its XML serialization does not use the classic tree relation). Its decentralized web-based concept allows and assumes redundancy of data published by multiple subjects.
Three data metrics
The metrics designed in this paper are marked as follows: • amount of structuredness -( ), • amount of hierarchicallity -ℎ( ), • amount of information -( ), while ( ), ℎ( ), ( ) ∈ 〈0; 1〉. is a data object -a bounded set of data that can consist of partial child data objects, or it can be elementary (the smallest possible object that still allows calculating the values of metrics). The computational formulas also use ( ) as a bit size of the data object and 1 … (or a string length, if all data is textual), which represent the elementary child objects of data object .
The triad of data metrics can be visualized in a "three-dimensional data metrics space" -a cube with edge length 1 (all values ∈ 〈0; 1〉), where one corner is the point [0; 0; 0] and the three adjacent edges correspond to these three data metrics (see further Figure 3).
Amount of structuredness
As mentioned in the Background section, this paper, compared to the three-level structuredness concept (structured, semi-structured and unstructured), considers structuredness as a continuous value.
The requirement for this metric is its ability to respond to two aspects of the examined data -its ability to break down into smaller parts and the degree of structure of individual maximally distributed parts of data (child elementary data objects). For elementary data objects, a significantly simpler way of determining the value of this metric can be expected. Both aspects should increase the resulting value of the degree of structuring of the examined data. This is also the reason why a simple weighted average calculation cannot be used -it ignores the ability of data to distribute.
Calculating the amount of structuredness of a large data object with multiple levels of child objects is based on splitting the objects into elementary child objects, while all middle level (non-elementary) child objects are ignored.
This amount is calculated by Formula 1, where ( ) is the result amount of structuredness, is a bit or text size function, is the data object and 1 … are its the elementary child objects. The weight of a data elementary child object is Such a setting of the formula means that the amount of structuredness of, for example, a data object 2 consisting of two equally large child objects (one structured and one unstructured, both having a weight of 1 2 ) is not 1 2 , as would be the case with a simple weighted average calculation, but the following: The reason is that the object is divisible into two child objects; therefore, the unstructured child data object reduced the value by only 0.25.
This method is based on the assumption that the structuredness ( ) of an elementary data object is obvious and its determination can be automated -an elementary object is usually a small single value (e.g. a number or a single-word string in a database etc.) or a larger structure that cannot be split without losing the information it represents (e.g. a sentence, any longer text as well as media files). The principle of this method is that the unstructured elementary object decreases the structuredness of object proportionally to its share of the object.
For example, a single long text is an unstructured elementary object. By organizing thousands of such texts into a list, database etc., we get a much more structured superior object. Therefore, the limit of ( ), as the number of data objects increases, is one; the organization of objects next to each other and the ability to split their superior data object makes it more structural.
Amount of hierarchicallity
The amount of hierarchicallity expresses how a data object fulfills the definition of hierarchical structure -a structure that can be expressed as a tree with a root node and its children on multiple levels (e.g. a web page or a XML file) without creating a redundancy. Formula 2 shows the definition of entity having a hierarchical property is the following (based on the definition from (Musca et al., 2011)).
Formula 2. Hierarchical property definition.
Relation ( , ) means that is 's parent -for example in a database, a row in a table refers by a foreign key to another "parent" row, which creates 1: relation, in XML a child element has a parent element etc. An entity referring to multiple others (having for example : relation) breaks the definition of hierarchical property -such structure cannot be represented by a tree without repeating information from another entity. More examples of this problem are given in (Vodňanský, 2016).
The amount of hierarchicallity is calculated by Formula 3.
This method is also based on splitting a data object into elementary child objects, while ( ( )) is the size of an elementary child data object that is not a part of any data object breaking the definition of hierarchical property and is not a child object of any object breaking the definition.
Amount of information
As the background stated, Shannon's concept of information entropy is not usable for complex data structures. This paper applies AIT by measuring the compression rate of the whole data object with a given compression algorithm. This method can be computationally demanding on large objects and no algorithm can achieve the lowest possible size of compressed data, but using AIT is possible on data of any type and structure with a clear result.
The amount of information is calculated by Formula 4.
Formula 4. Amount of information.
Function is the compression function -( ( )) is the size of the compressed data object and ( ) is the size of the original object.
The characteristics of data structure types listed in the previous section are summarized in Table 1, which includes the intervals of data metrics values of an optimal data storage -a storage filled by data with the most suitable characteristics.
Data transformations classification
The triad of data metrics allows classifying a transformation of data, based on the changes of metrics values. Data transformation is a process in which data object 0 is transformed into data object 1 , while both objects represent the same information.
Before the process of transformation, the system of three metrics helps to calculate the compatibility of data object 1 and data structure type 1 . Formula 5 shows the compatibility calculation (function ).
Formula 5. Amount of compatibility of data object 1 and data structure type 1 .
Data object is defined by a triad of data metric values -amount of structuredness ( 1 ), amount of hierarchicallity ℎ( 1 ) and amount of information ( 1 ). Data structure of type is defined by three intervals or values of these metrics, as shown in Table 1. The number √3 is the longest possible distance (geometrically, it is the inner diagonal of the cube). Geometrically, compatibility is based on the distance between a point ( 1 ) and a block ( 1 ) inside a three-dimensional data metrics space (estimated transformation length) -the shorter this distance is, the higher is the compatibility.
The real data transformation length is the distance of two points (scaled on the interval 〈0; 1〉) representing the first object 0 and the transformed object 1 . This number has the opposite function than compatibility -the higher the compatibility, the lower the transformation length should be; both numbers can be compared with each other. The calculation is shown in Formula 6.
Formula 6. Data transformation length between data objects 0 and 1 .
The transformation can be classified into 6 basic types (one data metric value is changing) of transformations and multiple other combined types (more metrics values are changing). The basic types are described by the following parts.
Structuralization
The process in which the amount of structuredness increases is typical to many job roles and disciplines, including data modelling and filling a database, accounting, semantic annotation (explained in more detail by (Oren et al., 2006)).This process is difficult to automate; one of such methods is FRED (Gangemi et al., 2017), a tool that transforms natural language text to RDF/OWL format, OCR (optical text recognition from images) and other natural language processing methods. In this context, (Šperková, 2014) mentions the process of "sorting, cleaning and transformation of unstructured data into matrixes".
Structuralization is neutral to the other transformation types -it allows increasing and decreasing both other data metrics.
Destructuralization
This inverse transformation type is typical for most interpersonal communication based on using structured data, e.g. a simple presentation. This process can only keep or increase the amount of hierarchicallity -by using less structured data, the definition of hierarchical property cannot be broken. The amount of information typically decreases but keeping this amount or even normalizing the data while being destructuralized cannot be excluded.
Hierarchization
This process, in which violations of hierarchical property definition are being removed, is typical for using data with a strict hierarchical structure -creating a webpage, XML document or even this paper itself (if inner cross-references are excluded). This process is described in more detail in (Vodňanský, 2016).
Hierarchization is neutral with respect to structuredness. It can very probably increase the level of redundancy when non-hierarchical structures are transformed and decrease the amount of information, while it cannot normalize the data and increase it in any way.
Dehierarchization
This process creates breakings of the definition of hierarchical property. It can be assumed that this creation of double reference from a single entity is intentional (a reference cannot occur randomly). It is typical for processes creating graphs and diagrams -data models, ontologies, mind maps etc.
This inverse transformation can increase the amount of structuredness. This metric cannot be decreased just by organizing data non-hierarchically. It also very typically occurs together with the normalization process, while denormalization is practically excluded -a non-hierarchical reference does not create redundancy, only if it is redundant itself.
Normalization
The process, in which redundancies are removed and the amount of information increases, is very popular in RDBs and its principles are described in multiple papers, e.g. (Codd, 1990;Halpin, 2001). A more detailed description and examples are superfluous for this paper. This paper understands normalization as a process applied on existing data, not just its schema and not just RDBs.
It is typical for the same activities as the strongly related dehierarchization and it excludes hierarchization due to reasons set out in that section. This process is typical for structured data and Structuralization, but it is admissible even for unstructured data, which can be less redundant, therefore it is neutral to the amount of structuredness.
Denormalization
This inverse process, which creates redundancy, occurs when data is being prepared for easier or faster access for a machine (a data cache storage) or a human.
It can probably decrease structuredness but allows its increase (for which there is probably no practical example) and it also probably increases hierarchicallity; dehierarchization cannot denormalize data.
Combined transformation
The discussion in the paragraphs above described the basic transformation types and their relation to each other. These relations can be combined into more complex types, in which two or three metrics change. By excluding the combinations described above and inverse pairs, 18 types exist and can be illustrated by the following Venn diagram (Figure 1), where a letter represents the initial letter of the transformation type name and an added letter D represents the inverse type.
Metadata warehouse
The metadata warehouse principle is based on the popular OLAP (online analytical processing) concept used in BIT (Business intelligence technologies). The metadata warehouse does not only serve as a repository, it is also an instrument of calculation itself -the values for parent objects are calculated from already stored values for lower-level objects.
There are 3 main differences from the OLAP concept based on tables of facts and tables of dimension. The first difference is that it uses the three metrics together, which can serve as dimensions as wellhierarchicallity can be studied on multiple levels of structuredness and vice versa. The second difference is that values cannot be aggregated -they must be computed separately for an object on all levels (multiple data objects may contain no redundancy but can be redundant to each other inside the parent data object, they can break the hierarchy together or be organized more structurally). The third difference is that the structure of the warehouse is universally set for various data, while the OLAP cube is usually designed for a specific usage.
The metadata warehouse can be built once (to study a specific situation ad-hoc) or continuously (to monitor data changes in time). The main principle of storing metadata in a warehouse is its referencing to the unique data it represents (in a format depending on data structure type). Metadata warehouse is a relational database with structure shown in Figure 2. The first table shows an object that has its primary key, foreign key referring to a parent data object inside which this object is contained (the main, "root" object has null value and there must be no cycle), bit size (or string length for fully textual data), three amounts of metrics, reference to the represented data explained above and two datetimes defining the time range when the data object is changeless -each change in any parameter creates a new data object inside the warehouse with a new time range which can express a data transformation.
The second table shows the : relation, storing the information about the data mutual referencing of data (foreign keys in database, links in HTML, predicates in RDF data), when the referrer and rereferred objects are distinguished (they usually are in relation with 1: cardinality). This table is used for hierarchicallity calculation.
Metric-based data visualization
The triad of metrics allows three-dimensional visualization of most situations described in this paper. This visualization method uses a cube with the amount of hierarchicallity on axis , the amount of information on axis and the amount of structuredness on axis . While all metrics belong to the interval 〈0; 1〉, the three-dimensional space is represented by a cube shown in Figure 3. The axis from top to bottom is the amount of information, the axis from left to right is the amount of hierarchicallity and the oblique axis from bottom left to the middle is the amount of structuredness. Inside this space, a data object is shown as a point with three metric values as coordinates. Data structure types, based on Table 1, with intervals or single values are shown as blocks, rectangles, lines or pointstheir intersections show areas compatible with multiple structure types. Data transformations are shown as arrows, representing a vector from one data object (point) to another. Webpage (Vodňanský, 2020) proposes an open-source application for 3D visualization of data object metrics and data structure types.
Method evaluation by example case study
This part shows three practical examples of using the principles on two purposefully created demonstration data sets with minimal size to illustrate the calculation. Both data sets represent the same information. Both are used to show the principle of data metrics and compatibility calculation. They are also used to show an example of data transformation, expressed by transforming the first data set into the second one.
Relational database
The first data set consist of a RDB with 3 tables, based on the topic of organizing sport courses. The database table are shown in Table 2, Table 3 and Table 4. The metadata warehouse is built in six steps.
Step 1: This step creates the "data_object" table, while id, creation ("existence_from") and location specification is set (in the format "/table/id/column"). This process starts from the main "root" object to fill out the reference to the parent object.
Step 2: This step fills the "data_reference" table, while in RDBs, only objects at RDB row level are practically referring using their foreign keys.
Step 3: This step fills the size of the objects. Due to the text format of the data, a simple string length can be used instead of bit size.
Step 4 Step 5: This step calculates the amount of structuredness. All elementary objects are structured except the description of Area 1 (data object 7), which contains a sentence. This data object decreases the metric value for all parent objects in proportion to its share, therefore the amount for object 3 is (1−1))+54(1− 54 68 (1−0)) 68 =0.3546, where 13 is the size of child object 6, 54 is the size of child object 7, and 68 is the size of the entire data object 3. Other data object values (especially 2 and 1, which are the parent of this one) are computed analogically.
Step 6: This step calculates the amount of information. This applies the AIT principle and uses bzip2 algorithm to compress the data (in this case, a simple text representation of the data objects, in which the child objects are split by a single space). The value of root data object 1 is 204 238 =0.8571, while 204 is the length of the compressed text and 238 is the original length. For very small data objects, the compressed text can be longer than the original; then the value is 1.
The final metadata warehouse is shown in the Appendix A. The compatibility of this (root) data object can be calculated as Analogically, the compatibility with the other data type structures can be calculated, which results in Table 5. It shows that even though this data set is a RDB, it is compatibility with RDBs is decreased by an unstructured child object, but it is still the most appropriate data structure type together with RDF.
Compatibility with XML and JSON is decreased mostly by the data set's hierarchicallity, which is only 0.5278.
JSON
The second data set consist of the same information as first one, but it is organized hierarchically in accordance with the JSON format: The metadata warehouse is built in six steps again. The following steps list only the differences from the first data set.
Step 1: The location specification is based on sub-object names and numbers in square brackets (in case of array items), separated by a slash.
Step 2: The JSON data objects have no explicit mechanism of interval referencing, therefore the "data_reference" table is empty and omitted.
Step 3: The size of the data object is calculated from the text length without delimiting parentheses or brackets but including all whitespace inside. Therefore, the size is not always equal to the sum of all children object sizes.
Step 4: As mentioned in step 2, the amount of hierarchicallity in the JSON data structure type is always 1.
Step 5: The amount of structuredness is calculated analogically to the previous dataset. The JSON format includes a lot of whitespace and delimiting characters. A sum of children data object sizes instead of the size of the data object itself is used.
Step 6: This step applies AIT the same way, including inner whitespaces and delimiters.
The compatibility is shown in Table 6. The metadata warehouse in the Appendix B. It shows, that even though this data set is JSON, it is more compatible with XML (due to unstructured child object occurrence) and therefore the selected data structure type is not optimal. Also, the very low amount of information 0.2426 (high redundancy) of the data makes it very incompatible with RDBS data structure type.
Data transformation
The transformation study uses results from the first and second data set as the initial and final state of transformation. The data from the relational database is transformed into a JSON document, which is a common process, e.g. in web-based technologies. The changes of metrics values can classify this transformation as a slight structuralization (number of structured data objects and their levels has increased due to the created redundancy), hierarchization and denormalization (S, H, DN transformation type area in Figure 1). This can be written as the following vector: From this, the transformation length can be calculated ( 0 is the first row and 1 is the second row in Table 7): ( 0 , 1 ) = 1 √3 √0.1179 2 + 0.1033 2 + (−0.6145) 2 = 0.3661.
Figure 4. Data transformation in a three-dimensional space.
The length is approximately one third, which means a lower level of complexity and therefore a less expensive (demanding) data transformation process. This corresponds to the compatibility of the first data set and JSON type, which is 0.6974 (approximately two thirds), as shown in Table 5, which is inverse to the transformation length. Also, guidelines for points 0 and 1 have been added to help to imagine the position of the points (this paper does not allow inserting an animation of a rotating space, which would be more illustrative).
Discussion
This paper opened the topic of data structure type and its corresponding data heterogeneity and calculation possibilities. It proposed a basic theoretic method of characterizing data and their structure types by a triad of metrics. These metrics values can be used to compare data with each other and to analyze data transformation process complexity.
Transformation process complexity can also be predicted by calculating the compatibility between data and data structure type. The calculated values can be stored in a metadata warehouse and allow a further application of statistic methods. The transformation process can be classified into six basic types and multiple combined types.
The triad of values and their changes have shown the ability to visually represent the state of a data object, the optimal area of data structure type and how the values change direction in three-dimensional space. The calculated compatibility can detect a non-optimal storage design.
Using this approach can help to answer the question asked in the Introduction: 1. Data can be stored the more easily by XML or JSON the more hierarchical it is. The higher the amount of structuredness is, the more suitable it is for the JSON format. The example case study has shown the hierarchicallity difference. 2. Data objects with the highest amount of hierarchicallity are most compatible with the XML format.
When the example case study hierarchized the data from the relational database, the amount of information strongly decreased -a redundancy was created. 3. Data unsuitable for relational databases are data with a high redundancy (low amount of information) and a low amount of structuredness. The process of data normalization and structuralization can require the input of an expert. 4. The most dynamic parts (child objects) of a data object can be detected by simple size volatility.
The triad of metrics can explain the reason for size change (redundancy, low structuredness etc.). 5. This is partially explained in answer 4 -size is measured alongside the three metrics. The source of the size increase can be traced inside the metadata warehouse from root to smaller child objects. 6. The most redundant data parts can be detected by combinations of child data objects whose parent data object has a low amount of information despite the fact that the amount of information of the child objects is significantly higher -such objects might be mutually redundant. This can be seen in parent object 5 and child objects 6 and 10 in Appendix B.
The example case study demonstrated usage on simple data sets, which experimentally evaluated the usability of the described methods, which was the objective of the paper. It showed a very typical situation, when non-hierarchical data (RDBs) are hierarchized into JSON, which creates redundancy.
The main limitation of these methods is their dependency on an expert, who needs to understand the data and to design the calculation and data warehouse storage process -no full or partial automatization method has been found, but it probably can be further designed for typical data structure types. Also, using algorithmic information theory on very large data can be computationally demanding. The metrics also showed that there can exist some mutual dependency between them (especially the amount of information and hierarchicallity), but a high level of hierarchicallity is not the only reason for redundancy and normalized data can also be hierarchical.
The designed methods can be used in data modelling as well as in part of the data mining process -it can help to select an optimal data structure type or evaluate the optimality of an existing data storage and provide feedback. It can help to reveal problematic areas where data volume increases due to redundancies caused by inappropriate hierarchization or unstructured storage of structured information.
A continuous and automated data analysis applying the described method can be used to monitor any base of data (not only a database system itself) for a longer time. This process has a similar principle to OLAP storage, but it focused mostly on database engineers, developers, and data curators rather than business.
Most processes in software development perform some kind of data transformation. Calculated compatibility can help to select the optimal structure type. Transformation length indicates the real complexity of the transformation. This can help to select the direction of transformation and to decrease the complexity and the overall cost of the transformation process. Despite possible external influences on the selection of data structure type, this method can bring calculation-based arguments to the discussion about structure selection.
Appendix B
Metadata warehouse of the second data set. Columns with "existence" are omitted due to space requirements and same values as first data set. | 8,390.2 | 2021-01-01T00:00:00.000 | [
"Computer Science"
] |
The nightjar and the ant: Intercontinental migration reveals a cryptic interaction
Abstract Birds and ants co‐occur in most terrestrial ecosystems and engage in a range of interactions. Competition, mutualism and predation are prominent examples of these interactions, but there are possibly many others that remain to be identified and characterized. This study provides quantitative estimates of the frequency of toe amputations resulting from ant bites in a population of migratory red‐necked nightjars (Caprimulgus ruficollis) monitored for 15 years (2009–2023) in S Spain, and identifies the attacker(s) based on taxonomic analyses of ant‐mandible remains found on injured toes. Less than 1% of examined adults (N = 369) missed one or more toes. The analysis of ant remains identified African army ants (Dorylus sp.) as the primary cause of toe amputations in nightjars and revealed that body parts of the attacker may remain attached to the birds even after intercontinental migration. No cases of severe damage were observed in juveniles (N = 269), apart from the mandible of a Messor barbarus – a local ant species – attached to one of the teeth of the characteristic comb of the medial toe of nightjars. The incidence of ant‐bite damage may appear unimportant for nightjar populations, but this might not be true if only birds that manage to survive their injuries and potential complications (e.g. severe bleeding and sepsis from opportunistic infections) return from the tropics. More field studies, ideally in tropical areas, that incorporate routine examination of ant‐induced injuries into their protocols are needed to understand the true incidence and eco‐evolutionary implications of antagonistic ant‐bird interactions.
| INTRODUC TI ON
The interactions between birds and insects have been the subject of extensive research (Morse, 1971;Willis & Oniki, 1978).Bird-insect interactions in the form of predation, parasitism, and mutualism have received the most attention because of their ubiquitous occurrence and importance as study systems for understanding coevolution (Holmes et al., 1979;Maziarz et al., 2021;Price, 1997;Quinn & Ueta, 2008).Other less conventional or more cryptic interaction types also occur along the antagonism-commensalism gradient, such as competition (e.g. for food resources; Smith & Balda, 1979) or 'anting' (i.e. the use of ants as a treatment against parasites; Camacho & Potti, 2018), but they often go undetected or underreported and their true nature and biological significance remain to be elucidated in most cases.
Of all insects, ants (Formicidae) are the most abundant in most terrestrial ecosystems (Del Toro et al., 2012).There is ample literature documenting antagonistic ant-bird interactions, primarily in the form of bird predation on ants (reviewed in Avilés, 2024).There are also studies reporting negative effects of ants on aspects such as bird abundance, behaviour and reproduction, primarily through predation on eggs and chicks (Álvarez-Blanco et al., 2020;Davis et al., 2008;Haemig., 1999).However, assessing the impact of ants on more mobile segments of a population (e.g.juveniles and adults) is challenging and our understanding of the nature and magnitude of these effects is therefore limited.
This study takes advantage of a long-term capture-recapture program to examine the nature and frequency of injuries (toe amputations) of a migratory insectivorous bird -the red-necked nightjar (Caprimulgus ruficollis) -in their breeding area in S Spain.The study was motivated by observations of nightjars missing toes, particularly of an adult bird recorded in 2015 that had what appeared to be the mandible of an ant embedded in the medial toe of one of their feet.
Six years later, in 2021, we found another adult that had the full head of an ant embedded in a central toe.Since then, we engaged in a systematic revision of the feet of all captured nightjars to evaluate the frequency and severity of toe injuries.Observations of insect remains attached to injured toes in three nightjars provided an opportunity to identify the agents of damage based on a taxonomic analysis.We close by discussing the incidence of such injuries on nightjar populations and identify further research needs.
| Study system
Data on toe damage were collected during three breeding seasons Nightjars spend much of their time on the ground: they are ground-nesting birds and both members of the pair share incubation and chick-rearing duties for over 35 days, the female taking the larger share (Camacho, 2013a).During the day, they remain motionless on the ground in the shade of trees or shrubs (Camacho et al., 2014).During the night, they use open spaces on the ground as perches to facilitate detection of aerial insects (mainly moths), although they can also occasionally pick up terrestrial insects that are running around (Jackson, 2003).In spending so much time on the ground, nightjars are exposed to accidental contact with all sorts of diurnal and nocturnal ground-dwelling creatures.
| Field and lab procedures
Nightjars are attracted to roads at night (Jackson, 2003), so road transects are one of the most commonly used methods for nightjar monitoring (Camacho, 2013b;De Felipe et al., 2019).From late March to late October, we conducted nocturnal surveys for red-necked nightjars on a regular (3-7 days) basis along a 24-km road transect by driving a car at slow speed (30 km h −1 ).Nightjars found on roads were captured using a torch and a butterfly net (Jackson, 1984) and marked with numbered metal rings (if not ringed already) for individual recognition.Birds were sexed and aged as first-year (juvenile), second-year (approx. 1 year old), or older using plumage characteristics (Camacho, 2013a;Forero et al., 1995), and measured according to the standard protocol for long-term monitoring of nightjars (see details in Hidalgo-Rodríguez et al., 2021).Many individuals' ages are known because they were first captured as juveniles (Camacho et al., 2022).
To estimate the frequency of damage on limbs, we carefully examined the feet and toes of all captured individuals, noted anomalies (e.g.breaks in skin, and full or partial amputations; Figure 1a) and checked for remains of the attacker (e.g.mandibles or full head; Figure 1b-d).The remains found during the 3 years of systematic record were removed from the bird and preserved in 96° ethanol for taxonomic identification in the laboratory.The remains, specifically a left mandible found in 2021 and a full head found in 2023, were closely examined under an 80× stereomicroscope to identify the attacker to the genus or species level (Borowiec, 2016;Fisher & Bolton, 2016).
We calculated the proportion of juveniles and adults missing feet or toes (or in the process of losing them, as determined from the severity of damage) as a measure of the incidence of damage, and used the Fisher's exact test as implemented in R (https:// cran.r-proje ct.org/ ) to test the null hypothesis that the incidence of damage does not differ between age classes.Non-systematic data collected from birds trapped during ringing operations in previous years (2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019)(2020) provided additional useful information about the nature of damage.
| RE SULTS
Our dataset consisted of a total of 855 captures from 588 individual nightjars that had been examined for limb damage and loss at least once during 97 nocturnal surveys conducted in 2021-2023 (Table 1).Sampling effort, measured as either the total number of captures (range: 282-287) or as the number of different individuals examined (range: 207-225), was uniform across the 3 years of systematic study (Table 1).No evidence of severe damage (i.e.toe amputation) was observed in any of the 219 juveniles examined, although one of the juveniles caught in 2023 had a full (ant) head attached to the claw of the central toe (Figure 1c).Of the 369 different adults captured during 2021-2023, only three missed a toe: a 1-year-old male captured for the first time in 2021 (toe still attached to the foot by a thin piece of skin; Figure 1b), a 5-year-old male captured in both 2021 and 2022 (first caught as a secondyear bird in 2017), and a 1-year-old female captured in 2023 (first caught as a non-injured juvenile the year before).The estimated frequency of damage in adults across the 3 years of study was 1:123 (0.81%), although annual estimates differed as much as 2-fold (Table 1).Four additional adults missing one or more phalanges in one or more toes were recorded during non-systematic checks in previous years (Table 1).Despite the lack of records of toe amputations in juvenile nightjars, there is insufficient statistical support to conclude that there is a greater incidence of damage in adults (Fisher's exact test, p = .299).
Damage generally consisted of total amputation (80% of cases) or incomplete excision (20%) of one or two toes.The most severe damage was recorded for an adult female captured in 2015, which missed three toes of the right foot and one lateral toe and the claw of the medial toe on the left foot (Figure 1a).(Fabricius, 1793).Some 20 species of Dorylus ants have been listed for Mali and Guinea (antsm ap.org), where the captured nightjars spend the winter.It is currently impossible to identify the attacker at the species level, even if we had retrieved the entire individual, and the entire genus is currently under taxonomic revision (Kiko Gómez, personal communication).
The full head attached to the juvenile (year 2023) caused no apparent damage to the bird, apart from nicking one of the teeth of the characteristic pecten (comb) of the medial toe of nightjars (Figure 1c).The full head was identified as belonging to a worker of Messor barbarus, an easily identifiable species that can be distinguished from other species of the genus in the Iberian Peninsula by the unique reddish colouration on the head (Lebas et al., 2017).
| DISCUSS ION
We provided quantitative estimates of the frequency of toe amputations in juvenile and adult red-necked nightjars, evaluated the severity of damage to the birds, and identified the body parts of the ants that remained attached to nightjar toes.Less than <1% of individuals missed one or more toes.Toe amputations were only observed in adults that had performed at least one round-trip migration from Europe to Africa and, although ant-nightjar encounters also occurred in juveniles, these resulted in minor injuries.The taxonomic analysis of mandible remains identified army-ant bites as the most likely cause of toe amputations in adult nightjars.
Our estimate of the incidence of ant bites is based on the proportion of toe-mutilated birds, but the cause of damage could only be confirmed for those individuals that had remains of ant mandibles or the full head on their toes.None of the nightjars examined in this population (>4500 captures) exhibited typical poxvirus-like lesions on the eyes, feet or toes, suggesting that poxvirus infection does not play a major role in the observed toe amputations.Besides amputations, over the 15 years of study we have observed structural injuries to the sternum (n = 7), tarsus (n = 1), mandibles (n = 4), and toes (n = 5) of nightjars, affecting <1% of all trapped individuals.
These injuries are hypothesized to reflect healed fractures based on the presence of bone callus formation and/or angular deformity.
Damaged toes due to traumatic fracture differ from confirmed antbitten toes in that fractured toes often become distorted in position, pointing obliquely across the other toes, but the skin remains intact.
By contrast, army ant bites result in a break in the skin of the affected phalanx and, even though the compressive force applied to the toe is not enough to fracture the bone, prolonged compression might cause ischemia and subsequent necrosis of the bone tissue, ultimately resulting in the excision and loss of one or more of the phalanges or the entire toe.The anatomical similarities of mutilations among birds missing toes and the confirmed ant-bite injuries in birds in the process of losing a toe suggest that ant bites, rather than laceration from trauma (e.g.road traffic or building collisions) or predator attacks (e.g.carnivores), are the most likely cause of the observed toe amputations in adult nightjars.
No trace of severe damage from ant bites was seen in juveniles despite their sample size being even larger than that of adults ( 1145juveniles vs. 1104 adults considering all years of the long-term study).Nevertheless, the total number of injured birds is too small to support the statement that injuries from ant bites are restricted to adults.The observation of the full head of a M. barbarus attached to the pecten of the medial toe of a juvenile demonstrates that they are not entirely free from ant attacks, although ant species occurring in south Spain are likely too small for their mandibles to grasp on the relatively thick toes of a nightjar.M. barbarus is very common on the Mediterranean coast (Lebas et al., 2017), but it may not pose as much of a threat to nightjars as army ants because M. barbarus feed mainly on seeds and their mandibles are relatively small in comparison to those of army ants.
The genus Dorylus ranges from sub-Saharan Africa across North Africa and Asia Minor to Borneo in Southeast Asia.The Afrotropics harbour the largest number of species (116 so far described), including subterranean-, leaf litter-and surface-dwelling species.Of these, surface-dwelling species have, among other characteristics, the greatest mandible length and mandibular aperture (Schöning et al., 2005).Dorylus species are generalist predators that take any type of prey, from immature individuals of other insects to vertebrate carrion (Schöning & Moffett, 2007).Unlike leaf litter-dwelling and subterranean species, Dorylus ants spend much of their time foraging on the soil surface, and given their mobile, predatory lifestyle and aggressive behaviour, attacks on the birds they encounter on the ground should be common.
Nineteenth-century naturalists assumed that birds eat army ants but, as evidence accumulated about the foraging patterns of birds, it became clear that birds peck only at ants attacking their feet or plumage (Willis & Oniki, 1978).No mention of army ants is made in the published diet studies of Afrotropical nightjars (Jackson, 2000).
In addition, although army ants are major predators of vertebrates (e.g.birds, sea turtles), most documented cases of predation on birds took place during the nesting stage (Ikaran et al., 2020;Robinson & Robinson, 2001).None of these organisms therefore appear to serve as a trophic resource for the other, suggesting that ant-nightjars encounters are coincidental and detrimental to at least one and possibly both parties.
Missing and damaged toes have also been observed in other migratory and sedentary bird species and, as in the red-necked nightjar, these injuries were hypothesized to result from ant attacks (e.g.Ibáñez et al., 2007;Sugg, 1974).Sugg (1974) that it might be used as a comb for removing parasites during preening, or to straighten out the rictal bristles (Cleere, 2010).Moreover, as evidenced by at least one of the injured nightjars, multiple ant bites on the same bird might cause multiple amputations, potentially compromising locomotion on a chronic basis and, ultimately, longterm survival.Breaks in the skin from ant bites, even if affecting only one toe, might increase the risk of fatal complications, such as severe bleeding and sepsis from opportunistic infections, so some individuals may die before migrating and being captured on their breeding grounds.That is, only the fraction of birds that survive ant attacks and subsequent damage might be recorded by us.Hence, the true incidence of ant bites on birds could be much higher than it appears to be. and evolution (Hendry et al., 2022).
(2021-2023) as part of a long-term (2009-present) study of rednecked nightjars in the Doñana National Park, S Spain (see Camacho et al., 2014, 2022 for details on the study site and sampling protocols).Most nightjars arrive in the study area in early May and generally postpone their migration departure until mid-October (i.e.approx.5½-month stay in the breeding area; Camacho, 2013a).GPS tracks of adult nightjars from the study population indicate that most individuals migrate through Morocco and Mauritania to spend the non-breeding season (November-April, approx.5½ months) in S Mali and SE Guinea (C.Camacho and P. Sáez-Gómez, unpublished data).
The encountered left mandible (year 2021) measured 2.3 mm in length by a maximum width of 0.8 mm, was dark brown in colour F I G U R E 1 Observed ant-induced injuries on Nightjars (a-c), and detail of Dorylus sp.mandible (d) collected from one of them.Note that image (c) does not correspond to a Dorylus ant, but to a Messor barbarus found on a juvenile nightjar.Since bird-ant interactions can occur at any time, individuals are included for as many years as they have been checked.Note that for this reason the summation of annual numbers of recorded adults does not match the grand total of different adults recorded in 2021-2023.TA B L E 1 Number of captures, number of individuals examined, and proportion of injured birds recorded during the 3year period of systematic study and the previous 12 years.and ended in a bifurcation, with a robust central tooth.It was identified as belonging to a worker ant of the army ant genus Dorylus reported on the specific bite effects by Dorylus sp. on 10 out of 343 (2.9%) pied kingfishers (Ceryle rudis) from a breeding colony on the Kenyan shore of Lake Victoria.Toe amputations and damage from ant bites appear to be more common in pied kingfishers (3%) than in red-necked nightjars (<1%) despite the latter spending more time on the ground.The estimated proportion of toe amputations in another population of red-necked nightjars monitored for 6 years in E Spain is 0.4% (three cases out of 751 different adults examined in 2017-2022; J.M. Zamora, personal communication).The field protocols in this study, as in the first years of our long-term study, did not include routine screening for toe or foot anomalies, so estimates are based on opportunistic observations made during ringing sessions.The proportions of toe amputations in both populations, as estimated from opportunistic observations, are identical (0.4%) and half the magnitude of that based on a systematic examination of both feet (0.8%), suggesting that routine screening for anomalies is needed to obtain unbiased estimates.Caution is therefore required in extrapolating our results to other populations, species and geographic regions.Different estimates could be obtained if we study nightjars in their wintering area (where ants occur).The frequency of ant-bite damage in breeding nightjars appears almost negligible.However, this might be an underestimate if only birds that manage to survive their injuries migrate successfully from their non-breeding grounds to our study area.Most injuries observed on nightjars, as in the pied kingfishers, consisted of single-toe amputations that should not necessarily impair the birds' functional capacities and compromise performance in tasks such as predator avoidance or food acquisition.One of the injured nightjars actually survived from one breeding season to the next.Nonetheless, the loss of the medial toe and, therefore, of the pecten could be detrimental to nightjars because, even though the exact purpose of the pecten remains unclear, it has been suggested
Further
research is required to understand the true incidence and eco-evolutionary implications of antagonistic ant-bird interactions.Being accidental, the factors promoting ant-nightjar encounters are difficult to determine, and consequences different from injuries (e.g. a change in behaviour, systemic symptoms) are difficult to assess and require more detailed study.Fortunately, migratory nightjars are becoming popular model organisms in movement ecology studies across the globe, including capture programs conducted in tropical areas during the non-breedingseason (e.g.Cockle et al., 2023;Evens et al., 2023).Thus, it is possible that studies similar to ours could be attempted in other populations, species and geographic regions.More of these studies should encompass areas of the distribution range of large, tropical ants and incorporate routine examination and systematic recording of injuries into their protocols to increase the chances of encountering clear evidence of the agent of damage and to obtain reliable estimates of the true frequency of injuries and their origin (stochastic or not).Moreover, studies that consider aspects of performance (e.g.mating success, foraging success) and components of fitness (e.g.survival, reproductive success or lifespan) in injured birds relative to uninjured ones are encouraged as an essential step to understand the influence of injuries on natural selection | 4,557 | 2024-05-01T00:00:00.000 | [
"Environmental Science",
"Biology"
] |
Forecasting Methane Emissions from Hard Coal Mines Including the Methane Drainage Process
: With regard to underground mining, methane is a gas that, on the one hand, poses a threat to the exploitation process and, on the other hand, creates an opportunity for economic development. As a result of coal exploitation, large amounts of coal enter the natural environment mainly through ventilation systems. Since methane is a greenhouse gas, its emission has a significant impact on global warming. Nevertheless, methane is also a high-energy gas that can be utilized as a very valuable energy resource. These di ff erent properties of methane prompted an analysis of both the current and the future states of methane emissions from coal seams, taking into account the possibilities of its use. For this reason, the following article presents the results of the study of methane emissions from Polish hard coal mines between 1993–2018 and their forecast until 2025. In order to predict methane emissions, research methodology was developed based on artificial neural networks and selected statistical methods. The multi-layer perceptron (MLP) network was used to make a prognostic model. The aim of the study was to develop a method to predict methane emissions and determine trends in terms of the amount of methane that may enter the natural environment in the coming years and the amount that can be used as a result of the methane drainage process. The methodology developed with the use of neural networks, the conducted research, and the findings constitute a new approach in the scope of both analysis and prediction of methane emissions from hard coal mines. The results obtained confirm that this methodology works well in mining practice and can also be successfully used in other industries to forecast greenhouse gas and other substance emissions.
Introduction
Methane is emitted from a range of natural and anthropogenic sources. The main anthropogenic sources of methane emissions worldwide include coal mining, agriculture, waste, biomass, and the incomplete combustion of fossil fuels or natural gas distribution [1].
In Poland, nearly 29% of total methane emissions into the atmosphere are related to the direct hard coal mining process. They include ventilation emissions and emissions from methane drainage systems. Hard coal mining in Poland is the second main source of methane emissions into the atmosphere, just behind agriculture (about 34% of total methane emissions into the atmosphere in Poland) [2].
Since methane is one of the most harmful greenhouse gases, such a large share of mining industry in its emissions is a serious problem for this industry. In addition, due to its flammable and explosive properties, methane also poses a significant threat to the safety of mining processes [3,4]. The occurrence of either fire or explosion of methane in underground mining excavations is associated with huge material losses and great perils to the life and the health of the crew. The flammability and the explosiveness of methane are due to its high energy potential, which is increasingly being used in the energy sector. The process of methane extraction from coal deposits reduces its emission into the atmosphere and the risk of explosion or fire, and it allows additional economic benefits for mining enterprises [5]. That is why the ongoing development of the methane drainage process is being observed, which is an opportunity for these enterprises to diversify their production and to make effective use of this gas.
Methane is a gas accompanying the hard coal seams. Once emitted into the atmosphere, it can survive for nine to 15 years (average perturbation life 12.4). Given the amount of greenhouse gas emissions, methane ranks second behind carbon dioxide [6,7]. Methane is a powerful greenhouse gas with a global warming potential (GWP) 34 times greater than carbon dioxide over a 100 year time period and 86 times greater than carbon dioxide over a 20 year time period according to the Intergovernmental Panel on Climate Change [8].
This gas is classified as both flammable and explosive due to the possibility of rapid ignition in the atmosphere with a minimum concentration of 12% oxygen, which may result in local fires and explosions [9].
On the other hand, out of all fossil energy resources, methane is the cleanest source of energy [10]. To be more precise, it has the lowest carbon dioxide (CO 2 ) emission factor (almost two times lower when compared to coal) [11,12].
Methane emissions are inextricably linked to hard coal exploitation processes. Therefore, this article focuses on the analysis of methane emissions reported in the hard coal production process. All methane released during and after mining operations is called coal mine methane. The main sources (associated with coal exploitation processes) from which it enters the atmosphere are as follows [13,14]: − Degasification systems at underground coal mines (also commonly referred to as drainage systems); − Ventilation air from underground mines, which contains dilute concentrations of methane; − Abandoned or closed mines, from which methane may seep out through vent holes or through fissures or cracks in the ground; − Fugitive emissions from post-mining operations, in which coal continues to emit methane as it is stored in piles and transported.
In Poland, the main source of methane emissions is mined coal, the pores of which contain this gas. The mining process opens those pores from which methane enters the mine atmosphere. Then, through the ventilation system, it is emitted into mining excavations to later reach the natural environment. The unused part of methane extracted in the methane drainage process also leaks into the atmosphere. This situation is particularly unfavorable and incomprehensible, as this system is quite expensive and methane itself is considered a very good energy fuel [15,16].
Despite the undoubted benefits of the methane capture process used in mines, related activities are not yet very common in Polish mines.
In 2018, the total methane emissions reported for the Polish hard coal mines amounted to 919.1 million m 3 CH 4 , of which only 317 million m 3 CH 4 (34.6%) was covered by methane drainage systems. Unfortunately, this gas was not used entirely for economic purposes, since as high as 113.9 million m 3 was emitted into the atmosphere. In other words, over 30% of gas was not utilized whatsoever. The use of methane for economic purposes was only 22.2% in 2018 when compared to the total methane emissions. This means that almost 80% of methane from hard coal was emitted into the atmosphere [17].
This results in a significant deterioration of the atmosphere and a waste of valuable energy raw material. In addition, capturing only about 30% of methane released in the exploitation process also causes a high risk of its ignition and explosion in mining excavations. The methane-related hazard is one of the most dangerous risks in underground coal mining. The significance of this problem is also evidenced by the fact that methane is present in virtually all coal mines, and its amount in the exploited seams was reported to have significantly increased in recent years due to their deeper location.
It should also be noted that, in the last twenty years, the number of mines in Poland has significantly decreased. The mass (expressed in millions of tons) of mined coal has also decreased ( Figure 1) [17]. Despite this, the amount of methane released during mining is on the rise. This is the result of the exploitation of coal seams with high and increasing absolute methane content [volumetric amount of methane released into excavations (the so-called methane content in ventilation air) and captured by methane drainage systems per unit of time]. As a consequence, despite the decline in hard coal exploitation in Polish mines, the methane threat is increasing.
Energies 2019, 12, x FOR PEER REVIEW 3 of 28 exploited seams was reported to have significantly increased in recent years due to their deeper location. It should also be noted that, in the last twenty years, the number of mines in Poland has significantly decreased. The mass (expressed in millions of tons) of mined coal has also decreased ( Figure 1) [17]. Despite this, the amount of methane released during mining is on the rise. This is the result of the exploitation of coal seams with high and increasing absolute methane content [volumetric amount of methane released into excavations (the so-called methane content in ventilation air) and captured by methane drainage systems per unit of time]. As a consequence, despite the decline in hard coal exploitation in Polish mines, the methane threat is increasing. Figure 1. The amount of methane released and captured by the methane drainage system in Polish hard coal mines (own study based on [17]).
The amount of methane released from coal seams is also due to a high concentration of coal output, which is done by increasing longwall lengths and mining speed. This combined with a large depth of exploitation and the increasing saturation of coal seams with methane causes its elevated emissions. The average annual increase in the depth of exploitation in Polish hard coal mines is about 8 m/year [18,19]. In addition, at great depths, dynamic release of free methane occurs, leading to adverse effects. Here, methane is trapped in tectonic disturbance zones (faults, crevices) under high hydrostatic pressure of overlaying rocks. These phenomena also lead to a significant increase in methane emissions.
In 2018, the total methane emissions of Polish hard coal mines amounted to 916.1 million m 3 CH4. When considering the amount of coal extracted this year (63.4 million tons), one ton of coal emitted 14.45 m 3 CH4. For comparison, in 1994, one ton of extracted coal resulted in the emission of 5.76 m3 CH4 (with extraction of 132.7 million tons/year and 69 active hard coal mines) [17]. These figures show how significantly the amount of methane per ton of extracted coal changed. It also indicates the importance of the problem that the Polish mining industry needs to face in the coming years. Moreover, it should be noted that the Polish economy depends to a large extent on energy obtained from coal combustion. Therefore, it is impossible to either cease or significantly reduce exploitation in the near future, despite the increasing pressure from the European Union and other organizations concerned with environmental protection. Undoubtedly, both the process of coal The amount of methane released from coal seams is also due to a high concentration of coal output, which is done by increasing longwall lengths and mining speed. This combined with a large depth of exploitation and the increasing saturation of coal seams with methane causes its elevated emissions. The average annual increase in the depth of exploitation in Polish hard coal mines is about 8 m/year [18,19]. In addition, at great depths, dynamic release of free methane occurs, leading to adverse effects. Here, methane is trapped in tectonic disturbance zones (faults, crevices) under high hydrostatic pressure of overlaying rocks. These phenomena also lead to a significant increase in methane emissions.
In 2018, the total methane emissions of Polish hard coal mines amounted to 916.1 million m 3 CH 4 . When considering the amount of coal extracted this year (63.4 million tons), one ton of coal emitted 14.45 m 3 CH 4 . For comparison, in 1994, one ton of extracted coal resulted in the emission of 5.76 m 3 CH 4 (with extraction of 132.7 million tons/year and 69 active hard coal mines) [17]. These figures show how significantly the amount of methane per ton of extracted coal changed. It also indicates the importance of the problem that the Polish mining industry needs to face in the coming years. Moreover, it should be noted that the Polish economy depends to a large extent on energy obtained from coal combustion. Therefore, it is impossible to either cease or significantly reduce exploitation in the near future, despite the increasing pressure from the European Union and other organizations concerned with environmental protection. Undoubtedly, both the process of coal exploitation and its combustion are not environmentally friendly, yet for the time being, the Polish economy is not prepared for rapid changes in terms of transition to other, more ecological energy sources [20]. That is why, at the current stage of economic development, it seems reasonable to develop more efficient methods of coal exploitation and usage, which would also reduce harmful emissions into the atmosphere.
In this regard, it seems appropriate to carry out an analysis in terms of predicted amounts of methane that will be emitted by coal mining in the coming years.
Forecasts for methane release usually include mining excavations in terms of exploitation safety. This type of forecast can be exemplified by the dynamic prediction of absolute methane content of bearing capacity of longwalls [18], the prediction of methane concentration distribution in mining excavations using the Computational Fluid Dynamics (CFD) methods [21,22], or the prediction of methane concentration using artificial neural networks [23,24]. This approach is justified when considering safety of the exploitation process. As already mentioned, methane is considered an immensely hazardous gas.
Nevertheless, from the point of view of sustainable development of the country, other elements are also of key importance. They include the effects methane has on the natural environment and the provision of an additional-in this case, ecological-energy raw material, which methane undoubtedly is. The so-called sustainable economy model, which is becoming effective in the European Union, enforces ecological, socially acceptable, and economically effective measures in this respect.
The exploitation of methane coexisting with coal can be an important factor when improving the economic condition of mining enterprises. With regard to the decarbonization policy strongly supported by the European Union [25,26], effective methane extraction can be crucial for improving the image of coal mining. This process enhances exploitation safety, reduces the negative impact on the environment, and can improve the economic condition of these enterprises.
In this context, knowledge of the future amounts of methane emissions, including its acquisition, may become the basis for making strategic decisions in the field of energy management both in the country and by mining enterprises.
Little research is devoted to the issue of predicting methane emissions from Polish mines. The problem of methane emissions in the atmosphere was studied by [27]. Here, the authors focused on the prediction of methane emissions from a closed mine. In a study by [28], the concept of predicting the amount of methane emitted from closed mines was presented. In study [29], Kirchgessner et al. presented an equation based on the multi-linear regression technique between coalbed methane contents, coal productions, and mine emissions (with R 2 = 0.59). The form of the equation that they developed is a standard multi-linear function with a slope and intercept.
Thus far, no method has been developed to predict the amount of methane (ventilation methane emissions) released to the atmosphere from active mines or the amount of methane captured by methane drainage systems and emitted from these systems into the atmosphere when considering historical data or the emission trend.
Therefore, this paper presents a method for predicting the amount of methane (ventilation methane emissions) emitted into the atmosphere by Polish mines and captured by methane drainage systems. This method enabled the analysis that allowed for the forecast of methane emissions from these mines until 2025. Moreover, this forecast also takes into account the amount of methane captured by methane drainage systems and the amount emitted into the natural environment. The analysis was carried out based on the data on methane emissions from Polish coal mines between 1993 and 2018.
In order to predict gas emissions, statistical methods are most commonly used, including Autoregressive Integrated Moving Average (ARIMA) models for the time series analysis [30,31] and regression analysis [32]. However, for the selected parameters related to methane emissions in the mining production process, the authors decided to use artificial neural networks. The existing literature failed to apply these networks to predict methane emissions into the atmosphere, especially from hard coal mines.
Despite the small number of studies covering the use of intelligent methods to predict emission phenomena, methods based on artificial intelligence should be utilized to a greater extent for this purpose.
The study discuses both the subject area and the developed research methodology. The analysis involved five parameters typical of methane emissions (coal output, total methane emissions, ventilation methane emissions, the amount of methane captured by methane drainage systems, and the amount of methane used) in individual years. Then, the parameters were referred to one ton of extracted coal. Correlations between these parameters were also determined for the analyzed period. Afterwards, a tool developed for and used in the analysis based on the artificial neural network model was discussed. The paper also presents the forecast results together with an error analysis and conclusions. Selected statistical methods were also used to analyze both the data and the findings.
Study Area
The main area where hard coal and coking mines are located and which emits the highest amount of methane in Poland is the area of the Upper Silesian Coal Basin ( Figure 2). In these mines, methane is treated as a raw material accompanying hard coal deposits, which since 1962 has been subject to documentation principles specified in geological and mining law [33]. Since the 1990s, methane has also been treated as the useful gas [33]. Despite the small number of studies covering the use of intelligent methods to predict emission phenomena, methods based on artificial intelligence should be utilized to a greater extent for this purpose.
The study discuses both the subject area and the developed research methodology. The analysis involved five parameters typical of methane emissions (coal output, total methane emissions, ventilation methane emissions, the amount of methane captured by methane drainage systems, and the amount of methane used) in individual years. Then, the parameters were referred to one ton of extracted coal. Correlations between these parameters were also determined for the analyzed period. Afterwards, a tool developed for and used in the analysis based on the artificial neural network model was discussed. The paper also presents the forecast results together with an error analysis and conclusions. Selected statistical methods were also used to analyze both the data and the findings.
Study Area
The main area where hard coal and coking mines are located and which emits the highest amount of methane in Poland is the area of the Upper Silesian Coal Basin ( Figure 2). In these mines, methane is treated as a raw material accompanying hard coal deposits, which since 1962 has been subject to documentation principles specified in geological and mining law [33]. Since the 1990s, methane has also been treated as the useful gas [33]. These models clearly show that the greatest methane emissions are found in seams located at a depth of less than 1000 m. Currently, the vast majority of exploited seams are located at these depths [35].
Data
Data from the State Mining Authority [17] and the Annual Report on the state of basic natural and technical hazards in hard coal mining [36] were used to analyze the current state of methane emissions, methane drainage, and methane emissions from methane drainage systems into the atmosphere, as well as to make predictions. Data on the amount of methane emitted from hard coal mines in Poland into the atmosphere (including methane drainage systems) come from measurements conducted by ventilation services of hard coal mines. They are required to both register and pass this information to the State Mining Authority. In order to determine ventilation methane content, a complex measurement procedure needs to be be carried out in each mine. It is based on continuous measurements of the speed of the air stream flowing through mining excavations, methane air concentrations, and the geometry of excavations. Measurements are carried out using stationary air flow velocity and methane concentration meters, which are built-in in mining excavations. In turn, the total methane content is determined by adding up the amount of methane captured by methane drainage systems and the amount of ventilation methane. An advanced ventilation parameter registration system (in most cases automatic) supervised by specialized ventilation services guarantees the reliability of data.
Data on both the extraction and the emission of methane from Polish hard coal mines for the years 1993-2018 are presented in Table 1. The volume of extraction is specified in megatons (Mt = million Mg; mln = million), while the total methane and ventilation content, methane drainage, and methane usage are in millions of m 3 . These models clearly show that the greatest methane emissions are found in seams located at a depth of less than 1000 m. Currently, the vast majority of exploited seams are located at these depths [35].
Data
Data from the State Mining Authority [17] and the Annual Report on the state of basic natural and technical hazards in hard coal mining [36] were used to analyze the current state of methane emissions, methane drainage, and methane emissions from methane drainage systems into the atmosphere, as well as to make predictions. Data on the amount of methane emitted from hard coal mines in Poland into the atmosphere (including methane drainage systems) come from measurements conducted by ventilation services of hard coal mines. They are required to both register and pass this information to the State Mining Authority. In order to determine ventilation methane content, a complex measurement procedure needs to be be carried out in each mine. It is based on continuous measurements of the speed of the air stream flowing through mining excavations, methane air concentrations, and the geometry of excavations. Measurements are carried out using stationary air flow velocity and methane concentration meters, which are built-in in mining excavations. In turn, the total methane content is determined by adding up the amount of methane captured by methane drainage systems and the amount of ventilation methane. An advanced ventilation parameter registration system (in most cases automatic) supervised by specialized ventilation services guarantees the reliability of data.
Data on both the extraction and the emission of methane from Polish hard coal mines for the years 1993-2018 are presented in Table 1. The volume of extraction is specified in megatons (Mt = million Mg; mln = million), while the total methane and ventilation content, methane drainage, and methane usage are in millions of m 3 . Based on the data contained in Table 1, the total and the ventilation methane emissions as well as the amount of methane drainage and the economic use of methane per one ton of extracted coal were determined.
A statistical analysis was performed for the data presented in Tables 1 and 2. It was based on the Pearson correlation coefficients between the studied variables and the correlation matrix between these coefficients. The results are summarized in Tables 3 and 4, respectively.
The Pearson correlation coefficients were calculated from the relationship: The correlation analysis between the variables was carried out for the level of statistical significance at p < 0.05. The analysis of the results showed that the studied variables were characterized by different values of the correlation coefficient (Tables 3 and 4).
For the variables presented in Table 1, the highest value of this coefficient occurred between total methane emissions and methane drainage (0.9686), methane drainage and methane emissions from methane drainage (0.8749), methane drainage and methane utilization (0.8666), and between total and ventilation methane emissions (0.8700).
For the variables presented in Table 2, high values of the correlation coefficient (over 0.9) occurred for all studied variables except for methane drainage per ton of coal and emissions into the atmosphere from methane drainage systems per ton of coal. For these variables, the correlation coefficient value was 0.8810.
Graphic relationships along with the parameters selected for the statistical analysis between the studied variables are shown in Figures 4-7 for the data presented in Table 1 and in Figures 8-11 for the data presented in Table 2. The correlation analysis between the variables was carried out for the level of statistical significance at p < 0.05. The analysis of the results showed that the studied variables were characterized by different values of the correlation coefficient (Tables 3 and 4).
For the variables presented in Table 1, the highest value of this coefficient occurred between total methane emissions and methane drainage (0.9686), methane drainage and methane emissions from methane drainage (0.8749), methane drainage and methane utilization (0.8666), and between total and ventilation methane emissions (0.8700).
For the variables presented in Table 2, high values of the correlation coefficient (over 0.9) occurred for all studied variables except for methane drainage per ton of coal and emissions into the atmosphere from methane drainage systems per ton of coal. For these variables, the correlation coefficient value was 0.8810.
Graphic relationships along with the parameters selected for the statistical analysis between the studied variables are shown in Figures 4-7 for the data presented in Table 1 and in Figures 8-11 for the data presented in Table 2. Based on the statistical analysis, a strong positive correlation can be observed in most cases. A negative correlation can be noted only in the relationship between coal output and the total methane emissions.
This negative correlation should not be interpreted in such a way that a decrease in coal output causes an increase in methane emissions. The increase in methane emissions with a simultaneous decrease in coal output is associated with the phenomenon of exploitation at increasingly greater depths in Poland, where coal seams are more saturated with methane. Therefore, despite the noticeable decrease in coal output, methane emissions are still on the rise.
Methods
The further stage of the analysis of methane emissions from Polish mines involved the main part of the work, i.e., forecast. It aimed at determining the amount of methane emissions from Polish mines into the atmosphere with ventilation air (ventilation methane emissions), the amount of methane captured by methane drainage systems, and the amount of methane emitted from methane drainage systems into the atmosphere. Artificial neural networks were used for this part.
The main purpose was to determine the absolute and the relative values (per ton of coal extracted) of ventilation methane emissions, methane drainage, and the amount of methane emissions from methane drainage systems into the atmosphere by 2025.
Artificial neural networks are alternative calculation techniques that are used in many areas [37,38] for the prediction [39,40] and the classification of large data sets and their analysis (e.g., in the context of finding cause and effect relationships between data) [41], data matching (especially in the event of information overload), and optimization [42,43].
Artificial neural networks (ANNs) were developed for the first time by McCulloch and Pitts [44]. The definition of an ANN frequently recurs to parallelism with biological paradigms, its structure resembling the brainʹs architecture and the human learning procedures [45].
The structure of a simple artificial neural network consists of three layers of neurons, of which the first layer is the input layer, the second is the hidden layer (there can even be several of them), and the last is the output layer ( Figure 12). Based on the statistical analysis, a strong positive correlation can be observed in most cases. A negative correlation can be noted only in the relationship between coal output and the total methane emissions.
This negative correlation should not be interpreted in such a way that a decrease in coal output causes an increase in methane emissions. The increase in methane emissions with a simultaneous decrease in coal output is associated with the phenomenon of exploitation at increasingly greater depths in Poland, where coal seams are more saturated with methane. Therefore, despite the noticeable decrease in coal output, methane emissions are still on the rise.
Methods
The further stage of the analysis of methane emissions from Polish mines involved the main part of the work, i.e., forecast. It aimed at determining the amount of methane emissions from Polish mines into the atmosphere with ventilation air (ventilation methane emissions), the amount of methane captured by methane drainage systems, and the amount of methane emitted from methane drainage systems into the atmosphere. Artificial neural networks were used for this part.
The main purpose was to determine the absolute and the relative values (per ton of coal extracted) of ventilation methane emissions, methane drainage, and the amount of methane emissions from methane drainage systems into the atmosphere by 2025.
Artificial neural networks are alternative calculation techniques that are used in many areas [37,38] for the prediction [39,40] and the classification of large data sets and their analysis (e.g., in the context of finding cause and effect relationships between data) [41], data matching (especially in the event of information overload), and optimization [42,43].
Artificial neural networks (ANNs) were developed for the first time by McCulloch and Pitts [44]. The definition of an ANN frequently recurs to parallelism with biological paradigms, its structure resembling the brain's architecture and the human learning procedures [45].
The structure of a simple artificial neural network consists of three layers of neurons, of which the first layer is the input layer, the second is the hidden layer (there can even be several of them), and the last is the output layer ( Figure 12). In order to predict both the absolute and the relative values (calculated per ton of extracted coal) of ventilation methane emissions, methane drainage, and the amount of methane emitted from methane drainage systems into the atmosphere by 2025, the multilayer perceptron (MLP) network was used. It was trained on real data.
In an MLP model, each xj variable from the input layer links with each neuron of the hidden layer through the connections with the weights wij. These values are then summed up, resulting in the si signal: The differentiable activation function inside the hidden layer converts this signal and sends the result yi (yi = f(si)) to the output layer. The number of neurons in the hidden layer is determined by a predefined approximation error, with the relationship between the number of neurons in the hidden layer and the approximation accuracy being directly proportionate. The weights of a given model are modified by means of an optimization algorithm, and the process is referred to as network learning. Since the error backpropagation algorithm uses the steepest descent method, this algorithm requires a large number of iterations, which may affect the speed of calculations [24,[46][47][48][49].
Results and Discussion
Data sets presented in Tables 1 and 2, characterizing coal output, methane emission volume, methane drainage, and methane utilization. They were divided into two subsets (due to the limited size of the data set): the training data set (80% of cases) and the test data set (20% of cases).
In order to predict absolute and relative values by 2025, networks consisting of a single hidden layer were used. The process of testing this network was carried out for the minimum number of neurons in the hidden layer, which was three. In order to predict both the absolute and the relative values (calculated per ton of extracted coal) of ventilation methane emissions, methane drainage, and the amount of methane emitted from methane drainage systems into the atmosphere by 2025, the multilayer perceptron (MLP) network was used. It was trained on real data.
In an MLP model, each x j variable from the input layer links with each neuron of the hidden layer through the connections with the weights w ij . These values are then summed up, resulting in the s i signal: The differentiable activation function inside the hidden layer converts this signal and sends the result y i (y i = f(s i )) to the output layer. The number of neurons in the hidden layer is determined by a predefined approximation error, with the relationship between the number of neurons in the hidden layer and the approximation accuracy being directly proportionate. The weights of a given model are modified by means of an optimization algorithm, and the process is referred to as network learning. Since the error backpropagation algorithm uses the steepest descent method, this algorithm requires a large number of iterations, which may affect the speed of calculations [24,[46][47][48][49].
Results and Discussion
Data sets presented in Tables 1 and 2, characterizing coal output, methane emission volume, methane drainage, and methane utilization. They were divided into two subsets (due to the limited size of the data set): the training data set (80% of cases) and the test data set (20% of cases).
In order to predict absolute and relative values by 2025, networks consisting of a single hidden layer were used. The process of testing this network was carried out for the minimum number of neurons in the hidden layer, which was three. Table 5 summarizes the structures of neural networks that obtained the highest correlation coefficients between actual and predicted quantities in the conducted training tests. During the optimization process of neural network parameters used for forecasting, various architectures were created, changing the number of neurons and studying various activation functions used in these networks. The network architectures presented in the paper turned out to be most advantageous (the highest values of the correlation coefficient) for individual studied variants. As can be seen, for each of the studied parameters, networks with different structures obtained the best correlation coefficients.
The presented correlation coefficients for the training data set were at a satisfactory level, especially when taking into account the number of data adopted for the forecast (25 values determining the studied variables). Neural networks have special properties. This means that the more data there are, the better the network quality is and the more accurate the forecast will be. The lack of data on the amount of methane emitted by Polish hard coal mines before 1993 prevented the analysis of a larger data set.
The tests allowed the selection of neural networks (structure) that were used to carry out prognostic studies.
Based on the analyses, the predicted absolute and relative values were calculated (per ton of extracted coal) with regard to ventilation methane emissions, the amount of methane drainage, and the amount of methane emitted from methane drainage systems into the atmosphere by 2025. The time horizon of the forecast (ex-ante) covered the years 2019-2025. Figures 13-18 present the actual and the calculated (by the neural network) absolute and relative values of ventilation methane emissions, the amount of methane drainage, and the amount of methane emitted from methane drainage systems into the atmosphere. The findings show that the actual course of methane ventilation emissions was most accurately mapped ( Figure 13). The actual amount of methane drainage ( Figure 14) and the amount of methane emissions from methane drainage systems into the atmosphere were slightly less accurately mapped ( Figure 15).
As mentioned earlier, the amount of data used for the analysis had a significant impact on the results. Since the number of records used for the forecast was small and amounted to only 25, the expost forecast was made only from 2001.
Based on the results, it can also be stated that the predicted values of ventilation methane emissions, methane drainage, and methane emissions from methane drainage systems into the atmosphere were most often overstated for the lowest actual values and understated for the highest The findings show that the actual course of methane ventilation emissions was most accurately mapped ( Figure 13). The actual amount of methane drainage ( Figure 14) and the amount of methane emissions from methane drainage systems into the atmosphere were slightly less accurately mapped ( Figure 15).
As mentioned earlier, the amount of data used for the analysis had a significant impact on the results. Since the number of records used for the forecast was small and amounted to only 25, the expost forecast was made only from 2001.
Based on the results, it can also be stated that the predicted values of ventilation methane emissions, methane drainage, and methane emissions from methane drainage systems into the atmosphere were most often overstated for the lowest actual values and understated for the highest The findings show that the actual course of methane ventilation emissions was most accurately mapped ( Figure 13). The actual amount of methane drainage ( Figure 14) and the amount of methane emissions from methane drainage systems into the atmosphere were slightly less accurately mapped ( Figure 15).
As mentioned earlier, the amount of data used for the analysis had a significant impact on the results. Since the number of records used for the forecast was small and amounted to only 25, the ex-post forecast was made only from 2001.
Based on the results, it can also be stated that the predicted values of ventilation methane emissions, methane drainage, and methane emissions from methane drainage systems into the atmosphere were most often overstated for the lowest actual values and understated for the highest actual values. This was due to the approximation system based on which neural network models work.
In general, however, it can be said that the adopted architecture of the MLP network allowed the authors to accurately predict methane emissions into the atmosphere and methane drainage in the adopted time horizon. It seems reasonable to conduct further research to determine whether and how much the change in the forecast horizon will affect the accuracy of the calculations.
It is clear that, in the following years, a reduction in the amount of methane emitted into the environment should be expected (Figure 13), and by 2021, a decrease in methane emissions from methane drainage systems will have been reported. In this case, an increase in this value after 2021 is disconcerting (Figure 15). This slight predicted decrease in the amount of methane drainage in subsequent years is also not very optimistic (Figure 14). In this regard, decisive actions should be taken to improve the situation.
In order to reduce methane emissions into the environment, it is necessary to improve the efficiency of methane drainage systems. The amount of methane captured by these installations should be increased, and its emission into the environment from these systems should be reduced. It is unacceptable that the captured methane (as an energy raw material with very good properties) should again enter the atmosphere. The results indicate that methane drainage systems need to be modernized and their efficiency needs to be improved by ensuring methane collection from these systems. Failure to meet these conditions may lead to a significant increase in methane emissions into the atmosphere after 2021.
On the other hand, the analysis of forecast results in the scope of studied quantities related to one ton of coal are not very optimistic either. In terms of methane emissions into the environment, these values will be slightly increasing ( Figure 16). The forecast regarding the amount of methane drainage per ton of extracted coal is also unsatisfactory, which shows only a slight decrease ( Figure 17).
The graphic dispersions of the actual and the predicted values of studied parameters are shown in Figures 19 and 20. actual values. This was due to the approximation system based on which neural network models work.
In general, however, it can be said that the adopted architecture of the MLP network allowed the authors to accurately predict methane emissions into the atmosphere and methane drainage in the adopted time horizon. It seems reasonable to conduct further research to determine whether and how much the change in the forecast horizon will affect the accuracy of the calculations.
It is clear that, in the following years, a reduction in the amount of methane emitted into the environment should be expected (Figure 13), and by 2021, a decrease in methane emissions from methane drainage systems will have been reported. In this case, an increase in this value after 2021 is disconcerting ( Figure 15). This slight predicted decrease in the amount of methane drainage in subsequent years is also not very optimistic ( Figure 14). In this regard, decisive actions should be taken to improve the situation.
In order to reduce methane emissions into the environment, it is necessary to improve the efficiency of methane drainage systems. The amount of methane captured by these installations should be increased, and its emission into the environment from these systems should be reduced. It is unacceptable that the captured methane (as an energy raw material with very good properties) should again enter the atmosphere. The results indicate that methane drainage systems need to be modernized and their efficiency needs to be improved by ensuring methane collection from these systems. Failure to meet these conditions may lead to a significant increase in methane emissions into the atmosphere after 2021.
On the other hand, the analysis of forecast results in the scope of studied quantities related to one ton of coal are not very optimistic either. In terms of methane emissions into the environment, these values will be slightly increasing ( Figure 16). The forecast regarding the amount of methane drainage per ton of extracted coal is also unsatisfactory, which shows only a slight decrease ( Figure 17).
The graphic dispersions of the actual and the predicted values of studied parameters are shown in Figures 19 and 20. When analyzing the data presented in Figures 20 and 21, it can be seen that the results for methane emissions from methane drainage systems into the atmosphere per one ton of extracted coal showed the greatest dispersion. For this forecast, the amount of methane emissions showed the least accuracy.
Unfortunately, despite the attempts to search for networks with different configurations, no better results were obtained for this case.
In order to better visualize the results, Figures 21 and 22 present histograms of the actual and the predicted values of studied parameters together with marked density functions. When analyzing the data presented in Figures 20 and 21, it can be seen that the results for methane emissions from methane drainage systems into the atmosphere per one ton of extracted coal showed the greatest dispersion. For this forecast, the amount of methane emissions showed the least accuracy.
Unfortunately, despite the attempts to search for networks with different configurations, no better results were obtained for this case.
In order to better visualize the results, Figures 21 and 22 present histograms of the actual and the predicted values of studied parameters together with marked density functions. Based on the results, it can be observed that, in many cases, we were dealing with the asymmetrical course of both the actual and the predicted values of studied parameters. Also, the determined density functions of real and predicted distributions showed certain differences, which was confirmed by the reported dispersion. Based on the histograms, the distribution of the values of studied parameters could be determined depending on the number of observations (in this case, the number of years in which these values occurred). It can also be seen that, when including the predicted values of studied parameters in these distributions, the range of these values was only slightly broadened. The predicted values most often coincided with the values reported in the studied years. However, they tended to concentrate more within the area of higher values of studied parameters (which was reflected in the shift in the density function of these distributions).
In addition, an error analysis was made between the actual values of studied parameters and their values obtained from the neural networks. The final error values for the studied parameters are summarized in Table 6. When examining the prediction errors determined in terms of absolute and relative values of ventilation methane emissions, methane drainage, and methane emissions from methane drainage systems into the atmosphere, they were found to be at an acceptable level.
The highest mean absolute percentage error (MAPE) error value was 9.9% for methane emissions from methane drainage systems into the atmosphere per one ton of extracted coal. The lowest MAPE error value was 3.8% for ventilation methane emissions. Based on the results, it can be observed that, in many cases, we were dealing with the asymmetrical course of both the actual and the predicted values of studied parameters. Also, the determined density functions of real and predicted distributions showed certain differences, which was confirmed by the reported dispersion. Based on the histograms, the distribution of the values of studied parameters could be determined depending on the number of observations (in this case, the number of years in which these values occurred). It can also be seen that, when including the predicted values of studied parameters in these distributions, the range of these values was only slightly broadened. The predicted values most often coincided with the values reported in the studied years. However, they tended to concentrate more within the area of higher values of studied parameters (which was reflected in the shift in the density function of these distributions).
In addition, an error analysis was made between the actual values of studied parameters and their values obtained from the neural networks. The final error values for the studied parameters are summarized in Table 6. When examining the prediction errors determined in terms of absolute and relative values of ventilation methane emissions, methane drainage, and methane emissions from methane drainage systems into the atmosphere, they were found to be at an acceptable level.
The highest mean absolute percentage error (MAPE) error value was 9.9% for methane emissions from methane drainage systems into the atmosphere per one ton of extracted coal. The lowest MAPE error value was 3.8% for ventilation methane emissions.
In the case of root mean square error (RMSE) errors, the highest values were observed for the amount of methane drainage and the ventilation methane emissions from methane drainage systems.
It can therefore be claimed that, despite a relatively small amount of data, the results are satisfactory and allow the inference process to be carried out as to the direction of actions that can and should be taken in the future to reduce methane emissions into the natural environment and increase its economic use.
Based on the results, the proposed forecast, as with all kinds of forecasts, is subject to some uncertainty/inaccuracy. These inaccuracies also occur when predicting emissions of other greenhouse gases, e.g., carbon dioxide [50] or greenhouse gases in total [51]. Also, they result from the quality of input data and the tools used for analysis. However, more and more advanced systems for registering ventilation parameters are used, which means that these inaccuracies, especially in terms of input data, are becoming smaller. However, in the context of methane emission forecasting methods used thus far, the use of artificial neural networks seems fully justified, as the results obtained are most accurate.
In the previous publications [29,52,53], methane emissions into mining excavations were specified, but no forecasts of atmospheric emissions or methane drainage were included.
In a study by [53], a method to estimate methane emissions from hard coal mines in Poland is presented. This method is based on the algorithm shown in [53] and involves the assumption that the methane emission index (EF) is calculated on the basis of coal output from methane mines and its total methane content. A similar subject is discussed by [29]. All these studies concern the estimation of methane emissions taking into account "a priori" the volume of extraction in a given year.
In this context, the use of neural networks gives much greater opportunities to forecast methane emissions, taking into account a larger number of factors affecting this emission.
From the mining point of view, it seems reasonable to use the developed method to forecast methane and other gas or substance emissions (e.g., harmful dust) not only for the entire industry but also for individual mines. This approach would allow the development of emission reduction programs dedicated to specific enterprises.
The method presented in this article is universal and, based on real data, also allows for methane emission forecasts from both coal mines and other economic sectors, taking into consideration factors that affect the volume of this emission.
Conclusions
This paper presents the method for predicting the state of methane emissions from underground coal mines. Both artificial neural networks and selected elements of statistical analysis were used for this study. Their role was to process data on methane emissions in recent years to later obtain as much information and knowledge as possible with regard to the direction of changes in methane emissions from mining operations in recent years and to determine their forecast for the next few years.
The study utilized data on coal output, ventilation methane emissions, and the amount of methane emitted from the methane drainage system into the environment. The amount of emissions with regard to these factors reported in the past helps forecast the amount of methane emissions into the atmosphere and the amount of methane drainage for the future. However, the presented forecasting method fails to take into account changes in market dynamics, operational practices, government oversight/regulations, or mitigation technologies, which may affect the amount of methane emissions.
Based on the results, it can be stated that the adopted MLP neural network architectures allow the forecast of methane emissions into the atmosphere and the amount of methane drainage in the adopted time horizon with satisfactory accuracy. At the same time, forecast errors in the absolute and the relative values of methane ventilation emissions, the methane drainage amounts, and the methane emissions from methane drainage systems into the atmosphere were at an acceptable level. It can therefore be concluded that, despite the relatively small amount of data, the findings are satisfactory.
The forecast results indicate that the amount of methane emitted into the environment will decrease in subsequent years; however, the pace of these changes seems to be too slow.
Thus, it appears reasonable to increase the intensity of methane capture from both mining excavations and coal seams. From a technical point of view, this course of action is most justified. However, the amount of methane captured in the methane drainage process needs to be significantly reduced, as such situations should not occur in this process. Ultimately, methane emissions from methane drainage systems need to strive for zero.
A serious problem was reported in terms of safety and the chance for economic extraction of large amounts of methane. It is related to increasing methane emissions per ton of extracted coal. Deteriorating coal mining conditions increase the amount of this gas in the exploited seams. This was also confirmed by the distribution of gas emissions shown on the histograms. This information should be used to take measures to decrease methane emissions into the environment and increase its usage.
The following approach to solving the problem of methane emissions during the mining production process should ideally fit into the implementation of the concept of sustainable development economy. Further development of the mining industry in the era of European decarbonization must also take greater account of both social and ecological aspects. In the case of methane, this seems to be achievable because this gas is very attractive in terms of energy and ecology (clean fuel).
It is therefore indisputable that methane, as a high-energy gas, should not be emitted into the atmosphere but used for economic purposes. Therefore, it is necessary to increase the efficiency of methane capture by methane drainage installations and to use it for electricity and heat production. This energy can be used by the mines themselves, and its surpluses can be an additional source of income.
One of the aims of the presented work was to show how small amounts of methane are currently being used in the energy sector. However, it is completely incomprehensible to emit methane captured by methane drainage systems into the atmosphere. In this respect, it is necessary for the mining enterprises to take decisive actions, as there are technologies that allow the economic use of methane, such as cogeneration and trigeneration systems.
To sum up, the developed method and the obtained results can effectively support decisions made to decrease methane emissions into the atmosphere and increase the amount of methane utilized.
The neural networks and the statistical methods used for the analysis enabled the acquisition of a substantial amount of important information and knowledge in the field of methane emissions in the mining production process. | 12,146.4 | 2019-10-11T00:00:00.000 | [
"Environmental Science",
"Engineering"
] |
A FAHP-FUZZY Approach of Evaluating Banking Service Quality
With the development of the interest rate liberalization in China, the narrowing of spreads has great influence on China's commercial banks operating income, the commercial banks "innovation" pressure is increasingly urgent. The electronic bank, as the main means of innovation, was being developed rapidly in recent years, which played an important role in reducing the constantly increasing intermediary business income during operation. The electronic banks promote the business innovation, and it is very convenient for the customer service. The purpose of innovation is to improve the quality of service. Bank service quality is an important factor to affect the bank's reputation and an important means of competition among banks. Based on the model of service quality of commercial banks, according to Chinese current situation, the paper puts forward an evaluation method using the FAHP-FUZZY model of service quality evaluation, in order to promote the continuous improvement of commercial banking service quality.
Introduction
Since the year of 2012, because of the economic downturn and the People's Bank of China (PBC) adjustment of the floating range of interest rate by midyear, one research has shown that the Interest Expense is the most important operating expense for banks (about 85%), the narrowing of spreads creates a huge challenge for bank management, also the interest rate liberalization will impact on the profitability of commercial banks.By the end of April, the 2012 annual report of listed banks had showed that the Industrial and Commercial Bank of China, the Agricultural Bank of China, the Bank of China, the China Construction Bank and the Bank of Communications those assets scale was nearly 50% of the total Chinese banking, and the five banks had realized a net profit of 774.6 billion Yuan, increased 14.9 percent at the same period of last year.The average growth rate of 2012 fell behind the year of 2011 by more than 30 percent.A number of phenomena show that banking has turned into the market-oriented and low-growth industry from the semi-monopolistic and high-growth industry.In the process of declining growth, banks are facing the functional restructuring and the business transformation, that is, banks whose customers are mainly large enterprises begin to serve small businesses.Under the constraints of competition intensifies in the domestic and international interbank, banks will have to adjust customer structure and improve service quality in order to seek for new sources of profit.Banks' service quality plays an important role in their reputation, which is a main approach during competition among banks as well as the core competence of banks' sustainable development.Therefore, how to improve banks' service quality and the structure of asset quality is an urgent subject in the background of market-oriented interest rate in China.Grouroos, the representative of Nordic service quality school, embodies the idea of customers on Service Quality, and thus it can be called "market-oriental quality of service".Grouroos also suggested that the quality of service should be classified into Outcome or Technical Quality and Process Quality (Gronroos, C., 1984).The former highlights the objectivity of the evaluating while the latter emphasizes the subjectivity of the evaluating.Fomell classified the quality of service into Customization (the degree of suppliers' meeting the customers' demands) and Reliability (the stability and accessibility of service from service providers) (Fornell, C., Michael, D. J., et al, 1996), which are similar to Ravald Grouroos's Outcome or Technical Quality and Process Quality.According to Ms. Wen Yanbi's and Mr. Wang Xiaochun's researches, the overall quality of service not only includes hardware quality and software quality but also contains contact fairness, result fairness and procedural fairness; the above-mentioned three fairness and two service qualities equally make up the sub-factors of the overall quality of service (Biyan, W., & Chunxiao, W., 2005).However, the evaluating process of service quality is the process of the subjective perception of value, during which contact fairness and procedural fairness have a high correlation with Process Quality or Software Quality while result fairness with Technical Quality or Hardware Quality.Hence, the evaluating of service quality remains two main means of Process Quality (or Software Quality) and Technical Quality (or Hardware Quality).
Brands and distribution are very important in the purchase and consumption of tangible products because in most cases consumers cannot have an acquaintance with manufacturers and can only infer from brands and distribution for what manufacturers are like, whereas corporate images (rather than tangible products or brands) are very important to service enterprises in service industry where consumers can see service enterprises themselves as well as the sources of service businesses and the way of operation without the shelter of brands and distribution.Berry found that the corporate image was the driving factor of the successful service enterprises through researching 14 high-performance service businesses (including banks); he also suggested that positive images should be shaped by means of increasing customers' experiences of service and put forward the ways of constructing corporate images by boosting the reputation enterprises, establishing emotion relationship with customers, internalizing brands (which is to deepen the idea of brands in the staff) and so on (Berry, L. L., 2000).Mr. Liu Jun found that the images and service quality of Chinese-funded banks and foreign-invested banks equally affected the cognition of service value.Corporate images, which affect Technical Quality and Process Quality, are accumulated consequences after customers' long-term service experiences (Jun, L., 2004).For this reason, corporate images and quality cognition affect each other.
Scholars construct the measure of service quality largely from the above-mentioned three dimensions (that is Technical Quality, Process Quality and Corporate images).For instance, in the 10 variables of SERVQUAL by Parasuraman, Competence (the staff's knowledge and technique as well as their ability of organization, research and development) and Tangibility (service entities) are relevant to Technical Quality; Credibility(the reliability of service enterprises and the consideration of customers) is associated with Corporate Images; Process Quality relates to Professionalism (the efficiency of the staff's completing service), Responsiveness(the willingness of the staff's providing service), Accessibility (the accessible and convenient contact), Courtesy (politeness, respect, consideration and kindness), Communication (the effective expression and listening), Security (zero risk and privacy) and Empathy (the ability of understanding customers) (Parasuraman, A., Zeithaml, V. A., & Berry, L. L., 1985).The excellent quality of service can bring customer satisfaction and customer loyalty.Economists like Fornell classify customer satisfaction into overall satisfaction, the satisfaction resulting from the comparison between experience and expected quality as well as the satisfaction resulting from the comparison between experience and ideal quality.Therefore, customer satisfaction results from customers' psychological comparison whose benchmark is expectation.The formation of expectation originates from the previous experience of the application and from the judgment of competitors and ideal status.Customer intentions or behavior loyalty can be measured by repeated purchase.Economists like Gremler think customer satisfaction has a close relationship with the four kinds of customer loyalty (the loyalty of cognition, the loyalty of emotions, the loyalty of intentions and the loyalty of behavior) (Gremler, D. D, Brown, S.W., Bitner, M. J., & Parasuraman, A., 2001).In order to make a quantitative measurement of the bank's service levels, Mr. Xu Jun and his colleagues made an evaluating from the perspective of factor analysis.They conducted a survey of service levels targeting CMBC and CCB as research objects and preliminarily designed the comparatively overall evaluating system of commercial banks including 8 primary indicators, which are Tangibility, Sensibility, Trust, Reactivity, Assurance, Professionalism, Humanity and Accessibility, and 30 secondary indicators and so on (Jun, X., & Jiong, Y., 2008).
Constructing the Evaluating Indicator System
Constructing the comprehensive evaluation indicator system is the vital foundation of the fuzzy comprehensive evaluation meanwhile the choice of the evaluating indicator directly affects the evaluating conclusion.In this paper, the author refers to the above-mentioned relevant literature, employs the 10 variables of SERVQUAL by Parasuraman, reconstructs Mr. Xu Jun's theory and establishes the evaluating indicator system, which includes three layers, that is to say, the target layer, the main criteria layer and the sub-criteria layer.The target layer is "the comprehensive evaluation of listed banks' service quality"; the main criteria layer is the primary evaluating indicator including technical quality, corporate images and process quality; the sub-criteria layer is the secondary and tertiary evaluating indicator, which is the detail of the main criteria layer.It is shown in Table 1.
Determining the Weight of Each Indicator by Fuzzy Analytic Hierarchy Process (FAHP) Method
Analytic Hierarchy Process (AHP) is a system analysis method of combining qualitative analysis with quantitative analysis, which was put forward by Professor A.L.Saaty in 1970s, who was an American operational researcher of the University of Pittsburgh.The key process of AHP is to establish judgment matrix, whether the judgment matrix is reasonable or not, it will directly affect the effect of APH.This method can effectively analyze non-sequential relationship between the level of objective criteria system and comprehensive measurement by the decision maker's judgment and comparison.It is widely used in the fields of society, economy and management etc. because the system is simple and practical (Jibin, L., Yang, X., Liangan, H., & Jiazhong, L., 2006).However, it has the following disadvantages (Jijun, Z., 2000): Firstly, it is difficult to test and judge the consistency of judgment matrix; secondly, there is a significant difference between the consistency of judgment matrix and that of human beings; thirdly, the criteria of the consistency of judgment matrix: 0.1 CR lacks the scientific evidence; fourthly, when the judgment matrix is not consistent, its adjustment will be very difficult.Nevertheless, FAHP can overcome the above-mentioned disadvantages and it is simpler and more scientific than the traditional AHP.
FAHP firstly constructs a fuzzy consistent matrix by pairwise comparison of the elements of the same layer.In this paper, if Table 2 is used to scale and is true, the constructed judgment matrix ( ) ij R r will be fuzzy consistent matrix, that is to say, it is unnecessary to test the consistency of matrix.Then according to the character of fuzzy consistent matrix, the weight of the element in each layer i w can be measured as follows. 1
Evaluating the Object by Fuzzy Comprehensive Evaluation Method
The advantage of AHP is that it enables the evaluating indicators which are difficult to quantify by other means to be quantitatively analyzed by pairwise comparison, under the condition of the complex structure of judgment targets and the shortage of necessary data.Then it can take the complex evaluating indicators into a clear and easy hierarchical structure which will effectively determine the relative importance of various factors in the evaluating of multi-factor, and to make a further evaluating.However, the disadvange of APH is the shortage of a unitive and specific indicator quantifying method in the process of judging the targets as a whole.Therefore, in practice, the fuzzy analytic hierarchy process and the fuzzy comprehensive evaluation should be combined together to evaluate the bidding units.Namely, first of all, use the fuzzy analytic hierarchy process to calculate the various indicator weights, and then use the comprehensive evaluation in fuzzy mathematics to make a comprehensive evaluation.
The fuzzy comprehensive evaluation is to make a comprehensive evaluation of something using the tool of fuzzy mathematics during taking the effect of various factors into consideration (Tao, Z., 2006).We assume that is the set of m elements of describing the evaluated object, and is the set of n judgments of describing the status of each element.
The evaluating indicator to determine the target are decided by n factors, and can be uniquely defined by U , ,the influence degree of each i u is difference for determining the level of critical objects.i.e., their weights are difference, Which distribute the weight is one fuzzy subset of U , and can be uniquely defined by W , ,where i w is the weight of i u , 0 . By evaluating the each single fuzzy factor, the fuzzy comprehensive evaluation matrix can be obtained as follows.
Where 1 2 ( , , , ) is the ith single factor evaluating of i u ,so ij r indicates that the frequency distribution of ith factor (1 ,and generally make it normalization to satisfy with The results of comprehensive evaluation can be obtained by the complex calculating as follows.
Where j G indicates that the level of evaluated subject for the set of judgments.i.e., the membership degree of j v for fuzzy sets G .Then starting from the principle of maximum membership degree, in Takes its maximum value as the level of evaluating object, and also may according to the formula of fuzzy vector or the principle of weighted averages, each grade were assigned to a certain score, and normalization (Yonghong, H., & Sipin, H., 2001).
Determining the Weight of Each Indicator by FAHP
According to the hierarchy structure model in Table 1, make a level's indicator which is relative to previous level's indicator to pairwise comprise by the importance degrees, obtain the fuzzy consistent judgment matrix, the importance degrees of indicator pairwise comprise to determine by the 0.1~0.9Quantity Scale in Table 2.
Then make a level's indicator that relative to previous level's indicator to sort by the importance degrees, obtain the relative weights among all indicators as follows(table 3~table 16).
In a similar way, the other of one-stage fuzzy comprehensive evaluation about C hierarchy as follows.12 (0.17,0.40,0.20,0.
(0.217,0.317,0.467 ) As above which indicated that 16% of the experts determined the bank's QOS are best; 40% are better, 22% are good, 15% are bad, the rest of 6% are worse.According to the maximum membership principle, the bank's QOS should be concluded as better.
Conclusion
From the analysis and examples of this paper, it is convenient and suitable for banks to apply the AHP and the fuzzy comprehensive evaluation to the evaluation of banks' QOS, which can help banks to make the evaluating more reasonable and largely reduce the effect of the judges' personal subjective factors.Thus the judges can make a more objective evaluating, which is of great significance to promote the fairness, the openness and the impartiality of evaluating.
Banks'QOS is the customers' value judgment.Through analyzing the indices and the factors of affecting banks' QOS, the paper indicates how banks improve their quality of service.Apart from the evaluating indices of banks' QOS, banks should systematically improve their QOS from the bank's product and service through constant innovation, so that banks' service can really meet customers' expectation and constantly enhance the attraction of customers in the increasingly fierce market competition.
can obtain the result of fuzzy comprehensive evaluation by complex calculating as follow,
Table 1 .
The comprehensive evaluation indicator system of listed banks' service quality C )1 Service staffs pay special attention to customers in the principle of customer orientation.(371P )ank's staffs enable to inform customers of preventative measures and the handling of some similar issues.(372P)
Table 3 .
The fuzzy consistent judgment matrix and weight of first evaluating indicator A
Table 4 .
The fuzzy consistent judgment matrix and weight of second evaluating indicator 1
Table 5 .
The fuzzy consistent judgment matrix and weight of second evaluating indicator 2
Table 6 .
The fuzzy consistent judgment matrix and weight of second evaluating indicator 3
Table 7 .
The fuzzy consistent judgment matrix and weight of third evaluating indicator11
Table 8 .
The fuzzy consistent judgment matrix and weight of third evaluating indicator12
Table 10 .
The fuzzy consistent judgment matrix and weight of third evaluating indicator
Table 12 .
The fuzzy consistent judgment matrix and weight of third evaluating indicator
Table 13 .
The fuzzy consistent judgment matrix and weight of third evaluating indicator
Table 16 .
The fuzzy consistent judgment matrix and weight of third evaluating indicator Using Fuzzy Comprehensive Evaluation Method to Evaluate the Listed Banks' Service Quality4.2.1 Determing the Estimation ScalesThe evaluation scales were decided by the evaluation committee that includes 10 experts of technical and economic aspects, and given the set of judgments V = {best, better, good, bad, worse}.4.2.2 Determing the Relation Matrix between the Criteria Layer of C and VAs we have calculated the relative weights of A , B ,C hierarchies by FAHP method before, so we can obtain the One-stage fuzzy comprehensive evaluation about 11 C as follows.
,we can calculate the Comprehensive Evaluation of the criteria layer of B as follows. | 3,766 | 2013-06-25T00:00:00.000 | [
"Business",
"Economics"
] |
Diet of Lontra longicaudis ( Carnivora : Mustelidae ) in a pool system in Atlantic Forest of Minas Gerais State , southeastern Brazil
The present study aimed to investigate the feeding habits of Lontra longicaudis in a pool system within the Private Reserve of Natural Patrimony (RPPN) Usina Maurício, located in Paraíba do Sul river basin, Atlantic Forest of southeastern Minas Gerais State. The diet composition was determined based on the identification of items present in 212 scats sampled between July 2008 and October 2009 in a 4.1 km stretch of the pool system. The found items and its respective percentages of occurrence were: mollusks (0.5%), insects (16.5%), spiders (1.4%), crustaceans (3.3%), fish (96.7%), amphibians (0.9%), snakes (3.8%), birds (2.8%), mammals (8.5%) and fruits (0.5%). Among fish, the identified families and respective percentages of occurrence were: Loricariidae (65.4%), Pimelodidae (42.9%) Cichlidae (22%), Characidae (7.3%), Erythrinidae (3.9%), Synbranchidae (2.4%), Anostomidae (2%). Therefore fish make up the most consumed item in the study area, with the predominance of benthic siluriformes (families Loricariidae and Pimelodidae).
Introduction
The Neotropical river otter Lontra longicaudis (Olfers, 1818) is a semi-aquatic mustelid, inhabitant of continental water bodies and even marine systems.It is distributed from northeastern Mexico to Uruguay to Buenos Aires Province, Argentina (CHEIDA et al., 2006;EMMONS;FEER, 1997).
Acta Scientiarum.Biological Sciences Maringá, v. 34, n. 4, p. 407-412, Oct.-Dec., 2012 In the State of Minas Gerais, southeastern Brazil, information on diet of L. longicaudis is scarce.Passamani and Camargo (1995) analyzed eight scats in Furnas reservoir, in the western Minas Gerais.Moreover, there no available data on feeding habits of L. longicaudis in the State.In face of scarcity about information on diet composition of the species in this region, the present study presents data on L. longicaudis feeding habits in Private Reserve of Natural Patrimony (RPPN) Usina Maurício, Atlantic Forest of southeastern Minas Gerais State.
The RPPN Usina Maurício is inserted in Atlantic Forest biome (IBGE, 2004).The climate in the region is classified as Cwa (tropical humid) of Köppen.The study was conducted in 4.1 km of the stretch with reduced flow of the Usina Maurício hydroelectric power plant, composed by a system of 19 pools with varied lengths, widths and depths.The three largest pools present the following measures (length x width): 150 x 52 m, 86 x 37 m, 63 x 28 m (Ricardo A. Silva, pers.com.).The pools remain isolated most of the year, connecting during the raining season (November to February).Pool bottom and banks are predominantly rocky, and vegetation in the area is composed by Stational Semidecidual Forest.
Data sampling
The diet composition of L. longicaudis was determined based on identification of scats.From July 2008 to October 2009 biweekly samplings were performed, except for the period from the second half of December and early March (summer season), when heavy floods did not allow access to the study area.All the scats found were collected, stored in individual labeled plastic bags and then frozen.Afterwards, scats were washed through a 1mm sieve and solid remains were dried in an oven at 36°C.Screening was realized with the aid of a stereoscopic microscope.The consumed taxa were determined based on the identification of remains such as scales, vertebrae, hairs, carapaces and other structures, which were compared with specimens stored in a reference collection of the study area.The frequency of occurrence of each prey item was calculated by the rate of the number of scats containing such item over the total number of analyzed scats.
Discussion
Benthic siluriformes (Loricariidae and Pimelodidae species) represented the major prey of L. longicaudis in the pool system investigated in the present study.Loricariidae was also the most recorded fish prey in the works of Pardini (1998) and Kasper et al. (2004Kasper et al. ( , 2008)).Meanwhile, estimative of fish availability made by Pardini (1998) indicated a high numerical predominance of Characidae species in the system investigated by the author, suggesting selectivity on fish predation.Kasper et al. ( 2004) also found a high preference for Pimelodidae family.These authors mentioned a preference of L. longicaudis for benthic and sedentary fishes, whose habits facilitate their capture.In RPPN Usina Maurício Hypostomus affinis was the most captured species in the sampled pools.The utilized gillnets present a wide range of sampling, capturing species at surface, midwater and bottom, which indicate that H. affinis may be a numerically well representative larger-sized species in the study area.In this context, a higher abundance associated to a benthic sedentary habit reflected in the predominance of the loricariid H. affinis in diet of L. longicaudis in the studied area.
Other recorded fish families (Gymnotidae, Anostomidae, Erythrinidae, Synbranchidae, Characidae and Cichlidae) occurred with relatively low frequencies.Except for the characids and cichlids, it was possible to identify the other fish remains until species level.Scales of S. brasiliensis differ in size and thickness from the scales of other characid species sampled in the pools (Astyanax bimaculatus and Myleus micans).In this case, the consumption of young individuals of S. brasiliensis could be underestimated considering the morphological similarities with the other smaller characid species.Among cichlids, Geophagus brasiliensis, Cichla monoculus, Crenicichla lacustris and Oreochromis niloticus were sampled in the pools.The exotic O. niloticus has cycloid scales, which allowed a differentiation from G. brasiliensis, C. lacustris and C. monoculus, characterized by having ctenoid scales (pers.obs.).Hoplias malabaricus, Leporinus mormyrops, Gymnotus carapo and Synbranchus marmoratus were also the only species of their respective families sampled in the pool system.However, recorded 67 fish species in Pomba river, including 17 characids, eight loricariids, six cichlids and the erythrinid Hoplias lacerdae.Thereby, possibly, other fish species not captured in our samplings could occur in the pool system, and the dietary spectrum of L. longicaudis in the area could be wider than determined in the present study.
Among invertebrates, insects were a well representative group in the analyzed samples, present in 16.5% of total scats (n = 35).The occurrence of insects seems to vary among investigations about L. longicaudis diet, present in higher (e.g.PASSAMANI; CAMARGO, 1995;PARDINI, 1998) or lower frequencies (e.g.HELDER-JOSÉ; ANDRADE, 1997; QUADROS; MONTEIRO-FILHO, 2001; ALARCON; SIMÕES-LOPES, 2004;KASPER et al., 2004KASPER et al., , 2008;;QUINTELA et al., 2008).Furthermore, insects, spiders and other arthropod remains found in L. longicaudis scats can also represent digestive tract content of predated fish, as pointed out by Quadros and Monteiro-Filho (2001) for terrestrial arthropods.In the present study, however, arthropods may be an important component in the otter diet once insects not associated to fish remains were found in six samples, one of those with only megalopteran Corydalus sp.larvae.Megalopteran larvae were also recorded by Pardini (1998) and Kasper et al. (2004Kasper et al. ( , 2008) ) while belostomatids were registered by Colares and Waldemarin (2000), Kasper et al. (2004) andQuintela et al. (2008).Coleopterans were found with low frequency by Quintela et al. (2008) while orthopterans are herein recorded by the first time in L. longicaudis diet.Orthopterans (Gryllotalpa gryllotalpa, Gryllotalpidae) were also verified in the diet of the Eurasian otter Lutra lutra in southeastern Bulgaria (GEORGIEV, 2006).
Mammals represented the third item regarding frequency (18%), a higher value than those reported in earlier studies (e.g.PARDINI, 1998;QUADROS;MONTEIRO-FILHO, 2001;KASPER et al., 2004KASPER et al., , 2008;;QUINTELA et al., 2008).Other terrestrial vertebrates (amphibians, reptiles and birds) were less representative, which is in agreement with all abovementioned studies.Quadros and Monteiro-Filho (2001) also emphasize the difficulties to recognize amphibian' remains in scats due to the lack of keratinized structures.In this context, amphibian occurrence may have been underestimated, considering that samples contained very fragmented bones, which does not allow its identification.Also Weber (1990) found a relationship between the presence of amphibian remains in Lutra lutra scats and the availability of this item in environment.Anuran amphibians are abundant in the studied pool system (especially the cycloramphid Thoropa miliaris) and the utilization of this food resource could be higher than determined.
Unidentified fruits occurred in one sample (0.47%).The consumption of vegetal items is rarely reported in studies on Neotropical otter diet.Quadros and Monteiro-Filho (2000, 2001) found fruit remains of three arboreal species in 2.97% of analyzed samples while Kasper et al. (2008) identified grass remains in a mean frequency of 0.7%.Quadros and Monteiro-Filho (2000) consider that fruits consumption is opportunistic and complementary in L. longicaudis diet, considering the high availability of this food item in the environment and low frequency in fecal samples.
Interestingly the fruits were found in one single sample containing remains of unidentified bird, and therefore can represent content of the digestive tract from the predated bird.
We observed a varied diet of L. longicaudis in the studied area, with a prominent dominance of benthic fish.Our data, therefore, contributes to the knowledge on Neotropical otter feeding habits in pools systems of the still poorly investigated Minas Gerais Atlantic Forest, highlighting the importance of RPPN Usina Maurício for the species conservation.
Conclusion
The neotropical otter presented a varied diet in the study area.Fish, however, represented the most consumed item, as observed in previous investigations on the species diet.Among fishes, the predominance of Loricariidae was already determined in L. longicaudis trophic ecology studies conducted in Southeastern and Southern Brazil.Molluscs, insects, spiders, crustaceans, amphibians, reptiles, birds and mammals can also be considered minor preys in the studied system.
Figure 1 .
Figure 1.Location of Minas Gerais State (A) and Private Reserve of Nature Patrimony Usina Maurício, Usina Maurício (B and C).
Table 1 .
Food items identified in 212 scat samples of Lontra longicaudis in Private Reserve of Nature Patrimony Usina Maurício, Minas Gerais State, Brazil.N = number of scats containing each item, % = percentage of occurrence in total scats.Percentage of occurrence of food items in 212 scat samples of Lontra longicaudis in the Private Reserve of Nature Patrimony Usina Maurício, Minas Gerais State, Brazil. | 2,237.2 | 2012-03-23T00:00:00.000 | [
"Environmental Science",
"Biology"
] |
Modelling the impact of raising tobacco taxes on public health and finance
Abstract Objective To investigate the potential for tobacco tax to contribute to the 2030 agenda for sustainable development by reducing tobacco use, saving lives and generating tax revenues. Methods A model of the global cigarette market in 2014 – developed using data for 181 countries – was used to quantify the impact of raising cigarette excise in each country by one international dollar (I$) per 20-cigarette pack. All currencies were converted into I$ using purchasing power parity exchange rates. The results were summarized by income group and region. Findings According to our model, the tax increase would lead the mean retail price of cigarettes to increase by 42% – from 3.20 to 4.55 I$ per 20-cigarette pack. The prevalence of daily smoking would fall by 9% – from 14.1% to 12.9% of adults – resulting in 66 million fewer smokers and 15 million fewer smoking-attributable deaths among the adults who were alive in 2014. Cigarette excise revenue would increase by 47% – from 402 billion to 593 billion I$ – giving an extra 190 billion I$s in revenue. This, in turn, could help create the fiscal space required to finance development priorities. For example, if the extra revenue was allocated to health budgets, public expenditure on health could increase by 4% globally. Conclusion Tobacco taxation can prevent millions of smoking-attributable deaths throughout the world and contribute to achieving the sustainable development goals. There is also potential for tobacco taxation to create the fiscal space needed to finance development, particularly in low- and middle-income countries.
Introduction
Although the use of a tobacco tax to reduce smoking is still relatively new in many countries, there is a long history throughout the world of governments implementing such a tax to generate revenue. In many forums, tobacco taxation has also been highlighted as a means of mobilizing domestic resources to finance health and other development programmes. [1][2][3] The recent setting of the Addis Ababa Action Agenda 4 and the 2030 Agenda for Sustainable Development 5 have further heightened interest in tobacco taxation.
In July 2015 the United Nations General Assembly endorsed the Addis Ababa Action Agenda. In this agenda, which was an outcome of the Third International Conference on Financing for Development, the United Nations recognized that "price and tax measures on tobacco can be an effective and important means to reduce tobacco consumption and health-care costs, and represent a revenue stream for financing for development in many countries. " 4 Subsequently, in September 2015, the 2030 Agenda for Sustainable Development was also adopted in a United Nations General Assembly. 5 This agenda includes 17 sustainable development goals (SDGs) that all Member States have agreed to achieve by 2030. SDG 3, which is to "ensure healthy lives and promote well-being for all ages", includes target 3.4 -to reduce premature mortality from noncommunicable diseases by one third -and target 3.a -to strengthen country-level implementation of the World Health Organization's (WHO's) Framework Convention on Tobacco Control (FCTC). 5,6 The FCTC is an international treaty with 180 Parties who have committed to protecting public health through the implementation of comprehensive measures of tobacco control. Article 6 of the FCTC recognizes price and tax measures as effective means to reduce the demand for tobacco, and the guidelines for Article 6's implementation encourage the use of taxation in comprehensive strategies for tobacco control. 7,8 It seems likely that tobacco taxation will be an important evidence-based intervention to help many countries achieve their development objectives. As tobacco tax rates in many low-and middle-income countries are currently low and demand for tobacco products is relatively inelastic, many countries could increase government revenues substantially through tobacco taxation. 9 By creating the fiscal space to finance development programmes while, at the same time, reducing tobacco use, tobacco taxation could be a win-win policy for governments.
In the first year that the so-called sin-tax reforms were implemented in the Philippines, the tax on low-priced brands of cigarettes was increased by 341% and this led to a 114% increase in annual excise revenue. 10 Under the reforms, 85% of the extra revenue is being used to subsidize universal health care for 14 million families and upgrade medical facilities. There are at least 30 other countries that dedicate a certain amount of their tobacco taxes to health. 10 Although such dedicated allocations are not always feasible, the reforms in the Philippines have shown that substantial increases in tobacco taxation can lead to improvements in public health finance.
Retrospective studies have shown the importance of tobacco taxation in public health outcomes. For example, in the United States of America, it has been observed that a 10% increase in cigarette taxes could decrease the number of deaths from respiratory cancers by 1.5%. 11 The French Government increased cigarette taxes substantially from the mid-1990s, with cigarette prices tripling in real terms by 2005. Among French males, rates of death from lung cancer fell by 50% during the same period. 12,13 Research Tobacco taxes and health Mark Goodchild et al.
We wished to demonstrate the potential for tobacco taxation to reduce tobacco use, save lives and generate tax revenues globally. We therefore developed a model of the global cigarette market using data for 181 countries that together represented 98% of the world's smokers. We used the model to quantify the impacts of the increase in excise on the retail price of cigarettes, cigarette excise revenue, cigarette consumption, the number of daily cigarette smokers, and the future number of smokingattributable deaths averted among the world's adult population in 2014.
Data sources
Data on taxes and prices per 20-cigarette pack of the most popular brand of cigarette in each study country in 2014 were sourced from WHO's Report on the global tobacco epidemic. 10 In this data set, the amount of excise and other taxes on cigarettes is calculated on the basis of each country's actual tax system. Excise is a tax imposed on selected commodities such as cigarettes and is the main fiscal instrument that governments use to generate extra tax revenue from those commodities.
The quantity of cigarettes sold in each country was calculated using data from two market survey companies -Canadean 14 and Euromonitor International 15 -and from WHO's work with Member States. 16 The numbers of daily cigarette smokers were calculated using the United Nations Population Division's country-specific estimates of the adult population in 2014 and WHO's estimates of the prevalence of daily cigarette smoking among adults. 17,18 Online databases of the International Monetary Fund, the World Bank and WHO were used to source macroeconomic data on inflation, government health expenditure and purchasing power parity exchange rates. 19,20 Our findings are reported in international dollars (I$) to provide an accurate comparison of cigarette prices between countries -after taking into account differences in the purchasing power of countries at different levels of income and development.
Taxes and prices
Countries apply different kinds of excise systems to cigarettes. Excise may be a fixed amount per pack or a percentage of the pack's value or a combination of the two. For each included country, the database of WHO's Report on the global tobacco epidemic reports the amount of taxes on the most popular brand of cigarette. We took these amounts as the baseline levels of tobacco tax and then simulated the effects of increasing excise, by I$ 1.00 per 20-cigarette tax, over the next year. This level of intervention was chosen because it reduces the affordability of cigarettes in all Member States, particularly in low-and middleincome countries where cigarette taxes and prices are relatively low. 9, 10 We allowed general consumption taxes -e.g. value-added tax or sales tax -to rise as normal, on the basis of the retail or wholesale prices of the cigarettes. Although we fixed the per-pack values of other kinds of taxes -e.g. import duties or surcharges -this simplification would have had little impact on our main findings since excise and value-added tax are the most important taxes on cigarettes in most countries.
The retail price that consumers pay includes all applicable taxes plus the producer or industry price net of taxes. The retail price of a pack of cigarettes, P R , can be calculated as: where P p is the producer price net of taxes, T E is the excise amount per pack, T VAT is the amount of value-added tax per pack and T O is the other taxes -e.g. import duties.
For the model, we used the standard assumption of full pass through of taxes onto the retail price of cigarettes. 21 In addition, the producer price net of tax was assumed to increase in line with the global inflation rate -reflecting, for example, the maintenance of industry cost and profit margins in real terms. The new retail price that we modelled, P R *, was calculated as: where P p * is the new industry price per pack after adjusting for inflation, T E * is the new excise amount per pack -i.e. T E plus I$ 1.00 -and T VAT * is the new VAT amount per pack. Total excise revenue was calculated as the excise per pack multiplied by the quantity of cigarette packs sold in the retail market (S).
Consumption and use
The extent to which higher cigarette prices reduce consumption is governed by the price elasticity of the demand. For example, a price elasticity of −0.3 means that a 10% increase in cigarette prices will reduce cigarette consumption by 3%. Studies in high-income countries have revealed price elasticities that range from −0.25 to −0.5 while studies in low-and middle-income countries Tobacco taxes and health Mark Goodchild et al.
have revealed corresponding elasticities between −0.2 and −0.8. 22 It appears that cigarette consumers in low-and middleincome countries are generally more price-sensitive that their counterparts in high-income countries. In this study, the price elasticities of cigarettes in high-, middle-and low-income countries were assumed to be −0.3, −0.4 and −0.5, respectively. The number of packs sold in response to the price increase (S*) was calculated as: where ΔP is the percentage change in the retail price and ε p is the price elasticity of demand. As the elasticities are short-term parameters, we assumed that the full impact of the price increase on consumption would occur within one to three years. Table 1 shows the key assumptions used in the modelling. The price elasticity of demand reflects a combination of conditional demand -i.e. the amount or intensity of smoking -and smoking prevalence. 23 Global evidence suggests that, for cigarettes, half of the impact of higher prices comes from a reduction in smoking prevalence. [22][23][24] Consequently, for our model, we assumed that the prevalence elasticity was half of the price elasticity -i.e. −0.15, −0.2 and −0.25 in high-, middle-and low-income countries, respectively. We used these prevalence elasticities to estimate the reduction in the number of smokers in the current adult population that would result from our modelled increase in excise. The prevalence elasticities we used are the same as those previously used to assess the global impact of tobacco control policies. 25
Public health outcomes
We used a single cohort approach 23,25 to measure the impact of tobacco taxation on the expected number of smokingattributable deaths among the world's adults who were alive in 2014. In this approach the impact of tobacco control policies was measured first in terms of the reduction in the number of smokers among the current adult population and then in terms of the future health outcomes for the same population cohort over the course of their remaining lives. We defined anyone older than 15 years as an adult and we used a medium to long-term time horizon to cover the
Research
Tobacco taxes and health Mark Goodchild et al.
remaining lives of the current cohort of adult smokers. For each study country, we estimated the baseline number of adult daily cigarette smokers from the size of the adult population and the prevalence of daily cigarette smokers among the adults.
Epidemiological studies over the past 50 years have shown that tobacco ultimately kills a third to half of all people who use it. 26,27 By applying a relatively low risk of a smoking-attributable death -of 33% -to the adult daily smokers in our model, we aimed to produce a conservative estimate of the number of smoking-attributable deaths that could be averted by the tax intervention.
We estimated the positive impact of tobacco taxation on health as the expected decrease in the number of smoking-attributable deaths -after accounting for those current smokers who will cease smoking before they die. The benefits of quitting are many and occur early for several serious diseases. 28 Overall, adults who cease smoking before they reach middle age avoid almost all the excess hazards of smoking. 26 Nonetheless, some adjustment is required to account for the fact that not all smokers who quit can avoid early death. National studies typically use a mortality adjustment factor of 70% for smokers who quit. 29,30 A global mortality adjustment factor has been calculated on the assumption that 95%, 75%, 70%, 50% and 10% of those who cease smoking when aged 15 to 29, 30 to 39, 40 to 49, 50 to 59 and at least 60 years, respectively, will avoid an early death. 23 We applied the same percentages to the age profile of the world's population in 2014, leading to a mean adjustment factor of 67%. Thus, we assumed that 67% of the adult daily smokers in 2014 who would otherwise have suffered an early death from a disease caused by smoking would avoid such a death if they ceased smoking. We estimated the number of smoking-attributable deaths averted as a result of the tax increase as 67% of 33% of the reduction in the number of daily adult smokers resulting from the increase in cigarette prices.
baseline
In 2014, the mean amount of excise was estimated to be I$ 1.37 per 20-cigarette pack. This represented 43% of the mean retail price of I$ 3.20 per pack ( Table 2). As the total annual cigarette consumption was calculated to be 294 billion packs, the total excise revenue generated globally from the sale of cigarettes was estimated to be I$ 402 billion -or about 328 billion United States dollars (US$). Although we estimated that there were 740 million adults who were daily cigarette smokers worldwide in 2014, almost 320 million (43%) of these smokers lived in just four middle-income countries: Brazil, China, India and the Russian Federation. These numbers exclude smokers of other forms of tobacco -e.g. bidi smokers in south-east Asia. We estimated that, under the baseline scenario, at least 247 million daily smokers from among the adult population in 2014 will ultimately die from a smokingattributable disease.
Tax simulation
Raising excise by I$ 1.00 per 20-cigarette pack in all countries would generate a substantial increase in cigarette tax yields in all countries. Excise per pack would increase by 80% globally ( Table 2). Tax yields would increase the most in the Eastern Mediterraneanpartly because many countries in this region did not levy any cigarette excise in 2014 (Table 3). The mean retail price of cigarettes would increase by 42% globally. Cigarette prices would increase by a mean of 63% in low-income countries but only by a mean of 25% in high-income countries ( Table 2).
Global cigarette consumption would decrease by 18% -representing 53 billion fewer cigarette packs compared with 2014 (Table 2). Cigarette consumption would decline most in the Western Pacific -reflecting this region's large consumption base. The amount of cigarette excise revenue generated throughout the world would increase by I$ 190 billion -or about US$ 141 billion. All income groups and regions would see substantial growth in excise revenues. The African continent would expand excise revenue from cigarettes by as much as 85% ( Table 3).
The extra excise revenue from cigarettes would help create the fiscal space needed by countries to meet their development priorities. For example, if all of the extra revenue from raising cigarette excise was allocated to government health budgets, then public expenditure on health could increase by 4% globally (Fig. 1). A third of all low-and middle-income countries would be able to increase public health expenditure by more than 10% in this manner.
In terms of health outcomes, the prevalence of daily cigarette smoking among adults would decline by 9% in relative terms -i.e. from 14.1% to 12.9% of the adult population (Fig. 2). This decrease translates into 66 million fewer smokers. The expected number of smoking-attributable deaths from among the world's adult population in 2014 would decrease by 15 million -reflecting a decline of about 6% in smoking-related mortality among this cohort ( Table 2). The majority of the smoking-attributable deaths averted would be in low-and middle-income countries.
Discussion
WHO has been working with its Member States to implement the FCTC. For example, it has been collaborating with ministries of finance to help them adopt better policies on tobacco taxation. Detailed, country-level tax modelssimilar to the one described here -have helped to frame discussions on the policy objectives of tobacco taxation. A frequent and important precondition, from the perspective of public finance, is the need for reforms to generate higher tax revenues sustainably -at least over the short to medium term. Among the other concerns of government officials that are being addressed in these collaborations is the threat of illicit trade. In this present study, we do not address illicit trade directly. However, betweencountry differences in cigarette taxes and prices would be narrowed -not widened -by the tax increase that we modelled. This might be expected to reduce the incentive for illicit trade.
In reality, illicit trade occurs in low-tax jurisdictions as well as high-tax ones and there is no direct correlation between rates of tobacco taxation and tobacco smuggling. 31 Factors other than taxes and prices serve to motivate or enable illicit trade. The administrative capacity of many tax and customs departments needs to be strengthened. 22,31 Given the transnational nature of the illicit trade in cigarettes, it is clear that a coordinated international response is needed. In November 2012, the Conference of Parties to the FCTC adopted the Protocol to Eliminate Illicit Trade in Tobacco Products. 32 While negotiating this protocol, Member States have agreed to a set of control measures that should help address the critical administration and transnational issues.
The single cohort approach that we used in this study fails to incorporate the dynamic aspects of changing demographics and smoking prevalence.
These considerations require the use of a structural model such as that used in the global burden of disease projections. 33,34 However, structural models are relatively sophisticated and data-intensive and beyond the intended scale of our study. In addition, when another study compared their single cohort analysis with the results of using dynamic models in nine countries, they found that the dynamic aspects of policy change did not substantially change their main findings. 25 Another limitation of the present study is that we applied the same mortality risk in all countries. The relative risks have been found to be lower in low-and middle-income countries than in high-income ones 25 -possibly because of population differences in age at initiation of smoking, smoking intensity and/or the background risk from other causes of death. Structural models have included smoking impact factors that indirectly measure the accumulated risk. Some studies report sensitivity analyses based on a range of 33% to 50% mortality risk. 25,30 In this study we applied conservative assumptions to ensure that the results were also conservative. Therefore, our estimate of the number of smoking-attributable deaths averted could well be an underestimate -especially in Europe and North America where the tobacco epidemic is currently strongest.
Conclusion
Tobacco taxation can prevent millions of smoking-attributable deaths throughout the world and contribute to the achievement of global health objectives, such as SDG target 3.4. There is substantial potential for tobacco taxation to create the fiscal space needed to finance development, particularly in low-and middle-income countries. ■ | 4,674.6 | 2016-02-12T00:00:00.000 | [
"Economics",
"Medicine",
"Political Science"
] |
Quality of Life in Patients with High-grade Non–muscle-invasive Bladder Cancer Undergoing Standard Versus Reduced Frequency of Bacillus Calmette-Guérin Instillations: The EAU-RF NIMBUS Trial
Take Home Message This study did not find better quality of life with a reduction in the number of bacillus Calmette-Guérin instillations in patients with high-grade non–muscle-invasive bladder cancer. This result together with the previous finding that a reduced frequency schedule is inferior underlines the use of a standard bacillus Calmette-Guérin instillation schedule.
Introduction
Urothelial bladder cancer (UBC) carries a large global disease burden, being the 11th most common cancer, with approximately 550 000 new cases annually [1].Nearly 75% of all primary UBC patients are diagnosed with nonmuscle-invasive bladder cancer (NMIBC).Patients with high-grade NMIBC have increased risks of recurrence, progression, and metastases [2].Intravesical bacillus Calmette-Guérin (BCG) instillations following a transurethral resection of the bladder tumor (TURBT) are the standard of care to reduce these risks.The European Association of Urology (EAU) guidelines recommend a weekly instillation for 6 wk as an induction phase, followed by a maintenance phase of 1 yr (three times 3 weekly instillations at 3, 6, and 12 mo) after TURBT for intermediate-risk and up to 3 yr for high-risk patients [3][4][5].Adverse events, however, are significant during the long-term administration of BCG, often leading to treatment discontinuation [6,7].
The European Organization for Research and Treatment of Cancer (EORTC) trial (EORTC 30962) concluded that BCG dose reduction did not affect toxicity level and led to higher recurrence rates [8].The European EAU-RF NIMBUS trial evaluated whether a reduced instillation frequency during both the induction and the maintenance phase is noninferior to EAU guideline standard of care [9].Unfortunately, safety analyses showed the reduced approach to be inferior to the standard approach for the risk of recurrence, leading to early cessation of patient recruitment to avoid further harm in the reduced BCG frequency arm.The current post hoc analysis of the EAU-RF NIMBUS trial evaluated whether patients with reduced BCG instillation frequency in both the induction and the maintenance phase experienced lower toxicity and consequently better quality of life (QoL) than patients receiving the standard BCG instillation frequency.
Patients and methods
The EAU-RF NIMBUS trial was a European randomized controlled trial that assessed whether a reduction in the BCG instillation frequency is noninferior to the standard BCG frequency in patients with high-grade NMIBC (Ta-T1) [9].Recruitment took place between December 2013 and July 2019 at 51 study sites spread across Germany, The Netherlands, France, Belgium, and Spain.Patient recruitment was ceased on July 1, 2019, after a data review and safety analysis by the Independent Data Monitoring Committee (IDMC) showed the reduced BCG instillation arm to be inferior to the standard BCG instillation arm with regard to the risk of recurrence.
The trial had been approved by all the relevant institutional review boards and independent ethics committees, and had been performed according to the Declaration of Helsinki [10], Good Clinical Practice, and local regulatory requirements.
Inclusion and exclusion criteria
BCG-naïve patients who had been clinically diagnosed with primary or recurrent high-grade NMIBC (Ta or T1), with single or multiple urothelial papillary bladder carcinoma(s), and with or without concomitant carcinoma in situ (CIS) were eligible.A routine repeated TURBT (re-TUR and/or re-re-TUR) had to be performed to confirm the absence of muscle-invasive cancer.High-grade Ta patients were allowed to be included without a re-TUR in case a biopsy specimen confirmed the complete removal of the tumor and included detrusor muscle tissue.
The exclusion criteria were having had previous systemic or multiinstillation intravesical chemotherapy within the preceding 3 mo, having any type of tumor(s) in the upper urinary tract or prostatic urethra at any time, having any immunodeficiency, and having any other type of malignancy besides basal cell carcinoma of the skin or localized prostate cancer under active surveillance.
Randomization
After enrolment, patients were allocated using a validated randomization program (EAU-RF website) according to the minimization method with a random element as described by Pocock [11].Stratification factors included center, Ta versus T1, concomitant CIS versus no CIS, single versus multiple tumors, and BCG strain (Connaught, Medac, or Tice).The patients were randomized to either one of two treatment groups: 1.The standard frequency (SF) arm.Induction: once a week BCG instillations at weeks 1-6; maintenance: once a week instillations at weeks 1-3 at months 3, 6, and 12 (15 planned instillations).
Follow-up was conducted through cystoscopy and urine cytology every 3 mo during the first 2 yr and every 6 mo thereafter.Histological confirmation had to be provided in case of CIS, or if there was a suspicion of disease recurrence.
Patients' participation in the study was ended in case of a recurrence in the bladder, a urothelial carcinoma in the upper urinary tract or prostatic urethra, or presence of distant metastases, or in case systemic chemotherapy was indicated.
The remaining items evaluate any additional symptoms that are commonly perceived in cancer patients (dyspnea, appetite loss, sleep disturbance, constipation, and diarrhea).Paper questionnaires on QoL were handed out during an outpatient visit at the right time points.The questionnaires were completed prior to the first and the last instillation of each BCG cycle, leading to a total of eight measurement points (T0-T7; see Fig. 1).The endpoint of the NIMBUS trial was time to first recurrence.
Consequently, QoL questionnaires were not filled out anymore if patients experienced a recurrence.
In addition, treating physicians were responsible for carrying out side effect (SE) evaluations by means of a form that included known local and systemic SEs (World Health Organization grading of toxicity: grade 1, mild; grade 2, moderate; grade 3, severe; and grade 4, life-threatening toxicity) prior to the first and the last instillation of each BCG cycle [13].
Endpoints
The primary endpoint for the analysis was QoL.Additionally, toxicity incidence and severity were recorded.
All the five functional and three symptom scales plus the individual symptom items of the questionnaires were transformed to a 0-100 score.A high scale score represents a higher response level.Thus, a high score for a functional scale represents a high/healthy level of functioning, a high score for the global health status/QoL represents high QoL, but a high score for a symptom scale/item represents a high level of symptoms/problems.Differences in the mean QoL between the two treatment arms were evaluated using linear regression at T1, T5, and T7 while adjusting for T0 (baseline measurement, ie, prior to induction week 1).Differences between the trends in QoL of the two treatment arms were tested for significance by performing a linear mixed model using time as the fixed factor with eight levels (T0-T7).Chi-square or Fisher exact tests were used to test for significant differences in the number of SEs between the two treatment arms.
After performing the ITT QoL and SE analyses, supplementary perprotocol (PP) QoL and SE analyses were performed.Patients were excluded from the PP analysis if they had incomplete treatment due to missed instillations, had extra BCG instillation(s), switched treatment arm after the study's premature stop, or stopped treatment for other reasons besides SEs or recurrence.
Results
A total of 359 patients were randomized to one of the two treatment arms.The SF arm contained 182 patients, while the RF arm contained 177 patients.At baseline, there were no significant differences in characteristics between the two treatment arms (Table 1).At the time of study discontinuation, 52% (n = 94) of the patients in the SF arm received all 15 planned instillations.In total, 48 (26%) patients in this arm received nine or fewer instillations.In the RF arm, 45% (n = 79) received all nine planned instillations at the time of study stop.In the SF arm, 24 patients developed a recurrence or new CIS within 1 yr and went off study.In the RF arm, this number was 46 (Fig. 2).In total, 30 and 55 patients in the SF and RF arms, respectively, developed a recurrence.
QoL analyses
The QLQ-C30 questionnaires were completed by 304 (84.7%) patients at T1, 226 (63.0%) patients at T5, and 168 (47.2%) patients at T7. Detailed results of the questionnaires can be found in Table 2.A summary of the results is depicted in Figure 3. Aside from the physical functioning at T5 (p = 0.05), we found no differences in the means of any QoL scale between the two treatment arms (p > 0.05; Table 2).Moreover, the linear mixed model, which was adjusted for T0, did not show any statistically significant temporal changes in any QoL domain for both the SF and the RF arm (91% in the SF arm and 94% in the RF arm completed the QoL assessment at T0).
Toxicity
SE evaluations were completed in 57.7% of patients at T1, 44.5% of patients at T5, and 34.6% of patients at T7 (Table 3).
For patients for whom an SE form was not filled out, we conducted an enquiry among the participating urologists.Thirty-three of 51 sites responded; 26 out of the 33 responding urologists (79%) stated that there were no SEs when the SE form was not filled out.In patients recruited by the remaining seven sites (21%), there might have been SEs, but the grading was not assessed.
In Table 3, we present a best case scenario assuming that there were no SEs in patients in whom the SE form was not filled out.Globally, the treatment toxicity did not exceed grade II in the majority of the patients.Grade III and IV local SEs were more frequent in the SF arm (n = 7; 3.8%) than in the RF arm (n = 1; 0.6%; p = 0.07; data not shown).The Overall, local SEs were reported more often than systemic SEs at all time points for both arms.Urination problems (frequency, urgency, dysuria, and incontinence) were the most commonly reported local SEs, whereas fever and general malaise were the most frequent systemic SEs.Although mostly insignificant, the numbers of recorded local and systemic SEs were generally higher in the SF arm.We found the SF arm to have a significantly higher frequency of total local SEs at T7 (p 0.001).Moreover, the total number of patients with local SEs was significantly higher in the SF arm than in the RF arm at T5 (p = 0.01) and T7 (p 0.001).Specifically, there were significant differences in the incidence of urgency (p = 0.05) and general malaise (p = 0.03) at T1, and frequency (p = 0.002), urgency (p = 0.02), dysuria (p = 0.001), and chemical cystitis (p = 0.03) at T7 favoring RF arm patients.
Follow-up
Side effects (after previous instillations; n = 14) Side effects (after previous instillations; n = 5) When a worst case scenario is assumed in which all patients without an SE grading form had SEs, we see a prevalence of local SEs at T2 of 42% and 35% in the SF and RF arms, respectively.At T5 and T7, the prevalence was 41.2% versus 31.1% and 39.6% versus 27.1%, respectively.The prevalence of systemic SEs at T2, T5, and T7 was 22.5% versus 18.1%, 24.7% versus 19.8%, and 20.3% versus 17.5%, respectively.Overall, we see the same pattern of fewer SEs in the RF arm than in the SF arm, but differences are small.
PP analyses
After excluding patients according to the PP criteria, a total of 249 patients remained, of whom 123 (49.4%) were ran-
AppeƟte loss
Fig. 3 (continued) domized to the SF arm and 126 (50.6%) to the RF arm.No significant differences in the baseline characteristics were observed between the two arms (Supplementary Table 1).Supplementary Table 2 and Supplementary Figure 1 summarize the results obtained from the QoL analyses.For the largest part, similar results to those in the ITT analysis are seen.The difference in physical functioning in the ITT analysis at T5 is no longer present in the PP results.However, we found a higher mean score of diarrhea in the RF arm at T1 (p = 0.01), which was not the case in the ITT analysis.Again, the linear mixed model did not display statistically significant temporal changes in any QoL domain for both the SF and the RF arm (p > 0.05).The PP analysis of SEs was also largely consistent with the ITT analysis (Supplementary Table 3).However, unlike in the ITT analysis, no significant differences were found in the total number of patients with local SEs at T5 (p = 0.15), urgency at T1 (p = 0.39), and general malaise at T1 (p = 0.06).We additionally found the number of patients with bacterial cystitis in the SF arm to be significantly higher than that in the RF arm (p = 0.03), which was not the case in the ITT analysis.
Discussion
An analysis of EAU-RF NIMBUS study data did not show better QoL in patients undergoing an RF BCG instillation regimen.However, there were significant differences in the incidence of general malaise at T1, and of storage symptoms of frequency, urgency, and dysuria at T7 favoring RF arm patients.Previous studies showed contrasting results in terms of the QoL and toxicity experienced after a dosage reduction in BCG.Yokomizo et al. [14] found a lower BCG dose (40 mg) to be associated with lower toxicity and better QoL than the standard BCG dose (80 mg).This study focused primarily on an eight-instillation induction phase, while QoL was assessed only once after the induction phase had ended.The EORTC 30962 trial analyzed the efficacy of one-third BCG doses compared with the standard dose.They did not report any difference in toxicity between the reduced and full-dose arms [8].This trial, however, was mainly designed to analyze the toxicity after the maintenance phase and did not focus on the induction phase.Nonetheless, these studies focused on the effect of a reduced dose of BCG instillations, whereas our study focused on a full dose but RF of BCG instillations.A direct comparison of these studies is therefore difficult.
In accordance with literature, our study reported urinary SEs, general malaise, and fever as the most frequent SEs caused by BCG instillations [15,16].These SEs were predominantly mild to moderate, reflecting generally good BCG tolerability.
A reduced BCG instillation frequency however significantly decreased the number of overall SEs.This difference was significant for general malaise (p = 0.03) at T1, and for frequency (p = 0.002), urgency (p = 0.02), and dysuria (p = 0.001) at T7.In fact, three times more patients in the SF arm did not complete the instillations due to SEs (14 vs 5 patients).
Interestingly, the higher toxicity reported in the SF arm did not translate into worse QoL.This unexpected result may be explained by the instrument used to measure QoL (QLQ-C30), which is not optimal for this group of patients.Although the QLQ-C30 questionnaire has been validated internationally, it does not focus directly on (non-muscleinvasive) bladder cancer (BC).Literature suggests that the questionnaire fails to assess finer BC-specific details/domains, which reduces the responsiveness to changes.Domains such as sexual functioning, self-consciousness, embarrassment, and psychological distress are of greater importance in BC patients but are not assessed thoroughly by the QLQ-C30 questionnaire [17].Several BC-specific questionnaires have been designed to offer an instrument that closes these gaps, such as the EORTC QLQ-NIMBC24 questionnaire, which has shown excellent measurement properties with regard to validity, reliability, and responsiveness [18], but this did not exist at the time of the original trial design.
The randomized setting of our study is a strength.In addition, the eight time point QoL evaluations over a year should have been able to pick up temporal QoL changes.Nonetheless, there are limitations to this post hoc analysis.In addition to the suboptimal QoL questionnaire, QoL was not measured anymore after the endpoint (a recurrence) was reached so that the influence of, for example, extra TURBTs could not be studied.Moreover, the large number of unanswered questionnaires resulted in a smaller number of evaluable patients (again raising further questions about the suitability of the instrument used to measure QoL changes).Lastly, the patients could not be blinded to RF instillation, which may have induced a response bias.This may thus have instigated some sort of placebo effect in patients to indeed experience better QoL with RF BCG instillations and vice versa.
Conclusions
An analysis of the EAU-RF NIMBUS study data did not show better QoL with EORTC QLQ-C30 v3.0, in patients undergoing an RF BCG instillation regimen despite lower storage symptoms at T7 in favor of RF.This finding may possibly be explained by the insensitivity of the EORTC QLQ-C30 questionnaire for small QoL domain changes.Our study, together with the previous finding that an RF schedule is inferior, supports the use of a standard BCG instillation schedule.
Fig. 1 -
Fig. 1 -Overview of the two treatment arms where each block represents a BCG instillation.The crossed out blocks in the RF arm represent the instillations that had not been performed.The different time points represent the moments the QLQ-C30 questionnaire and the side-effect evaluations had been completed.BCG = bacillus Calmette-Guérin; MM = maintenance month; QLQ-C30 = Quality of Life Questionnaire Core 30 version 3.0; RF = reduced frequency; SF = standard frequency; T = time point; W = week.
0) BCG = bacillus Calmette-Guérin; CI = confidence interval; CIS = carcinoma in situ.a Three patients had CIS only: 1Â treatment completed (15 instillations), patient included in follow-up, no recurrence; 1Â treatment completed (14 instillations), patient included in follow-up, first recurrence, and tumor in prostatic urethra at month 36; 1Â consent withdrawn after six instillations, patient included in follow-up until that time point, no recurrence.b Patient did not receive BCG and was not included in follow-up.c One patient was previously treated with BCG.This was a protocol violation.The patient was kept in the analyses for consistency with the original paper on the NIMBUS trial (by Grimm et al.[9]).E U R O P E A N U R O L O G Y O P E N S C I E N C E 5 6 ( 2 0 2 3 ) 1 5 -2 4 number of grade III and IV systemic SEs were similar to that of the local SEs (3.8% and 0.6%, respectively).
Fig. 2 -
Fig. 2 -Intravesical treatments received and reasons to stop.a Examples of ''Other'' are consent withdrawn, lost to follow-up, and patient not compliant.
C30 = European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 version 3.0; QoL = quality of life; RF = reduced frequency; SD = standard deviation; SF = standard frequency.E U R O P E A N U R O L O G Y O P E N S C I E N C E 5 6 ( 2 0 2 3 ) 1 5 -2 4
Fig. 3 -
Fig. 3 -Summary of the results from the EORTC QLQ-C30 where the X axes represent the points and the Y axes represent the mean QoL of the different EORTC scales and items (all scales have a range 0-100; for QoL scales, a higher score means better QoL; for symptom scales, a higher score means more symptoms).EORTC = European Organization for Research and Treatment of Cancer; QoL = quality of life; QLQ-C30 = Quality of Life Questionnaire Core 30 version 3.0; RF = reduced frequency; SF = standard frequency.
Calmette-Guérin; RF = reduced frequency; SE = side effect; SF = standard frequency; WHO = World Health Organization.a The following grade 3 or grade 4 side effects were observed: (1) grade 3 local side effects: one event in the RF group (M3W1) and seven events in four patients in the SF group (3Â M2W6, 2Â M6W1, and 2Â M6W3); (2) grade 4 local side effects: none; (3) grade 3 systemic side effects: one event in the RF group (M3W1); and (4) grade 4 systemic side effects: seven events in four patients in the SF group (M6W1, M6W3, and M12W2).b For patients for whom a side-effect form was not filled out, we assumed that there were no side effects.c Calculated using Mann-Whitney U test based on the average number of side effects per patient.E U R O P E A N U R O L O G Y O P E N S C I E N C E 5 6 ( 2 0 2 3 ) 1 5 -2 4
Table 3 -
Incidence of WHO grade I-IV side effects a by treatment groups at time points T1 (induction week 6), T5 (maintenance month 6, week 3), and T7 (maintenance month 12, week 3) | 4,690 | 2023-09-12T00:00:00.000 | [
"Biology"
] |
Candidate Gene Approach Identifies Multiple Genes and Signaling Pathways Downstream of Tbx4 in the Developing Allantois
Loss of Tbx4 results in absence of chorio-allantoic fusion and failure of formation of the primary vascular plexus of the allantois leading to embryonic death at E10.5. We reviewed the literature for genes implicated in chorio-allantoic fusion, cavitation and vascular plexus formation, processes affected in Tbx4 mutant allantoises. Using this candidate gene approach, we identified a number of genes downstream of Tbx4 in the allantois including extracellular matrix molecules Vcan, Has2, and Itgα5, transcription factors Snai1 and Twist, and signaling molecules Bmp2, Bmp7, Notch2, Jag1 and Wnt2. In addition, we show that the canonical Wnt signaling pathway contributes to the vessel-forming potential of the allantois. Ex vivo, the Tbx4 mutant phenotype can be rescued using agonists of the Wnt signaling pathway and, in wildtype allantoises, an inhibitor of the canonical Wnt signaling pathway disrupts vascular plexus formation. In vivo, Tbx4 and Wnt2 double heterozygous placentas show decreased vasculature suggesting interactions between Tbx4 and the canonical Wnt signaling pathway in the process of allantois-derived blood vessel formation.
Introduction
The chorio-allantoic placenta of eutherian mammals is critical for fetal development and growth during gestation. The allantois first appears at embryonic day (E) 7.5 as a bud of mesoderm that emerges from the posterior end of the primitive streak [1,2], and then grows into the exocoelomic cavity, cavitates between E7.5 and E8. 25 and undergoes chorio-allantoic fusion [3]. Formation of endothelium occurs de novo within the allantois at the headfold (HF) stage, beginning in the distal allantois with the appearance of Flk1-positive angioblasts, which are precursors of endothelial cells (ECs) [4,5]. Specification of angioblasts and their morphogenesis into endothelial tubes (ETs) then proceeds proximally to the base of the allantois, where nascent allantoic blood vessels fuse with those of the embryo to create a continuous vasculature throughout the embryo and yolk sac [1,5]. The allantois vessel network, also known as a vascular plexus, is ultimately remodeled into an umbilical artery, umbilical vein and the fetal vessels of the placenta. The vascular plexus of the allantois presumably promotes the growth of mural cells to provide structural support for the vascular walls, similar to the yolk sac vascular plexus [6]. Extracellular matrix (ECM) is also present, and although there have been some reports of the presence of specific matrix components in the allantois [7,8,9,10], the presence of mural cells and the composition of the ECM is largely unknown.
Following chorio-allantoic fusion, the chorion forms villi into which the allantois vasculature grows, ultimately forming the labyrinthine layer of the placenta. Other components of the placenta include the outermost, maternally-derived decidual layer, the giant cell layer derived from trophectoderm and the spongiotrophoblast layer derived from polar trophectoderm. Defective development of any of these layers can lead to placental insufficiency and, in severe cases, embryonic death [11].
Mutation of the T-box transcription factor gene Tbx4 results in abnormal vascular development in the allantois, loss of cavitation, apoptosis and lack of chorio-allantoic fusion leading to embryonic death at E10.5. Pecam, a marker of ECs, is abundantly expressed in cells of Tbx4 mutant allantoises but these ECs do not coalesce into a primary vascular plexus [12]. Comparison of Tbx4 RNA and Pecam protein localization as well as lineage tracing using a Tbx4-cre allele suggests that neither the ECs of the umbilical vessels nor their precursors express Tbx4 [13]. In spite of this, Tbx4 null mutants show a defect in allantois EC organization suggesting that Tbx4 plays a non-cell-autonomous role in formation of the vascular plexus.
Thus, we took a candidate gene approach to find Tbx4 target genes expressed in the mesenchyme that could explain this noncell-autonomous effect. Candidates were chosen if their loss either leads to chorio-allantoic fusion defects, cavitation defects or results in a vascular phenotype [14] similar to the Tbx4 mutant allantois vascular phenotype. We analyzed expression of ECM components: Hyaluronic acid synthase2 (Has2) [15], because hyaluronic acid is involved in allantois cavitation [10,16], which is disrupted in Tbx4 mutants [12]; Versican (Vcan) [17], which is a chondroitin sulphate proteoglycan binding partner for hyaluronic acid; a5 integrin (Itga5) and Fibronectin (Fn), ECM genes known to be essential for development of endothelial tubes [18]. We examined components of several signaling pathways: Bmp2, Bmp7 and Bmp5, which are implicated in chorio-allantoic fusion [19,20]; Wnt2, one of the canonical Wnt family members, which is expressed in the allantois and important for placentation [21]; Notch2, which is expressed in the allantois [22]; Delta like 4 (Dll4) and Jagged1 (Jag1), ligands for Notch receptors, implicated in vascular development [23,24]. Among transcription factors, we analyzed expression of Twist and Snai1. Twist has been shown to be important for Snai1 expression [25]; epiblast-specific deletion of Snai1 leads to a vascular phenotype in which the ECs of the embryo and allantois express Pecam but fail to coalesce to form ETs [26], similar to Tbx4 mutant allantoises.
We found that expression of multiple ECM genes, signaling molecules and transcription factors is affected in Tbx4 mutants, some of which have conserved T-box binding sites in their promoters. We further show that canonical Wnt signaling contributes to vessel-forming potential of the ECs of allantoises ex vivo and growth of the umbilical vessels into the placenta in vivo.
Tbx4 Mutant Allantoises Fail to Form a Vascular Plexus ex vivo
To investigate their vessel-forming potential, Tbx4 mutant and control allantoises were cultured ex vivo for 24 hours starting from the late head fold (LHF) stage to the 6 somite stage (E8-8.5). Wildtype or Tbx4 heterozygous allantoises from the earliest developmental stages give rise to clusters of ECs, whereas explants from later stages spread out and give rise to a network of interconnected ETs that form a plexus [27,28], as shown by Flk1 antibody staining ( Figure 1A-F). Tbx4 mutant allantoises, on the other hand, had clusters of ECs or ETs but failed to form a vascular plexus of interconnected ETs even in explants from the most advanced embryos ( Figure 1G-I). Methylene blue nuclear staining ( Figure 1B,E,H) shows the extent of allantois outgrowth. The proportion of explants of each genotype that proceeded to form EC clusters, ETs or a vascular plexus is shown in Figure 1J. Because Tbx4 mutant allantoises show increased apoptosis [12], we tested whether decreased cell numbers affected plexus formation. When split in half, control allantoises isolated at 2-4 somites (E8.25) still formed a vascular plexus upon culture, although the plexus was smaller than that from whole allantoises (data not shown) suggesting that reduced cell number alone was not the cause of lack of plexus formation in Tbx4 mutants. As Tbx4 is not expressed in ECs of the allantois and cells derived from Tbx4expressing cells never contribute to the endothelium of the umbilical vessels [13], these data suggest that Tbx4 plays a noncell-autonomous role in the development of allantois vasculature. Thus, Tbx4 could either regulate the developing mesenchyme surrounding the ECs or the ECM through which the ECs migrate in order to form ETs and a primary plexus.
Tbx4 is Upstream of ECM Molecules, Signaling Molecules and Transcription Factors
To explore the role of Tbx4 in the development of the allantois we analyzed the expression of various ECM markers, signaling molecules and transcription factors in a candidate gene approach.
Expression of both Bmp2 and Bmp7 was undetectable in Tbx4 mutant allantoises ( Figure 2F,G). Wnt2 is expressed throughout the process of growth and vascularization of the allantois starting at the LHF stage [21]. Wnt2 expression was undetectable in Tbx4 mutant allantoises ( Figure 2H and Figure 3). On the other hand, non-canonical Wnt5a was expressed in mutant allantoises ( Figure 2I). Dll4 was expressed in Tbx4 mutant allantoises in EC clusters, which fail to form vessels ( Figure 2J). Jag1 and Notch2 were expressed in control allantoises but were not detected in Tbx4 mutant allantoises ( Figure 2K,L). Similarly, Snai1 and Twist were not detected in Tbx4 mutant allantoises ( Figure 2M,N).
Agonists of the Canonical Wnt Signaling Pathway Rescue and a Wnt Signaling Antagonist Phenocopies the Tbx4 Mutant Phenotype ex vivo
Because canonical Wnt2 was not expressed above background in Tbx4 mutant allantoises (red arrows Figure 3A,B) although it was expressed in hearts of Tbx4 mutant embryos (yellow arrows Figure 3A,B), we examined the contribution of the canonical Wnt signaling pathway to the vessel-forming potential of the allantois under the control of Tbx4. The canonical Wnt signaling pathway was activated in culture using LiCl, an inhibitor of GSK3-b and thus an activator of b-catenin-mediated transcription [29]. When treated with LiCl, Tbx4 mutant allantois explants formed ETs which interconnected and formed a vascular plexus ex vivo ( Figure 4B) unlike mutant allantoises cultured in the absence of LiCl, which never formed a vascular plexus ( Figure 4A). The total number of interconnections of the ETs in the vascular network was compared in rescued and control mutant allantoises as a measure of the extent of plexus formation ( Figure 4C) and was shown by notched box plots to be significantly different between control and LiCl treated mutants [30] (Mann Whitney U test p = 0.0016). Similar rescue of ET and plexus formation was seen when Wnt2 conditioned media was added to mutant allantois cultures (control mutants, n = 2, number of interconnections 1, 10; mutants with Wnt2 conditioned media, n = 2, number of interconnections 54, 63) ( Figure 4D, E).
IWR1, a small molecule inhibitor, was used to block the canonical Wnt signaling pathway at a concentration found to be non-toxic to cells in organ culture [31]. Allantoises from wildtype embryos were isolated at EHF-LHF ( Figure 4F,G,H) and LHF-1 somite stages ( Figure 4I,J,K) and IWR1 in DMSO or DMSO vehicle alone was added to the cultures. In the presence of inhibitor, wildtype allantoises failed to form a vascular plexus but showed formation of EC clusters positive for Flk1. Taken together, these results suggest that canonical Wnt signaling contributes to the vessel-forming ability of allantoises ex vivo, and in the absence of this signal the ECs fail to form a primary vascular plexus.
Tbx4 and Wnt2 Genetically Interact during the Formation of the Chorio-allantoic Placenta Since canonical Wnt signaling is necessary for formation of allantois vessels ex vivo we tested whether Tbx4-mediated control of canonical Wnt signaling alone was responsible for vessel formation in allantoises in vivo. Wnt2 is the only canonical Wnt known to be expressed in the allantois and thus we analyzed Tbx4 +/2 ;Wnt2 +/2 and Tbx4 +/2 ;Wnt2 2/2 allantoises. Allantoises from both these genotypes fused with the chorion normally at E8.5 and formed normal umbilical vessels at E10.5 (data not shown) indicating no genetic interactions between Tbx4 and Wnt2 during allantois vessel formation in vivo. We hypothesized that lack of a phenotype in the Tbx4 +/2 ;Wnt2 2/2 allantoises could be due to activation of b-catenin by another pathway. Thus, in order to create a mesenchymal deletion of b-catenin, we used a Tbx4-cre [13] and a b-catenin conditional allele [32]. Tbx4 cre/+ ; bcatenin cond/cond embryos did not show an allantoic vascular phenotype (data not shown). To determine if Tbx4 and Wnt2 interact later during allantoisderived vessel formation, chorio-allantoic placenta development of Tbx4 +/2 ;Wnt2 +/2 embryos was analyzed. At E11.5, there is a decreased amount of perivascular stroma in the chorio-allantoic plate of double heterozygous placentas ( Figure 5A-D). The extent of vascularization of the chorio-allantoic plate was analyzed using H & E sections of controls and double heterozygote placentas ( Figure 5A,B). Quantification of the vessels shows a significantly lower number of vessels present in double heterozygotes compared to controls (Mann Whitney U test, p = 0.001). Identification of vessels was confirmed by Pecam staining (Fig. 5C,D). At E13.5, double heterozygotes showed increased interdigitation of the spongiotrophoblast and labyrinth (Figure 5E,F,G,H). Additionally, at E13.5 there was a significant reduction in the labyrinthine layer as seen from the ratio of area of labyrinth to the combined area of labyrinth and spongiotrophoblast in double heterozygotes ( Figure 5I, Mann Whitney U test, p = 0.002). These results indicate that Tbx4 and Wnt2 interact to form the allantois-derived vasculature of the chorio-allantoic plate and the labyrinthine layer of the placenta. Tbx4 and its closely related paralogue Tbx5 share 94% amino acid identity in their T-box domains [33]. Although the DNA binding sequence for Tbx4 has not been characterized, we tested whether Tbx4 protein is capable of binding to the TBX5 consensus T-box binding element (TBE) in the ANF promoter [34]. When incubated with cell lysates containing either mouse Tbx4 or human TBX5, probe containing the ANF TBE showed a mobility shift which was lost when the core sequence of the TBE was mutated ( Figure 6A) confirming that Tbx4 can bind the TBX5 TBE. We then used ConTra [35] to examine 2 Kb of the promoter region of the genes affected in Tbx4 mutant allantoises for this element and, where it was present, analyzed its conservation. Four out of the ten genes affected in the Tbx4 mutant allantoises -Vcan, Has2, Twist and Bmp2have TBE's conserved in at least mouse and human ( Figure 6B, yellow circles) or more than 5 mammalian species including mouse and human ( Figure 6B, red circles). An example of site conservation between multiple mammalian species including mouse and human upstream of Vcan is shown in Figure 6C and an example of site conservation between human and mouse upstream of Twist is shown in Figure 6D.
Discussion
Tbx4 affects multiple processes important for allantois development including chorio-allantoic fusion, cavitation, cell proliferation, apoptosis and vasculogenesis [12]. Here we show evidence that Tbx4 plays a role in these processes by regulating expression of a variety of ECM genes, signaling molecules and transcription factor genes either directly or indirectly. We show that expression of ECM molecules Vcan, Has2, and Itga5, transcription factors Snai1 and Twist, and signaling molecules Bmp2, Bmp7, Notch2, Jag1 and Wnt2 is undetectable in Tbx4 mutant allantoises. Further, we show that Tbx4 could potentially directly regulate the expression of Vcan, Has2, Twist and Bmp2, as they have conserved T-box binding sites in their promoters. Additionally, individual mutations in some of these genes show that their absence alone would be enough to explain the EC phenotype of Tbx4 mutant allantoises. For example, an epiblast-specific mutation in Snai1, which is absent from Tbx4 mutant allantoises, results in the lack of a vascular plexus in the embryo proper and in allantois explants, although explants show the presence of differentiated Pecam-positive ECs [26]. Itga5 [18] and Wnt2 mutant embryoid bodies [36] that are differentiated along the vascular pathway show Pecam-positive ECs but do not form a vascular plexus; neither of these genes is detected in Tbx4 mutant allantoises. The cell adhesion molecule Vcam1 is lacking in Tbx4 mutant allantoises, which potentially explains their lack of chorioallantoic fusion [12].
Four of the genes absent in Tbx4 mutant allantoises, Vcan, Jag1, Snai1 and Twist, are downstream targets of the canonical Wnt signaling pathway in different systems (http://www.stanford.edu/ group/nusselab/cgi-bin/Wnt/target_genes), indicating a role for canonical Wnt signaling in development of allantois vasculature downstream of Tbx4. Additionally, since Tbx4 is not expressed in ECs of the allantois but its absence still causes a drastic effect on the assembly of ECs into a vascular plexus, loss of a secreted molecule like Wnt2 could explain this non-cell-autonomous effect. There is evidence that Wnt2 plays an important role in EC proliferation and network formation specifically in hepatic sinusoidal ECs (HSEC's). Wnt inhibitors lead to reduced proliferation and reduced endothelial-tube-forming ability of HSEC's on matrigel [37]. Furthermore, Wnt2 null embryoid bodies, which fail to form a Pecam-positive vascular plexus, can be rescued by the addition of Wnt2 conditioned media [36]. We were able to rescue the vessel-forming ability of Tbx4 mutant allantoises by activating the Wnt signaling pathway using LiCl or by culturing in Wnt2 conditioned media. We were also able to disrupt vascular plexus formation in wildtype allantoises by addition of a tankyrase inhibitor IWR1. Tankyrase destabilizes axin and prevents bcatenin from degradation, thus in the presence of the inhibitor, bcatenin is degraded, blocking the canonical Wnt signaling pathway [31]. Taken together, our results suggest for the first time that canonical Wnt signaling contributes to the vessel-forming ability of allantoises ex vivo, and in the absence of this signal the ECs fail to form a primary vascular plexus.
Tbx4 and Wnt2 do not show a genetic interaction during allantoic vessel formation in vivo. One possible explanation is that other, as yet unknown, canonical Wnts are expressed in the allantois. Alternatively, b-catenin could be activated downstream of Tbx4 by other pathways that are still active in Tbx4 +/2 ;Wnt2 +/2 and Tbx4 +/2 ;Wnt2 2/2 allantoises. It has already been shown that conditional excision of b-catenin in ECs does not affect allantois vascular development [38]. Thus, neither an endothelial [38] nor a mesenchymal deletion of b-catenin (our study) reproduces the Tbx4 mutant phenotype, suggesting other pathways downstream of Tbx4 are involved. The presence of conserved TBEs in the promoters of Vcan, Has2, Twist and Bmp2 make these genes good candidates for direct regulation by Tbx4 during the development of the allantois vasculature.
It is also possible that Tbx4 and Wnt2 may interact in formation of allantois vasculature but a phenotype is not evident until later in development in the allantois-derived placental vasculature. This is the case with the HoxA genes -HoxA10, HoxA11 and HoxA13all of which are expressed in the allantois. The expression of the HoxA genes is essential during a brief window of allantois development but the mutant phenotype is only evident at later stages as a disruption of placental vasculature [39]. Thus, it may be a common theme for some genes expressed in the allantois to manifest a phenotype in the placenta by regulating the development of the allantois. We show that indeed, Tbx4 and Wnt2 interact in double heterozygotes in the formation of placental vasculature. By itself Wnt2 is important for the proper formation of the different placental layers and Wnt2 null mutants show disruptions in the placental labyrinthine layer [21]. In addition, canonical Wnt signaling has been shown to be important for placenta development; deficiency in a number of genes involved in Wnt/b-catenin signaling show mid-gestation lethality due to placental insufficiency. For example, conditional inactivation of b-catenin in ECs shows a phenotype similar to Wnt2;Tbx4 double heterozygotes, where the labyrinthine layer is smaller and less vascularized [38]. Rspondin3, a secretory molecule whose proposed function is to promote the canonical Wnt signaling pathway, is expressed in the allantois. Fetal vessels of Rspondin3 null embryos fail to penetrate the chorion leading to lethality during midgestation [40]. Similarly, the Wnt receptor Fzd5 is expressed in the labyrinthine layer and null mutants for Fzd5 die due to lack of penetration of fetal vessels into the chorion [41]. We show that the Wnt2;Tbx4 double heterozgygous placentas have decreased vascular coverage in the chorioallantoic plate, again suggesting the importance of Wnt2 in the vessel-forming potential of ECs derived from the allantois. Furthermore, double heterozygous placentas show increased interdigitation of placental layers and reductions in the labyrinthine layer of the placenta. Thus, although Wnt2 does not genetically interact with Tbx4 in the process of vessel formation in the allantois, these genes do interact in formation of the vasculature of the chorio-allantoic plate and the labyrinthine layer of the placenta.
Mice
Mice carrying a Tbx4 null allele, Tbx4 tm1 . 2Pa [12], hereafter referred to as Tbx4 2 , a Wnt2 null allele, Wnt2 tm1 . 1(rtTA)Eem [42], hereafter referred to as Wnt2 2 , a Tbx4-cre allele, an insertion into the endogenous Tbx4 locus resulting in a bicistronic allele that expresses both cre and Tbx4 and has been shown to be expressed in all areas of Tbx4 expression [13] and a b-catenin loss-of-function conditional allele [32], hereafter referred to as bcatenin cond , were genotyped as described previously.
Histology and Vessel Counts in the Chorio-allantoic Plate
Embryos and placentas were collected from timed matings and dissected out of the decidua; yolk sacs were removed for genotyping and the placentas were fixed in Bouin's fixative (Sigma). After dehydration in ethanol, placentas were embedded in paraffin wax, sectioned at 10 mm and stained with hematoxylin and eosin (H & E). From cross sections through the center of the E11.5 placenta, every 5 th section (for a total of 15-20 sections/ placenta) was used to quantitate vasculature by counting the number of vessels per section present in the chorio-allantoic plate. The total number of vessels per placenta was used as a representation of the amount of vasculature in the chorio-allantoic plate. Three to five placentas were analyzed for each genotype. At E13.5, quantitation of the area of labyrinthine and spongiotrophoblast layers of placenta was done on H & E stained sections using the software NIS Elements (Nikon). For antibody staining, placentas were fixed in paraformaldehyde, equilibrated in 30% sucrose, embedded in OCT (Tissue-Tek) and cryosectioned at 10 mm.
Ex vivo Allantois Culture
Allantoises were aspirated using glass pipettes to obtain the full length from distal tip to the site of attachment to the amnion and yolk sac [28] and transferred to 24 well dishes containing 0.5 ml of rat serum (Pel-Freeze Biologicals) and DMEM (Invitrogen) in a 1:1 ratio. The allantoises were scored for developmental progression based on formation of EC clusters, ETs or a vascular plexus at the end of 24 hours. For rescue experiments, lithium chloride (LiCl) was added to the media at a final concentration of 5 mM. Wnt2 producing CHO cells were used to make Wnt2 conditioned media and vector transfected CHO cells were used to make control conditioned media [37]. For conditioned media rescue experiments, allantoises were cultured in 50% rat serum, 25% control or Wnt2-conditioned media and 25% DMEM. For inhibition of canonical Wnt signaling, 100 mM IWR1 (Cayman Chemicals) prepared in DMSO, a concentration shown to be non-toxic to cultured cells [31] or DMSO alone was added to cultures and allantoises were cultured for 36-40 hours. For antibody staining, the cultures were fixed overnight in 4% PFA, washed in PBT, and treated for IHC for Flk1. For each experiment n was $2.
Generation of Fluorescent Probes Containing TBE Upstream of ANF
To generate probe, 12 bp linker sequence (acg agt ctc tac) was added to the 39 end of the reverse complement of the TBX5 TBE in the promoter region upstream of the ANF start site (2275 to 2246) to generate a wildtype T-box binding sequence, 59gct ccc act tca aag gtg tga gaa gag taa acg agt ctc tac 39. To generate a mutant version of this site the core sequence of the T-box binding element was mutated to get the sequence 59gct ccc act tca aag gga tga gaa gag taa acg agt ctc tac 39 (the residues mutated have been underlined and the linker is italicized). A fluorescently labeled oligonucleotide complimentary to the linker sequence, Cy5-59 gta gag act cgt 39, was annealed to both these sequences and filled using Klenow polymerase to generate the fluorescently labeled ANF wildtype and mutant double stranded probes, which were used for the electrophoretic mobility shift assay (EMSA).
Preparation of Cell Lysates Containing Tbx4 and Tbx5 Protein and EMSA
To generate Tbx4 and Tbx5 protein, reactions were carried out using a transcription translation coupled reticulolysate system (TNTH T7 Quick Coupled Transcription/Translation System, Promega) according to the manufacturer's protocol. 1ug of human TBX5 cDNA [34] and mouse Tbx4 cDNA [45] was incubated in a 50 ml reaction, for 1 hour at 30uC. Cell lysates containing Tbx4 or Tbx5 protein were incubated with approximately 0.3 ng of fluorescently labeled probe in binding buffer (20 mM HEPES, pH 7.5, 50 mM KCl, 5 mM MgCl 2 , 10 mM ZnCl 2 , 6% glycerol, 200 mg of bovine serum albumin per ml, and 50 mg of poly(dI-dC)?poly(dI-dC) per ml [46]) for 20 minutes at room temperature. This reaction was then loaded onto a 4% polyacrylamide gel, samples were run for 3 h at 120 V at room temperature, the gel was vacuum dried and imaged using Typhoon TRIO variable mode imager (Amersham Biosciences). Probe alone or probe with empty lysates (without cDNA) were used as negative controls. | 5,626.6 | 2012-08-28T00:00:00.000 | [
"Biology"
] |
Categorical Tori
We give explicit and elementary constructions of the categorical extensions of a torus by the circle and discuss an application to loop group extensions. Examples include maximal tori of simple and simply connected compact Lie groups and the tori associated to the Leech and Niemeyer lattices. We obtain the extraspecial 2-groups as the isomorphism classes of categorical fixed points under an involution action.
Acknowledgments
It is a pleasure to thank David Roberts for very helpful conversations and correspondence. The idea for Construction 2.1 came from a conversation with him, and I understand that he will also write about it elsewhere. The idea to look for crossed module extensions of tori, picked up in Section 6.2, is also due to David. Many thanks go to Matthew Ando for very inspiring conversations, to Konrad Waldorf for patiently answering a long list of emails full of technical questions and to Arun Ram, who helped with some of the references.
Constructions of categorical tori
To build a categorical torus, we need a finite dimensional lattice 1 , which we denote Λ ∨ , and an integer-valued bilinear form J on Λ ∨ . Up to isomorphism, our constructions will depend only on the even symmetric bilinear form I(m, n) = J(m, n) + J(n, m).
We will also write J for the bilinear form J ⊗ R on t. The exponential map is and we make the identification Λ ∨ = ker(exp) ⊆ t.
We write t/ /Λ ∨ for the action groupoid of Λ ∨ on t. We have a canonical equivalence of Lie groupoids 2 where T is viewed as groupoid with only identity arrows. We will give three equivalent constructions of the categorical torus associated to (Λ ∨ , J). The first is as a strict monoidal Lie groupoid.
So, T has objects t and arrows t × Λ ∨ × U (1), the source of (x, m, z) is x, and the target of (x, m, z) is x + m. We equip T with the following multiplication x • y = x + y on objects (x, m, z) • (y, n, w) = (x + y, m + n, z · w · exp(−J(m, y))) on arrows.
The unit object is 0, and associativity and unit isomorphisms are identities. This makes (T, •) a strict monoidal Lie groupoid, i.e., a group object in the category of Lie-groupoids.
The second construction re-interprets the data of Construction 2.1 as a compact Lie 2-group in the sense of Schommer-Pries [SP11].
Construction 2.2. Let T be the Lie groupoid i.e., T has objects T and arrows T × U (1), source and target are projection onto the first factor and composition of arrows is multiplication in the second factor. Then we have an equivalence of Lie groupoids Note that p × id does not possess a continuous inverse equivalence. So, one interprets the data of Construction 2.1 as those of a 2-group object in a suitable localization of the bicategory of Lie-groupoids, namely the bicategory of bibundles Bibun. Different communities have different language for the 1-morphisms in Bibun. Depending on your taste, you may think of multiplication on T as the zig-zag, span, orbifold map, or anafunctor Our third construction is as a multiplicative bundle gerbe in the sense of [Bry08] and [CJM + 05].
Construction 2.3. Let I be the trivial bundle gerbe over T . Recall that a multiplication on I is a stable isomorphism µ : pr * 1 I ⊗ pr * 2 I −→ m * I of bundle gerbes over T ×T , where the pr i are the projections to the factors and m is multiplication in T . Such a stable isomorphism of trivial bundle gerbes is the same as a line bundle, and we take µ to be for (m, n) ∈ Λ ∨ × Λ ∨ . Finally, we need to specify the associativity isomorphism where pr ij is the projection onto factors i and j and m ij is multiplication of these two factors, e.g., m 12 = m × id. We take α to be the canonical isomorphism resulting from the fact that its source and target have identical multipliers Remark 2.4. Construction 2.3 takes the sum of two bilinear forms, J 1 + J 2 to the tensor product of multiplicative bundle gerbes.
By [Wal12,Thm.3.2.5], the data of a multiplicative bundle gerbe over T are equivalent to those of an extension of T by pt / /U (1). In the case of Construction 2.3, we may think of the objects of I as pairs (t, L) with t ∈ T and L a hermitian line, and of the arrows (t, L) → (t, L ′ ) as the unitary isomorphisms from L to L ′ . Then I is a monoidal groupoid with multiplication (s, L 1 ) • (t, L 2 ) = (s · t, L J s,t ⊗ L 1 ⊗ L 2 ). The associativity isomorphisms for • are encoded in α, α r,s,t : L J r·s,t ⊗ L J r,s ∼ = L J r,s·t ⊗ L J s,t . Proposition 2.5. The three constructions yield equivalent 2-group extension of T .
Proof. It is clear that Construction 2.1 and Construction 2.2 are equivalent. To see their equivalence with Construction 2.3, let F be the functor To make F a monoidal equivalence, we need a natural isomorphism This is a map of the form or, equivalently, a trivialization of L J over t × t. We have such a trivialization by construction of L J . One checks that (F, φ) is a monoidal equivalence from T to I.
Definition 2.6. We will write J t for the bilinear form J t (m, n) = J(n, m) on Λ ∨ . We say that J is symmetric if J = J t and that J is skew symmetric if J = −J t . A symmetric bilinear form is called even if Proposition 2.7. The bilinear forms J and J t yield equivalent 2-group extensions of T .
Corollary 2.8. (i) If I is an even symmetric bilinear form on Λ ∨ , then the multiplicative bundle gerbe associated to I possesses a square root. (ii) If B is a skew symmetric integral bilinear form on Λ ∨ , then B yields a trivial 2-group extension of T .
Proof. (i) follows from the fact that every even symmetric bilinear form I can be written in the form I = J + J t for an integer-valued bilinear form J. For instance, fix a basis (b 1 , . . . , b r ) of Λ ∨ and set else.
(ii) follows from the fact that, similarly, every skew symmetric bilinear form B can be written in the form B = J − J t for an integer-valued bilinear form J.
Corollary 2.9. Let J be an integer-valued bilinear form on Λ ∨ . Then, up to equivalence over T , the 2-group (T, • J ) only depends on the even bilinear form Proof. Let J 1 and J 2 be two integer-valued bilinear forms on Λ ∨ , and assume that Then J 1 − J 2 is skew symmetric. By Corollary 2.8, the multiplicative bundle gerbe obtained from J 1 − J 2 is trivial. Using Remark 2.4, we conclude that the multiplicative bundle gerbes obtained from J 1 and J 2 are isomorphic.
The example of the circle
Let t = R and Λ ∨ = Z. Any even symmetric bilinear form on Z is an integer multiple of I(m, n) = 2mn.
For this I, there is a unique choice of J, namely J(m, n) = mn.
The basic circle extension U (1) of the circle group U (1) consists of the following data: for (m, n) ∈ Z 2 , (iii) the canonical isomorphism For k ∈ Z, the kth circle extension U (1) k of U (1) is obtained by replacing the multipliers with exp(kmy).
Here are a few words of explanation about these choices: gerbes on U (1) are classified, up to stable isomorphism, by their Dixmier-Douady class in So, any bundle gerbe over U (1) is trivializable, and we might as well start with the trivial bundle gerbe I. Line bundles on U (1) × U (1) are classified, up to isomorphism, by their first Chern class in This is isomorphic to the group of skew symmetric bilinear forms on Z 2 , which is infinite cyclic, generated by the determinant. To construct a line bundle L with Chern class we use Chern-Weil theory: Figure 1. Chern-Weil theory: the multipliers can be read off from this cocycle in the truncatedČech-Deligne double complex for U (1) × U (1).
Note that this argument builds L as a line bundle with connection, given by the 1-form Similarly, the bundle L ⊗k with multipliers exp(kmy) has connection kxdy and Chern class k ·det.
The classification
Let T be a compact torus with Lie algebra t and coweight lattice Λ ∨ = ker(exp). Up to equivalence, the 2-group extensions of T by pt / /U (1) are classified by , where the left-hand side is Lie group cohomology as in [SP11] [WW13]. There are a number of different, but equivalent definitions of Lie group cohomology. We choose to work with thě Cech-simplicial double complexČ * (BT • ; U (1)) as in the classification of multiplicative bundle gerbes in [Bry08] and [CJM + 05, Prop 5.2]. The goal of this section is to analyze the degree four part of the composite of isomorphisms where S * (Λ) is the symmetric algebra of the weight lattice Λ = Hom(Λ ∨ , Z).
Weights are given degree 2, so that the degree four part is We may think of elements of S 2 Λ as homogeneous polynomials of degree 2 in the weights, and we have the symmetrization map identifying S 2 (Λ) with the group of even symmetric bilinear forms on Λ ∨ . So, (2) establishes a group homomorphism from the even symmetric bilinear forms on Λ ∨ to the isomorphism classes of multiplicative bundle gerbes over T .
Theorem 4.1. Let I be an even symmetric bilinear form on Λ ∨ , and let J be an integral bilinear form on Λ ∨ satisfying I = J + J t . If we apply Construction 2.3 to (Λ ∨ , J), then the resulting multiplicative bundle gerbe is classified by I.
Proof. Let λ ∈ Λ be a weight of T with character e λ , and write for the line bundle on BT classified by Be λ . The first isomorphism in (2) goes back to Borel, and is defined as where on the left-hand side, the weights have degree 2.
To define the second isomorphism in (2), let ET be a contractible free T -space, BT = ET /T , and and recall that ET × BT · · · × BT ET ∼ = ET × T × · · · × T.
So, the maps
form a hypercover of BT whoseČech double complex can be identified withČ * (BT • ; Z). Under this identification, the cup product becomes where r and s are theČech degrees of f and g, pr 1 is the projection onto the first deg simp (f ) factors and pr 2 is the projection onto the last deg simp (g) factors.
Note that these first two isomorphisms are actually isomorphisms of graded rings. To determine the image of λµ in H 4 (BT • ; Z), we determine the images of λ and µ and then take the cup product. The first Chern class of L λ inČech hypercohomology is given by the multipliers Hence, λ maps to the degree (1, 1) cocycle in theČech-simplicial double complex. Given two weights, λ and µ, their cup product is represented by the cocycle inČ 2 (BT 2 ; Z). The last isomorphism in (2) is the inverse of the connecting homomorphism for the short exact sequence of presheaves It is easy to read off that Construction 2.3 associates to J = µ ⊗ λ the multiplicative bundle gerbe corresponding to the U (1)-valuedČech-simplicial 3-cocycle (1, λ(m)µ(y), 1, 1).
The image of this cocycle under the connecting homomorphism is indeed the integral 4-cocycle
(1, 1, λ(k)µ(n), 1, 1), see the picture: This concludes the proof. in the Leray-Serre spectral sequence for the universal bundle EG → BG. For a compact, connected Lie group G, this map was calculated (in all degrees) by Chern and Simons. In the relevant degree, their result is summarized by the commuting diagram where ν is the Cartan 3-form 3 associated to I, ν(x, y, z) = I([x, y], z).
Since the Lie bracket on a torus is zero, it follows that the Dixmier-Douady class of the underlying bundle gerbe of any multiplicative bundle gerbe on T vanishes.
Connections
If J = λ ⊗ µ, then the Chern-Weil discussion analogous to that of Section 3 produces the connection ∇ on L J with connection 1-form ω t×t = λdµ and curvature 2-form For arbitrary J, we introduce the maps Then ∇ is defined by the 1-form The pair (L J , ∇) makes a multiplicative bundle gerbe with connection (in the sense of [Wal10]) out of the trivial bundle gerbe I on T (with the remaining data trivial).
Symmetries
In many interesting examples, we have a finite group Γ of linear isometries of (Λ ∨ , I). In this case, I may be interpreted as a Γ-invariant cohomology element, and the action of Γ on T preserves the multiplicative bundle gerbe classified by I up to isomorphism. As a consequence, we obtain a 2-group extension of Γ, namely the automorphism 2-group of the categorical torus. There are two variations worth exploring, depending on whether or not we require our symmetries to preserve the connection. We will come back to this topic at a different occasion.
Maximal tori
Let G be a compact connected Lie group with maximal torus T and Weyl group W . Then we have The formula [CS74, (3.10)] is often cited in its original form where ω is the right-invariant Maurer-Cartan form on G. A look at the definitions on page 50 of [CS74] identifies I(ω ∧ [ω, ω]) with the bi-invariant 3-form on G whose restriction to g = T1G equals 12 ν.
and we can choose I as a multiple of the Killing form. If G is simple and simply connected, then we have The positive definite generator I bas is called the basic bilinear form on Λ ∨ . The form kI bas ∈ H 4 (BG; Z) classifies the 2-group denoted String k (G) in [SP11]. The restriction of String k (G) to T is equivalent to the categorical torus constructed from (Λ ∨ , kI bas ). This property determines String k (G) uniquely. In other words, if we are given a 2-group extension G of G by pt / /U (1), and wonder which of the String k (G) it is equivalent to, it suffices to identify its restriction to T . This recognition principle promises to be useful for a comparison of the many equivalent constructions of the String 2-groups.
The Leech lattice
Another interesting example is given by the Leech lattice Λ ∨ = Λ Leech ⊆ R 24 , together with the symmetric bilinear form on R 24 making Λ ∨ an even unimodular lattice. Then T = R 24 /Λ Leech is the Leech torus, and the group of linear isometries of (Λ ∨ , I) is the Conway group Co 0 . So I can be interpreted as an element I ∈ H 4 (BT, Z) Co 0 .
Niemeyer lattices
Similarly, we can choose Λ ∨ as one of the Niemeyer lattices, e.g., A 24 1 or A 12 2 , and I as the symmetric bilinear form on R 24 making Λ ∨ an even unimodular lattice. Then T = R 24 /A 24 1 (respectively, T = R 24 /A 12 2 ). The group of linear isometries of (Λ ∨ , I) is the Mathieu group M 24 or M 12 , and we have I ∈ H 4 (BT, Z) M 24 respectively I ∈ H 4 (BT, Z) M 12 .
Conway, Mathieu and Weyl 2-groups
To return to our remarks in Section 4.2, the symmetries of categorical tori form 2-group extensions of Weyl and Mathieu groups as well as Conway's group Co 0 . The low dimensional (co)homology of the Mathieu groups was calculated in [DSE09], the generator of the cohomology group 6. Other things classified by (Λ ∨ , I) Besides our categorical tori, the degree 4 cohomology class classifies a number of important objects. In this section, we list a few of these and offer some comments on their relationship to T .
From categorical groups to loop groups
I classifies a central extension LT of the loop group of T by U (1). If T is a maximal torus of G and I is invariant under the Weyl group, as in Section 5.1, then I classifies a central extension LG of the loop group of G.
The relationship between loop groups and 2-groups is well studied, see for instance [BSCS07] and, more recently, [Wal13] and [Wal12].
We will work in the context of Waldorf's transgression-regression machine. The term 'transgression' in this context does not refer to the map τ mentioned above, but to a recipe for turning multiplicative bundle gerbes with connection into central extensions of loop groups. Applied to (I, (L J , ∇), α), transgression yields the central extension of LT with trivial underlying principal bundle and 2-cocycle where Hol stands for holonomy, and (f, g) is a lift of (ϕ, γ) to t × t. Proof. Pressley and Segal denote our I by −, − , their b can be taken to be our J, and they write Λ for the cocharacter latticeŤ = Hom(U (1), T ). We first note that c describes the correct extension of this lattice namely c Λ ∨ (m, n) = (−1) J(m,n) .
The commutators in LT are
(integration by parts). Over the identity component of LT , our extension is completely described by its Lie-algebra cocycle, (5) (6), it sends (g, r) ∈ Lt to Waldorf's regression machine reconstructs a Lie 2-group extension G of G by pt / /U (1) from a central extension LG equipped with an extra piece of data, called fusion product. Let P 1 G be the space of paths in based at 1 in G, and π the evaluation map π: P 1 G −→ G γ −→ γ(1).
Then the arrows in theČech groupoid P 1 G × G P 1 G P 1 G are identified with the space ΩG of based loops in G, and replacing ΩG with the central extension ΩG, one obtains the pt / /U (1) extension of G ΩG). G := (P 1 G The composition of arrows in ΩG is the fusion product mentioned above. Let us study this in detail for our torus. The construction of c in (4) as holonomy implies that c is a fusion map, i.e., for triples (ϕ 1 , ϕ 2 , ϕ 3 ) and (γ 1 , γ 2 , γ 3 ) with identical start and end points, ϕ 1 (0) = ϕ 2 (0) = ϕ 3 (0), and ϕ 1 (1) = ϕ 2 (1) = ϕ 3 (1), and likewise for the γ i . Therefore, we have a canonical fusion product on LT . To compare the regression of LT to Construction 2.1, we consider the commuting diagram Here exp [0,x] To compare this expression with the factor turning up in Construction 2.1, we note that and δ simp exp • J 2 = 0. Let T be the groupoid of Construction 2.1, and let ⊙ be the monoidal structure on T obtained like •, but with the factor exp(−J(m, y)) replaced by c(γ (x,m) γ (y,n) ). Then the map makes the identity functor a monoidal equivalence from (T, •) to (T, ⊙).
Extraspecial 2-groups
Given an even symmetric bilinear form I on Λ ∨ , we have the integer-valued quadratic form on Λ ∨ . The form I can be recovered from φ by the identity i.e., I(m, n) = φ(m + n) − φ(m) − φ(n).
Let T [2] be the F 2 -vectorspace of points of order 2 in T . Then φ induces a quadratic form on which we will also denote φ: T [2] −→ F 2 .
Such a φ classifies an extraspecial 2-group 5 , i.e., a central extension An explicit 2-cocycle for this extension is given by J ⊗ F 2 , for any integer-valued J on Λ ∨ with I = J + J t . In the situation of Section 5.2, there is a prominent subgroup of the Monster, isomorphic to where Co 1 is the Conway group Co 1 = Co 0 /±1. This subgroup is typically the first step in the construction of the Monster, see for instance [Tit85] Not unexpectedly, it turns out that the Looijenga line bundle also plays a key role in the character theory of representation of categorical tori. This leads to a theta function formalism for 2-characters, resembling that of loop group characters, but developed directly from the categorical picture, without any mention of loops. We will come back to this topic at a different occasion. | 5,056.8 | 2014-06-27T00:00:00.000 | [
"Mathematics"
] |
Flexible , compact WDM receivers using cascaded optical and electrical down-conversion
We propose a super-channel flexible wavelength division multiplexing (WDM) receiver architecture. The receiver, which requires no optical filtering, only a pair (I and Q phases) of coherent optical detectors, and an electrical receiver system, can simultaneously recover multiple wavelength-multiplexed channels using cascaded optical and electrical down-conversion. The receiver data capacity increases in proportion to the number of electrical sub-carrier channels. The proposed receiver concept has been described using a six-channel WDM receiver, and a two-channel ( ± 25GHz) receiver IC, which is a key block of the WDM receiver, has been successfully demonstrated with two and three 2.5Gb/s binary-phase-shiftkey (BPSK) modulated channels. ©2013 Optical Society of America OCIS codes: (060.1660) Coherent communications; (250.5300) Photonic integrated circuits; (350.4010) Microwaves. References and links 1. Infinera, “Super-channels: DWDM Transmission at 100Gb/s and Beyond” WP-SC-10–2012 (2013). 2. B. Zhang, C. Malouin, and T. J. Schmidt, “Towards full band colorless reception with coherent balanced receivers,” Opt. Express 20(9), 10339–10352 (2012). 3. J. Renaudier, O. Bertran-Pardo, H. Mardoyan, P. Tran, G. Charlet, S. Bigo, A. Konczykowska, J.-Y. Dupuy, F. Jorge, M. Riet, and J. Godin, “Spectrally efficient long-haul transmission of 22-Tb/s using 40-Gbaud PDM16QAM with coherent detection,” OFC'2012 Conference, OW4C2 (2012). 4. X. Liu, S. Chandrasekhar, P. J. Winzer, T. Lotz, J. Carlson, J. Yang, G. Cheren, and S. Zederbaum, “21.5-Tb/s Guard-Banded Superchannel Transmission over 56x100-km (5600-km) ULAF Using 30-Gbaud Pilot-Free OFDM-16QAM Signals with 5.75-b/s/Hz Net Spectral Efficiency,” in Proceedings of ECOC’2012, post deadline Th3C5 (2012). 5. T. Zami, “What is the Benefit of Elastic Superchannel for WDM Network?” in Proceedings of ECOC’2013, 2226 Sept. 2013 (2013). 6. O. Gerstel, M. Jinno, A. Lord, and S. Yoo, “Elastic optical networking: a new dawn for the optical layer?” IEEE Commun. Mag. 50(2), s12–s20 (2012). 7. J. Renaudier, R. Rios-Muller, L. Schmalen, M. Salsi, P. Tran, G. Charlet, and S. Bigo, “1-Tb/s Transceiver Spanning Over Just Three 50-GHz Frequency Slots for Long-Haul Systems,” in Proceedings of ECOC’2013, post deadline PD2D5 (2013). 8. H. Ito, T. Furuta, S. Kodama, and T. Ishibashi, “InP/InGaAs uni-travelling-carrier photodiode with 310 GHz bandwidth,” Electron. Lett. 36(21), 1809–1810 (2000). 9. M. Urteaga, R. Pierson, P. Rowell, V. Jain, E. Lobisser, and M. J. W. Rodwell, “130nm InP DHBTs with fτ >0.52THz and fmax >1.1THz,” Device Research Conference (DRC), 2011 69th Annual, 281–282 (2011). 10. B. Heinemann, R. Barth, D. Bolze, J. Drews, G. G. Fischer, A. Fox, O. Fursenko, T. Grabolla, U. Haak, D. Knoll, R. Kurps, M. Lisker, S. Marschmeyer, H. Rücker, D. Schmidt, J. Schmidt, M. A. Schubert, B. Tillack, C. Wipf, D. Wolansky, and Y. Yamamoto, “SiGe HBT technology with fτ/fmax of 300GHz/500GHz and 2.0 ps CML gate delay,” Electron Devices Meeting (IEDM), 2010 IEEE international, 30.5.1–30.5.4 (2010). 11. H. Li, B. Jagannathan, J. Wang, T.-C. Su, S. Sweeney, J. J. Pekarik, and Y. Shi, “Technology Scaling and Device Design for 350 GHz RF Performance in a 45nm Bulk CMOS Process,” VLSI Technology, 2007 IEEE Symposium on, 56–57 (2007). 12. J. B. Hacker, M. Urteaga, M. Seo, A. Skalare, and R. H. Lin, “InP HBT Amplifier MMICs Operating to 0.67 THz,” 2013 IEEE International Microwave Symposium, (2013). 13. H. Park, M. Piels, E. Bloch, M. Lu, A. Sivananthan, Z. Griffith, L. Johansson, J. Bowers, L. Coldren, and M. Rodwell, “Integrated Circuits for Wavelength Division De-multiplexing in the Electrical Domain,” in Proceedings of ECOC’2013, Mo4C3 (2013). 14. E. Bloch, H. Park, M. Lu, T. Reed, Z. Griffith, L. A. Johansson, L. A. Coldren, D. Ritter, and M. J. W. Rodwell, “A 1-20 GHz All-Digital InP HBT Optical Wavelength Synthesis IC,” IEEE Trans. Microw. Theory Tech. 61(1), 570–580 (2013). 15. Z. He, W. Wu, J. Chen, Y. Li, D. Stackenas, and H. Zirath, “An FPGA-based 5 Gbit/s D-QPSK modem for Eband point-to-point radios,” 2011 European Microwave Conference, 690–692 (2011) 16. M. Urteaga, R. Pierson, P. Rowell, M. Choe, D. Mensa, and B. Brar, “Advanced InP DHBT process for high speed LSI circuits,” IPRM 2008. 20th International conference, 1–5 (2008). 17. E. Bloch, H.-C. Park, Z. Griffith, M. Urteaga, D. Ritter, and M. J. W. Rodwell, “A 107 GHz 55 dBΩ InP broadband transimpedance amplifier IC for high-speed optical communication links,” in Proceedings of IEEE CSICS, 2013 (2013), pp. 1–4.
Introduction
High spectral efficiency and high channel capacity wavelength division multiplexing (WDM) systems such as Tb/s super-channels [1][2][3][4] are needed to meet future demands for high data rate transmission.For these systems, simple hardware and low DC power consumption are both desirable.Moreover, WDM systems with elastic and flexible channel allocations have been proposed to aid in adapting to variations in network traffic and conditions [5,6].
A conventional coherent WDM receiver consists of a fixed-or flexible-grid optical demultiplexing filter and an array of coherent receivers, as is shown in Fig. 1(a) [1][2][3][4].The photonic IC (PIC) is complex, containing the optical filter and an array of coherent receivers, each coherent receiver containing a local oscillator (LO) laser, a 90 degree optical hybrid, and a pair of balanced photodiodes (PDs).Each pair of PICs and electronics requires its own highspeed electrical interface [1,7].The large number of components and interfaces can increase manufacturing costs and decrease manufacturing yield.The number of PIC components can be reduced by high-speed and high-resolution analog-digital-converters (ADCs) and digital- signal-processors (DSPs), but this imposes a large power penalty and comes at the cost of reduced optical reach [7].
In this work, we propose and demonstrate a new coherent WDM receiver architecture (Fig. 1(b)) which can scale toward Tb/s operation while using only a pair (I and Q phases) of optical detectors to recover multiple wavelength channels.The proposed receiver has a simple circuit configuration, can provide high spectral efficiency, and accommodates a flexible range of channel spacing and of data modulation formats.The complexity of the required PIC is greatly reduced, removing the optical filter and leaving only one set of balanced detectors and an LO laser.Most signal processing is instead performed in electrical integrated circuits (EICs).Sets of optical WDM channels are recovered in the EIC using single-sideband (SSB) mixers and their associated LOs as frequency down-converters.The EIC requires only a single broadband trans-impedance amplifier (TIA).
The channel capacity in the proposed receiver is ultimately limited by the bandwidth of the photodiodes (PDs) and of the de-multiplexing EIC.Very wide bandwidths are now feasible for these components.Uni-travelling carrier (UTC) PDs with 3-dB bandwidth >300GHz [8] and high-speed InP hetero-junction bipolar transistors (HBTs) having powergain cutoff frequencies f max >1THz [9] have both been reported.Even silicon-based transistors have attained 500 GHz f max [10,11].Amplifier EICs operating above 670GHz have been reported [12].Given such components, the proposed receiver architecture can scale beyond 1Tb/s capacity using >125GHz bandwidth PDs and EICs together with dual optical polarizations and 16-QAM modulation.Using 300 GHz photodiodes [8] and transistor amplifiers [12], a single EIC receiver could process 24 WDM channels at 25 GHz channel spacing.Further, in the proposed receiver, the optical channel wavelength allocation can be adjusted dynamically by tuning LO frequencies within the electrical frequency conversion circuits.
In this paper, an extension of [13], we first describe (section 2) the coherent multi-channel WDM receiver concept using cascaded optical and electrical down-conversions.We then describe (section 3) the component IC blocks of a two-channel ( ± 25GHz) implementation of the WDM receiver IC.Finally (section 4) we describe demonstrations of the two-channel receiver IC, including measurements of image rejection and of adjacent channel interference.
A single-chip multi-channel WDM receiver concept
Figure 2 shows a more detailed block diagram of the WDM receiver using, as an illustration, a six-channel configuration with 25GHz channel spacing.The receiver can de-multiplex these six channels simultaneously using a single PIC with only a pair (I and Q) of detection channels and a single EIC.The PIC is a standard coherent balanced detector.The EIC includes a broadband TIA, clock and RF signal distribution networks, and broadband analog SSB mixers.Use of single-sideband (SSB) mixers permits the receiver to recover, without crosstalk, WDM channels lying both above and below the frequency of the optical LO.This SSB mixing therefore doubles number of the optical WDM channels which can be recovered within a given EIC and PD analog bandwidth.De-multiplexing of the six WDM carriers proceeds as follows: 1) An optical hybrid combines the six modulated WDM channels with a relatively strong optical LO (f 0 ) mixing these together on balanced photodiodes.In this optical downconversion, the WDM channels become six electrical RF sub-carriers at 25 GHz spacing ( ± 12.5 GHz, ± 37.5GHz, and ± 62.5GHz).
2) Each RF sub-carrier is then down-converted to a baseband using a SSB mixer.This recovers both the I and Q components of the optical modulation.Independent SSB mixers independently recover the ( ± 12.5 GHz, ± 37.5GHz, and ± 62.5GHz) subcarriers.
3) Low-pass filters in the I and Q outputs suppress interference from adjacent channels.Data can then be recovered through standard ADC and DSP.Note that if the optical wavelength separations in the WDM transmitter are set by optical wavelength synthesis [14], then the WDM channel frequency separations are determined by the precision of the microwave synthesizers used to set the optical frequency offsets.In this case, the SSB mixers can use stable but asynchronous local oscillators with the resulting LO phase and frequency errors corrected by electrical differential phase-shift-keying demodulation [15].
Designs for two-channel ( ± 25GHz) receiver IC
To demonstrate the proposed WDM receiver concept, a two-channel receiver IC, a key block of the WDM receiver, was designed using Teledyne Science Company's 500nm HBT process, using transistors with 300GHz f τ and f max [16], and tested using two and three wavelength-modulated channels.Figure 3 The PIC I' and Q' output signals are amplified with DC-107GHz broadband amplifiers [17] at the EIC inputs.In the present demonstration, an electrical LO is generated off-chip; internal to the EIC, polyphase filters and limiting amplifiers generate the quadrature LO signals needed for the Weaver-configuration analog SSB mixers.RF signals within the EIC are routed using 50Ω transmission-lines and buffer amplifiers.The present ICs were designed without attention to DC power consumption; power consumption of future versions of the IC could be greatly decreased by minimizing the number of such internal 50Ω interconnects.The recovered I and Q components of each down-converted carrier are available at the EIC outputs; these I and Q signals are low pass filtered and amplified off-chip.Since the full amplitude and phase information of the input WDM carrier is preserved during downconversion, the receiver is in principle compatible with any higher-order modulation formats (e.g.QPSK or 16-QAM).Using this receiver EIC, the two modulated channels with 50GHz channel spacing can be de-modulated simultaneously on f 0 ± 25GHz optical channels.
System demonstrations and experimental results
To demonstrate the WDM receiver system, using the two channel receiver IC (Fig. 3) we have measured image rejection for the two modulated carriers at f 0 ± 25GHz, and thus ± 25GHz electrical sub-carriers, and have measured adjacent channel interference using three modulated carriers with channel spacing varying from 5GHz to 20GHz.
Image rejection experiment
To double the number of WDM channel processed, the SSB mixers were used in the receiver IC, independently recovering signals both below and above the optical LO frequency.Figure 4 shows a test configuration of the image rejection experiment.Three free-running distributed feedback (DFB) lasers, which have a narrow linewidth and low relative intensity noise (RIN), were used: L 1 as a transmitter carrier laser #1 at optical carrier frequency f 0 + 25GHz, L 2 as a transmitter carrier laser #2 at frequency f 0 -25GHz, and L 3 as the LO laser at frequency f 0 .The carrier lasers L 1 and L 2 were modulated using 2.5Gb/s binary-phase-shift-keying (BPSK) data with two pseudo-random-binary-sequence (PRBS) pattern generators and Mach-Zehnder modulators (MZMs).PRBS of lengths 2 31 -1 and 2 15 -1 were applied to lasers L 1 and L 2 in this test.The modulated channels are combined in a 50%/50% optical directional coupler and coupled to the optical signal port of the receiver, with the LO laser L 3 connected to the receiver LO port.
In this experiment, we used a free-space 90° optical hybrid and balanced PDs contained in an Agilent optical modulation analyzer (OMA) N4391A.The balanced PDs have >33GHz 3-dB-bandwidth.The relative delays of the I'-channel and Q'-channel cables between the OMA and the two-channel receiver IC were carefully matched within 50ps by adjusting RF cable lengths and using phase shifters (PSs).This delay mismatch limited the bit rate to no more than 2.5Gb/s, but this mismatch issue can be solved by the hybrid integration of PIC and EIC.The EIC was contacted using multi-finger GGB RF probes.The two-channel receiver IC has I and Q outputs for both the ± 25GHz channels.One pair of I ch+ and Q ch+ outputs was low-pass filtered, amplified, and stored in an Agilent real-time oscilloscope (DSA-X 92004A).Frequency and phase errors were corrected by DSP (in this work, we used a simple phase error estimation technique using a Matlab code), thereby recovering the I-Q transmitted data Fig. 4. Test setup for image rejection using two wavelength channels.streams, as shown in Figs.5(a)-5(d).The second pair of I ch-and Q ch-outputs was connected to an electrical spectrum analyzer (ESA) to monitor the suppression of adjacent-channel interference in the frequency domain.Figure 5(e) shows the spectra measured from the EIC output, comparing the cases when the lasers L 1 and L 2 are both active.
Figures 5(a) and 5(b) show the recovered data outputs on ( + ) and (-) channels using a single 2.5Gb/s BPSK modulated carrier (L 1 is on and L 2 is off) without the 50%/50% directional coupler.After the post signal processing, Fig. 5
Adjacent channel interference experiment
To measure receiver crosstalk from adjacent channels and to determine the maximum spectral efficiency given the minimum channel grid spacing, three-channel experiments were performed using the test configuration shown in Fig. 6.Three DFB lasers and a Koshin (LS-601A) tunable laser, which also have a narrow linewidth and low RIN, were used: L 1 as a transmit carrier laser #1 at an optical frequency variable from f 0 + 5GHz to f 0 + 20GHz, L 2 as a fixed transmit carrier laser #2 at an optical frequency f 0 + 25GHz, L 3 as a transmit carrier laser #3 at an optical frequency variable from f 0 + 30GHz to f 0 + 45GHz, and L 4 as a fixed LO laser at frequency f 0 .Again, three channels of 2.5Gb/s BPSK data were modulated on the carrier lasers of L 1 , L 2 and L 3 using differential outputs of a PRBS 2 31 -1 pattern generator and two MZMs as shown in Fig. 6.The two patterns were de-correlated using a path length difference of about one meter.The three modulated channels are combined using 50%/50% directional couplers and the power of each channel is equalized and monitored through an APEX high-resolution optical spectrum analyzer (OSA).Figure 7 shows the measured spectrum of the three optical modulated channels with channel grids of (a) 20GHz, (b) 10GHz and (c) 5GHz, the final case having no frequency guard band.
In this experiment, the qualities of the recovered data were measured and compared in Fig. 8 for the different channel spacing and differing combinations of input and output low-pass filter bandwidths.Here filter #1 is low pass filters (LPFs) located before the optical modulators (MZMs) to suppress spectral side lobes in the transmitted optical spectrum, while filter2 is LPFs located at the two-channel receiver IC outputs to suppress adjacent channel interference.With 2.2GHz LPFs in the transmitter and 3GHz LPFs in the receiver, even with a 5GHz channel spacing, corresponding to no frequency guard band, a signal-to-noise including interference ratio (Q-factor), extracted from eye diagram outputs, of approximately 10:1 is measured.Note that Q = 6 corresponds to 10 −9 bit error rate given Gaussian statistics.
As Fig. 8 demonstrates, the receiver IC operates with high signal/crosstalk ratios even when channels carrying 2.5Gsymbol/second modulation are spaced in frequency by only 5 GHz; the proposed WDM receiver system provides high spectral efficiency.
Conclusion
We have demonstrated WDM receiver using cascaded optical and electrical down-conversion to recover multiple optical wavelength channels.Given the demonstrated bandwidths of modern photodiodes and EICs, the receiver can readily scale to a large number of WDM channels using only a single (I and Q) pair of optical coherent detection channels.Initial demonstrations indicate high image rejection and low adjacent channel interference.
Fig. 1 .
Fig. 1.Two different WDM receiver concepts: (a) conventional WDM receiver using optical filters, and (b) a new proposed WDM receiver.
Fig. 2 .
Fig. 2. A concept schematic diagram for a coherent single-chip multi-channel WDM receiver and its de-multiplexing flows for six modulated channels.
(a) shows the schematic of the 2-channel receiver EIC, while Fig.3(b)shows a die photograph.The six-channel receiver of Fig.2was also designed and fabricated; to date, six-channel optical system experiments have not yet been completed, and we describe only results with the 2-channel unit.
Fig. 3 .
Fig. 3. (a) An EIC schematic for a two-channel ( ± 25GHz) receiver IC, and (b) a photograph of the fabricated and mounted EIC on the test AlN board.
Fig. 5 .
Fig. 5. Experimental results for image rejection measurement: (a)-(b) the eye diagrams for the activated ( + ) channel and the suppressed (-) channel with a single modulated carrier, (c)-(d) the eye diagrams for the ( + ) and (-) channels with two modulated carriers, and (e) the measured output spectra when the signal and the adjacent (crosstalk) channels are active.Crosstalk suppression is ~25dB.
Figure 5 (
e) indicates that there is about 25dB (18:1 in voltage) image rejection ratio between the activated and suppressed channels.This is typical of that expected for welldesigned SSB mixers.Regarding two 2.5Gb/s BPSK modulated carriers (L 1 is on and L 2 is on), both ( + ) and (-) channels show open eye diagrams, as shown in Figs.5(c) and 5(d).The slightly degraded eye diagrams in this test are most likely because the input power to the twochannel receiver IC is 2:1 lower than that in the experiment involving only a single modulated carrier.
Fig. 8 .
Fig. 8. Measured eye diagram qualities for the different channel spacing and filter combinations (filter #1 -before the optical modulators, and filter #2 -after the EIC). | 4,239.2 | 2014-01-13T00:00:00.000 | [
"Physics"
] |
Biomimetic Self-Assembled Chiral Inorganic Nanomaterials: A New Strategy for Solving Medical Problems
The rapid expansion of the study of chiral inorganic structures has led to the extension of the functional boundaries of inorganic materials. Nature-inspired self-assembled chiral inorganic structures exhibit diverse morphologies due to their high assembly efficiency and controlled assembly process, and they exhibit superior inherent properties such as mechanical properties, chiral optical activity, and chiral fluorescence. Although chiral self-assembled inorganic structures are becoming more mature in chiral catalysis and chiral optical regulation, biomedical research is still in its infancy. In this paper, various forms of chiral self-assembled inorganic structures are summarized, which provides a structural starting point for various applications of chiral self-assembly inorganic structures in biomedical fields. Based on the few existing research statuses and mechanism discussions on the chiral self-assembled materials-mediated regulation of cell behavior, molecular probes, and tumor therapy, this paper provides guidance for future chiral self-assembled structures to solve the same or similar medical problems. In the field of chiral photonics, chiral self-assembled structures exhibit a chirality-induced selection effect, while selectivity is exhibited by chiral isomers in the medical field. It is worth considering whether there is some correspondence or juxtaposition between these phenomena. Future chiral self-assembled structures in medicine will focus on the precise treatment of tumors, induction of soft and hard tissue regeneration, explanation of the biochemical mechanisms and processes of its medical effects, and improvement of related theories.
Introduction
Chirality is an intrinsic property of the natural world, ranging from that which characterizes the DNA double helix at the nanoscale to that which characterizes the macroscopic spiral arms of galaxies [1]. In the past two decades, our understanding of chirality has rapidly evolved from encompassing only simple chiral organic compounds to encompassing the spontaneous organization of metal or ceramic nanoparticles induced by chiral organic compounds as templates [2]. The wide array of research systems and dimensions has greatly expanded the functional categories of inorganic particles [3]. Excluding the fields of chiral optics and chiral catalysis, the functionality of chiral structures has been fully exploited and utilized to replace applications of traditional materials. In the medical field, the influence of chiral assemblies is gradually increasing [4,5]. Crucial trace metals in chiral assemblies show unusual special properties that contribute to their role in physiological processes [6,7]. There exists the need to define the mechanisms of action of these chiral structures and the metals themselves that are present on both cells and microscopic matrix microstructures. Characterizing these mechanisms will require determining both the physiological activity of the metal itself and the induced selection effect of chiral polarization [8,9]. It should be noted that the self-assembly process is crucial for the formation and regulation of chiral structures, and the use of a system of electrostatic force or other non-covalent bonding forces to achieve spontaneous organization from the nanoscale to the microscale is a typical biomimetic process [10,11]. The process is equivalent to the self-planning and organization of the formation of microstructures that is inherent in all physiological activities in nature [12][13][14][15]. Self-assembled nanostructures in nature often have a mechanical efficiency exceeding 90%, whereas self-assembly by artificial intervention, while less efficient, can be harnessed to better control the self-assembled structures and their accuracy, while their corresponding energy efficiency is also much higher than that of non-self-assembled processes [16]. The reported diversity of chiral self-assembly structures-polyhedron [17], tube [18], petal [19], flower [20], helix [21], rod [22] and hedgehog chiral particles [23]with multiple levels of a complex structure (Figure 1a-c), is the basis for their significant functional advantages across multiple domains. Such advantages capitalize on the effects of chiral self-assembly, which impact basic cellular behaviors, such as cell adhesion and proliferation, the differentiation-induction of stem cells, and the regulation of tumor cell fates [24,25]. Various forms of chiral self-assembly, including simple structures and the complex assembly of topologies, have shown surprising potential in the field of biomedical science [26]. Here, we review these chiral self-assembled structures at different scales and morphologies as obtained by diversified synthetic methods, and we outline their distinctive synthetic approaches. We focus on the application of chiral self-assembled structures in the medical field, especially the influence of chiral self-assembled structures on the basic behaviors of cells, including cell adhesion and proliferation, the absorption of surface proteins, and osteogenic differentiation. Meanwhile, we also summarize the progress in their applications for molecular probes, such as enantiomeric amino acid and reactive oxygen species (ROS) recognition and tumor therapies through inducing the autophagy and apoptosis of tumor cells and the photothermal effect [27,28]. This review also looks forward to the existing boundaries limiting their future applications and the challenges posed by using chiral self-assembled structures in the biomedical science field. Synthesis and morphology of the outspread petal-like structure (reproduced with copyright permission from [19], American Chemical Society). (c) SEM images of Au-L-Cys and Au-D-Cys coccolith-like particles and Au-DL-Cys kayak particles with low magnification, and enlarged SEM images of Au-L-Cys, Au-D-Cys, and Au-DL-Cys kayak particles (reproduced with copyright permission from [23], The American Association for the Advancement of Science).
Synthesis of Biomimetic Self-Assembled Chiral Structures
Organic small molecules, inorganic particles, and nanoparticles can form complex and orderly structures through self-assembly. Inspired by this micromechanical behavior in nature, artificially inducing the spontaneous organization of microcomponents by regulating the interaction of particles in chemical or physical reactions represents an effective synthesis method to mimic nature's self-assembling behavior [29,30]. Recent research on chiral self-assembly has focused on inorganic nanoparticles and their assembly bodies, whereby chiral building blocks such as DNA, amino acids, lipids, and sugars play an important role in inducing metal ion assembly [31][32][33]. Typical chiral self-assembly structures are ubiquitous from the nanoscale to the micron scale. While typical nanoparticles have symmetrical structures (such as spheres and cubes), asymmetric defects, vertices, and voids cause nanoparticles to exhibit chiral geometry, which is the chiral nature of nanoparticles. For example, screw dislocation-mediated growth is responsible for the
Synthesis of Biomimetic Self-Assembled Chiral Structures
Organic small molecules, inorganic particles, and nanoparticles can form complex and orderly structures through self-assembly. Inspired by this micromechanical behavior in nature, artificially inducing the spontaneous organization of microcomponents by regulating the interaction of particles in chemical or physical reactions represents an effective synthesis method to mimic nature's self-assembling behavior [29,30]. Recent research on chiral self-assembly has focused on inorganic nanoparticles and their assembly bodies, whereby chiral building blocks such as DNA, amino acids, lipids, and sugars play an important role in inducing metal ion assembly [31][32][33]. Typical chiral self-assembly structures are ubiquitous from the nanoscale to the micron scale. While typical nanoparticles have symmetrical structures (such as spheres and cubes), asymmetric defects, vertices, and voids cause nanoparticles to exhibit chiral geometry, which is the chiral nature of nanoparticles. For example, screw dislocation-mediated growth is responsible for the chiral polyhedral shape formation of tellurium nanocrystals grown from solution [34]. Further, it was found that two Au nanorods can form strongly chiral nanoscale systems. Their chiral properties originate from the small dihedral angle between two adjacent nanorods, which breaks the centrosymetric nature of the two parallel identical cylinders [35]. The second type of chiral nanostructure is the induced chirality of chiral ligand-modified inorganic nanoparticles. There have been many studies about the chiral synthesis related to various nanostructures, especially nanorods. For example, the cholic acid-induced anisotropic epitaxial growth of chiral CdSe/CdS nanorods (CCCNs) and their self-assembly into chiral nematic-like films (CNFs) via an evaporation-induced self-assembly route [36] has been reported. Discrete gold nanorods (Au NRs) with strong chiroptical responses in the visible and near-infrared region were synthesized by a seed-mediated approach in the presence of L-or D-cysteine by Zheng et al. [37]. In addition, nanoclusters formed by the specific arrangement of nanoparticles can also form chiral structures. Previous research has reported the 2D self-assembly of ligand-capped Au15 nanoclusters into mono-, few-, and multi-layered sheets in colloidal solution. This 2D self-assembly is caused by the 1D dipolar attraction common in nanometer-sized objects and produces an asymmetric van der Waals attraction [38]. Most existing chiral nanoparticles are formed from chiral transfer reactions in which the metal nanoparticles are modified by chiral ligands. Common methods include the reduction of metal salts and the replacement of metals by chiral complexes. For example, undecagold cluster compounds were synthesized by the chemical reduction of the corresponding precursor complex, Au2X2 (BINAP) [39]. Additionally, ligands such as L-or D-aspartic acid (Asp), L-or D-proline (Pro), and cysteine are used for the transmission of chiral signals [40,41].
Besides, DNA origami is another tactic used to induce chirality, taking advantage of its programmability, flexibility, and high assembly efficiency. This strategy mainly involves binding DNA to AuNPs through the prethiolation of strands [42] or combining it with other metal via an electrostatic force [43]. Researchers have found that self-assembly can be controlled through the integration of gold nanoparticles [44]. Similarly, Kuzyk et al. used DNA origami technology to synthesize a gold nanoparticle helical structure [45]. By designing the 'X' pattern of the arrangement of DNA capturing strands (15 nt) on both sides of a two-dimensional DNA origami template, AuNRs functionalized with the complementary DNA sequences were positioned on the origami and were assembled into AuNR helices, with the origami intercalated between neighboring AuNRs ( Figure 2) [46]. Other structures were also reported, including a DNA pyramid containing four AuNPs linked with DNA stands that can produce a stronger CD response compared to chiral metal clusters or NPs [47]. DNA-bound NPs offer another platform to create new conjunction sites. Lan et al. successfully combined K21 cyanine dyes with DNA-assembled AuNPs, which also presented a stronger CD signal [48]. The basic chiral components in chiral structures beyond the nanoscale can accurately transmit asymmetric polarization into topologies due to the action of structural guides, and the corresponding isomeric self-assemblers can accurately form structures with mirror symmetry under the same synthesis conditions. Unlike chiral nanoparticle formation, the formation of chiral topological self-assemblies involves not only the expansion of ligand-directed chiral metal nanoparticles, but also the amplification of chiral signals [53,54]. This must capture the balance between the self-sealing end of the structure and its selflimiting point [55]. The differentiated structures of chiral self-assemblies are characterized by their chiral ligands, solvents, composition, synthesis temperature, and growth time [56][57][58]. Both left-handed and right-handed chiral cysteines can self-assemble with stable CdTe to produce chiral mesoscale helices. These helical structures contain multi-level features ranging from the nanometer scale to the micrometer scale, and the termination of the length and diameter of the helix demonstrates the self-limiting nature of the structure, which is related to the directional attachment of inorganic nanoparticles [59]. On this basis, Jiao et al. obtained a chiral helix structure through the self-assembly of semiconductor nanoparticles and then studied the resulting structural changes of the chiral helix by adjusting the solvent composition, pH, and ligand density. The modulation of the coordination and hydrogen bonds enabled controllable chirality deviation and greatly improved assembly yields [60]. The efficient assembly of chiral penicillamine and ZnS nanoparticles may have formed primary chiral supraparticles, which were then further co-assembled by glutathione-modified gold nanoparticles and ZnS supraparticles, thereby finally forming two types of composite assemblies of ZnS-Au supraparticles [61]. Jiang et al. created a hierarchically organized chiral biomineral structure of calcium carbonate whose chirality could be switched by a single L-enantiomer of an amino acid, resulting in a chiral biomineral structure that closely resembled the pathological inner ear stone [62]. Ribbons of stacked, board-shaped cadmium selenide (CdSe) nanoplatelets (NPLs) were twisted upon It is also feasible to obtain chiral metal nanoparticles by treating chiral organometallic precursors with light, sound, and magnetism. Srivastava et al. produced cadmium telluride (CdTe) nanoparticles to form both left and right chiral helical structures using illumination with visible light, while Xu et al. irradiated precursor particles with circularly polarized light (CPL) and linearly polarized light [49,50]. Liquid exfoliation by ultrasonic treatment and modulation by magnetic fields have also been used to treat chiral organometallic precursors [51,52].
The basic chiral components in chiral structures beyond the nanoscale can accurately transmit asymmetric polarization into topologies due to the action of structural guides, and the corresponding isomeric self-assemblers can accurately form structures with mirror symmetry under the same synthesis conditions. Unlike chiral nanoparticle formation, the formation of chiral topological self-assemblies involves not only the expansion of liganddirected chiral metal nanoparticles, but also the amplification of chiral signals [53,54]. This must capture the balance between the self-sealing end of the structure and its self-limiting point [55]. The differentiated structures of chiral self-assemblies are characterized by their chiral ligands, solvents, composition, synthesis temperature, and growth time [56][57][58]. Both left-handed and right-handed chiral cysteines can self-assemble with stable CdTe to produce chiral mesoscale helices. These helical structures contain multi-level features ranging from the nanometer scale to the micrometer scale, and the termination of the length and diameter of the helix demonstrates the self-limiting nature of the structure, which is related to the directional attachment of inorganic nanoparticles [59]. On this basis, Jiao et al. obtained a chiral helix structure through the self-assembly of semiconductor nanoparticles and then studied the resulting structural changes of the chiral helix by adjusting the solvent composition, pH, and ligand density. The modulation of the coordination and hydrogen bonds enabled controllable chirality deviation and greatly improved assembly yields [60]. The efficient assembly of chiral penicillamine and ZnS nanoparticles may have formed primary chiral supraparticles, which were then further co-assembled by glutathione-modified gold nanoparticles and ZnS supraparticles, thereby finally forming two types of composite assemblies of ZnS-Au supraparticles [61]. Jiang et al. created a hierarchically organized chiral biomineral structure of calcium carbonate whose chirality could be switched by a single L-enantiomer of an amino acid, resulting in a chiral biomineral structure that closely resembled the pathological inner ear stone [62]. Ribbons of stacked, board-shaped cadmium selenide (CdSe) nanoplatelets (NPLs) were twisted upon the addition of an oleic acid ligand, leading to chiral ribbons that reached several micrometers in length and displayed a well-defined pitch of~400 nm [63]. Singh et al. found that under carefully controlled conditions, cubic nanocrystals of magnetite self-assembled into arrays of helical superstructures in a template-free manner. Depending on the density of the nanocrystals, the researchers identified different types of self-assembled superstructures, including one-dimensional belts and single, double, and triple helices [64]. More complex chiral self-assembled structures known as hedgehog-like chiral superstructures were also successfully synthesized and the perfect symmetry of the enantiomers was maintained at the micrometer scale [65]. The core mechanism underlying both the structure formation and stabilization stems from the electrostatic repulsion and elastic confinement. In addition, Kotov and his colleagues developed graph-theoretic methods to assess the complexity of chiral self-assemblies and provided guidelines for evaluating the intrinsic correlation between the optical asymmetry factors of chiral structures and their particle complexity [66]. In conclusion, the abundance of the synthesis methods used to produce chiral self-assembled structures has paved the way for the synthesis of diverse chiral structures and their subsequent applications in biomedical fields.
Medical Applications of Biomimetic Chiral Self-Assembled Particles
Most of the physiological functions of living organisms are highly dependent on chiral assemblies (such as DNA and RNA) rather than on monochiral biomolecules (such as nucleotides) [67][68][69]. Increasingly more studies have shown that films and nanoparticles modified with enantiomers play important regulatory roles in biological processes, such as cell adhesion, differentiation, protein and DNA adsorption, cytotoxicity, and genotoxicity, among other biological processes [70][71][72].
Effects on Cell Adhesion, Proliferation, and Differentiation
Numerous studies have shown that cells can interact with chiral surface molecules and exhibit different effects on enantiomers. Zhao et al. designed monolayer chiral Au nanoparticle (NP) films modified with L-and D-penicillamine (Pen), respectively (Figure 3a) [73]. The L-Pen-NP films could promote cell adhesion and accelerate cell proliferation compared with the D-Pen-NP films. Meanwhile, there was a greater degree of differentiation of the cells grown on the surface of the L-Pen-NP films. L/D PEFG (phenylalanine-based gelators) self-assemble to form cross-scale nanofibrous hydrogels with enantiomers of molecular chirality and supramolecular chirality [25]. Through studying the effects on the behaviors of three different cells, Dou found that the shift away from an unordered to ordered assembly and the increase in roughness with left-handed chiral aggregates led to increased cell proliferation, while right-handed chiral aggregates caused the opposite outcome, which indicates that left-handed self-assemblies play a key role in cell adhesion and proliferation. Through the bottom-up process of self-assembly, molecular chirality can be amplified to supramolecular chirality, thereby regulating cell behaviors with finer control, and this control is thought to be coordinated by the adsorption of proteins on the surface of the material. In addition to its effects on cell adhesion, proliferation, and protein adsorption, the L-chiral self-assembled structure with molecular chirality, as well as nanofibers with supramolecular chirality, can promote the expression of osteogenic proteins on the surfaces of dental pulp stem cells, which shows a positive effect on their differentiation. However, Dong et al. found that asymmetry in the rotation direction of the cytoskeleton caused by external chiral geometry can cause different levels of contractility in human mesenchymal stem cells, especially the right-handed geometry, which performs significant roles in cell proliferation, stem cell maintenance, differentiation, and migration [74]. Two different shapes of AuNPs, i.e., gold nanocubes (AuNCs) and gold nanooctahedras (AuNOs), were synthesized using L/D-valine as the chiral center (Figure 3b) [75]. Corresponding studies have shown that both the shape and chirality of nanoparticles can affect cellular uptake and migration, and cell diffusion is also chirality-associated, while cytotoxicity was found to be mainly affected by the amounts of ingested nanoparticles and was positively correlated with ROS (reactive oxygen species) levels.
Molecular Probes
Enantiomerically pure molecules play an important role in medicine [76,77], pharmacology [78,79], and chemical engineering [80][81][82][83]. In this context, the identification, detection, and separation of chiral compounds have long been the subjects of research. Previously, the chiral identification of enantiomers mainly relied on chromatographic separation [84][85][86], but the detection conditions involved were harsh and the process was time-consuming [87]. Fortunately, enantioselective fluorescent identification is a new method for chiral identification that features high sensitivity and a simple operation [88,89]. At present, a series of fluorescent probes composed of nanomaterials and supramolecules are widely used in chiral recognition. Valmik et al. synthesized a novel solid luminescent material with high thermal stability, photostability, and color-tunable properties by fusing the 2-benzoylfuran component with an aromatic unit with solid-state luminescence properties [90]. One of the forms features many strengths: it had good biocompatibility, a high fluorescence quantum yield, a large Stokes shift, was found to be well-internalized and uniformly dispersed in the cytoplasm of MDA-MB-231 cancer cells, showed high fluorescence intensity, and had potential as a chiral fluorescent probe. Jiang et al. proposed an electrochemical molecular probe based on penicillamine-modified small gold nanoparticles for the chiral recognition of 3,4-dihydroxyphenylalanine, and such enantioselectivity can be improved by adjusting the size of the penicillamine-gold nanoparticles [91]. Nanocones formed by Au-Cu9S5 nanoparticles, Ag2S, and upconverting nanoparticles can be used for the ultrasensitive quantitative detection of microRNAs in living cells, as well as their corresponding bioimaging in vivo [92]. Du et al. synthesized a BINOL-based chiral aldehyde containing hydrophilic polyethyleneglycol (PEG) groups, which can be used as a fluorescent probe for amino acids in aqueous solutions due to its highly enantioselective fluorescence effect on various amino acids after being combined with zinc ions [93].
Reactive oxygen species (ROS) are continuously produced by the various metabolic activities of the body and play a vital role as important signaling molecules at low concentrations [94,95]. However, the oxidation of biomolecules like proteins, lipids, and nucleic acids can severely damage cells [96,97], and high levels of ROS may cause the loss of mitochondrial cardiolipin and lipid peroxidation [98], inducing autophagy. By regulating beclin1 and Atg7, two proteins known to regulate autophagy, H 2 O 2 can mediate autophagy [99,100]. Another pathway that induces autophagy is the activation of AMPK, which inhibits the activity of mTOR [101][102][103][104]. Chen et al. synthesized L/D-Co(OH) 2 NPs and modified them with fluorescent molecules (Alexa Fluor 568 (AF568)) to achieve the quantitative detection of ROS levels in vivo under the dual signal of CD and MRI [105]. The study found that the D-Co(OH) 2 NPs performed with a lower cytotoxicity and higher cellular uptake compared to the L-Co(OH) 2 NPs, and because of the denser distribution of the D-Co(OH) 2 NPs in the cellular microenvironment, they also had a stronger ROS detection ability at low concentrations.
Tumor Treatment
Many self-assembled chiral structures are used in anti-tumor therapies, whereby their primary mechanisms of action are through inducing the autophagy and apoptosis of tumor cells. Some anti-tumor effects can also be achieved through the photothermal effect. L/Dcysteine-modified Cu2-xS nanocrystals (prepared by sacrificial template-ligand exchange) can induce the production of a large number of reactive oxygen species in tumor cells, thereby inducing autophagy to achieve anti-tumor effects, and autophagy induced by D-Cu2-xS, in particular, is extensive [106]. This nanostructure may also have a killing effect on tumor cells via photothermal action. By preparing single crystals of chiral tetranucleate copper(II)-based complexes (TNCu-A), Hou et al. found that these could induce apoptosis in human triple-negative breast cancer MDA-MB-231 cells in a mitochondria-dependent manner in vitro [107]. This effect was achieved by inhibiting the anti-apoptotic protein Bcl-2 while upregulating the expression of the pro-apoptotic proteins caspase-9 and Bax. Furthermore, these could inhibit tumor angiogenesis, thereby inhibiting tumor development and progression. Chiral folic acid (FA)-conjugated CdTe/CdS quantum dots could be specifically recognized and taken up by breast cancer cells to subsequently induce the release of the apoptotic factor caspase-9 into the cytoplasm through mitochondrial membrane potential depolarization, and then trigger a cascade reaction to activate caspase-3, which further promotes cellular apoptosis (Figure 3c) [108]. In addition, FA-Cys-CdTe/CdS QDs can selectively induce cancer cell apoptosis by disrupting the α-helical structures in cancer cell proteins. In addition to killing tumors via the mechanism involving chiral self-assembly and the chiral effect, the photothermal effect has been widely reported to kill tumor cells and inhibit tumor cell metastasis based on noble metal-based nanocomposites. The photothermal effect of chiral self-assembled structures comprises an active area of research. Miao et al. synthesized a chiral molecule-induced molybdenum oxide nanoparticle, applied its photothermal properties, and found it to act as a chirality-dependent visible and near-infrared (NIR) photosensitizer for HeLa cell photothermal therapy (PTT) in an in vitro setting [109]. Circularly polarized NIR light radiation also exerts a significant effect on tumor ablation in vivo, and this tumor ablation effect may be closely related to tumor cell apoptosis. These results provide relevant strategies for the application of chiral optics and chiral optogenetics in clinical medicine. Chiral molybdenum (Cys-MoO3-x) nanoparticles (NPs) can be actively taken up by oral squamous cell carcinoma cells and show great cytotoxicity on cell viability under 808 nm laser irradiation [110]. In addition to inducing autophagy and the apoptosis of tumor cells via the photothermal effect as part of tumor therapies, there exist some anti-tumor effects of chiral self-assembled structures that focus on immune-induced responses. For example, Wang et al. obtained prismatic cube-shaped chiral nanoparticles by adding chiral molecules to gold nanoparticles and found that L-NPs could significantly inhibit tumor growth in C57BL6 mice and improve their survival rate [111]. The study also found that the mechanism underlying this phenomenon was that the L-NPs stimulated innate immunity by activating natural killer (NK) and CD8 + T cells and enhanced acquired immunity. In addition to the above effects, the nanoparticles, especially the L-type, also enhanced the cytotoxicity of NK cells and promoted the apoptosis of tumor cells to achieve their anti-tumor therapeutic effects.
Conclusions and Outlook
The array of methods used for the synthesis of multidimensional chiral self-assembled structures and the corresponding research interest for their applications in the biomedical field has grown dramatically [112]. We have reviewed their synthetic pathways including chiral self-assembled nanostructures and superstructures, as well as important advances that have been made in the production of chiral self-assembled structures for the control of cell behavior, fluorescent probes, and tumor therapies. Studies have found that the application of chiral self-assembled structures in medicine primarily depends on chiral selectivity functioning to regulate biological processes. However, many challenges remain to be solved, such as the lack of clarity that still exists regarding the formation mechanism employed by topological chiral self-assembly isomers, as well as the rules and amplification mechanisms underlying chiral signal transmission. In the existing studies, the biological effects of self-assembled structures of different systems (such as the results of tumor treatment) are quite different, and there are certain conflicts in their mechanisms. For example, in some studies, left-handed chiral self-assemblies can effectively inhibit tumors, which is attributed to the induction of autophagy; conversely, one study found a more substantial tumor-killing effect of the right-handed self-assembled structure. It should be noted that some recent chiral photonics studies have presented results indicating the existence of a chirality-induced selection effect of chiral structures on photons [113,114]. This poses the following question: is this effect related to the core mechanism of the underlying phenomenon applied in the biomedical field? In addition, research is currently lacking on the biological effects of chiral structures at larger scales, such as micron-scale chiral self-assembled structures. The in vivo toxicity, immunogenicity, and metabolic pathways of inorganic nanoparticles, especially for chiral self-assembled structures containing chromium and gold, will restrict their clinical applications. To enrich the basic concepts and theoretical models regarding the mechanisms of chiral self-assembly, to clarify the precise biological effects and molecular mechanisms employed by chiral self-assembly structures, and to seek new biologically safe and stable chiral self-assembly structures in vivo will comprise the focus and primary direction of future research.
Institutional Review Board Statement: Not applicable.
Data Availability Statement: Data sharing is not applicable to this article as no new data were created or analyzed in this study.
Conflicts of Interest:
The authors declare no conflict of interest. | 6,017.2 | 2022-10-14T00:00:00.000 | [
"Chemistry",
"Materials Science"
] |
A Multi-view Image Matching Method for Feature Points Based on the Moving Z-plane Constraint
Focusing on the serious occlusion problem in city images, this paper makes full use of the advantage of multi-view image matching, and proposes a reliable multi-view image matching method based on the moving Z-Plane constraint. It supposes a fictitious plane in the object space, and the plane is divided to regular grid cell (small plane element) by a certain interval (image resolution). By moving the plane to different elevation positions, this algorithm makes feature point projection ray in overall images intersect with the plane, and constrains the candidate points by grid cells in the plane. Feature points which come from different images projection ray in the same grid cell on the plane may be regarded as the matching candidates. It selects the images which matching candidate points by gray similarity constraint to avoid the effect from occlusion image. According to the number of projection ray in the grid cell, this algorithm adopts hierarchy matching strategy of " the best candidate will be matched in the first instant " , and uses initial matching results as constraint condition in the latter matching process. Finally, the validity of the algorithm proposed in this paper is verified by the experiments using four UltraCamX (UCX) digital aerial images and the algorithm is shown to have reliable matching results.
INTRODUCTION
Image matching is one of the key technologies to obtain 3D space information from 2D airborne/spaceborne image by digital photogrammetry.Along with the increasing use of new digital sensors, it becomes more and more easier to acquire large overlap digital images covering the same area, the multiview image matching approach has attracted wide interests in both photogrammetry and computer vision.The traditional image matching is restricted to the imaging abilities of stereo sensors, which based on the matching of "single stereo-pair", and therefore is a challenging and "ill-posed" problem.Multiimage matching translates the single-stereo mode matching in image space to the multi-image matching by combining the image space and object space.Comparing with the single stereo-pair matching, multi-view image matching has advantages in below: Firstly, it improves the reliability of matching utilizing the redundant multi-view image information to effectively solve mistake matching with repetition texture, broken feature etc; Secondly, it maximally reduces the information blind area of image, and solves the occlusion problem in matching.The algorithm of multi-view image matching can be divided into two categories by matching modes: multi-image can be matched in pairwise mode or in simultaneously multiplematching mode.The former is based on the traditionally pairwise stereo matching.It separately matches the stereo pairs one by one, integrates all of the matching results in object space, and then obtains the correct matching results.For example, Pateraki (2005) proposes the algorithm of adaptive multi-image matching (AIM), which is used for ADS40 image.It separately establishes stereo pairs with reference image and searches image at first.Then the algorithm makes quality check to matching result in single stereo pair, and makes the least square matching with multi-image according to corrected matching result by the former matching.Yuan and Ming (2009) introduce a multi-image matching algorithm by integrating image and space information.This algorithm realizes the integration of the matching results in each pairwise in the object space through the multi-ray intersection with the function of gross error detection by iterative predictive approach.Matching result in pairwise mode is uncertain because it does not combine multiimage redundancy information in the process of matching.So it needs to be consistently constraint in object space or filter method to be integrated into the matching results of multiple stereo pairs in the object space, which simultaneously increases the complexity of algorithm.Matching to all images simultaneously by adopting object geometry constraint mode, which obtains the corresponding point and the space coordinate of the corresponding at the same time.Zhang (2005) and Zhang, et al. (2008) propose the geometrically constrained crosscorrelation (GC 3 ) algorithm, which chooses the nadir-viewing image of linear array image as the reference image, and extracts features from the reference image, and then searches the corresponding feature in the search image.It comes through the course of the from image space (reference image) to object space and to image space (search image).Because it is restraint to features extracted in the reference image, it does not adapt to the image obtained by the center projective in the areas with large gurgitation of the earth surface.If a certain space area is not imaged in the reference image because of occlusion but imaged in the other images, this area will not participate in matching.This problem may be solved to take turns to select each image as the reference image, but it increases the computation.Vertical line locus (VLL) is used by image correlator in DSR-11 mix digital photogrammetry workstation (Zhang & Zhang, 2002).VLL utilizes "ground element" in the object space as matching primitives, which have been used multi-image matching (Zhang, et al., 2007;Ji, 2008).It is uncertain because it integrates the similarity measure function in all stereo pairs in the course of matching, which reduces the effect from incorrect matching caused by occlusion and repetitive texture by right matching.Focusing on the serious occlusion problem in city images, this paper makes full use of the advantage of multi-view image matching, and proposes a reliable multi-view image matching algorithm supported by the moving Z-Plane constraint (MZPC).It introduces geometric constraint used in the Space-Sweep method (Collins, 1995), and makes multi-image matching simultaneously for feature points.Based on Space-Sweep method, the MZPC proposes the multi-image selective matching strategy under the grey similarity constraint.This algorithm can simultaneously carry on the matching of multiimage feature points under "the best candidate will be matched in the first instant" matching strategy and plane grid height constraint.
MULTI-VIEW IMAGE MATCHING FOR FEATURE POINTS UNDER THE MOVING Z-PLANE CONSTRAINT
The multi-view image matching algorithm under the moving Z-Plane constraint is based on the basic photogrammetry principle of forward intersection that the corresponding points in the different images will always intersect to the same object point in the object space.In this paper, supposing that the position where intersected by different image projection rays of feature points may be the space position of the imaged feature, one new object constraint mode to multi-view image simultaneously matching based on feature points can be established according to this hypothesis. .Using the inverse solution of collinearity equation, it makes the point projection rays in all image intersect with the plane, and obtains the object point ( , ,
Constraint by the Moving Z-Plane
in the plane, i=1,2,...,N, N is the total number of projection feature points.Then it separately statistics the number of projection rays in the each grid cell, and if in a certain cell (T is the threshold), the position of this grid cell will be regarded as the feature point position in the object space.Finally, it records all the grid cells which with in the plane of this height, and considers them as the grid cells to be matched.
T T
In the height plane, for each grid cell to be matched, its corresponding feature points in different images will be simultaneously recorded, which means that the projection rays pass through the same grid cell.This algorithm selects images having feature points, and performs the grey correlation matching.If the computed correlation coefficient is beyond a certain threshold, it regards the position of grid cell as the object feature point position, and the corresponding feature points in different image will be the corresponding points.Then the grid cell is evaluated with height value of the plane position, which will not participate in the latter matching.This process is called as grid cell in the plane matching.After that, this algorithm deletes the feature points successfully matched, and moves the plane to the next height position.Then it initializes the number of projection rays pass through all grid cells in the plane , and repeats the process above until reaches the lowest height value position.0 number = By moving the plane to different height positions, this algorithm utilizes the positions of grid cells in the plane to constraint the range of projection rays for feature points in different images, and determines the matching candidates.This algorithm is called as the moving Z-Plane constraint.The flow chart of overall algorithm is shown in Fig. 2, involving the key technology as follows:
Hierarchy Matching Strategy
Considering the occlusion problem in multi-view image matching, and some feature points in some images are not extracted in the process of feature extraction, our algorithm matches all the grid cells whose number of projection rays are above 1.The more the projection rays passing through a grid, the more reliable matching result will be obtained.According to the matching principle of "the best candidate will be matched in the first instant", this paper adopts a hierarchy matching strategy.First, it matches the best grid cells in the plane at each height position, namely matching the grid cells with , and is half of the number of image.
Second, it matches the second-best grid cells in the plane at each height position, namely matching the grid cells with .In order to enhance the reliability of matching results, this process utilizes the matching results of the best grid cells to constraint the matching results of the second-best grid cells.
Grey similarity constraint
For the feature points in different images whose projection rays pass through the same grid cell in the plane, this paper selects images that have the same feature points, and adopts the grey similarity constraint to match these feature points.This process involves three key problems as follows: ⑴ Correlation window transform strategy of from object space to image space.First, the object window with the center of grid cell to be matched in the plane is determined, and is marked by W. In this process, it assumes that the ground is flat, and all points in this window are having the same height values, the same as the height value of plane location.Then, this paper projects the four corners of the object window onto all images to be matched according to the mathematical model of the sensor imaging, and obtains four corner points of corresponding region correlation windows in different images.Finally, according to the sizes of correlation windows, it calculates the number of pixels within the correlation windows, and inserts the pixel gray values in the corresponding positions.⑵ Reference image selection.For each grid cell to be matched, the MZPC algorithm selects the image having a nearest distance from projection center in the plane to the grid cell to be matched as reference image.
⑶ Similarity measurement calculation.The MZPC algorithm takes correlation window in the reference image as a standard.
Then it separately calculates the normalized cross-correlation (NCC) based on the grey similarity measure in correlation windows between the reference image and other images, and finally obtains average of the NCC (ANCC) according to calculate the average of all the NCC.
Constraint by plane grid cell height
The MZPC algorithm initializes the height of all grid cells in the plane as zero, and records the grid cell height using a matrix having the same dimension with the grid cells in the plane, marked by grid height matrix m n R × , where m and n separately mean the number of column and the number of rows of grid cell in the plane.Each value in the matrix means the height of corresponding position grid cell in the plane.After matching the best grid cells, the MZPC algorithm assigns the height values to every successfully matched grid cells to constrain the latter matching.In the process of matching with the second-best grid cell, for the grid cells meeting the grey similarity constraint, it assigns the plane height to this grid cell, and compares with the heights of other grid cells in a certain neighborhood range.According to the surface smooth principle in the local range area, if there are some grid cells having the similar height value within this neighborhood, it will be considered as the correct matching result, otherwise will be abandoned.
EXPERIMENTS ANALYSIS
In the experiments, this paper selectes three UCX digital aerial images in the same strip, whose pixel size is 7.2 m μ , the corresponding ground resolution is 0.049m, and the along-track overlapping is more than 80%.The precise orientation elements of each image are obtained by the triangulation using the VirtuoZo.In these images, there are some high buildings, which produce different form occlusions to those surrounding surfaces.
Determine the grid plane
The coverage area of the image is determined as the range of moving Z-Plane.The height range of the area in image as show is 3 m ~ 93 m, and the plane moving step is 1m.
Feature point matching under the moving Z-Plane constraint
Focusing on experimental images, this paper adopts the Forstner operator to extract feature points in all images.The number of feature points extracted in image L1, image L2, and image L3 are 34343, 32901, and 28053, respectively.⑴ Initially matching of the best grid cells.It moves the plane from high to low, and matches the grid cells with 3 number = at each height position, the matching results of grid cells on Z=31 position is shown as Fig. 3.The threshold of ANCC is 0.85 in the grey similarity measurement calculation.The result of initially matching of the best grid cells is shown as Fig. 4, and the number of the homologous points is 3009, which is denoted by red cross, and the corresponding positions in the grid cell is shown as Fig. 5. ⑵ Matching of the second-best grid cells.After the matching of the best grid cells, this algorithm matches the grid cells with 2 number = , and the number of successfully matched grid cell is 5853, which corresponds to the homologous points in different images (Fig. 6).This paper adopts different colors to express the matching results of different images combination, for example, the yellow dots express the matching result of image L1~L2, the blue dots express the matching result of image L2~L3, and the red dots express the matching result of image L1~L3.The number of the homologous points obtained by matching with different image combination separately are 2871, 2566, and 416.Obviously, the longer baseline among multi-view images, and the bigger intersection angle, the more difficult of the matching.The ellipse area in image L1 and image L2 in Fig. 6 shows that this area does not match the homologous points because of the occlusion in the image L3 in the initial matching process.Through the matching of the second-best grid cells, this algorithm automatically selects the image L1 and image L2 to match according to the projection rays, and obtains the correct matching results, as shown in the homologous points by yellow color in the ellipse area, which effectively avoids the influence of the occlusion area in image L3.After the matching of the best grid cells and the second-best grid cells, the total number of the successfully matched grid cells is 8862 (Fig. 8), which corresponds to the homologous points in different images (Fig. 7).The meaning of different colors and different signs is the same as the front.It can be see that the second-best grid cells matching enhances the dense of initial matching results, more abundantly expresses the object details, and gets the homologous points in two overlapping area and occlusion area in the multi-view image.
Figure 8. Grid cells of final matching results
Result evaluation
This paper makes forward intersection with the homologous points, and obtains the discrete 3D points (Fig. 9).From this figure it can be seen that there are a few mistakes.As shown in Fig. 9, the ellipse regions includes the three error matching points obtained in initially matching of the best grid cells, and the seven error matching points obtained in matching with the second-best grid cells.So it needs to adopt a reliable filter to reject the mismatching results.
Contrast analysis with different constraint conditions
The existing multi-view image matching object constraint models, such as GC 3 algorithm (Zhang, 2005)
CONCLUSION
Focusing on the buildings occlusion problem in matching with city image, as well as the multiple solutions in matching due to repetitive texture, this paper summarizes the existing multiview image matching algorithms, and proposes a multi-view image matching algorithm for feature point supported by the moving Z-Plane constraint.This paper selects three UCX digital aerial images of a typical building area in the same strip for matching experiments, and verifies the validity of the proposed algorithm.The conclusions are as follows.⑴ The matching algorithm proposed in this paper can simultaneously match with any number of multi-view image, and obtains the matching results in any overlapping areas in multi-view image ; ⑵ Based on the selective matching, the MZPC algorithm effectively avoids the effects by occlusion image, and improves the reliability of the matched results; ⑶ The MZPC algorithm does not need the iterative calculation of height, and avoids the appearance of multiple peaks in the cross-correlation curve caused by similar texture, and reduces the probability of mismatches.⑷ The MZPC algorithm is entirely based on the matching with feature points.The uniformity and the density of feature points distribution directly determine the density of matching results.It needs to further research the dense matching with other matching primitives.⑸ The experimental results also have a few mismatched homologous points.It needs to research a high reliable method to reject the mistakes.
Figure 1 .
Figure 1.Sketch map of moving Z-Plane constraintSuppose a fictitious plane in the object space, the direction of this plane is vertical to the direction of vertical line in the object space, and the size of the plane contains the areas in the all images in the object space with .In Fig.1, the plane is divided into regular grid cells (small plane elements) by a certain interval ( ≥ image resolution), which can be seen as a
Figure 2 .
Figure 2. Flow chart of the overall algorithm
Figure 3 .
Figure 3.The matching results of cells as number=3 on Z=31 position Figure 4.The matching results of the best grid cells as Number=3
Figure 6 .
Figure 6.(left) The matching results of the second-best grid cells as number=2 Figure 7. (right)The final results
Figure 9 .
Figure 9. Discrete 3D points , modified vertical line locus algorithm (MVLL)(Ji, 2008), constrain the image space search range by points to be matched moving in the range of an approximate height value along a certain linear direction (projection ray locus, vertical line locus) in the object space, and reduces the search range from two-dimensional to onedimensional.The traditional standard normalized crosscorrelation (NCC) is expressed as a function of height value Z, algorithm calculates the sum of NCC (SNCC) values iteratively, and selects the height value Z corresponding to the maximum SNCC value as the object correct height value.The formulas are, | 4,359.6 | 2012-07-31T00:00:00.000 | [
"Computer Science"
] |
In-core dosimetry for the validation of neutron spectra in the CROCUS reactor
The present article describes the preliminary validation study of simulated in-core and reflector n eutron spectra in preparation of oncoming experimental programs in the zeropower reactor CROCUS at EPFL. For this purpose, a set of activation foils were irradiated at three characteristic positions in the CROCUS reactor, and the subsequent activities were analyzed via γ spectrometry. The experimental setup was then modeled with the Monte Carlo neutron transport code Serpent2 and associated with an analysis tool to include the effect of the reactor power history during experiments. The comparison of calculated and measured reaction rates (C/E) indicates a general consistency (at 2σ) between calculated and measured spectra. However, offsets of C/E values were observed in (n,γ) reactions, up to 18% for 115In and 8% for 63Cu dosimeters. This could be caused by an unexpected isotopic composition, uncertainties in nuclear data, or the spectrometry analysis. In addition, a 100-groups spectrum unfolding was performed using the experimentally determined reaction rates and the Serpent2 spectra as the prior knowledge. The unfolded spectra were mainly adjusted in the thermal and fast ranges, while few modifications w ere m ade i n t he e pithermal r egion d ue t o the low contribution of epithermal neutrons in activation processes. Moreover, within energy groups where the capture reactions show resonant behavior, flux depletion (up to 38% as compared to the prior spectra) is observed due to the absence of self-shielding effect in the unfolding process. For this purpose, an unfolding method based on energy groups weighting is developed and tested. Keywords—Neutron spectrum measurement, in-core dosimetry, gamma spectrometry, spectrum unfolding
I. INTRODUCTION
The development of the knowledge on nuclear data have largely contributed to the advances of reactor technology, while continuous improvement is still required for further optimization, new designs and waste management for example.
The nuclear data of stainless steel components (Fe, Cr and Ni) are of great interest for the performance assessment of the Gen-III EPR reactor. Validation of the iron nuclear data was accomplished in the EOLE reactor for this purpose [1]. In continuation of this work, the PETALE program [2]- [4] will be performed in the zero-power reactor CROCUS. For such program, the interpretation of experimental data would require a precise knowledge of neutron energy spectra.
In this paper is described the experimental study conducted in CROCUS, for the validation of simulated in-core and reflector neutron spectra by means of activation dosimetry.
II. METHOD A. Experimental setup
CROCUS is a an experimental reactor operating at a maximal power of 100 W. The reactor is light water-moderated with two interlocked fuel zones in a square lattice, separated by a water gap, as shown in fig. 1. The power of the reactor is monitored by two ex-core fission chambers and two γ compensated ionization chambers. Further information on the Figure 1: CROCUS core configuration with irradiation positions for the present study (after [5]) reference experimental configuration of the reactor can be found in [5].
For the current study, disk-form dosimeters are placed at three different positions (indicated in fig. 1) within the reactor: (i) at the center of the core, (ii) in one of the empty control rod tubes, (iii) in two empty aluminium rods placed in the near periphery of the core (surrounded by the water reflector). Dosimeters are prepared and laminated in a thin plastic film to avoid possible deformation and contamination during experiments. These dosimeters are then attached to Plexiglas plates for positioning.
At the end of the experiments, safety systems are activated to shut down the reactor. The activation of dosimeters are then analyzed via γ spectrometry, using High Purity Germanium (HPGe) detectors [6], [7].
B. Experimental design
The experimental design consists of selecting dosimeters, reactor power and operation time. It was accomplished with simulated activation experiments using the Monte Carlo neutron transport code Serpent2 [8], under criticality configuration based on the modeling in [9].
For illustration purpose, some simulated flux profiles (in 100 energy groups) at different experimental positions are shown in fig. 2, while the shaded areas represent the 1σ uncertainty of the simulation. The calculation is made in a 0.005 cm 3 detector, which is representative of the size of dosimeters. (1) resonant dosimeters presenting capture reactions (n,γ) (2) threshold dosimeters with threshold reactions such as (n,p) or (n,n').
The dosimeters are selected according to the contribution of the calculated reaction rates (RR) in different energy groups between 1 × 10 −5 eV and 20 MeV. Figure 3 illustrates the sensitivity of different dosimeters in terms of distributed and integral RR. Resonant dosimeters are sensitive to the thermal flux a nd c ertain g roups o f e nergy w here t here a re resonant peaks in the cross section. On the other side, threshold dosimeters are meant to target the fast flux since the threshold energy of the reaction is generally above 0.5 MeV. Additionally, resonant dosimeters covered with neutron poison (cadmium boxing) are also used to target epithermal contribution. The set of dosimeters (reactions and associated decay data) used for spectrum characterization is listed in Table 1.
C. Analysis with simulations
An effort has been made to accurately reproduce the experimental configurations in simulation for comparison purpose. In the Serpent2 calculations, dosimeters are modeled with their exact geometry and position within the reactor during experiments. The isotopic composition of the materials are assumed to be as the natural abundance. The activity A i (t) for a certain isotope i is estimated using the calculated RRs R i and the reactor power history P (t).
For precise activity estimation of short-lived isotopes, the power history is of importance. Consistently, the influence of power history is checked for all isotopes of interest. This is implemented in a numerical code based on a simplified Bateman equation, by neglecting reaction rates in isotopes once activated: where λ i is the decay constant, ∆t the elementary time step and C 0 the normalization constant. This normalization is achieved with a reference 197 Au (n,γ) dosimeter in each experiment.
A direct comparison can therefore be made between the measured dosimeter activity A E and the simulated one A C . These quantities are proportional to the measured and calculated RRs R E and R C . The ratio C/E between RRs in calculation (C) and experiment (E) can be estimated as: In addition to the validation based on C/E, the calculated neutron spectrum can be adjusted by the experimental feedback of measurements (R E ). This is accomplished via an unfolding procedure that compares the measured RR and that for an initial guess spectrum (the one calculated in Serpent2 in the present study). The best approximated spectrum is obtained by minimizing the residual on RRs in the χ 2 sense. The unfolding is done with a reference code MAXED [10] based on a maximum entropy criterion, and two non-linear optimization methods with the CVXPY package [11] in Python. The first method uses augmented response matrix (of cross sections) and the information of the spectrum (referred as the unweighted method hereafter). The other method uses the spectral distribution of RRs calculated in Serpent2 to weight the contribution of the response in each energy group. Thus, it will be referred as the weighted method. The unweighted method uses a response matrix of constant cross section values, therefore the self-shielding is not taken into account. The weighted method, by contrast, aims at introducing implicitly the self-shielding effect into the unfolding by assigning different importance to energy groups.
III. RESULTS AND DISCUSSION
In the present part, the comparison of RRs in simulations and experiments is summarized. The coherence between calculation and measurement indicates that the Serpent2 modeling using a common purpose JEFF 3.1.1 library [12] is in general consistent with outcomes of the activation dosimetry. The result of the comparison also allowed a study on spectrum unfolding. However, discrepancies between calculated and measured RRs are observed for some dosimeters. Possible sources of these discrepancies will be discussed.
A. C/E in thermal and epithermal ranges
The overview of C/E values for resonant dosimeters, therefore sensitive to thermal and epithermal ranges, is presented in fig. 4.
Figure 4: C/E values for resonant dosimeters at different positions
For each reaction at the three experimental positions, one representative and consistent value throughout a number of experiments is shown with its uncertainty. The C/E values indicate the agreement between simulations and measurements either with the absolute value or the relative one. It should be noted that uncertainties associated to resonant dosimeters are mostly small as their high activation allows a precise determination of the activity. The relatively large uncertainty for 64 Ni is partly due to its low isotopic proportion in nature (< 1%), partly due to its short half-life (2.5 hours). Discrepancies for 63 Cu, 65 Cu and 115 In dosimeters were observed to be an offset, regardless of the position of the dosimeter across different experiments. For 63 Cu and 65 Cu, the uncertainty in cross section data and branching ratio of γ decay emission is of the same order of magnitude [12]- [15]. In the case of 115 In the source of the offsets could be a possible isotopic composition that is different from natural abundance, or biases and uncertainties in γ spectrometry analysis not taken into account in this study.
It is worth noticing that the activity measurement of 63 Cu is made on a particular γ ray: the 511 keV photon created through the annihilation of a positron over an electron. It is however difficult to confirm that the background annihilation in the acquisition system during measurements is fully subtracted from the target decay emissions. Therefore, the applicability of spectrometry analysis is questionable by definition. It was also observed that the acquired peak shape in γ spectrum, illustrated in fig. 5, resembles that of a Gaussian (Poisson distribution of a large number of events), but the width of the peak is generally two times larger than ordinary peaks measured in this work.
B. C/E in the fast range
Despite the general agreement between simulation and experiment for threshold dosimeters as shown in fig. 6, large In addition to the dosimeters presented in fig. 6, the inelastic scattering reaction (n,n') of 115 In was also measured in experiments, through the 336 keV photon emission during the transition from the first metastable level to its fundamental level. However, the results show important discrepancies between calculation and measurement: the calculated RRs is about 6% of the measured one. This issue was also reported by [16]. A possible explanation is that, in general purpose nuclear data libraries, inelastic scattering cross section is defined for different metastable levels while in γ spectrometry the activity is measured as the sum of all transitions that would contribute to the metastable level of interest. The complexity of the decay scheme makes inconceivable to compare calculations with measured inelastic scattering RR through common libraries. fig. 7, the unfolding results are generally in agreement with the initial Serpent2 spectra. Sharp depletion (up to 38% as compared to the prior spectrum) or augmentation of the flux is observed in the epithermal range, which is caused by the large value of resonant capture cross sections. Energy groups with strong flux variation are identified to be associated with resonance peaks of different reactions ( fig. 7). In reality, the self-shielding limits the actual RR in the resonance region, while in the unfolding this mechanism has to be integrated numerically. This situation can be illustrated through the comparison of unfolded spectra by different methods: The unweighted method takes the cross section values as constant and independent of the spectrum, therefore the unfolded spectra show serious flux variation ( fig. 7a). The weighted method shows positive results for the self-shielding correction. With a relatively high cost associated to flux adjustment in resonance regions, the self-shielding effect is represented to a certain extent. Nevertheless, in the case of the spectrum at the control rod position ( fig. 7b) the flux depletion is important regardless of the method, possibly caused by numerical issues. Additionally, calculated and measured uncertainties in the RR of threshold dosimeters lead to large uncertainty in the unfolded fast spectra. Further improvement on the experimental procedure and more accurate reaction rate estimation in simulation are expected to optimize the unfolding results.
IV. CONCLUSION
This paper summarizes the used methodology to validate simulated neutron spectra in the zero-power reactor CROCUS for dosimetry applications. A set of activation dosimeters were chosen, so that the energy spectrum from thermal to up to 20 MeV is covered as much as possible. In-core experiments were designed by Monte Carlo modeling, and performed at different locations in CROCUS and the consequent activation is measured via γ spectrometry. The experimental setup were then accurately modeled to compare calculations with experimental results in terms of C/E. The general agreement on C/E indicates the consistency between calculated and measured spectra, and thus the validation of spectra and experimental methodology.
Spectrum unfolding was performed using experimental feedbacks. Two unfolding methods were implemented, which show larger uncertainty in the solution spectrum, but are consistent with the MAXED reference solution.
Inconsistency and offsets in C/E were observed in the analysis, which could be caused by different sources of biases in simulation, experiments and nuclear data. Further improvements are to be made on the design of activation experiments, along with a dedicated model for corrections in spectrometry analysis. | 3,136 | 2020-01-01T00:00:00.000 | [
"Physics",
"Engineering"
] |
Evaluation and Comparison of Different Machine Learning Methods to Predict Outcome of Tuberculosis Treatment Course
Tuberculosis treatment course completion is crucial to protect patients against prolonged infectiousness, relapse, lengthened and more expensive therapy due to multidrug resistance TB. Up to 50% of all patients do not complete treatment course. To solve this problem, TB treatment with patient supervision and support as an element of the “global plan to stop TB” was considered by the World Health Organization. The plan may require a model to predict the outcome of DOTS therapy; then, this tool may be used to determine how intensive the level of providing services and supports should be. This work applied and compared machine learning techniques initially to predict the outcome of TB therapy. After feature analysis, models by six algorithms including decision tree (DT), artificial neural network (ANN), logistic regression (LR), radial basis function (RBF), Bayesian networks (BN), and support vector machine (SVM) developed and validated. Data of training (N = 4515) and testing (N = 1935) sets were applied and models evaluated by prediction accuracy, F-measure and recall. Seventeen significantly correlated features were identified (P <= 0.004; 95% CI = 0.001 0.007); DT (C 4.5) was found to be the best algorithm with %74.21 prediction accuracy in comparing with ANN, BN, LR, RBF, and SVM with 62.06%, 57.88%, 57.31%, 53.74%, and 51.36% respectively. Data and distribution may create the opportunity for DT out performance. The predicted class for each TB case might be useful for improving the quality of care through making patients’ supervision and support more case—sensitive in order to enhance the quality of DOTS therapy.
Introduction
Over recent years, Tuberculosis has been considered, particularly in low and middle-income countries as a global public health concern with the estimated two million deaths annually [1,2].Up to 50% of all patients with TB do not complete treatment and fail to adhere to their therapy [3].It has been estimated that in industrialized countries non-completion of treatment is around 20% [4] and according to the Centre of Disease Control and Prevention in the United States, 25% of patients fail to complete their chemotherapy [5]; in other words, the proportion of patients with active disease who completes therapy under standard conditions ranges from as little as 20% -40% in developing countries to 70% -75% in the USA [6].Noncompliance is a significant factor leading to the persistence of tuberculosis in many countries and the consequences of this well recognized fact are prolonged infectiousness, relapse, prolonged and more expensive therapy due to multidrug resistance TB and death [7].It has been revealed that noncompliance is associated with a 10-fold increase in the incidents of poor results from treatment and accounted for most treatment failures [8].It is the most serious problem hampering tuberculosis treatment and control as patients with active disease who are non-compliant, sputum conversion to smearnegative will be delayed and relapse rates will be 5 -6 times higher and drug resistance may develop [6].That is, TB patients remained in the pool of active cases will in- crease the development of TB among latent cases who are prone to be infected or affected.This threats public health and makes huge costs that can be spent to improve public health improvement and promotion instead.DOTS (directly observed treatment, short course) which is current international control strategy for TB control involves the case detection and completes the entire course of treatment successfully.In 2006, in order to improve DOTS quality, World Health Organization (WHO) has designed "Stop TB" Plan [2].In this plan, it has been highlighted that health care services should identify and concentrate on interruption factors that halt TB treatment.Supervision which plays important role in patient treatment adherence and drug resistance prevention must be carried out in a context-specific and patient-sensitive manner.
Although WHO has highlighted the necessity of improving the quality of DOTS in terms of supervision and patient support in the "Stop TB" plan, there is no specific way to measure how intensive health providers' support and supervision should be and to whom of TB cases it should be provided.To make this clear, we may require a tool to predict the patient destination regarding TB treatment course completion.The tool may identify high risk TB patients for treatment course non-compliance as it is impossible to serve all TB patients with active supervision and support due to the cost consideration and limited resources.This may be used to define the level of supervision and support each patient needs based on the predicted outcome by an accurate predictive model.Currently, no system is available to estimate TB treatment course through using features of TB patients and designing a systematic method to predict the given outcome.
For the prediction of tuberculosis treatment course completion, the defined outcome related to each record of TB patient contains five potential classes: cure and completed treatment (desirable outcomes), failure, quit, and death (undesirable outcomes) [9].Patients who are classified with undesired outcomes need more supervision and support.Public health interventions for TB control need to be set according to how many of TB patients would shift to unwelcome outcomes.
Making sure that TB patients finish the treatment course entirely is a main step to TB control and public health promotion.Actually, here the main question of concern is "can new TB cases at risk of failing in treatment course completion identified from their early registration"?For this purpose, machine learning methods have been already applied and they worked properly according to previous studies [10,11].They infer a model through being trained by a set of historical examples in frame of data categories; thus, new examples with unknown classes could be assigned to one or more classes by pattern matching within the developed valid model [12].
This work used machine learning algorithms to achieve six predictive models of a crucial real-world problem which offer the medical research community a valid alternative to manual estimation of risky TB patients to quit or fail treatment course completion.This recognizes TB patients whom supervision in frame of DOTS therapy needs to be more active as they are risky for treatment course interruption.
Correlation Analysis
In the case of a large dataset, learning the dataset is not useful unless the unwanted features are removed since an irrelevant and redundant feature does not add anything positive and new to the target concept [13].To identifying influential predictors, a bivariate correlation is used which is a correlation between two variables including independent and dependent parameters by calculating the correlation coefficient "r" [14].As there is no specific prediction and hypothesis two-tailed test is carried out.
Suppose we have got two variables X (list of features) and Y (class), an optimal subset is always relative to a certain evaluation function.To discover the degree to which the variables are related, correlation criteria are applied [15].It reflects the degree of linear relationship between two variables ranging from +1 to −1.A perfect positive linear relationship between variables is shown by +1; however, −1 implies an entire negative linear association.The following formula (1) is used to calculate the value of r: where there are two variables X and Y and their means X, Y and standard deviations including S X and S Y respectively; n is the number of TB instances.The correlation coefficient can be tested for statistical significance using special t-test through following formula.
Degree of freedom for correlation coefficient calculation is equal to n − 2. From a t-table, we would find significant relationship between each of variables X and Y (P < 0.05).
Problem Modeling
To achieve the aim of this work, we evaluate a number of well-known classification algorithms on TB treatment course completion.In fact, prediction can be viewed as the construction and use of a model to assess the class of an unlabeled sample, or to evaluate the value or value ranges of an attribute that a given sample is likely to have [16].To train every of applied algorithms, we set a big matrices as follows: where X is a big matrice which is used to train the given algorithm by using train set in which i would be the ith number of samples and j is the jth number of predictor.i j x in our training set would be 17 4515 x and in applied testing set has been 17 1935 x .Y which is the dependent variable of TB treatment course in both training and testing sets addresses five classes where a label of "1" implies that TB case got cured and "2" means that s/he has completed the treatment course entirely; whilst a label of "3", "4", and "5" means that the patient belongs to the undesirable class such as quitting, failing the treatment course, or dead respectively.
Having compared the pros and cons of machine learning methods, to carry out the considered prediction task in this study six classifiers including decision tree (DT), Bayesian network (BN), logistic regression (LR), multilayer perceptron (MLP), radial basis function (RBF) and support vector machine (SVM) were selected.Next section is brief explanation about each of these selected classifiers.As shown in Figure 1 into training (two-third) and testing (the other one-third) datasets each containing seventeen significantly correlated attributes and the outcome variables for every record without any missing data.The six above named classifiers were applied to the training dataset to estimate the relationship among the attributes and to build predictive models.Afterwards, the testing dataset which was not used to model inference was utilized to calculate the predicted classes and compare the predicted values with the realones available in the testing dataset.The model which is the most fitted and accurate one will be selected to predict the outcome of Tb treatment course for new TB cases.
Decision Tree
The decision tree is a nonparametric estimation algorithm that input space is divided into local regions defined by a distance measure like the Euclidean norm; it is a flowchart-like tree structure where the internal node, branches, and leaf node means concepts associated with our training tuples.In this hierarchical data structure, the local region is identified in a sequence of recursive splits that in a smaller number of steps by implementing divideand-conquer strategy.This powerful classifier is famous for its intuitive explainability [17].In this research, C4.5 classification task-oriented algorithm has been applied through calculating: , , , log where i is the probability that an instance belongs to class i .Having calculated the entropy , which addresses the expected information according to partitioning by attribute i A we have: where and j s is the number of samples in subset .j s In the next step, the encoding information that would be gained by branching on i A is: The attribute i A with the highest information gain is chosen as the root node, the branches of the root node is formed according to various distinctive values of ij a .The tree grows until if all the samples are all of the same class, and then the node becomes a leaf and is labeled with that class.From the tree, understandable rules can be extracted in a quick processing [17].Here, the task of decision tree development is conducted using C4.5 classification algorithm where each tree leaf is allocated to a class of chemotherapy outcome along with number of misclassified cases.
Bayesian Networks
Generally, Bayesian classifiers are statistical approaches capable of predicting class membership likelihoods like the probability of the training set belonging to a specific class.It is based on Bayes theorem and well known for its high accuracy and speed when applied to a large data collection.Here, simple estimator Bayesian network has been used.
Let X be a sample whose class is to be defined and let H be a hypothesis that X belongs to class C. In this classification task, P H X needs to be defined which is the probability that the hypothesis H holds based on the data sample X.It can be calculated through: where P H and P X are the prior probability of H and X respectively. P X H is the posterior probability of X given on the H.In model training these three values are calculated and the probability for sample X to be in hypothesis H can be determined.There are two essential features that define a Bayesian network: a directed cyclic graph in which each node denotes a random variable, either discrete or continuous values, as well as a set of conditional probability tables.Let x be a data tuple related to the correspondent attributes 1 2 .Given that every attribute is conditionally independent on its non-descendent and parents in the network, the following equation is a representation of joint probability distribution., , y y , , , parents where the values for are related to the entries for ; is the probability of a specific combination of values of X. Probability distribution, with the probability of each class may be the output of this classification process [18].
Logistic Regression
Logistic regression is used primarily for predicting binary or multi-class dependent variables.This algorithm's response variable is discrete and it builds the model to predict the odds of its occurrence.This method's restrictive assumptions on normality and independence induced an increased application and popularity of machine learn-ing techniques for real-world prediction problems.In this study, multinomial logistic regression is applied which is an algorithm that constructs a separating hyperplane between two sets of data by using the logistic function to express distance from the hyperplane as a probability of dichotomous class membership: In this equation, i X symbolizes discrete or continuous predictor variables with numeric values; in the case of depending variable (Y) being dichotomous, we use this algorithm.The constants 0 1 2 , , , , n are the regression coefficients estimated from training data which are typically computed by using an iterative maximum likelihood technique.Normally, this formula's justification is that the log of the odds, a number that goes from to + , is a linear function.Particularly by using this model, stepwise selection of the variables can be made and the related coefficients calculated.In producing the LR equation, the statistical significance of the variables used to be determined by the maximum-likelihood ratio [19].
Artificial Neural Networks
Artificial Neural Network (ANN) is biologically inspired analytical method which is able of modeling extremely complex non-linear functions.Here, a common architecture named multi-layer perceptron (MLP) with learning by back-propagation algorithm is built.A neural network is a compound of linked input/output units in which every link has an associated weight.Adjusting the weights is the core phase for predicting the correct class label of input through iterative learning.This method is popularly used in classification and prediction tasks with high tolerance to noise and the ability to classify unseen patterns [19].Here, algorithm has conducted the learning process 500 times and the network structure is shown in a simple way when only two attributes have been applied as inputs.The structure of a two-input, one hidden layer neural network for our task with five outputs has been presented in Figure 2.
Radial Basis Function
A radial basis function network (RBFN) is an artificial intelligence network in which its activation function is simply radial basis in a linear combination.This type of network was designed to view a problem in curve-fitting (approximation) and high dimensional space.The real inspiration behind the RBF technique is finding amultidimensional function that offers the best fit to training tuple and then applies this multidimensional surface to interpolate the test data through regularization [20].Gaus- sian radial basis function is applied algorithm in this study.
Support Vector Machine
Support vector machine (SVM) is a new classification method for both linear and non-linear data.SVM applies nonlinear mapping to transform the original training tuple into a higher dimension.It seeks the optimal linear separation hyperplane which is the decision boundaries separating the tuple based on their class labels.Polykernel support vector machine has been applied in this investigation.Let the dataset Abe considered as x x y x A y D where i x is the set of learning data with correspondent class label i .For a two-class related training dataset, for instance, every i can take either +1 or −1.This could also be generalized to n dimensions and the SVM duty is to find the best dividing lines that can be drawn and divide all of the tuples of every class from the others.For multidimensional classes the hyperplanes should be found asdecision boundaries.This can be arranged by defining the maximum marginal hyperplane (MMH) since models with larger hyperplane are more accurate at classification [18].y y
Accuracy Measurements
In order to evaluate the prediction rate, the following related parameters need to be measured.Prediction accuracy percentage (model Accuracy), model fitness, recall, Precision, F-measure, and ROC area are considered criteria used to assess the models' validity.In confusion matrix (Figure 3) in which the columns denote the actual cases and the rows denote the predicted cases, the accu-racy of model obtained from training set (model fitness) and testing set (prediction accuracy) are calculated.The prediction accuracy is the percentage of correct prediction (true positive + true negative) divided by the total number of predictions.In machine learning, sensitivity is simply termed recall (r) and precision is the positive predicted value.F-measure is a harmonic means of precision and recall and the higher value of it merely reveals the better performance of a prediction task.Another comparative criterion is Roc curve which is a two-dimensional graph where the true positive (TP) rate on the Y axis is plotted on the false positive (FP) rate on the X axis [21].
TP TN TP TN FP FN
The Database
The available dataset has been driven from gathered records by health practitioners, nurses, and physicians at local TB control centres throughout Iran in 2005.In tuberculosis control centres, health deputies of each province in a network system collected data in every appointment.By using "Stop TB" software, more than 35 parameters were collected and through applying bivariate correlation method; those independent variables which are significantly correlated with target outcome are selected as predictors (P <= 0.05).They are presented and defined in Table 1.The distribution of corresponding outcomes of TB treatment course based on their frequency in training and testing sets are shown in Table 2.
Due to normal and non-normal distribution variables which were tested by both Kolmogorov-Smirnov test and visual shape of attribute's distribution, Pearson and Spearman rho methods were applied to find highly correlated factors.
Feature Selection
Significant Kolmogorov-Smirnov test confirmed with the shape of their distribution showed the distribution of our variables; Spearman rho was applied for those with non-normal distribution.
Having considered the results, it is revealed that males are more likely to not get cure and complete the course of TB treatment (r = −0.082,P < 0.0001).
As patients get older there is more probability to get undesirable result from treatment course (r = 0.158, P < 0.0001).
The more under-weight the patients are, they are more likely at the risk of unwelcome outcome (r = −0.056,P < 0.0001).
TB cases with extra-pulmonary TB are more likely at the risk of outcomes like quitting, failing or death (r = −0.066,P < 0.0001).
There is 0.127 more probability that immigrants from Iraq, Pakistan, and Afghanistan to quit, get failed in treatment course (r = 0.127, P < 0.0001).TB patients who are living in abroad as well as mobile cases are −0.027%more possible to get undesirable outcome compared with those who are living in urban and rural areas (r = −0.027,P < 0.0001).
The more time TB patients spend affected with this infectious disease, there is 0.073 times more chance of non-desirable outcomes of tuberculosis treatment course.
Also, the probabilities of having imprisonment history in his/her life, consuming drugs through intravenous, or having unprotected sex increase the likelihood of unwanted outcome with r = 0.157, 0.0172, and 0.16 respectively (P < 0.000).
Classification Algorithms Analysis
In this work, six classifiers including DT, BN, LR, MLP, RBF and SVM were applied to the patient dataset.Model development is conducted in two main steps including model fitness and model accuracy.To calculate the model fitness criteria we used the data of training set; however, to compute the model accuracy measurements, data of testing set is applied which is merely much more valuable to judge about our models accuracy.Related results of these experiments are demonstrated in Table 3.
Model fitness assessment by evaluating training accuracy is 84.45% for C4.5 decision trees; it is considerably less for Bayesian net where the value is 58.56%.The values of model fitness for logistic regression (56.5%),MLP (64.93%),RBF (50.65%), and SVM (53.04%) have been close.Having compared the Roc curves reveal that the area under curve for C4.5 has the most value for model fitness and accuracy with 0.96 and 0.97 respectively.This measurement is less for Bayesian net (0.85), logistic regression (0.82), MLP (0.81), RBF (0.79), and SVM (0.76) in terms of model accuracy.C4.5 decision tree has been able to build a model with greatest accuracy since the model fitness and prediction accuracy are 84.45% and 74.21% respectively.
Prediction accuracy for Bayesian networks has been calculated by 62.06%.Model accuracies obtained from other classifiers are different as this value for LR, MLP, RBF, and SVM have been 57.88%,57.31%, 53.74%, and 51.36% respectively.
Figure 4 is the comparative Roc curves based on the given outcome of tuberculosis treatment including cure, complete, quit, failed or dead.This figure shows six Roc curves for six developed models based on the given outcome.For the outcome "cure", C4.5 has outperformed than other classifiers with area under curve 0.958.Similarly, the most accurate result for outcomes 'completed' and "quit" obtained for C4.5 with 0.966 and 0.956 respectively.C4.5 has also performed the best for the outcome "failed" and "dead" by classifying 0.986 and 0.963 of cases correctly.Overall, these results of area under curve reveals better performance of C4.5 decision tree classification algorithm.Table 3 and Figure 4 present obtained results including model fitness and accuracy as well as produced area under ROC.
Discussion
Of the six investigated methods, decision tree has achieved the best performance while other classifiers have given relatively close results in lower ranks.According to previous studies [22,23], the technique with the best classification performance might behave differently from another one and there is no single best method for every circumstance.Decision trees that classify instances by sorting them based on feature values has outperformed other methods.C4.5 performs variously in different conditions.It has been reported that there is an association between the performance of applied tools and following issues including the type of problem we are analyzing, the type of input data (discrete or continuous), and finally emerging overlapping in outcome classes.
Due to the fact that our available dataset is a large volume of data with dimensional structure, normally it is expected that SVMs, neural networks, and decision tree outperform others.However, the dataset is mainly com- posed of fourteen discrete variables and three continuous attributions (age, weight, and length of disease); in this case, decision tree has produced the most promising results due to its dual ability to tackle both continuous and discrete/categorical predictors which is superior to other aforementioned techniques that are good at handling only continuous variables.
In the case of emerging relationship among attributions, BNs don't perform well to manage learning properly.There have been significant relationships between seventeen predictors and outcome class which may cause the weaker performance of Bayesian networks rather than C4.5; furthermore, discriminant algorithms like logistic regression also fail on this type of data with high correlation between the attributes.In this study, there are many correlations among variables, like weight and nationality (r = −0.052,P < 0.001), LBW and sex (r = −0.047,P < 0.001), imprisonment and sex (r = −0.156,P < 0.001), prison and weight (r = 0.065, P < 0.001), length and nationality (r = 0.099, P < 0.001).Those correlations in addition to applying fourteen discrete inputs might cause weaker results from BN and LR rather than C4.5. is good at coping with irrelevant data.This might be the case in this particular study since there are some variables with very low correlation coefficient that decision tree has not taken them very much to built the model and not at all in the main root nodes.For example, area (−0.027), prison (−0.026), diabetes (0.029) have low correlation coefficient, where this value for recent TB infection, imprisonment, IV drug using, sex are 0.250, 0.151, 0.172, 0.160 respectively.The variables with high correlation coefficient like recent TB infection, length, imprisonment, and treatment category have played a major role as root nodes whereas the variables with small correlation coefficient have been recognized as less important factors placed in very sub-nodes close to leaves which can be even pruned.Decision tree's ability of utilizing significant input factors in basis of their degree of contribution to estimate outcome of tuberculosis treatment course creates greater predictive model than those classifiers, such as MLP, RBF, and SVM which count every of input variable uniformly by weightening affecting the results' transparency.
The higher values for Kurtosis (>7) and skew (>1) denote that our variables are far from normality and decision tree and other symbolic methods (nonparametric schemes) tackle robustly with distributions with large kurtosis and skew [22].In this research's dataset, the average values of skew and kurtosis are 2.169 and 7.469 respectively (P < 0.05) confirming the non-normal distribution.Hence, the only available nonparametric symbolic learning algorithm in the current study is decision tree which performed well to partition the input space.In actuality, high skew (>1) or kurtosis (>7) along with the presence of binary/categorical variables, using relevant and correlated predictors without any missed instances or noised data have prepared the best opportunity for decision tree to predict more accurately than other applied algorithms.In view of the rank of other employed classifiers, BN outperforms others and four remaining classifiers work relatively similar with prediction accuracy percentage ranging from 53.74% to 57.82%.In a study [24], Tu has reviewed a number of researches concluding that LR & MLP perform closely; it is the case here where they performed with identical prediction accuracy (57.82).RBF is actually a type of neural network and it might be a postulation that based on their algorithm similarities and data type entity results are comparable.
Decision tree with flowchart-type structure is more likely method to be understandable for general users with low level of specialized knowledge about TB.Produced results of decision tree can be simply interpretable and applicable; their rules can be understood either by doctors or health practitioners who implement DOTS in rural area and make decision alone.
Conclusion
To sum up, available big body of real data related to TB patients has created an opportunity to generate accurate models which can predict outcome of DOTS therapy.This provides us information about outcome of treatment course for each patient and defines who needs high level of supervision and support; this is valuable as it is not possible to give every single of patients a full supervision and support distinctly.The decision tree model can be used to screen risky patients for fail in treatment course completion in population using general data gathering in routine general practice.This will help healthcare practitioners especially in rural regions to evaluate the risks of MDR-TB among their patients quickly, inexpensively, and noninvasively.TB control through totally implemented DOTS therapy is such a crucial stage in public health improvement and promotion.
Figure 2 .
Figure 2. The simplified structure of applied neural network only with two of our attributes "HIV" and "recent TB infection", one hidden layer and five classes in output layer.The structure is much simplified neural networks are popularly applied in classification and prediction, as they have advantages such as high tolerance to noise, and the ability to classify unseen patterns.
Figure 3 .
Figure 3.A simple confusion matrix, a two-by-two table, values in the main diagonal (TP and TN) represent instances correctly classified, while the other cells represent instances incorrectly classified (FP and FN errors).
Figure 4 .
Figure 4. ROC curve for C4.5 decision tree, logistic regression (LR), Bayesian networks (BN), radial basis function (RBF), multilayer perceptron neural networks(MLP), and support vector machine on the task of classifying outcome of TB treatment course based on the target outcome including cured, completed, quit, failed, and dead.In all cases decision tree has outperformed other methods with area under curve 0.92 for cured, 0.94 for completed, 0.88 for quit, 0.96 for failed, and 0.85 for dead.
Evaluation and Comparison of Different Machine Learning Methods to Predict Outcome of Tuberculosis Treatment Course 185
RBF ANN Figure 1. Schematic presentation of applied methodology of model development and validation process. In order to de- velop accurate predictive or classification models, firstly data of TB patients (in a huge matice with 6450 rows and 18 vectors) are divided as training and testing sets. Training set is used to learn data and build models through applying
6 algorithms including decision tree (DT), support vector machine (SVM), logistic regression (LR), Bayesian network (BN), radial basis function (RBF), and artificial neural network (ANN).To check the validity and accuracy of developed models, testing set is applied.
Table 3 . Comparison on model fitness and model accuracy of six various applied machine learning algorithms.
*training accuracy (%); ** prediction accuracy percentage (%). | 6,787.2 | 2013-07-31T00:00:00.000 | [
"Computer Science",
"Medicine"
] |
Inversion tectonics of the northern margin of the Basque Cantabrian Basin
– The northern margin of the Basque-Cantabrian Basin was analysed combining stratigraphic and structural data from both surface and subsurface together with reflectance of vitrinite data from oil wells. The use of cross-section balancing techniques in addition to thermal modelling enabled us to reconstruct the tectonic, burial and thermal evolutions of the basin margin as well as those of the Landes High to the N in two different periods. The section restoration at the end of the Cretaceous shows a northern basin margin structure influenced by evaporites related to south-dipping normal faults. The reconstruction in middle Eocene times yielded up to 1 800 m of Paleocene-middle Eocene deposits on top of the basin margin. Subsequent tectonic inversion related to the Pyrenean compression led to the north-directed thrusting of basement units and to the formation of thrust slices or inverted folds in the cover along the northern margin of the basin. Tectonic subsidence analysis together with maturity data provided evidence that oil was generated in the basin during the late syn-rift and post-rift stages in the Late Cretaceous and became overmature during the period of incipient inversion after 55 Ma. In the autochthonous Landes High, the oil was generated after the tectonic inversion period 37 Ma. Inversion tectonique de la marge nord du Bassin basque-cantabrique Mots clés.– Inversion tectonique, Analyse de la subsidence, Modélisation thermique, Pyrénées Occidentales, Bassin Basque-Canta-
INTRODUCTION
The Basque Cantabrian Basin, located in the northern margin of Spain, is a Mesozoic extensional basin which was inverted during the Pyrenean orogeny in Paleogene times (fig.1).From Triassic to Albian times, the extension was associated with the formation of the North Atlantic [Le Pichon et al., 1971;Montadert et al., 1979;Ziegler, 1988].This extension was particularly intense during the Aptian-Albian interval owing to transtensive movements related to the counterclockwise rotation of Iberia during the formation of the Bay of Biscay [Montadert et al., 1979;García-Mondéjar, 1996;García-Mondéjar et al., 1996].In this transtensional context, the subsidence was especially severe in the eastern part of the Basque Cantabrian Basin along the western branch of the Basque Arch.The onset of the Pyrenean orogeny led to the inversion of the thick early Cretaceous basin above the Keuper (Triassic) evaporitic level.The stratigraphy and sedimentology of the Mesozoic succession have been documented by García-Mondejar [1989].The large-scale geometry of the inverted basin has been described by Cámara [1997].However, the detailed evolution of this inversion tectonics has not yet been fully explained because of the relatively deep level of erosion affecting this inverted margin and because of the poor quality of the available seismic data.The lower Cretaceous rocks outcropping on the northern margin of the basin and imaged at the Cormorán-1 borehole present a high level of maturity of organic matter.These values contrast with the moderate values of the autochthonous platform deposits lo-cated to the north of the northern inversion front, which was a structural high during the Cretaceous.
The aim of this paper is to reconstruct the geometry of inversion tectonics of the northern margin of the Basque-Cantabrian Basin by integrating a balanced and a restored section with a backstripping analysis constrained by thermal and maturity data.Thermal modelling enable us to constrain the burial and exhumation of the northern margin of the basin.The two stages of restoration are (1) late Cretaceous, close to the transition from post-rift sag basin to foreland basin, and (2) the end of the middle Eocene close to the onset of the northern margin inversion [Soler et al., 1981;Cuevas et al., 1999].
THE BASQUE-CANTABRIAN BASIN STRATIGRAPHY
The study area is located on the northeastern side of the Basque Cantabrian Basin.This region is characterized by the change in the folding trend, from ENE-WSW to the east and WNW-ESE to the west, outlining the Basque Arch which approximately parallels the coastline (fig.1).Two rifting episodes are recorded in the pre-inversion stratigraphy of the Basque Arch : the early Triassic and the late Jurassic-early Cretaceous [e.g., García-Mondéjar, 1996].
The syn-rift stratigraphy corresponding to the early Triassic rifting consists of red continental clastics, evaporites and carbonates that crop out in the eastern branch of the Basque Arch, overlying and surrounding the Paleozoic massif of Cinco Villas [García-Mondejar, 1989].The upper Triassic Keuper facies crop out to the west, forming diapirs aligned in a NW-SE direction.These alignments suggest the position of pre-existent faults at depth [Serrano et al., 1994;García-Mondéjar, 1996;Cámara, 1997].
The post-rift stratigraphy above the Keuper consists of early to middle Jurassic platform carbonates, which are uncommon in the study area.The second syn-rift period embraces the late Jurassic and early Cretaceous infill (154-96 Ma), which was characterized by the presence of continental clastic sequences and carbonates infilling fault-bounded basins [Ramírez del Pozo, 1971;Pujalte, 1977;Pujalte, 1981;Rat, 1988].The Aptian-middle Albian interval was characterized by the deposition of fluvio-deltaic and shallow marine siliciclastic sediments, carbonate platforms and muddy intraplatform sequences known as the Urgonian Complex [Rat, 1988;García-Mondejar, 1989].Continued extension accompanied by the development of left-lateral strike-slip structures after the opening of the Bay of Biscay was coeval with the deposition of fluvial, shallow marine and siliciclastic flysch sequences (the Supra-Urgonian Complex, upper Albian-lower Cenomanian in age) [García-Mondejar, 1989].During the late Cretaceous there was deposition of carbonatic flysch sequences (Cenomanian-Santonian) and siliciclastic flysch deposits (Campanian-Maastrichtian) under the influence of the opening of the Bay of Biscay [Mathey, 1987].Submarine volcanism took place in the area during the Cretaceous especially in late Albian times [García-Mondéjar et al., 1996].These rocks crop out in the northern limb of the Biscay Synclinorium and are probably related to the faults of the early Cretaceous rifting period [Cámara, 1997] and/or to major crustal scale faults [Azambre and Rossy, 1976].
During the Maastrichtian-early Eocene, the onset of compressive movements related to the Pyrenean contraction led to the uplift and folding of the sedimentary infill.Concomitantly, around the Biscay Synclinorium a deep flysch trough developed, being infilled with fluvial, shallow-marine and thick turbidite sequences with an eastern provenance [Mathey, 1987;Pujalte et al., 1989;Pujalte et al., 1993].
GEOMETRY OF INVERSION TECTONICS Present geometry of the inverted basin
In this section, we describe the present onshore and offshore geometry of the northern margin of the Basque-Cantabrian Basin.We focus our attention on the western part of the Basque Arch, characterised by the existence of NW-SE trending structures.
The structure of the Basque-Cantabrian inverted basin was determined by the construction of a regional, 55-km long, SSW-NNE balanced section (fig.1), which was re-stored to the top of the Cretaceous (fig.2).The balanced section was constrained by two offshore exploratory oil wells and by published maps and sedimentological data from onshore [García-Mondéjar, 1985;Mathey, 1987;Robador and García-Mondéjar, 1987;García-Mondejar, 1989;Robador et al., 1991;Pujalte et al., 1993;EVE, 1994].Furthermore, a second cross-section located 12 km towards the NW shows the Cormorán-1 oil well and provides additional tectonic information on the basin inversion (fig.3).
The offshore structure is characterised by a set of north-directed thrust faults.The propagation of these thrusts results in the formation of the North Biscay Anticlinorium, whose structure varies along the strike.To the SE, a highly overturned fault propagation anticline made up of lower Cretaceous sediments and cored by Keuper evaporites developed (fig.2).These lower Cretaceous sediments show a significant reduction in thickness across the anticlinorium, being thicker in the upright section of the forelimb (1 400 m) and thinner in the overturned section (620 m).The associated tight syncline to the north is cored by Paleocene rocks located on top of thin upper Cretaceous platform sediments.To the NE, the frontal structure consists of a set of imbricated thrusts detached in FIG. 2. -Balanced section and section restored to the top of the Cretaceous across the inverted NE Basque-Cantabrian Basin.The restored section shows schematically the sedimentology and paleocurrents of the upper Cretaceous units.The location of the normal faults was deduced from the geometry of the restored section in agreement with regional data (see fig. the Keuper evaporites overthrusting basinal lower Cretaceous rocks on top of upper Cretaceous platform sediments (fig.3).
Along the regional section, the back limb of the North Biscay Anticlinorium crops out to the south of the NE prolongation of the Guernica diapir, where up to 1 500 m of evaporites were drilled in the Guernica-1 borehole (fig. 1 for location).In the section, the core of the Biscay anticlinorium was infilled with Keuper evaporites based on surface and subsurface information.To the south, the section shows the increasing thickness of the Cretaceous units, reaching a maximum (up to 7 500 m) in the Biscay Synclinorium.In this region a number of volcanic and volcanoclastic layers are interbedded with Cenomanian sediments (fig.2).
The Bilbao Anticlinorium, located to the South, is affected by an intense axial plane cleavage [Cámara, 1997], and is cored by late Jurassic-Neocomian thick sedimentary sequences.The presence of these sediments is evidenced by 2 exploration oil wells located to the SE of the regional section (fig. 1 for location).These wells imaged about 2 000 and 3 000 meters of Neocomian rocks [Lanaja et al., 1987].The thickness of these units diminishes progressively to the north, before petering out in the northernmost part of the structure (fig.2).The geometry and thickness of the upper Jurassic-Neocomian unit and the reduced thickness of the underlying lower Jurassic and Keuper sediments (as pointed out by Soler et al. [1981]) imply the presence of the basement in a relatively shallow structural position.The involvement of the basement in the structure of the area has been deduced from field data [e.g., Cuevas et al., 1999] and also inferred from geophysical data that suggest the NE-directed thrusting of Cretaceous intrusives together with lower crustal rocks [Aller and Zeyen, 1996].In contrast to Cámara [1997], who interpreted significantly thicker Keuper and late Jurassic-Neocomian units, the sections in figure 1 and figure 3 show that the basement is involved in the thrust system.
According to our interpretation, a main thrust carrying basement rocks propagated the deformation to the north, forming the Bilbao Anticlinorium-Biscay Synclinorium pair.
A new basement thrust moved to the north transporting the basin margin and forming the frontal structure (fig. 2 and fig.3).Onshore, under the frontal lower Cretaceous a frontal fault-related anticline developed, which trends parallel to the regional structures.The Vizcaya C-1 well imaged this anticline showing the involvement of the basement in the reverse faulting (fig.2).
The minimum shortening calculated in the SE sector of the study area is about 25 km and was mainly produced by the overthrusting of the basement units (fig.2).This amount of shortening does not take into account the deformation produced by the intense cleavage described in the Bilbao Anticlinorium and is thus a minimum value as pointed out by Cámara [1997].In front of the structure the deformation was accommodated mainly by the overturned anticline related to the Guernica diapir, which had developed during the extensional period [García-Mondéjar, 1987].The contrasting style of deformation in the frontal part of the two sections (highly overturned fold vs. north-directed thrust imbricates) could have been attributed to (1) differences in the pre-inversion geometry of the basin which are common in pull-apart settings [e.g., McClay and Dooley, 1995], and (2) to the long-lived Guernica diapir.Diapirs in this context are preferential sites for accommodating shortening during compression, inducing the formation of highly sheared folds at the basin margins in transpressive settings [e.g., Brun and Nalpas, 1996].
Restored section at the end of the Cretaceous
The geometry of the northern Basque-Cantabrian Basin at the end of the Cretaceous was determined by the restoration of the regional cross-section (fig.2).This restoration was undertaken by bed-length balance using the top of the Maastrichtian stratigraphic sequence as the reference.
The application of this method of restoration to the study area poses a number of problems : a) uncertainties in the geometry of the present-day section attributed to the exhumation of part of the section and also to the absence of good quality subsurface data; b) the existence in the present-day section of a large volume of evaporites hampers the area balance of the restored section; c) the strong lateral variations in the thickness of the lower Cretaceous units (mainly Urgonian and Supraurgonian, Aptian-lower Cenomanian units) complicate the construction of a palinspastic section valid for the whole study area.
In an attempt to minimise these problems, we used published additional sedimentological information to constrain the restored basin and we interpreted a significant accumulation of Keuper evaporites attached to the former normal faults, assuming diapirism during the extension.
After pinning the structure in the foreland, the northern margin of the Basque-Cantabrian Basin was positioned by restoring the overturned limb of the North Biscay Anticlinorium to its pre-inversion situation.The geometry of the lower Cretaceous units and the presence of coarse delta facies in the margin suggest that the margin was active, at least, during the Aptian-lower Cenomanian interval.To the south, the thick Cretaceous depocenter coincides with the location of the present Biscay Synclinorium.To the south of the basin margin, the complete Jurassic and Cretaceous cover attains a thickness of approx.5000 m.The thickness is about 3 500 m for the lower Cretaceous units and 4 000 m for the upper Cretaceous units along the main frontal flysch trough.This thickness is evidenced by the outcropping geometry of the upper Cretaceous units in the northern limb of the Biscay Synclinorium.The exact thickness of these units is difficult to estimate due to internal folding, and therefore the thickness represented in the restored section should be considered as a minimum.The maximum thickness for the upper Cretaceous units attains 2 500 m for the upper Cenomanian-Santonian unit and 1700 m for the Campanian-Maastrichtian unit [EVE, 1994].
The Cormorán-1 well was projected 12 km to the SE, parallel to the structures, in the balanced section.The restoration allowed us to locate the well stratigraphy in the reconstructed basin margin, resulting in an additional thickness of approx.1,000 m of upper Cenomanian-Campanian sediments on top of the measured 2,400 m of Aptian-lower Cenomanian infill (fig.2b).These upper Cretaceous sediments must be thinner than in the centre of the basin because of the situation of the stratigraphic section close to the basin margin.Soler et al. [1981] pointed out the existence of 750 m of upper Cretaceous in the North Biscay Anticlinorium, a value close to the 1,000 m projected in the restored section from the well data.
Further south, the restored section shows the incipient uplift and folding of the basement and lower Cretaceous infill, giving rise to a gentle anticline.The thinning towards the anticline crest of the syntectonic beds and the change in the direction of the paleocurrents (from approx.N-S during Albian-Cenomanian to approx.E-W during Maastrichtian-Eocene) provide evidence of folding activity [Pujalte et al., 1989;Robles et al., 1989;Pujalte et al., 1993;EVE, 1994].The formation of this anticline during the late Cretaceous can be related to the early stages of Pyrenean compression [Rat, 1988;Pujalte et al., 1989;Pujalte et al., 1993;EVE, 1994].The late Jurassic-Neocomian and lower Cretaceous successions are thicker to the south of the anticline than in the north.These units include platform and deltaic deposits sourced from the south, indicating the presence of an active southern margin [Robles et al., 1989].This tectonic inversion is, however, younger than inversions determined in the southern part of the Basque-Cantabrian Basin during the early Cretaceous [Malagón et al., 1994].
THERMAL EVOLUTION
Before the application of backstripping techniques, a thermal analysis based on vitrinite reflectance data was carried out to constrain a possible thermal evolution for both the northern margin of the Basque-Cantabrian Basin and the autochthonous Landes High and also to constrain the thickness and age of the stratigraphic reconstructed section of this northern margin.Maturity modelling is based on the linear relationship between depth and vitrinite reflectance (Ro) found in basins largely unaffected by major unconformities, young dip-slip faults and localized igneous activity.The percentage of vitrinite reflectance (%Ro) increases with depth and/or duration of burial, reflecting the response of the sedimentary section to the paleoheat flow [Dow, 1977].The Genex TM software calculates the increase in vitrinite reflectance, using a correlation curve between the given reflectance and the modelled transform ratio.This ratio is the amount of petroleum generated by primary cracking (which increases with temperature) with respect to the maximum amount generated in a complete petroleum evolution, the primary cracking being the transformation of kerogen to petroleum [Beicip-Franlab, 1998].
Northern margin of the Basque-Cantabrian Basin
The Cormorán-1 borehole cuts the inverted basinal deposits overthrusted on top of the platform deposits (see cross-section of figure 3).The reflectance of vitrinite shows very high %Ro values varying from 1.4 % in the shallow part of the section to 4 % at 2 400 m of depth (fig.4).The thermal modelling described in this section was performed by varying the values of heat flow and/or the thickness of the section.
As a first step, we used the present-day stratigraphic sequence from the Cormorán-1 borehole as depicted in figure 4, imposing a minimum constant value of heat flow at the bottom of the basement of 64 mWm -2 , which is relatively low for extensional basins, and a maximum constant value of 76 mWm -2 .Using these parameters that correspond to the present-day heat flow in the Cantabrian Basin [Marzán, 2000] With these constant values of heat flow, the only way to fit the observed data is to add a substantial thickness (between 3 000 and 4 000 m) of Tertiary pre-inversion sediments on top of the reconstructed section.This is unlikely in this region since the amount of Paleocene-Eocene sediments offshore only attains 2 000 m (Vizcaya C-1 borehole).
In the following steps the reconstructed section was kept constant up to the middle Eocene and a variable heat flow was introduced in accordance with the tectonic evolution of the basin.It is usually possible to adjust the heat flow of a rifted region by defining the duration and magnitude of the extension period (stretching factor).However, the Paleogene inversion of the Basque-Cantabrian Basin towards the north and south makes it difficult to estimate the late Jurassic-middle Albian rifting stretching factors.Regional data assembled by Vergés and García Senz [2001] indicates three main periods of tectonic evolution for the area : (1) rifting and strike-slip development from 120 to 93.5 Ma, (2) thermal relaxation from 93.5 to 55 Ma and (3) thrusting, foreland development and a final quiescent period from 55 Ma to the present.During the rifting and pull-apart development, a significant amount of heat flow between 80 mWm -2 and 90 mWm -2 was used (100 mWm -2 is a typical value for strike-slip settings [Allen and Allen, 1990]) and during the final stage of the development the present-day value of 64 mWm -2 was employed.In the interval the heat flow was progressively relaxed from 80-90 at 93.5 Ma to 64 mWm -2 at 55 Ma (fig.6).The thermal maturity shows the best fit, especially with the lower part of the observed vitrinite reflectance data using 85 mWm -2 .
The calculated initial heat flow is essential for producing the very high reflectance of vitrinites encountered in the Cormorán-1 borehole and in the lower Cretaceous rocks outcropping in the region.This is consistent with the emplacement of volcanic rocks in late Albian times, coeval with an important crustal thinning in the Basque-Cantabrian basin [Azambre and Rossy, 1976;García-Mondéjar et al., 1996].An important NW-SE magnetic anomaly, parallel to the main structures of the area, suggests that the presence of the volcanic rocks is a regional feature [Aller and Zeyen, 1996].The thermal calculations were made on the assumption of an additional reconstructed stratigraphy based on the regional geology that includes upper Cretaceous to middle Eocene rocks.However, the upper vitrinites are difficult to match; they are located in a different tectonic slice, close to the major normal fault bounding the basin (fig.3).When restored, this tectonic slice is positioned above deep basin deposits.The apparently higher %Ro of these vitrinites may be due to the original basin location of the rocks affected by a slightly different heat flow history.
The heat flow history as well as the reconstructed stratigraphic section determined from the thermal model of the Basque-Cantabrian Basin forms the basis for the backstripping analysis and geohistory discussed below.
Autochthonous Landes High
Three boreholes, located in the Landes High to the north of the Basque-Cantabrian Basin inversion front, show reflectance of vitrinite data sampled in the lower part of the boreholes in the basement Carboniferous coals.The three wells (Vizcaya B-1, Vizcaya C-1, and Vizcaya C-3 in fig. 1) display reduced Cretaceous thicknesses (between 200 and 400 m) unconformably overlying Paleozoic and Triassic sediments with slightly different %Ro values varying from 0.72 to 1.22.The deepest well is the Vizcaya C-3 cutting the most complete upper Cretaceous and Cenozoic stratigraphy of the Landes High.In this well, 500 m of eroded Triassic-Jurassic sediments were estimated by extrapolation from the more complete Aulesti-1 well.Interestingly, the thermal analysis of the three wells indicate a good fit with a constant heat flow of about 68-70 mWm -2 , thereby indicating that this amount can be used for geohistory.
RECONSTRUCTED STRATIGRAPHIC SECTION AT THE END OF THE MIDDLE EOCENE : BURIAL AND EXHUMATION
The results of the thermal analysis and the regional data allow us to reconstruct a complete stratigraphic sequence for the northern boundary of the Basque-Cantabrian Basin.This sequence includes approx. 1 000 m of upper Cretaceous sediments, about 1 000 m of Paleocene-early Eocene infill and about 800 m of middle Eocene deposits on top (fig.7).The inclusion of these middle Eocene sediments in the northern margin of the basin assumes that the inversion of the marginal area post-dated this age.However, an older age of deformation has been documented in the inner part of the inverted basin [Pujalte, 1989;Robles et al., 1989;Pujalte et al., 1993].The onset of Pyrenean-related compression accelerated the growth of an earlier, basement-involved Bilbao Anticlinorium, resulting in an increase in the tectonic subsidence and in the generation of a frontal trough in the centre of the basin.This led to a thick accumulation of flysch deposits in front of the anticline [Mathey, 1987;Rat, 1988;Robles et al., 1989] (fig.7).The assumed pre-tectonic latest Cretaceous to middle Eocene units were considered constant in thickness in the 2 selected wells but slightly thicker in the vicinities of the growing anticline.
With increasing shortening, the basin margin was uplifted and eroded and the lower Cretaceous rocks were exposed.A subsidence analysis performed in this scenario yielded information on the main pre-inversion tectonic events, which is difficult to deduce in the present highly eroded section.
BACKSTRIPPING ANALYSIS
Subsidence analyses were performed in two stratigraphic sections located in the northern margin of the Basque-Cantabrian Basin and in the more stable Landes High to the North.
The stratigraphic sequence in the northern basin margin is based on the reconstruction of the Cormorán-1 well (fig.7); and the northern section is based on the Vizcaya C-3 borehole, located 20 km offshore aligned with the regional section (see fig. 1 The subsidence analysis was performed by means of the backstripping method [Steckler and Watts, 1978] based on the decompaction of sediments, paleobathymetric estimations and assuming local isostasy.The utilized algorithm calculates the decompaction of sediments using porosity-depth relationships based on the following equation, 1/(z)=Lz+1/(0); where z is the given depth, L is the compaction factor, (0) is the surface porosity and (z) is the porosity at depth z.In the case of some lithologies, the algorithm uses empirical porosity-depth relationships based on different published sources [BEICIP-FRANLAP, 1998].The average porosity of a given formation results from the porosities of the different lithologies composing the formation.
The paleowater depth estimations were obtained from the literature and from the facies analysis for the northern Basque-Cantabrian Basin [Robador and García-Mondéjar, 1987;Robles, 1988;García-Mondejar, 1989;Robador et al., 1991;EVE, 1994].In the case of the Landes High, we used the paleobathymetric estimations of Brunet [1997] although they correspond to a section located approx.100 km to the north in the Parentis Basin.In the northern Basque-Cantabrian Basin the Urgonian sediments (Aptian-middle Albian) have been interpreted to be deposited at a depth not exceeding 200-300 m [García-Mondejar, 1989].However, we used a conservative water depth of 150 m for this period, taking into account the marginal situation of these sequences.Given the fact that the Supraurgonian unit (upper Albian-lower Cenomanian) records a deepening of the basin, from fan delta facies to turbidites, we progressively increased the water depth from 50 to 250 m for this interval.For the upper Cretaceous and lower Tertiary units we considered a paleobathymetry of 200 m, bearing in mind that the sedimentation took place in relatively deep troughs [García-Zarraga and Rodríguez-Lázaro, 1991].We slightly decreased the paleowater depth for the middle Eocene unit, taking into account the presence of contemporaneous coarse sediments in the basin [EVE, 1994].
In the Landes High, we assigned a paleobathymetry close to sea level for the Cretaceous, assuming that various unconformities determined for this period were produced in a shallow water context.The younger unconformity has been dated to 65 Ma (between Maastrichtian and middle Eocene).From 65 Ma to the present we assumed a fairly constant paleowater depth equivalent to that of the present day one (306 m).
The eustatic variations were not taken into account in the backstripping calculations owing to software limitations.However, their impact on the tectonic subsidence are relatively small.A sea-level curve of the Bay of Biscay shows a sea level rise from approx.75 m at the start of the Aptian to 150 m in the Santonian-Campanian interval [Brunet, 1997].The same curve shows a progressive sea level fall from the late Cretaceous to the present.The paleobathymetric estimations for the Cretaceous indicate a shallowing of the basin from Aptian to middle-late Albian times (120-105 Ma).The eustatic variation for the same period results in a deepening from 75 m to approx.110m, suggesting that the correct subsidence is slightly smaller than the one calculated.From middle-late Albian to lower Cenomanian times (105-97 Ma), the facies indicators suggest an increase in the paleobathymetry whereas the sea-level curve varies from approx.110m to 150 m.The impact of sea level fluctuations would result in a slightly stronger tectonic subsidence than that calculated for this interval (a maximum of few tens of meters).
The subsidence curves display a well-differentiated subsidence history for the Basque-Cantabrian Basin and its continuation to the N into the Landes High (fig.8).The northern margin of the basin shows a well-marked period of subsidence during the early Cretaceous and the lower part of the late Cretaceous from approx.120 to 85 Ma.The intensity of this subsidence event is, however, variable showing an early phase of moderate subsidence rate (120 to 95 Ma), followed by a sudden increase in the rate of tectonic subsidence (95 to 85 Ma), and ending with a moderate rate of subsidence (fig.8).These events can be related to the major geodynamic processes which occurred in the region during the opening of the Bay of Biscay.The early phase of moderate subsidence can be associated with rifting, whereas the rapid subsidence phase can be linked to the formation of pull-apart basins during the transtensional motion of the Iberian plate.The duration of this rapid period of subsidence lasted for approx.10 m.y. in this region.The slow rate of tectonic subsidence (85-55 Ma) could be a consequence of the transtensional tectonics as demon-strated by the regional data [e.g., Vergés and García Senz, 2001].To the E of this locality, the onset of this moderate phase corresponds to short phases of basin inversion within the general strike-slip movement between the Iberian and European plates.
After the Cenomanian-Santonian unconformity, at ~85 Ma, the tectonic subsidence was fairly constant and slow coinciding with the post-rift stage of the Basque-Cantabrian Basin.This subsidence pattern was interrupted by the onset of a slightly faster tectonic subsidence period during the early Tertiary at approx.55 Ma, which lasted for approx.15m.y.before the end of basin subsidence at about 37 Ma.The tectonic subsidence ended at ~40 Ma although the basin still subsided until 37 Ma, when uplift of the basin occurred by thrusting.
As discussed above, the Landes High displays a different vertical motion history.Uplift was the predominant process occurring at this rifted margin during both the early moderate basin subsidence stage and the early part of the rapid basin subsidence stage.This uplift culminated near the end of the rapid basin subsidence period linked to the pull-apart setting at approx. 100 Ma.This age roughly coincides with compressive events determined in the Mauléon-Lacq Basin and follows compressive events in the Organyà Basin in the southeastern Pyrenees [Berástegui et al., 1990;Vergés and García Senz, 2001].From late Albian to middle Eocene times the tectonic subsidence was slow and fairly constant ending with a stable period of several million years.This period preceded the onset of an abrupt increase in tectonic subsidence characterized by very high rates (fig.8).
The geometry of the Landes High subsidence curve shows a period of continuous tectonic subsidence from the late Cretaceous (100 Ma) to the late Eocene (37 Ma) followed by a period of accelerated subsidence that lasted until the late Oligocene (28 Ma) (fig.8).This subsidence pattern triggered the scarce and discontinuous deposition during the Cretaceous and early Tertiary (100-37 Ma), which contrasts with the rapid sedimentation in the late Eocene-early Oligocene interval (37-28.5 during the development of the foreland basin [Cámara, 1997].From the late Oligocene to the present the subsidence did not undergo significant variations, the curve showing an almost horizontal trend.
TIMING OF DEFORMATION
The tectonic inversion of the northern margin of the Basque-Cantabrian Basin has been dated as Eocene [Soler et al., 1981;Cuevas et al., 1999], latest Eocene-Oligocene [Derégnaucourt and Boillot, 1982;Rat, 1988] and late Cretaceous-Eocene [Cámara, 1997].The data presented in this work can help to clarify this important chronology of deformation, and its relationship to oil generation and migration.
The analysis of the tectonic subsidence of the northern side of the Basque-Cantabrian Basin and the southern part of the Landes High not only provides the timing of Tertiary inversion but also the progression of this deformation.During the early Eocene, at ~54 Ma the northern side of the thick Mesozoic Basque-Cantabrian Basin underwent a noticeable increase in tectonic subsidence after a long period of slow post-rift subsidence.This increase can be associated with the onset of significant compressional tectonic deformation affecting the southern and central regions of the basin.The NE directed Bilbao Anticlinorium, which involves basement units, represents the most suitable structure that led to an increase in the tectonic subsidence towards the foreland (fig. 2 and fig.7).This tectonic event lasted for about 14 m.y. and then, after a short quiescent period characterized by continuous deposition, the northern boundary of the basin was inverted above a NE directed thrust system.This inversion resulted in an uplift of the whole region after 37 Ma.
By contrast, the southern side of the Landes High was characterized by a significant increase in tectonic subsidence at 37 Ma in response to the tectonic inversion of the marginal areas of the northern Basque-Cantabrian Basin.The Landes High acted during this period as a foreland basin receiving most of its terrigenous supplies from the Pyrenean Range to the E.This foreland period lasted for about 8.5 m.y., ending at 28.5 Ma.From 28.5 Ma to the present, the Landes High has remained stable with low sedimentation rates.
In the Landes High, part of the increase in the subsidence was due to the considerable amount of upper Eocene and Oligocene deposits imaged by the offshore wells.These deposits are relatively poorly dated in the Vizcaya C-3 borehole.Although we do not have solid arguments to support a different dating, the internal thrusting in the Oligocene sediments and the markedly reduced section of the Miocene to Plio-Quaternary deposits on top suggest that the presence of about 2,000 m of Oligocene deposits is an overestimation.In such a case, the tectonic subsidence determined in this study would be a maximum.
The correlation of these two areas of the Basque-Cantabrian region suggests a clear shift in the tectonic activity migrating from the centre of the basin inversion of the Bilbao anticlinorium to its northern margin along the North-Biscay anticlinorium.Further deformation migrated towards the Landes High where small-displacement basement thrusts affect the lower part of the cover succession (fig. 2 and fig.3).
The two contrasting subsidence evolutions and tectonic chronologies in the northern Basque-Cantabrian Basin and the Landes High offer a number of possibilities for hydrocarbon exploration.On the basis of subsidence analysis and vitrinite data, we present the hydrocarbon maturity geohistory of the two studied sections (fig.9).
In the northern Basque-Cantabrian Basin the two periods of significant burial (1) in the late Cretaceous, with an elevated heat flow, and (2) in the early Eocene led to the generation of hydrocarbons at 100 Ma and to their overmaturity before Tertiary uplift.In the Landes High the generation of hydrocarbons started after the main period of subsidence and rapid sedimentary infill (37 Ma).This generation was partially coeval with the final stages of the growth of the foreland structures.It should be pointed out, however, that most of the oil in the upper Cretaceous and Eocene successions generated (and continues to generate) after the development of the thrust-belt.
CONCLUSIONS
A new balanced and restored section from the Bilbao anticlinorium in the S to the Landes High in the N based on surface and subsurface data shows the present structure of the inversion tectonics of the northern margin of the Basque-Cantabrian Basin.
The restored section at the end of the Cretaceous shows the post-rift geometry of the basin and the mild inversion of the ancestral Bilbao anticlinorium.
The shortening calculated from large structures is 25 km (33 %).This figure must be considered as a minimum owing to the significant tightening of folds and the widespread existence of a strong axial plane cleavage.The shortening resulted in the formation of thrust-slices or overturned folds cored by evaporites along the northern basin margin and the sedimentary cover.The geometry of the evaporitic bodies in the pre-inverted basin could influence the style of frontal deformation.
Thermal modelling indicated that the best fit of the vitrinite data is obtained by combining a variable heat flow in concordance with the major tectonic episodes in the re-gion, and the addition of approx.1,800 m of Paleocene-middle Eocene deposits on top of the basin margin (fig.10).
Tectonic subsidence modelling in combination with hydrocarbon maturity analysis suggests that oil generation started during the Mesozoic lattermost rift stage in the basin.Subsequently the basin underwent a thermal relaxation period, a stage of subsidence related to basement-involved folding in the hinterland and a final period of thrusting and exhumation.These processes led to the overmaturity of the lower Cretaceous outcropping sediments.In contrast, in the Landes High the oil generated after the Tertiary basin inversion and foreland basin development commenced at 37 Ma, resulting a present-day oil window which embraces upper Cretaceous to upper Eocene sediments.
FIG. 1
FIG. 1. -Simplified structural map of the NE Basque-Cantabrian Basin with the location of cross-sections and oil wells.FIG. 1. -Carte structurale simplifiée du Nord-Est du Bassin basque-cantabrique avec la localisation des coupes et des puits.
FIG. 2. -Balanced section and section restored to the top of the Cretaceous across the inverted NE Basque-Cantabrian Basin.The restored section shows schematically the sedimentology and paleocurrents of the upper Cretaceous units.The location of the normal faults was deduced from the geometry of the restored section in agreement with regional data (see fig. 1 for location).The sketch in the lower part of the figure shows the thrusting sequence.FIG. 2. -Coupe équilibrée et restaurée à la fin du Crétacé à travers le Bassin basque-cantabrique inversé.La coupe restaurée montre schématiquement la sédimentologie et les paléocourants des unités du Crétacé supérieur.La localisation des failles normales a été interprétée a partir de la géométrie de la coupe restaurée et des données régionales (cf.fig. 1 pour la localisation).Le schéma de la partie inférieure de la figure montre la séquence des chevauchements.
for location).The Vizcaya C-3 well attains a depth of 4 150 m, imaging 100 m of Carboniferous sediments at its base.This Paleozoic basement is unconformably overlain by upper Cretaceous limestones made up of 104 m of Cenomanian and 384 m of Santonian to Maastrichtian deposits, these two units being separated by an unconformity.The upper Cretaceous rocks are unconformably overlain by 760 m of middle and upper Eocene marls and limestones and by a 2 220 m thick unit of Oligocene marls.Finally, on top of the section, 324 m of sediments which were assigned to Miocene to recent, complete the stratigraphy.
FIG. 8. -Subsidence analysis of the northern margin of the Basque-Cantabrian Basin and of the Landes Platform High.FIG. 8. -Analyse de la subsidence de la marge nord du Bassin basquecantabrique et du plateau des Landes.
FIG. 10. -Geodynamic events for the northern margin of Basque-Cantabrian basin deduced from the subsidence analysis including proposed heat-flow values and timing of oil generation from Aptian to the present.FIG. 10. -Evénements géodynamiques pour la bordure nord du Bassin basque-cantabrique déduites de l'analyse de la subsidence incluant les valeurs de flux de chaleur proposées et la chronologie de la génération du pétrole de l'Aptien à l'actuel. | 8,528.8 | 2002-09-01T00:00:00.000 | [
"Geology"
] |
Pion nucleus Drell–Yan process and parton transverse momentum in the pion
We present a thorough analysis of unpolarized Drell–Yan (DY) pair production in pion–nucleus scattering. On the nucleus side, we use nuclear parton distributions along with parametrisations of the nucleon partonic transverse distribution available in the literature. Partonic longitudinal and transverse distributions of the pion are those obtained in a recent calculation in a Nambu–Jona Lasinio (NJL) framework, with Pauli–Villars regularization. The scale of the NJL model is determined with a minimisation procedure comparing NLO predictions based on NJL evolved pion distributions to rapidity differential DY cross sections data. The resulting distributions are then used to describe, up to next-to-leading logarithmic accuracy, the transverse momentum spectrum of dilepton pairs up to a transverse momentum of 2 GeV. With no additional parameters, fair agreement is found with available pion–nucleus data, confirming the virtues of the NJL description of pion parton structure. We find sizable evolution effects on the shape of the distributions and on the generated average transverse momentum of the dilepton pair. We furthermore discuss the possibility of gaining information about the behavior of the pion unpolarized transverse momentum dependent parton distribution from pion nucleus DY data.
Introduction
The non perturbative transverse structure of hadrons has attracted recently much attention and the issue of extracting transverse momentum dependendent parton distributions (TMDs) from data taken in different processes in present and forthcoming high-luminosity facilities represents an important goal of nowadays hadronic Physics. In particular, Drell-Yan (DY) pair production [1], discussed in this paper, and a e-mail<EMAIL_ADDRESS>semi-inclusive deep inelastic scattering are the main processes under investigation [2].
The cross section for DY pair production, differential in the transverse momentum of the pair, q T , is a particularly suitable observable for this kind of studies. In particular at small q T , where the TMD formalism is formulated, fixed order calculation of this process show large logarithmic corrections due to an incomplete cancellation of soft and collinear singularities between real and virtual contributions and need to be resummed to all orders to recover the predictivity of the theory [3][4][5].
The description of the q T DY spectrum in pp collisions has reached a high degree of sophistication [6]. On one side, theoretical improvements have increased the perturbative accuracy of the predictions [7][8][9][10][11]. On the other side, global fits of DY production at different energies have given access to the non perturbative proton transverse structure [12,13]. Both aspects have received increasing attention due to the formalisation of new and old concepts in the TMD language [14][15][16][17]. While there are differences between the language used in the modern and the older TMD approaches, physical results should not depend on it. A detailed comparison of the formalisms can be found in Refs. [18,19].
At high energy colliders, this improved knowledge aims to an increasingly better description of electroweak bosons production, with the Higgs q T spectrum being the highlighted case. Measurements of q T spectrum of the DY process, at lower centre of mass energies, are instead more sensitive to the hadronic non perturbative transverse structure.
DY pair production in pion-nucleus scattering is a unique probe of pion parton distribution functions (PDFs) and, as such represents a source of information on the pion parton structure. In particular for the q T spectrum this was realized long time ago by the authors of Ref. [20]. More recently, phenomenological analyses have appeared [21]. A fit to the q T spectrum of DY pairs produced in pion-nucleus collisions has been recently presented in Ref. [22].
Pion TMDs, which could be extracted in principle in a next generation of pion-nucleus DY experiments [23], have received recently considerable theoretical interest [21,[24][25][26][28][29][30][31]. In this paper we study the DY unpolarized pair production in pion-nucleus scattering, to next-to-leading logarithmic (NLL) perturbative accuracy, up to a transverse momentum of the produced lepton pair of 2 GeV. As non perturbative inputs, we use, for the bound nucleons, a longitudinal structure which takes into account nuclear effects and, for the transverse structure, a well established parameterization obtained through a phenomenological fit to proton-proton DY data (called, from now on, KN05 prescription) [12]. For the pion, we use TMDs obtained in a recent calculation [30], within a Nambu-Jona Lasinio (NJL) framework [32], with Pauli-Villars regularization. The corresponding RGE scale of the model is determined in a novel way by comparing the DY unpolarized cross section, integrated over q T , described by evolved pion PDFs evaluated in the NJL model to the data.
The aim of the present paper is to study the performances of the NJL model, widely used to describe the nonperturbative meson structure, against DY differential cross section data for the first time. We also analyze to what extent this process can be used to obtain information on the pion transverse structure in momentum space, as it happens for the proton in the corresponding process.
The paper is structured as follows. In the next section, we present the set-up of the calculation and introduce the ingredients used to describe the proton and pion structure. In the third section, we discuss the results of the calculation of DY cross sections in the kinematics of presently available data for pion-tungsten scattering. Eventually, we draw our conclusions in the last section.
Drell-Yan cross section
In the following we will be interested in the process of the type in which a virtual photon is produced with large invariant mass Q 2 and transverse momentum q T in the collisions of two hadrons at a centre-of-mass energy s = ( p 1 + p 2 ) 2 , with p 1,2 the four momentum of hadrons h 1,2 , respectively. When q 2 T becomes small compared to Q 2 , large logarithmic corrections of the form of α n s log m (Q 2 /q 2 T ) with 0 ≤ m ≤ 2n − 1 appear in fixed order results, being n the order of the perturbative calculation. These large logarithmic corrections can be resummed to all orders by using the Collins-Soper-Sterman (CSS) formalism [6]. In this limit, of interest for the present analysis and neglecting finite corrections in the q T ∼ Q region, the cross-section can be written as where b 0 = 2e −γ e , the symbol ⊗ stands for convolution and σ is the leading-order total partonic cross section for producing a lepton pair, σ (qq → l + l − ), and it is given by In Eq. (2), the a, b indices run on quark and gluons, J 0 (b q T ) is the Bessel function of first kind and f i/ h corresponds to the distribution of a parton i in a hadron h. The cross section in Eq. (2) is differential in τ = Q 2 /s and y, the rapidity of the DY pair. Momentum fractions appearing in parton distribution functions can be expressed in terms of these variables as Cross sections differential in x F = x 1 − x 2 = 2q / √ s, the longitudinal momentum of the pair in the hadronic centre of mass system, can be obtained from those differential in rapidity y by a suitable transformation. By defining A = x 2 F + 4τ one gets Momentum conservation further imposes that |x F | < 1 − τ . The large logarithmic corrections are conventiently exponentiated in b-space in the Sudakov perturbative form factor The functions C ab in Eq. (2) and A, B in Eq. (6) have perturbative expansions in α s , At present, the perturbative Sudakov form factor can be evaluated at next-to-next-to-leading logarithmic (NNLL) accuracy [11]. In the qq annihilation channel pertinent to Drell-Yan production, the evaluation of the Sudakov form factor at next-to-leading logarithmic (NLL) accuracy, the one reached in the present analysis, involves the coefficients which are the coefficient of the singular (1−z) −1 and δ(1−z) terms of the one-loop splitting function P (0) qq (z) and which is the coefficient of the singular term of the two-loop splitting function P (1) qq (z) in the z → 1 limit [34]. The general expression for C (1) ab are given by [11,35] Color factors in the previous equations are given by C A = 3, C F = 4/3, T R = 1/2 with n f being the number of active flavours. Together with the use of NLO pdfs, this guarantees the evaluation of the cross section at small q T at NLL accuracy. The last ingredient in Eq. (2) is the non perturbative form factor, S h 1 h 2 N P (b), which encodes the transverse structure of both the colliding hadrons. The latter is either fixed by comparison with data or parametrized with the help of hadronic models, as we shall do in this paper.
Proton structure
Predictions for the transverse momentum spectrum of DY pairs produced in pion-proton collisions do rely on the knowledge of the proton NP form factor. The latter is extracted from the transverse momentum spectrum of DY pairs produced in proton-proton ( pp) and proton-nucleus ( p A) collisions. Quite recent analyses [15,16] have appeared which address such an extraction. Since our aim here is to establish the possibility of studying the pion transverse non perturbative structure in pion-nucleus DY experiments, we here intend to minimize the uncertainity coming from the proton structure part of the calculation. We use the well known and widely accepted results of Konychev and Nadolsky (KN05) [12] obtained within the CSS formalism [6] where S pp N P (b) is extracted from global fit to Z -boson and low mass DY data, updating the results presented in Ref. [13]. The latter is parametrised as S pp N P (b) (12) = exp{−[a 1 + a 2 ln(M/(3.2 GeV)) + a 3 ln(100x 1 The a i parameters appearing in Eq. (12) are determined by a minimisation procedure against data and are given by [12] a 1 = 0.201 ± 0.011, a 2 = 0.184 ± 0.018, The fit is fully specified once a prescription for the treatment of the non perturbative, large-b, region both in the Sudakov form factor, Eq. (6), and the parton distributions is given.
The authors of Ref. [12] adopt the so-called b -prescription, and setting b max = 1.5 GeV −1 in the perturbative form factor. In principle, the same setting should be used in PDFs, which are evaluated at the factorisation scale However this choice for b max may imply a call to a specific PDFs parameterization below their lowest available scale, Q in . Since in Ref. [12] cross sections are evaluated with the NLO CTEQ6M PDFs [36], whose lowest Q accessible is Q in = 1.3 GeV, the b -prescription entering PDFs calls is used with b max = b 0 /Q in 0.86 GeV −1 which always guarantees μ F > Q in . It is important to remark that the non perturbative form factor is determined not only by fitting the parameters of the chosen functional form, but also by the specific regularisation prescription and its associated parameters adopted to deal with the infrared region. In general all these ingredients have been found to be highly correlated.
In order to present a benchmark of our code and to gauge how theory performs in extrapolation regions, we compare predictions from KN05 to the p A data of Ref. [37]. An additional ±25% normalisation error is assigned to the data [37]. In the original KN05 analysis, only the data at p lab = 400 GeV, q T < 1.4 GeV, 5 < M/GeV < 9 were included in the fit. In such a restricted region indeed the theory (solid lines) performs well offering a good benchmark of our code, as shown in the first row of Fig. 1. Since the π W data to be analyzed in the following are at p lab =252 GeV, it is important to check how well the theory performs in extrapolation regions at lower √ s and higher DY rapidity. Therefore we present in the second and third rows of Fig. 1 the KN05 benchmark (dashed lines) versus data [37] at p lab = 200 and 300 GeV, which were not included in the KN05 fit. By using Eq. (4) and Eq. (5) and assuming the invariant mass values indicated on the plots, the rapidity coverage of these data can be con- verted to the range 0 < x F < 0.3. In both cases we find good agreement between data and theory up to q T ∼ 2 GeV giving us confidence that the KN05 model can be successfully used in this (x F , q T ) range at the √ s of interest in this analysis.
Pion structure
A calculation of pion TMDs in a NJL framework, with Pauli-Villars regularisation, has been recently presented in Ref. [30]. Model calculations of meson partonic structure within this approach have a long story of successful predictions [38][39][40][41][42][43]. Collinear parton distributions obtained within a model have to be associated to a low momentum scale Q 2 0 and, in order to be used to predict measured quantities, have to be evolved to higher momentum scales according to perturbative QCD (pQCD).
In Ref. [30] the unpolarized NJL parton TMD has been obtained. Among its good properties, we stress that, upon integration over the intrinsic quark transverse momentum k T , the pion PDF q(x) is properly recovered with correct normalisation and the momentum sum rule is exactly satisfied. This is due to the fact that NJL is a field theoretical scheme and the correct support of the PDF, 0 ≤ x ≤ 1, is not imposed but arises naturally. In particular, the momentum sum rule reads dx x q (x) = 0.5, i.e. the fraction of momentum carried by each quark is one half of the total momentum, since at the scale of the model only valence quarks are present. The dependence on k T of the TMD obtained in Ref. [30] is very important for the present study. It is worth stressing that, in this approach, the k T dependence is automatically generated by the NJL dynamics and it is not imposed by using any educated guess. This is an important feature of the results of Ref. [30], not found in other approaches [21,27]. In this paper we will use the pion TMD obtained in Ref. [30] in the chiral limit, the latter allowing for a factorisation of the x and k T dependence at the low but undetermined scale Q 2 0 associated to the model: where one has (in π − , of interest here): The function T is given by which, due to a proper combination of the c i [30], behaves as k −6 T for asymptotic values of k T = |k T | and satisfies the normalisation Since the distribution Eq. (17) depends only upon k 2 T , its Fourier transform can be cast in the form where K 0 is the modified Bessel function of the second kind. The parameters used in Eq. (19) are given in Ref. [30] and read We notice that, beyond the chiral limit, the factorised expression (15) is violated only slightly due to a non trivial xdependence (see, e.g., Ref. [46]). We have therefore used the expression valid in the chiral limit to avoid further irrelevant complications in an evaluation which is already rather involved.
As noted above, the NJL pion model corresponds to a low hadronic scale Q 2 0 . Such a low scale has been determined previously by directly comparing the second moment of the pion PDF evaluated in NJL model with the results from the analysis of Ref. [48]. The procedure gives a value of Q 2 0 = 0.18 GeV 2 at NLO [45,46]. 1 In the present paper we use a different strategy: we consider Q 2 0 a free parameter of the NJL model which is then fixed with a minimisation procedure, outlined in the following, of the theoretical π − W DY cross sections, differential in √ τ and x F , against the corresponding experimental ones [44]. Theoretical cross sections are calculated according to where the partonic cross sections dσ i j are calculated at NLO accuracy by using the results of Ref. [48]. An additional correction takes into account the correct number of gluon polarisations in the MS in dimensional regularisation [49]. The NJL pion PDFs are evolved to NLO accuracy in the Variable Flavor Number Scheme, with the initial condition given in Eq. (16), with the help of the QCDNUM [50] evolution code. The QCD parameters are those of the NLO CTEQ6M parameterisation [36]. In particular we set the NLO running coupling to α (n f =5) s (M Z ) = 0.118 at the Z -boson mass, M Z . Since the data we are comparing to are obtained on a tungsten target, we take into account nuclear effects by using nuclear PDFs of Ref. [51]. We have carried out a χ 2 study to establish the hadronic scale of the model that describes the best the data at NLO in pQCD. Two cases have been considered: an evaluation of the χ 2 for the full range of x F and another one with a cut x F < 0.4, since the NJL model is expected to better reproduce the pion valence distributions, expected to populate the range of large and positive x F . The scales thus determined are 2 Drell-Yan pairs production in π − W collisions. Next-to-leading order cross sections obtained by using evolved NJL pion PDFs for three values of Q 2 0 are compared to data of Ref. [44] scale associated to the pion NJL model. The other two curves in Fig. 2, corresponding to Q 2 0 = 0.19 GeV 2 and Q 2 0 = 0.25 GeV 2 respectively, are added, in order to show the sensitivity to this particular choice of infrared Q 2 0 . It is worth noticing that the results show an acceptable agreement, both in shape and in normalisation. More in detail, a tendency of the theory to undershoot the data is identified in the range of small x F (−0.2 < x F < 0.2). This deficiency is not unexpected since, in the mentioned kinematic region, the dominant contribution to the cross sections involves sea quarks and gluons which are absent at Q 2 0 and are radiatively generated by QCD evolution. This is a typical drawback of models which contain only valence contributions at the hadronic scale. At this point we would like to mention that the theoretical description of the x F -spectra at large x F and the determination of pion parton distributions can be further improved employing resummation techniques presented in Refs. [49,52,53]. It is worth noticing that, as shown in those papers, threshold NLL resummation of the Wilson coefficients leads to larger cross sections at large x with respect to NLO ones. This, in turn, implies softer pion PDFs at large x. In the present context, this fact would imply a scale Q 2 0 for the NJL model lower than the one already determined by using NLO Wilson coefficients in Eq. (20).
Predictions for π W collisions data
Predictions for the π W Drell-Yan cross sections are obtained once appropriate modifications are implemented in Eq. (2). Evolved NJL pion parton distributions replace proton PDFs for hadron 1. Moreover the non-perturbative form factor S h 1 h 2 N P (b) depends on the particle species initiating the reaction. Therefore in π W collisions the latter is written as follows: where S π N P (b) is given in Eq. (19) and the square root on S pp N P (b), given in Eq. (12), takes into account that now only one proton is involved in the process. It is instructive to directly compare the proton and pion non perturbative transverse distributions used in the calculation. It is important to remark that the NJL pion transverse distribution in Eq. (19) differs from the corresponding proton factor in Eq. (12) in that it does not contain any explicit dependence neither on hard scale M nor on parton fractional momenta. Such a comparison is meaningful at the typical scale for which the transverse form factors and the longitudinal momentum part factorize. For the pion case this happens at the scale Q 2 0 determined in the previous section. For the proton TMD such a scale is ambiguously defined and, according to KN05 analysis, ranges between Q 2 in and (b 0 /b K N05 max ) 2 . Therefore we choose M = Q in = 1.3 GeV in Eq. (12) and fix the product x 1 x 2 = M 2 /s, see Eq. (4), exploiting the π − W kinematics with s calculated according to a beam energy of p lab = 252 GeV. This comparison is presented in Fig. 3, where our result for the pion non-perturbative form factor, S π N P (b), is also compared to the parametrisation of the nonperturbative pion form factor of Ref. [22] (called hereafter WLS). The approach of Ref. [22] is rather different from ours, both in the spirit and in the physical ingredients used. As a matter of fact, in that paper the proton non-perturbative form factor, S pp N P (b), has a structure similar to that of our Eq. (12) and for the pion the same form has been assumed, with the corresponding parameters obtained from a fit of the same cross section data used in the present paper. As a result, the pion non-perturbative form factor depends on both the hard scale and the parton momenta. A fit is then performed up to q T 3 GeV. We reiterate that our goal here is not to fit but rather to assume a well known structure for the proton nonperturbative form factor and to test the pure NJL predictions for the pion against the data. The purpose of the comparison with the WLS parametrisation is therefore mainly illustrative and quantitative conclusions can be hardly reached. All the distributions presented in Fig. 3 reduce to unity in the b → 0 limit, since they are all normalised to unity in transverse momentum space. In the top panel of Fig. 3 we compare the NJL transverse distribution to the pion parametrisation of Ref. [22] (called hereafter WLS) obtained from a fit of the same cross section data used in the present paper. One may notice that, for this model, the width of the distribution is smaller with respect to the NJL one, implying a larger average transverse momentum. In the bottom panel of Fig. 3 one may notice that the NJL pion transverse distribution develops a larger tail with respect to the gaussian drop of the proton distributions. Moreover the b-space width of the KN05 proton with M = 1.3 GeV is larger with respect to the pion one. When transformed back in k T space, this implies that the intrinsic transverse momentum in the pion is larger than the one in the proton, in agreement with the general expectations, since the pion is a much smaller system with respect to the proton. It is worth mentioning that both the KN05 and WLS non perturbative form factors have an explicit, althought slightly different, dependence upon the hard scale M, in both cases set equal to the invariant mass of the dilepton pair. Therefore we plot in each panels, as a representative case, the curves corresponding to both form factors evaluated with the scale set to M = 4 GeV. Comparing the latter curves to the ones with M ∼ 1 GeV, we conclude that the M-dependence generates a sizable non perturbative evolution of the form factor which is more pronounced for KN05 proton model than for the WLS pion model.
We now turn to the discussion of the perturbative part of the Sudakov form factor, Eq. (6). The latter, at variance with its non perturbative counter part, does not depend upon the type of initial state hadrons involved in the scattering process. In principle, the same regularisation procedure should be used both in the Sudakov and in the PDFs. This optimum indeed faces some technical problem, for example the call to PDFs to values outside the boundary of the grid in which they are defined and the different scales at which the transverse distributions are assumed to factorise on the proton and pion side, respectively. In order to accomodate all these different settings, we find useful to split the perturbative form factor in Eq. (6) in a form which allows to use distinct b max on the proton and pion side: For the proton parameters, we stick to KN05 settings since the a i 's in Eq. (12) optimized for b p max = 1.5 GeV −1 . On the pion side there is some freedom in adjusting b N J L max . However the pion TMD shows a x − k T factorised structure only at Q 2 0 , whose numerical value has been determined in the previous section. Therefore we can expect the b * -prescription to involve b π max values of the order b 0 /Q 0 ∼ 2.44 GeV −1 , which will be our default value to be used both in the Sudakov and in NJL pion parton distributions regularisation. We now turn to the comparison to lepton pair q T -spectra collected in tables D92-D97 of Refs. [44,54], measured in π W collisions. Such data actually refer to less differential cross sections with respect to the one appearing in Eq. (2). In this case differential cross sections are integrated over additional variables according to values specified in experimental analyses. We start presenting our results showing, in Fig. 4, cross sections differential in q T integrated in 0 < x F < 1 in various bins of the invariant mass of the pair, M. The comparison is performed up to a q T ∼ 2 GeV, where we have checked that the KN05 gives an adequate description of pp data. All three different predictions, to be discussed in the following, capture the normalisation of the data and share a tendency to slightly overestimate the data at very small q T and to underestimate them at larger q T . This effect progressively disappears increasing the mass of the lepton pair. Comparing the two curves corresponding to b N J L max = 2.44 GeV −1 and b N J L max = 1.5 GeV −1 , one may notice a substantial stability upon variation of the regulators on the pion side. On the same plot, in order to investigate the sensitivity to the pion transverse distribution, we additionally show the predictions obtained by substituting the pion transverse factor, Eq. (19), with S pp N P (b). As shown in Fig. 3, the non perturbative transverse distributions for the proton and pion differ at low scales. The corresponding curve, indicated with pp on the plot, is barely distinguishable from the other two. Such a comparison supports the hypothesis that the effect of the perturbative evolution, driven by Eq. (23), is to wash away differences in the non perturbative structure found at the hadronic scale. This result implies a reduced sensitivity to non perturbative structure. A quantitative analysis of the results yields that the χ 2 /d.o. f. of the shown distributions slowly decreases from values of order 4 to values of order 1 with increasing M from 4 GeV to 13 GeV. 2 These numbers are remarkably good if one considers that we are presenting pure model predictions without a fitting of parameters a posteriori. Besides, we observe that the agreement with data of the NJL distributions is slightly better than those obtained with a proton-like non-perturbative form factor for the pion, in any M bin, and, more importantly, that the difference in χ 2 /d.o. f. reaches 30 % for the lowest values of M. The region of low (but still perturbative) M is therefore selected as the most promising to access non-perturbative details of the pion transverse structure.
We proceed our discussion presenting in Fig. 5 the comparison between theory predictions against the same data, now integrated in the mass range 4 < M < 8.55 GeV in a number of x F bins. We remind the reader that we have verified that the KN05 model gives a satisfactory description of pp data up to x F ∼ 0.3. Up to this x F value, as shown in the first row of Fig. 5, the description π − W data is fair, as already observed in Fig. 4. Beyond that range, however, the width of the theoretical curves decreases more rapidly than observed in the data, with data substantially undershooted beyond q T ∼ 1 GeV. This effect is more pronounced as x F increases. In this region of relatively large pion fractional momenta it would be tempting to invoke, in order to describe the data, an x-dependent non perturbative structure. Such an interesting hypothesis, however, cannot be tested unless fixed order contributions at finite q T are included in the calculation. Moreover, given our working assumptions, the failure to agree with data at large values of q T was somehow expected. Whether it is due to the breakdown of the x −k T factorization, the inclusion of a more complex NP Sudakov form factor or the matching with the so called Y-term, further studies shall be pursued to answer that question.
On the theoretical side, we would like to mention that, in this range of quite large pion parton fractional momenta, the theoretical description of the q T -spectrum can be further improved employing joint resummation techniques described in Refs. [55,56]. In order to better appreciate how the width of theoretical predictions evolves with x F (and therefore with x π ) and the invariant mass of the lepton pair, we show in Fig. 6 Integration limits are provided by experimental conditions. For data, indicated by black lines in Fig. 6, the phenomenological parametrisation presented in Ref. [44] is used. For both theory and data, the q 2 T is calculated with a maximum value of q max T = 2 GeV. Theory predictions tend to undershoot the data but, overall, a good shape agreement is found. By comparing lines with and without TMD evolution (for the latter the perturbative Sudakov S q is removed from the evaluation of Eq. (2)) one can appreciate its large impact on the amount of generated q T . On the same plot, in order to investigate the sensitivity to the pion transverse structure, we additionally show the predictions obtained by substituting the pion transverse factor, Eq. (19), with S pp N P (b). As already seen in Fig. 4, differences are minimal, implying a reduced sensitivity to details of the non perturbative transverse factor. Therefore if one aims to better appreciate the strictly non perturbative form factor, one has to confine in corners where TMD evolution is minimised, but still in a perturbative range. These phase space regions can be identified by extrapolation from the right plot as the one at the lowest, but still perturbative, values of the invariant masses of the pair, as already noticed above while discussing Fig. 4.
Conclusions
A thorough analysis of DY pair production in pion-nucleus scattering has been presented. The main goal of our work has been the test of model predictions, obtained within the Nambu-Jona-Lasinio model for the transverse pion structure. In particular we have focused on the study of differential transverse momentum spectra of DY pairs produced in p A collisions calculated in the CSS framework at NLL accuracy borrowing from the literature the longitudinal and transverse proton structure. The pion is treated in the Nambu-Jona-Lasinio model. No further assumption has been made: even the momentum scale associated to the model is obtained via a minimization procedure of NLO theory to DY experimental longitudinal spectra. The latter turns out to be a low one, in line with that normally used, which could be predicted within the spirit of the model without fitting "a posteriori". The agreement found between our pion-nucleus theoretical cross sections and experimental data is rather successful, confirming the predictive power of the NJL model, for both the longitudinal pion parton distributions and its transverse structure. We notice that the theory tends to systematically undershoot the data on the higher end of the considered q T interval. All interpretations of this effect, however, are not conclusive without the inclusion of the finite, fixed order, contributions which populate the q T ∼ Q region and are neglected in our calculation. The possibility to distinguish between different non perturbative transverse momentum distributions in DY data appears instead more questionable. In this complicated sce- T as a function of M integrated in the range 0 < x F < 1. Averaged values are obtained integrating both predictions and the phenomenological parametrisation of the data up to q max T = 2 GeV nario, a possible strategy would be the measurement of DY pion-nucleus q T -spectra, in bins of x F , at low values of the mass of the pair, as the present study suggests to look into this kinematical window to emphasize the non-perturbative content of the pion. Further analyses of the pion non-perturbative form factor, as a function of the hard scale, should be pursued so we could progress on that point. In the very same window, new data could allow a deeper investigation of the dependence of the non perturbative form factor upon the hard scale of the process. | 7,873.6 | 2018-08-01T00:00:00.000 | [
"Physics"
] |
Triplet-Superconductivity in Triple-Band Crossings
Multi-band superconductivity in topological semimetals are the paradigms of unconventional superconductors. Their exotic gap structures and topological properties have fascinated searching for material realizations and applications. In this paper, we focus on triple point fermions, a new type of band crossings, and we claim that their superconductivity uniquely stabilizes spin-triplet pairing. Unlike conventional superconductors and other multi band superconductors, such triplet superconductivity is the novel phenomena of triple point fermions where the spin-singlet pairing is strictly forbidden in the on-site interaction due to the Fermi statistics. We find that two distinct triplet superconductors, characterized by the presence and absence of time-reversal symmetry, are allowed which in principle can be controlled by tuning the chemical potential. For the triplet superconductor with time-reversal symmetry, we show that topologically protected nodal lines are realized. In contrast, for time-reversal broken case, the complication of topologically protected Bogoliubov Fermi surfaces emerges. Our theoretical study provides a new guidance for searching triplet superconductivities and their exotic implications.
Multi-band superconductivity in topological semimetals are the paradigms of unconventional superconductors. Their exotic gap structures and topological properties have fascinated searching for material realizations and applications. In this paper, we focus on triple point fermions, a new type of band crossings, and we claim that their superconductivity uniquely stabilizes spin-triplet pairing. Unlike conventional superconductors and other multi-band superconductors, such triplet superconductivity is the novel phenomena of triple point fermions where the spin-singlet pairing is strictly forbidden in the on-site interaction due to the Fermi statistics. We find that two distinct triplet superconductors, characterized by the presence and absence of time-reversal symmetry, are allowed which in principle can be controlled by tuning the chemical potential. For the triplet superconductor with time-reversal symmetry, we show that topologically protected nodal lines are realized. In contrast, for time-reversal broken case, the complication of topologically protected Bogoliubov Fermi surfaces emerges. Our theoretical study provides a new guidance for searching triplet superconductivities and their exotic implications.
The discovery of topological semimetallic phases have realized various types of new quasiparticles, characterized by topologically non-trivial band crossings. These quasiparticles are particularly interesting as its low-energy effective theory can mirror relativistic elementary particles. The representative examples are Dirac and Weyl semimetals mimicking the relativistic massless spin-1/2 fermions.[1-4] They have gathered great interests due to the connection to the high-energy physics. Moreover, the condensed matter systems can realize even more exotic kinds of quasiparticle excitations that has no analogue of elementary particles in high energy physics. [5][6][7][8][9][10][11][12][13][14][15] Especially, recent studies show that the triple-band crossings are also realized at high-symmetry points [12,13] or on high-symmetry lines [15][16][17][18][19][20][21] stabilized by spatial symmetries. These quasiparticle excitations, referred to as triple point fermions, can carry the effective integer spin-1 since it is not constrained by the spin-statistics theorem. One representative low-energy theory, which captures triple point fermions, is characterized by linear band touching of two spin polarized bands with the Chern number ±2 and the existence of additional middle band with the trivial Chern number. [12] Such peculiar spin structures and energy dispersions of triple point fermions can have major impact on the nature of the correlated ground states in the presence of the many-body interactions. In particular, the possible unconventional superconducting states calls for concrete theoretical understanding. [22] The area of the unconventional superconductivity is characterized by the non-trivial pairing symmetries of the superconducting order parameters. Especially, it has recently been proposed that multi-band systems offer a new platform to achieve unconventional superconductivity. [23][24][25][26][27][28][29][30][31][32] One example of such multi-band system is spin-orbit coupled j = 3/2 system where possible realization of Cooper pair with total spin S = 2, 3 has been investigated. [32][33][34][35][36][37][38][39][40] Moreover the system is known to offer generic routes to achieve pairing instabilities towards such unconventional superconductivity employing inter-band pairing channels. [41,42] Despite growing interests in multi-band system, there have been few studies of superconductivity in j = 1 system where Cooper pair with S = 0, 2 are forbidden to have even spatial parity by Fermi statistics.
In general, the spatial parity and the total spin of superconducting order parameter are not independent of each other. The Fermi statistics constrains them to be anti-symmetric under the exchange of two identical electrons forming a Cooper pair. For instance, the conventional s-wave BCS superconductors must be spin-singlet pairing, while the triplet superconductivity can only be realized with odd-parity order parameters. However, this scenario can be drastically changed and, indeed, inverted if we consider the pairing of pseudospin j = 1 electrons. Specifically, the spin-singlet of the two composite spin j = 1 fermions is symmetric under the exchange of the two spins. Accordingly, the formation of the spin-singlet pairs with even-parity is strictly forbidden, but the spin-triplet pairing is only allowed. This unique property of pseudospin j = 1 electron motivates us the further investigations for the hunt of the new form of unconventional superconductivity.
In this work, we propose possible triplet superconducting ground states of triple point fermions. Using the Landau theory of the superconductivity, we discuss two distinct superconducting phases in the presence of SO(3) symmetry. They are characterized by distinct spin textures of the superconducting order parameters. Based on one-loop calculation, we find the time-reversal symmetric triplet pairing is energetically favored when the chemical potential lies far below the tripleband crossing point with the middle band having upward dispersion, which we refer to as '(s z )' state. In this case, the triplet superconductors contain topologically protected nodal lines. On the other hand, with the chemical potential lying near or above the band crossing point, the middle band participates to pair with the other bands. This state breaks the timereversal symmetry and the resulting spin texture of the order parameter resembles 3D chiral p x +ip y superconductor, so we refer to it as '(s x +is y )' state. In this case, multiple Bogoliubov Fermi surfaces with finite Chern numbers emerge. Such unusual triplet superconductor is the generic feature of triple point fermions and can be controlled by tuning the chemical potential.
To begin our discussion, we consider the SO(3) and timereversal symmetric low-energy theory of triple point fermions residing in two inequivalent valleys, which are time reversal partners. [12] Up to the quadratic order, the Hamiltonian can be expanded near the band crossing point as, (1) Here, we define the three spinor as ψ ±,k = (ψ ±,k,1 , ψ ±,k,0 , ψ ±,k,−1 ). The first subscript ± and the second subscript ±1 and 0 indicate the valley and the spin degree of freedom respectively. J = (J x , J y , J z ) represents the j = 1 angular momentum matrices. They are explicitly written as, I 3 is the three-dimensional identity matrix. v and µ are the effective linear velocity of the band crossings and the chemical potential respectively. vk · J term breaks inversion symmetry. c|k| 2 term represents possible bending of bands and we assume c > 0 without loss of generality. The Hamiltonian in Eq.
(1) has two bands with the dispersion, ±1 (k) = ±v|k|+c|k| 2 −µ, having opposite spins, and the middle band with the dispersion, 0 (k) = c|k| 2 − µ. These three bands can be characterized by the monopole charge of the Berry curvature, C ±1 = ∓2 and C 0 = 0 respectively. Prior to the description of the microscopic interactions, we first discuss the generic form of the allowed pairing order parameters. The pairing order parameters can be written as the sum of bilinear form, ψ † +,k g(K +k)M S γψ * −,−k , where the function g(K+k) describes the orbital part of the Cooper pair and γ = e −iπJy is the unitary part of the time-reversal operator T = γK (K is the complex conjugate operator). Here, K indicates the position of the valley + in momentum space. The matrix M S specifies the total spin, S, of the Cooper pair and is listed in Table I. Composite of two j = 1 spins can generate total spins up to spin-singlet (S = 0), triplet (S = 1) and quintet (S = 2). Fermi statistics forces the order parameter to satisfy According to this condition, the even-parity pairings (g(−K −k) = g(K +k)) only allow spin-triplet pairing while the odd-parity pairings (g(−K − k) = −g(K + k)) allow spin-singlet and quintet pairings. Such combinations of the spatial parity and the spin shows exactly the opposite pattern from superconductors which are comprised of spin half-integer electrons. This is the key observation of our work. As a result, the superconducting state driven by the on-site interactions must be spin-triplet pairing state. In addition, it is worthwhile to note that this behavior is different from the spin polarized superconductors described by an effective j = 0. For spin polarized case, the system requires dominant further-neighbor interactions to induce the triplet pairing, which is clearly distinct from our case.
Fermi statistics allow only the MS=1 locally (momentum independent). While MS=0,2 should be additionally multiplied by an odd power of momentum to satisfy Fermi statistics. The column "Even,Odd" indicates that the spatial parity of the superconducting order parameter. Γ matrices are Motivated by the above discussions, we now consider the following form of the generic on-site interactions, The interaction terms constitute a complete set of on-site interactions with SO(3) symmetry [43] and correspond to the interactions between p-wave-orbital densities since ψ † r J a ψ r transforms as p-wave orbitals. Here, we consider repulsive on-site interactions, g > 0. To rewrite the interaction into pairing channels, we use the Fierz identity for electrons with pseudospin j = 1. [41,[43][44][45][46] The particle-hole channel interactions in Eq. (3) are exactly rewritten into the form of the pairing channels as following, We now find that there exists superconducting instability even when the on-site interactions are all repulsive (g > 0). Based on the pairing interaction in Eq. (4), we now derive the Ginzburg-Landau(GL) free energy, F ( ∆, T, v, µ, g), as a function of the order parameter, ∆ = (∆ x , ∆ y , ∆ z ), where ∆ a = ψ T r (J a γ) † ψ r corresponds to the s-wave spin-triplet pair with total spin S = 1. By integrating out the electronic degrees of freedom, the free energy functional is written as, where the matrix elements of I a is given as (I a ) bc = i abc where abc is the Levi-Civita symbol. [27,47] We find that Eq. (5) can have the two possible superconducting ground states solely depending on the value of the coefficient q 2 . When q 2 > 0, time-reversal symmetric state with the order parameter ∆ = (0, 0, 1) is stabilized. For q 2 < 0, however, timereversal broken state is stabilized with the order parameter, ∆ = (1, i, 0). We note that the complex order parameters ∆ transformed under SO(3) rotation are physically equivalent to the ones mentioned above. From now on, we call ∆ = (0, 0, 1) and (1, i, 0) states as (s z ) and (s x +is y ) states respectively. The (s z ) and (s x +is y ) order parameters have distinct spin textures. The M S matrix of (s z ) state is explicitly given as, From the explicit form of the above matrix, we can observe that the (s z ) state forms Cooper pairs with opposite spin components, using inter-band pairing. On the other hand, the M S matrix of (s x +is y ) state is given as, For (s x + is y ) state, if we consider an electron with j z = 1, it makes inter-band pairing with a j z = 0 electron. Similarly, (s x −is y ) state pairs a j z = −1 electron with a j z = 0 electron. By explicitly investigating the sign of q 2 within the leading one-loop calculation, one can determine the energetically favored state (See Supplementary Material for details). Fig.1 shows the calculated phase diagram as a function of dimensionless parameters, µ/T and v/ √ T , while keeping a dimensionless parameter c = 1/10 with the momentum cutoff Λ = µ/v. First of all, when the chemical potential lies far below the band crossing point, we find that the (s z ) state is favored preserving time-reversal symmetry. However, when the chemical potential approaches to the band crossing point, the contribution of the middle band to the free energy become significant. We find that (s x +is y ) pairing is stabilized with the chemical potential lying near or above the band crossing point. In the limit where the middle band is perfectly flat (c = 0), our one-loop calculation with momentum cutoff Λ = µ/v shows that q 2 is always negative, favoring (s x +is y ) state (See Supplementary Material for details).
After constructing the Landau theory and phase diagram of the superconductivity, we now discuss the Bogoliubov-de Gennes (BdG) quasiparticle spectrum of the superconducting states. The BdG Hamiltonian reads where Ψ † k = (ψ † +,k , ψ −,−k ) andĥ(k) = (ck 2 − µ)I 3 + vk · J . Here,∆ = |∆|J z γ for the (s z ) state and∆ = |∆|(J x + iJ y )γ for the (s x + is y ) state where ∆ is a real constant. For the time-reversal symmetric superconductor with (s z ) pairing, H BdG (k) belongs to class BDI [48] and the spectrum is derived by the singular value decomposition of the following matrix,ĥ(k)+i|∆|J z = v(k x , k y , k z +i|∆|)·J +(ck 2 −µ)I 3 .
The corresponding eigenvalues are given as, For the momentum point k where λ s (k) = 0 is satisfied, the BdG energy spectrum become gapless. We find that λ ±1 (k) = 0 if k z = 0 and c|k| 2 − µ = ∓ v 2 |k| 2 − |∆| 2 . These conditions define the two nodal lines when µ > c∆ 2 /v 2 − v 2 /4c. Fig.2 (a) show the two nodal rings, which are represented by solid and dashed line. The nodal rings are topologically protected by non-trivial winding number, ω ∈ Z, thus they are stable against any symmetry preserving perturbations. The winding number can be calculated as ω = 1 2π s 2π 0 ∂ θ arg(λ s ), where the integration is taken along the loop that encircles each nodal line. [48,49] We immediately find that the winding number of each solid and dashed nodal ring is 1 and −1, respectively. Similarly, the condition, λ 0 (k) = 0, defines the nodal surface, often referred to as Bogoliubov Fermi surface. Unlike the nodal lines, the Bogoliubov Fermi surface, characterized by λ 0 (k) = 0, is topologically trivial. This can be seen by including the additional oddparity spin-singlet superconducting order parameter, which instantly gaps out the system. As a consequence, we expect topologically stable nodal lines for the time-reversal symmetric (s z ) phase.
We now consider the gap structure of the time-reversal broken (s x +is y ) state. In this case, our system belongs to class D [48] and the gapless region can be calculated by finding k points which satisfy det[H BdG (k)] = 0. This condition can be rewritten as, The above single condition generally defines the surface in the three-dimensional momentum space. It realizes nodal surface in the BdG energy spectrum, which is now referred to as Bogoliubov Fermi surface (See Fig.2 (b)-(d)). This Bogoliubov Fermi surface can be characterized by two distinct topological invariants. [48] The first is the Z 2 valued number of occupied BdG bands. Each Bogoliubov Fermi surface is non-degenerate since the time-reversal symmetry is absent. This indicates that these surfaces are all topologically protected, since Z 2 number always changes by 1 as the energy level cross the single Bogoliubov Fermi surface in the momentum space. The non-trivial Z 2 number means that the Bogoliubov Fermi surface are locally stable until the two Bogoliubov Fermi surfaces pair-annihilate. In addition to the Z 2 number, the Bogoliubov Fermi surface can be also characterized by non-trivial Chern number. In Fig.2 the Bogoliubov Fermi surface with the Chern number 1(-1) is colored blue(red). Rather than simply presenting the numerical results, we argue that the non-trivial Chern number is a necessary consequence of the well-known parity anomaly of two-dimensional Dirac fermion. [50,51] First of all, we consider the adiabatic change from (s x ) state to (s x +is y ) state by slightly turning on (s y ) pairing. In (s x ) state, k z = 0 plane can be viewed as nodal point superconductor with the two Dirac nodal points per each Bogoliubov Fermi surface. This Dirac nodal points are pinned at the zero energy states and they are the time-reversal partner to each other. As the infinitesimal time-reversal breaking (s y ) pairing is turned on, the two Dirac points gaps out and must carry the Chern number ±1/2, which is analogous to the parity anomaly in the two-dimensional Dirac fermion. Since the time-reversal symmetry is broken, the effective mass gap of the Dirac points must be opposite with each other, therefore k z = 0 plane must be characterized by non-trivial Chern number. As a consequence, in full threedimensional momentum space, each topologically protected nodal line is inflated into a couple of Bogoliubov Fermi surfaces possessing non-trivial Chern number ±1. In principle, the inflation of the nodal line to the nodal surface occur for each nodal line, and the total Chern number at k z = 0 plane can be canceled each other. However, each Bogoliubov Fermi surface must carry non-trivial Chern number until they pairannihilate.
In conclusion, we have studied the triplet superconductivity of triple point fermions described by pseudospin-1 representation. In the superconductor composed of pseudospin-1 electrons, the even-parity paring can occur only with the spintriplet pairing. Furthermore, we have shown that multiband interaction uniquely opens attractive triplet pairing channels. Based on the Landau theory, we find two distinct triplet superconducting phases depending on the chemical potential µ: the time-reversal symmetric (s z ) state and the time-reversal broken (s x +is y ) state. In particular, (s x +is y ) phase is being favored when the chemical potential lies near or above the triple-band crossing point in such a way that middle band plays a role in electron pairing. Moreover, we find that the two states can be distinguished by different dimension of nodes and topological characteristics. Hence, we suggest the triplet superconductor is naturally stabilized in the triple point fermions in the presence of on-site interactions. In general, the superconducting instability is not limited to the on-site interactions, and therefore one may expect odd-parity superconductivity from the neighboring site pairings. In this case, we may expect p-wave spin-singlet and quintet states. The investigations on the possible odd-parity superconductivities would be an interesting topic for future study.
Here K ≡ (k 0 , k) and k 0 = 2π(n + 1/2)T denotes Matsubara frequency. Then, the free energy is written as, where ∆ = a J a γ∆ a . Let F n ( ∆) be the contribution to the free energy that contains n-th power of ∆ a . We have Meanwhile, we can parametrize the general terms in F n ( ∆) accordingly.
(S15) Utilizing Eq.S13 and Eq.S15, we investigate the sign of q 2 varyingμ andv while keeping c = 1/10. Then we acquire the phase diagram as given in the main text.
In the limit where the middle band is perfectly flat(c = 0), q 1 and q 2 can be simplified as below with normalizing the field, ψ, such thatv becomes unity. (S17) In Fig.S1 we display the plot of q 1 T 3/2 and −q 2 T 3/2 using Eq.S16 and Eq.S17. We observe that the free energy is stable (q 1 T 3/2 > 0) and time reversal broken phase is energetically favored (q 2 T 3/2 < 0) when the middle band is perfectly flat (c = 0). | 4,713.8 | 2019-09-09T00:00:00.000 | [
"Physics"
] |
Evidence of robust 2D transport and Efros-Shklovskii variable range hopping in disordered topological insulator (Bi2Se3) nanowires
We report the experimental observation of variable range hopping conduction in focused-ion-beam (FIB) fabricated ultra-narrow nanowires of topological insulator (Bi2Se3). The value of the exponent (d + 1)−1 in the hopping equation was extracted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \sim \frac{1}{2}\,\,$$\end{document}~12for different widths of nanowires, which is the proof of the presence of Efros-Shklovskii hopping transport mechanism in a strongly disordered system. High localization lengths (0.5 nm, 20 nm) were calculated for the devices. A careful analysis of the temperature dependent fluctuations present in the magnetoresistance curves, using the standard Universal Conductance Fluctuation theory, indicates the presence of 2D topological surface states. Also, the surface state contribution to the conductance was found very close to one conductance quantum. We believe that our experimental findings shed light on the understanding of quantum transport in disordered topological insulator based nanostructures.
disorder/interactions), but under strong disorder, the topological properties show insulating characteristics 23,[26][27][28] . In a recent study, it was demonstrated that introducing strong disorder in nanoscale TI samples might suppress the bulk conduction, while TSS remains immune to disorder 29 . Another report observed the change in band structure and TSS relocation to lower quintuple layers in presence of impurity and adsorbate addition 20,21 . The Universal conductance fluctuation (UCF) in weakly disordered TI specimens has also been a subject of intense research in recent years and different studies were performed to show the UCF arising from SS, the role of underlying symmetries (time reversal symmetry and spin rotational symmetry) in characterizing the UCF in 2D TIs, etc 25,[30][31][32][33] . In another report, ion milling, which creates defect states (like vacancies) in semiconductors, was used to signify the role of hopping transport in the magnetoresistance (M-R) of TIs 34 . Hopping transport in disordered electronic systems is a manifestation of Anderson localization (AL) 35 . Recently, Liao et al. 36 reported the experimental demonstration of AL in ultrathin films of 3D TIs, where due to increasing the disorder, a crossover from WAL regime of diffusive transport to AL regime of Mott-type variable range hopping (VRH) transport was observed.
In this work, we present a comprehensive experimental analysis of VRH transport in highly disordered nanowires of Bi 2 Se 3 of different widths, fabricated using the focused-ion-beam (FIB) technique. The excitation current and temperature dependent M-R curves of the Bi 2 Se 3 nanowires were studied to know the robust nature of 2D transport in disordered TI nanowires. A very high resistance was observed for the nanodevices which showed Mott-type VRH mechanism. The VRH fitting of resistance versus temperature (R-T) curves, revealed a very less hopping potential barrier which is indicative of a high localization length. TSS contribution to conductivity, which is usually close to one conductance quantum (e 2 /h), was observed dependent on the width of the nanowire. But the factor p in the generalised VRH equation for ohmic regime [37][38][39] is observed very close to 1 2 for different widths of the nanowires indicating the no dependence on the dimensionality, i.e. the proof of the Efros-Shklovskii (ES-) VRH mechanism, which has been usually demonstrated in strongly disordered systems at low temperatures [39][40][41] .
Results
We have fabricated Bi 2 Se 3 nanowire devices using the FIB milling technique from micro-mechanically exfoliated thin flakes of Bi 2 Se 3 on cleaned SiO 2 /Si substrates 9,42 . The details of the fabrication process are given in the Methods section. The standard four probe low temperature electrical transport measurements were carried out in Physical Property Measurement System (PPMS, Quantum Design) with a 9 T magnet. In order to understand the effect of perpendicular magnetic field on the resistance of the FIB fabricated Bi 2 Se 3 nanowire, first we have removed the contribution of UCF (discussed in the next section) by smoothing the M-R curve. For such a purpose, choosing a smoothing filter with small window size (~20 points) 43 and less order polynomial (~1 or 2) is better, as this would easily smooth out any extra unwanted oscillatory/fluctuating behaviour from the actual background signal. The red coloured smoothed M-R curve can be seen along-with the actual M-R (in blue) at 2 K (100 nA current) for device W3 in Fig. 1a. A positive linear M-R with a WAL at zero magnetic field is observed, which is a characteristic of TI-based devices at low temperature 5,14,[44][45][46] . In the past, a linear M-R was observed in nanoplates and nanostructures of Bi 2 Se 3 and the M-R was found directly proportional to the carrier mobility (μ) and applied magnetic field (B), as The angular dependence measurements were performed on these Bi 2 Se 3 nanostructures which confirmed the origin of linear MR as a signature of 2D transport property [47][48][49] . The linear M-R behaviour is also well explained with the classical model proposed by Parish and Littlewood 50,51 . Note that, a sharp WAL dip is not distinguishable in our nanowire, which can be attributed to the presence of other kinds of electronic interference/interaction effects originating from the disorder/defect sites in our samples, which compete with the quantum interference effect causing WAL. The background curve (orange) in Fig. 1a shows the oscillations at higher magnetic fields. Earlier reports have assigned such oscillation in TI materials under perpendicular field to the Shubnikov de-Haas (SdH) effect 6,11,14 . In order to understand the origin of SdH oscillations in our TI-based sample, we performed the Landau level fan diagram analysis. It is a well-known fact that with increase in the magnetic field, successive emptying of Landau levels (LL) takes place, which manifests itself as oscillations in the resistance for a 2D Fermi surface 11 . The LL index (n) is related to the magnetic field (B) and Fermi cross-sectional area (S F ) by 2 , e = electronic charge and = reduced Planck's constant. The parameter of interest in the above equation is the factor γ, which is known as the Berry's phase and its value is predicted to be ± 0.5 for a TSS. The value of γ can be calculated from the Y-intercept in LL index versus B −1 plot as shown in Fig. 1b, where the red circles and olive squares represent the corresponding minima and maxima in M-R oscillations (inset in Fig. 1b), respectively. The value of Y-intercept in our case comes out to be -0.54, which is very close to the 1/2-shifted SdH oscillation predicted for TSS. This analysis indicates towards a very important property of TSS, i.e. robustness to disorder/deformation. The presence of large UCF superimposed on the background M-R did not affect the 1/2-shifted energy spectrum related to the TSS. Also, both the Landau level fan diagram and UCF (as discussed in next section) analysis suggests the presence of TSS in our samples. Fluctuations in the resistance as a function of magnetic field for device W3 (length ~1.846 µm, width ~110.3 nm) are clearly evident from Fig. 2. We performed the temperature and excitation current (at a particular temperature) dependent M-R analysis, in order to understand the origin of these fluctuations. Figure 2a , curves at 100 nA excitation current for temperatures 2, 4 and 6 K. The curves have been shifted for better clarity. The decreasing amplitude of these fluctuations with temperature, reproducibility in reverse magnetic field sweeps and the aperiodicity with respect to magnetic field suggested us to compare this effect with the UCF. It is a well known phenomenon that the electronic interference in disordered mesoscopic metallic samples with sizes smaller than the phase coherent length (L ϕ ) manifests itself as some noisy/ aperiodic patterns of the order of e 2 /h in the M-R curves, known as UCF 52,53 . When the sample size becomes larger than L ϕ , self-averaging effects come into play, which decreases the amplitude of fluctuations 53,54 . Inset of Fig. 2a shows the fluctuations in conductance, which has been extracted by subtracting the background conductance, obtained from a fit to a five order polynomial, from actual conductance. Although, complete reproducibility of fluctuations at different temperatures is not expected in our case due to very high disorder and deformations in the nanowire due to milling procedure, a good amount of reproducible features are present in the fluctuation curves of 2 K and 4 K (shifted for clarity in inset of Fig. 2a), which reflect the inherent fingerprint of the sample. Quantitative analysis of the amplitude of fluctuations in our sample was performed using the UCF theory proposed by Lee et al. 53 . The conductance fluctuations for UCF analysis can be obtained from the relation: where 〈G(B)〉 denotes the ensemble average of magneto-conductance. A correlation function of the fluctuations is used to calculate the amplitude of UCF, i.e., root mean square value of the conductance fluctuation (rms(δG)or δ ⟨ ⟩ G 2 ), at (0) 53 . The calculated UCF magnitude at different temperatures is around 0.01 e 2 /h and depicts a slow exponentially decaying dependence with temperature (T −0.47 ) as shown in the lower inset of Fig. 2b. Here, an important quantity of interest is the full width at half maximum of the auto-correlation function Further, it is known that in a 2D system, when L ϕ is the shortest length, we have where φ o is equal to the magnetic flux quantum, h/e 53 . Using equation (7), we calculated the L ϕ values and its temperature dependence (Fig. 2b) show an exponential decay with ϕ − .
L T 0 36 . Theoretically, the power law dependence of L ϕ for a 1D, 2D and 3D system is predicted to be proportional to T −1/3 , T −1/2 and T −3/4 , respectively 55 . For a 2D system with < where L ϕ is the phase coherent length and L x , L z are width and channel length of the sample 53 . Since ϕ − .
The temperature dependence of UCF amplitude in our case is T −0.47 , which is very similar to the theoretically expected characteristics 31,33,53,54 and suggests a 2D dominated system. Note that we did not observe the theoretically accurate estimate predicated for 2D or 1D transport, i.e. UCF amplitude T −0.5 or T −0.33 , respectively. Earlier studies demonstrating the 2D transport in TI materials have also reported small deviations in comparison to the theoretical values. Kim et al. 32
L
T 0 43 for Bismuth nanowires, which reflected that the system is a hybridization of both one and two dimensions. Since, the exponent of −0.43 was closer to the 2D system value of −0.5, therefore, it was considered that the SS plays a major role in transport. Trivedi et al. 33 reported slightly deviating values of T −0.44 and T −0.6 from rms(δG) and L ϕ analysis in the nanosheets of Bismuth Telluro-Sulfide, and compared the case with 2D system. Matsuo et al. 30 reported a dependence of T −0.65 for quasi-one-dimensional Bi 2 Se 3 nanowires, where again the transport was attributed to the 2D system. We believe that further investigations on ultra-narrow wires (widths < 100 nm) are required to shed more light on the nature of 1D transport.
The excitation current dependent UCF study at 2 K has been shown in Fig. 3a,b. The decrease in M-R and conductance fluctuations with increasing current is clearly evident from Fig. 3a, which shows the M-R change curves at different currents of 100 nA, 500 nA and 1.8 µA. The average conductance almost doubles itself from 0.7 e 2 /h to 1.3 e 2 /h, while increasing the current from 5 nA to 1.8 µA, shown in Fig. 3b (right Y-axis). This is indicative of the fact that the transmission probability of carriers between the electrodes and the nanowire is dependent on the current bias. A possible explanation for this could be the carrier confinement effects introduced due to the low dimensionality of the fabricated nanowire. Figure 3b (left Y-axis) shows the exponential decrease in rms(δG) with current, i.e. . The dependence of UCF on voltage (or current) bias in mesoscopic conductors is a highly complicated phenomenon and some studies in the past were conducted to understand it. Ludwig et al. 56 found that for non-interacting electronic systems, the variance 〈δG 2 〉 shows a monotonic increase with voltage, i.e., an enhancement in fluctuations in observed with increase in voltage, which is in good agreement with the pioneering work by Larkin and Khmel'nitskii 57 . Simultaneously, they also discussed about the systems with high EEI, where dephasing due to Coulomb interaction occurs and results into a non-monotonic dependence of conductance fluctuations with voltage. Further, it was shown that in such systems rms(δG)falls as V 1/ for higher voltages. In our case, the FIB fabricated nanowire resembles to a strongly disordered electronic system with sufficiently high localized states, leading to high EEI and dephasing effects, which may be a reason for the observed V 1/ drop in the amplitude of conductance fluctuations.
Discussion
We now discuss the non-trivial effects of disorder and localization induced hopping transport in the Bi 2 Se 3 nanowires fabricated by using FIB milling method. For this purpose, we performed the detailed analysis of the temperature dependence of conductance for device W3 and W6. The upper right inset in Fig. 4a shows the cooling curve for device W3. A huge monotonic increase in the resistance with decreasing temperature till 4.4 K is observed. Below this, the resistance increases very sharply from 18 kΩ (at 4.4 K) to 19.6 kΩ (at 2 K). The R-T behaviour in our sample clearly indicates a huge insulating behaviour. Previous TI-based transport studies have reported metallic or semiconducting R-T behaviour based on the bulk conduction band or shallow impurity band 5, 6, 11, 58 . In our case, none of these standard models properly fit to the experimental results. Therefore, we use the VRH model to fit our obtained results. Figure 4a shows the VRH fit (blue solid line) on conductance versus temperature plot (orange circles) for device W3. In addition to the standard Mott-type VRH conductance, a constant conductance is also used to correctly define the overall conductance of the sample 29 . Therefore, the total conductance (G) is given by the relation: where G TSS is the conductance contribution from TSS,G o is a pre-factor,T o is a characteristic temperature related to the hopping energy and d is the effective dimension of the system. The exponent (d + 1) −1 determines the type of conduction mechanism in the sample based on the shape of density of states (DOS) at the Fermi level 40,41 . A recent study on Bi 2 Te 3 nanotubes showed that the temperature dependent second term in Eq. (9) comes from the bulk channel 29 . A closer inspection of Fig. 4a suggests that the experimental conductance deviates from the VRH fit at ~42 K (shown in the blue shaded region). The deviation from VRH transport mechanism and the sharp decrease in conductance can be attributed to the high EEI in TIs at low temperatures, which is discussed after the hopping mechanism. The blue solid line is a fit to Eq. (9) using fitting parameters G TSS ~ 6 × 10 −5 S, G o ~ 1.68 × 10 −4 S, d ~ 1.3 and T o ~ 940.7 K for device W3. Figure 4b demonstrates that the TSS contribution is a constant slightly above one conductance quantum (e 2 /h ~ 3.874 × 10 −5 S) for most of the temperature range but decreases for temperatures below 6 K due to EEI effects. The TSS conductance was obtained by subtracting the bulk conductance contribution given by the second term in Eq. (9) from the experimentally acquired total conductance. Previous works have shown similar kind of TSS contribution close to one conductance quantum 29,59 . Also, a comparison between the total conductance and bulk conductance is plotted in inset of Fig. 4b, which clearly reflects that the bulk contribution is very low compared to the total conductance, which is evident from the low value of G o . A similar kind of analysis was performed on another nanowire device W6 having much greater width (~683 nm) than W3. FESEM image of the device is shown in the upper right inset of Fig. 4c. The lower inset shows the R-T characteristic of the device. Again, we observe a highly insulating behaviour for this device with Figure 4d indicates that the contribution from bulk conductance is very large and almost equal to the total conductance. Thus, the TSS contribution is very less than one conductance quantum and is shown by the fluctuating G TSS values of ~0.3 e 2 /h in inset of Fig. 4d. The above two results clearly indicate that the conduction through TSS is dominant in W3 nanowire, where the surface-to-volume ratio is very high as compared to W6. This result is very interesting as it demonstrates the FIB fabrication method as an efficient way to tune the conductance contribution from SS and bulk transport. We observed that there are two kinds of conductance channels, one from the TSS and another from the bulk. With increasing width of the nanowire the contribution from bulk also increases due to more conduction channels available in the bulk. Previously, ref. 29 stated that due to strong disorder in the system bulk conductance is suppressed and TSS contribution is enhanced significantly for TI nanotubes with very less cross-sectional area. A similar kind of phenomenon is observed in our case for smaller width nanowire W3.
The deviation from VRH model was observed in the conductance at very low temperature for device W3. Here we use the EEI theory to explain this phenomenon. The nanowire conductance falls off logarithmically with temperature below 5 K (lower inset in Fig. 4a). The correction in 2D conductance due to EEI is given by the relation 16,17,60 : where σ F is the electron screening factor and T o is the reference temperature from where the correction σ ∆ is measured. The values obtained for σ F and T o from the fitting curve (solid purple line in lower inset of Fig. 4a) are 0.2 and 2.1 K, respectively. Theoretically, it is predicted that the value of σ F lies in the range of 0 to 1, and a value close to zero implies the presence of strong EEI in the system. Many previous reports have exhibited the relevance of many-body interactions in low-dimensional TI samples. The anomalous insulating behaviour in TIs at low temperatures was proposed as a result of co-existence of both the strong electronic interactions and the topological delocalization (TD) 16,19 . The enhancement of Coulombic interaction could suppress the density of states at the Fermi level and thus, a logarithmic increase in resistance is observed for device W3. Also, the competition between the TD and EEI at low temperature modifies the quantum correction to conductance as is the enhancement in conductance by TD and ∆G T ( ) EEI is the suppression of conductance by EEI 16,18 . The short phase coherence length and less WAL observed in our nanowire can be attributed to the strong electronic interaction, which is enhanced in thinner samples due to disorder and reduced dimensionality. Previous reports have shown that in very thin TI samples an energy gap opens up at the Dirac point due to the coupling of top and bottom surfaces, which weakens the TD and reduces conductance at low temperature 61,62 .
We believe that our experimental findings are a consequence of the strong localization effects created in the sample by disorder/deformation, and FIB milling procedure is inherently well known to introduce such defects/ deformations/disorder. In presence of disorder, localized/defect states are created and electron transport occurs by hopping across these states which is a totally different behaviour than that of other conduction electrons 63 . The DOS around the Fermi level is an important parameter for understanding the hopping of electrons between the localized states. In our case, the VRH analysis for both the nanowires gives almost the same value of exponent (d + 1) −1 , which is close to 1 2 . There is no dependency of hopping transport on the width/dimensions of the nanowire. This is indicative of the fact that under high disorder regime, ES-VRH mechanism is followed in our nanowires, where the DOS near the Fermi level is not constant (in contrast with Mott-VRH where DOS is constant near the Fermi level) and vanishes linearly with energy, and the value of the exponent is 1 2 in all dimensions 39,40 . In a previous work on Bi 2 Se 3 thin films, the characteristic temperature T o in Eq. (9) was found to be ~10 6 K 64 . But in our case, it is 940.7 K and 23.42 K for W3 and W6, respectively. Since, the hopping potential is proportional to T o , we can conclude that the energy required by electron to hop from one localized state to another is very small in our case. Also, the probability of hopping, (12) where R is the spatial separation and W is the energy separation between two localized states 65 , is more for our case. The localization length was calculated using the expression where ε o and ε represent the permittivity of vacuum and the dielectric constant of the material 39,41 . The value of εε o is taken as 100 ε o for Bi 2 Se 3 66, 67 . High localization lengths of 0.5 nm and 20 nm for W3 and W6, respectively, indicate the presence of strong disorder in our nanowires. Due to high localization and confinement effects in our nanowire, L ϕ tends to be very small in comparison to the dimensions of the nanowire. Also, our analysis suggests that for a low dimensional disordered TI sample, voltage induced dephasing takes place, so that L ϕ is unstable and varies with the excitation current. It is very important to note that the conductance fluctuations in our case cannot be simply attributed to the noise due to current induced heating effects, since our nanodevice was perfectly placed in liquid Helium with a proper heat exchange setup. Also, the heating effects due to ramping up of magnetic field can be ruled out, as ramping rates of less than 30 Oe/sec were used. The UCFs have a totally different origin from other quantum interference effects like WAL. The amplitude of UCF provides an excellent way to study the quantum transport properties in a material, since UCF is related to the phase-coherent length, sample size and symmetry of the Hamiltonian inside the sample 68 . Recently, quantum oscillations in the form of UCF were not only used to reveal the topological character of the material but also to study the phase transition from Dirac-to-Weyl semimetal by breaking time reversal symmetry in Cd 3 As 2 nanowires 69 . In the case of 3D TIs, both the bulk and SS contribute to the amplitude of UCF 31 . The absence of backscattering in TIs suppresses the number of scattering paths and thus, leads to the scattering confinement, which plays a significant role in 2D UCF 70 . Previously, for TI samples a T −1/2 dependence of L ϕ were considered as an important signature of 2D TSS 45 . However, in our case the electronic transport seems to be originating from the hybridization of 1D and 2D system, as indicated by the exponent of −0.36. The unusual exponent close to 1D behaviour is of no surprise, since the fabricated nanowire in our case resembles a quasi-1D system. It is important to note that the quantum confinement effects due to low dimensionality of the system also affect the transport properties in a TI-based system, and many future studies need to be performed in order to fully understand the complex mechanism related to the interplay of 2D TSS and quantum confinement effects.
In summary. We have experimentally demonstrated the existence of ES-VRH mechanism in highly disordered/deformed nanowires of Bi 2 Se 3 . The device fabrication procedure used in our case inherently provides an excellent way to test for the robustness of TSS to strong disorder introduced in the system via Ga + implantation, deformation formed due to milling and other impurities deposited during in situ metal contacts formation. Our analysis revealed that the UCFs originate from the 2D-TSS, thus confirming the robust nature of TSS. Also, it was shown that the contribution of TSS and bulk can be studied in the FIB fabricated nanowires of Bi 2 Se 3 .
Methods
Device fabrication. The thin Bi 2 Se 3 nanowire based devices were fabricated by micromechanical cleavage technique using the standard scotch tape method and focused ion beam (FIB) milling. Firstly, SiO 2 / Si wafers (p-type highly Boron doped with ~0.001-0.005 Ω-cm) were cleaned with acetone, iso-propanol, methanol and de-ionised water, and additional oxygen plasma treatment was performed for ~10 min. The thin flakes of Bi 2 Se 3 (99.999% CAS#12068-69-8, Alfa Aesar) were deposited using exfoliation method on SiO 2 /Si substrates with pre-sputtered thick Au/Ti pads (~80/5 nm). This method produces thin and random sized flakes of Bi 2 Se 3 which were further localized under optical microscope (Olympus MX51) and field emission scanning electron microscopy (FESEM, Zeiss-Auriga). The located thin flakes were milled using FIB, with Ga + ions as a source material for the ion beam. The electrical contacts on the thin nanowire of Bi 2 Se 3 were made with metal electrodes of platinum through FIB based gas injection system (GIS). | 6,013 | 2017-08-10T00:00:00.000 | [
"Physics"
] |
A Low-Complexity Block Diagonalization Algorithm for MU-MIMO Two-Way Relay Systems with Complex Lattice Reduction
We design a scheme of precoding matrices for two-way multiuser relay systems, where a multiantenna relay station (RS) operating in an amplify-and-forward model simultaneously receives information from all multiple-antenna users. Considering the feasibility in mathematical analysis, users are distributed in two symmetrical groups. To reduce the complexity of proposed precoding scheme, we employ the QR decomposition and complex lattice reduction (CLR) transform to replace the two times singular value decomposition (SVD) of conventional BD-based precoding algorithm by introducing a combined channel inversion to eliminate the multiple users interference (MUI). Simulation and performance analysis demonstrate that the proposed LR-MMSE algorithm has not only a better bit error rate (BER) performance, a higher sum-rate, and simple architecture, but also 89.8% and 35.5% less complexity compared to BD- and MMSE-based scheme.
Introduction
Two-way relaying (TWR) cooperation has attracted considerable attention due to the high spectral efficiency. A typical model of the TWR protocol with multiple access channel (MAC) phase and broadcast channel (BC) phase has been investigated in [1]. In the MAC phase, two source nodes transmit their message to the relay node simultaneously. After processing the mixed received signal at the relay, a combined version of the received signal is broadcast to each source node in the BC phase.
Among the numerous two-way relaying protocols, such as amplify-and-forward (AF) and decode-and forward [2], AF protocol gets favored as one or more low-complexity relay nodes are adopted to assist the communication between sources and destinations without decoding the signals.
In wireless sensor networks (WSNs), a large number of static or mobile sensors cooperate to perceive, compute, and transmit message assistant by the relay [3]. Since sensor nodes are usually operated by lightweight batteries that are difficult to replace or recharge, energy becomes one of the most crucial resources in WSNs. Moreover, experimental measurements have shown that the energy consumption of a sensor node is dominated by communications [4]. Thus, a simple and low-complexity transceiver model for sensors is a powerful solution to solve this problem. Recently, a prominent research in [5] has proposed a multilevel physical layer network coding to rule the rate of sensors, which saves the power of networks potentially.
As we known, multiple-input multiple-output (MIMO) relays potentially obtain both spectral efficiency and link reliability by exploiting the multiple antennas, and, jointing the TWR technology, it can improve the system performance dramatically. Recently, the precoding design for TWR MIMO systems is extended to multiuser cases, which can be roughly divided into two categories: symmetric systems and asymmetric systems [6][7][8][9][10][11]. In symmetric system, all users are supposed to be distributed in two groups in the form of pairings [6][7][8]. In asymmetric systems, a based station exchanges messages with multiple users [9][10][11], which is a typical scenario of cellular networks.
Unlike the received signals in single-user MIMO (SU-MIMO) systems, the received signals of different users in multiuser MIMO (MU-MIMO) systems not only suffer 2 International Journal of Distributed Sensor Networks from the noise and intra-antenna interference but are also disturbed by the multiuser interference (MUI). However, when we consider two-way relaying multiuser MIMO (TWR MU-MIMO) systems, the MUI becomes more complicated, since there are two groups of users mixing in the received signals.
Channel inversion strategies based on zero forcing (ZF) and minimum mean squared error (MMSE) can be used to cancel the MUI but result in amplifying the noise [12,13]. Block diagonalization (BD) has been proposed in [14,15] to improve the system throughput and provide convenience for the power control. Literature [16] studies the precoding scheme of the singular value decomposition (SVD) and employs a technology of channels aligning referring to [17]. However, the complexity of SVD is very high when the number of users and the specification of the antenna are large. In order to reduce the complexity, generalized singular value decomposition (GSVD) has been applied in [18] to design the downlink of MIMO systems. Unfortunately, there is also a great performance loss at low signal to noise ratios (SNRs) when the noise is the dominant factor. In recent research, a systematic scheme is proposed in [7], which imposes a power constraint at the relay node and obtains an approximate result by fixing the transmitting and receiving matrices at the user nodes firstly. In [19], a low complexity precoding algorithm is proposed to reduce the condition number of the effective number by introducing a complex lattice reduction (CLR) transform in one-way relay MU-MIMO systems.
In this paper, we focus on AF TWR MU-MIMO systems and strive to find a lower complexity and better performance precoding scheme. By employing a combined channel inversion and replacing the SVD with QR decomposition, we can cancel the MUI remarkably. Then, the CLR transform is used to design the precoding matrices of users which locate in another group and is in pairs of receiving users.
The rest of the paper is organized as follows. Section 2 describes the conventional system model. Section 3 introduces the BD-based precoding algorithm. The scheme LR-MMSE is proposed in Section 4. Performance analysis and simulation are given in Sections 5 and 6. Finally, we conclude this work in Section 7.
Notation. Throughout this paper, for a matrix A, Tr{A}, A , A , A −1 , and A † denote the trace, transpose, complex conjugate transpose, inverse, and pseudoinverse, respectively. " " stands for the expectation of a random variable; the terms C × represent the ( × )-dimensional space with complexvalued elements. The notations ‖A‖ and ‖A‖ 2 denote Frobenius-norm and 2-norm of matrix A. I × is a -byidentity matrix; 0 is a -dimensional zero matrix.
General System Model
Investigate an uncoded TWR MU-MIMO system, with pairs of user equipment (UE) and one two-way relay station (RS) which operates in AF protocol. For simplicity, we consider a symmetric model in which users are divided into two groups: ∈ { , }. The th user in group and the relay has equipped , and antennas, respectively. A block diagram of such a system is shown in Figure 1.
Ruled that there is no direct link between two users and all the channels experience independent and flat MIMO fading. Let d , ( ) denote the data symbol at time for the th user in group (UE , ). The UE , performs transmit precoding beamforming with vector P , ∈ C × , and transmits the signal in the following form of matrix: where d , ( ) ∈ C , ×1 and satisfies [d , ( )d , ( )] = 2 n , ⋅ I , . We assume the average power of the transmit signal is 1. Then, users UE and UE exchange their signals by the assisted relay node RS in two phases. Firstly, in the multiple access channel (MAC) phase, UE and UE transmit their signals simultaneously to relay and the received signals at RS are given by where H , ∈ C × , is the channel coefficient coming from UE , to RS and n ( ) ∈ C ×1 is a zero-mean addictive white Gaussian noise (AWGN) at the relay and [n ( )n ( )] = 2 n I . In the broadcast (the broadcast channel) phase, the RS forwards and broadcasts the mixed signals s( ) ∈ C ×1 to all user by multiplying a forward matrix f ∈ C × , where s ( ) = fr ( ) . ( It is worth noting that the f is an assistant precoding matrix for decoding the signals at receiver sites. Since it is supposed that channel reciprocity (This assumption is reasonable when we consider time division duplex (TDD) systems. Channels in uplink and downlink are static and identical in the frequency domain.) is met within two phases, the received signals y , ( ) ∈ C , ×1 at the user UE , side are International Journal of Distributed Sensor Networks 3 where n , ( ) ∈ C , ×1 is also a zero-mean AWGN at the UE , and [n , ( )n , ( )] = 2 I , . The UE , combines its own received signal (4) by using a received decoding vector g , ∈ C 1× , to get the estimatê whered̂, ( ) denotes the information coming from UÊ, , the subscripts and̂∈ { , }, and ̸ =̂represents the user pairings. For notational convenience, the time index is henceforth wiped off. Refer to Figure 2(a), which shows the overall procedure for transmit-and receive-beamforming and relaying of multiuser two-way systems. Next, designs of the scheme of precoding matrices are given with the criteria of BD, ZF, and MMSE. Further, the lattice reduction is applied to suppress the bit error rate (BER) of the system. As a result, we can omit the decoding matrix to obtain a simple structure.
Generalized Design of BD-Based Precoding Algorithm
The BD-and SVD-based precoding algorithms are proposed in [10] and [12], respectively. Combining the two ideas, the BD-based scheme can be given as follows.
The received signals at the users UE , and UÊ, sides can be written asd̂, The detail components ofd̂, arê The 1st item can be dismissed by the perfect selfinterference suppression, and the key in the rest of processing is to eliminate MUI and obtain a good performance of the signal estimation.
Based on the system model, the combined channel matrix is given by In order to eliminate the MUI in (7), we impose the BD constraint in which Ĥ, f̂, = 0 ∀ ∈ {1, . . . , } .
Then, the second SVD operation is used to obtain the precoding and decoding matrix. We have Finally, the user UÊ, 's precoding matrix P̂, = V̂, Ω̂, and the decoding matrix is obtained as g , = Û, , where Ω̂, is a diagonal matrix and its elements stand for the transmit power allocation. To guarantee that the V , is not a null matrix, we must have > . Note that we omit the dimension matching matrices in (11) which will be mentioned in the next section.
The Proposed LR-MMSE Precoding Algorithms
In this section, we describe the proposed precoding algorithms based on a strategy that employs a combined channel inversion method, QR decompositions, and lattice reductions in detail. We can solve the questions with ZF and MMSE criteria. For a better BER performance, the complex lattice reduction is proposed to improve the MMSE performance.
ZF-Based Design.
The ZF-based design of precoding matrix termed as ZF can be performed in two steps.
Step 1. It is to obtain the relay forward matrix f , ,zf which eliminates the MUI completely by a QR decomposition of Ĥ, .
Firstly, by using the ZF inversion to the combined channel matrix H, we have where H † zf is the pseudoinverse of matrix H zf and Ĥ, ,zf ∈ Ĉ, × is the submatrix of H † zf . Note that H zf H † zf = I [12]. Thus, the off-diagonal block matrices of HH † zf are zero and Ĥ, ,zf is in the null space of Ĥ, ; that is, Imposing the QR decomposition on Ĥ, ,zf = Q̂, ,zf R̂, ,zf , we have where Q̂, ,zf ∈ Ĉ, × is an orthogonal matrix. Then, substituting f 1 , into (7), the 2nd and 3rd items will be equal to zeros.
Step 2. It is to obtain the user precoding and receive decoding matrices p̂, ,zf and g , ,zf by a SVD of the effective channel.
Note that the degree of freedom between H , and Ĥ, is × , we match the dimensions with the help of Π̂, ∈ C ×̂, , and ‖Π̂, ‖ 2 = 1. Then, substituting f 1 , and Π̂, into the 4th item ofd̂, , we have the effective channel H eff , ,zf = H , Π̂, Q̂, ,zf Ĥ, , where H eff , ,zf ∈ C , ×̂, . The ZF algorithm can be completed by applying the SVD operation to the matrix H eff , ,zf = Û, ,zf Λ̂, ,zf V̂, ,zf . Then, the user precoding and receive decoding matrices p̂, ,zf and g , ,zf are designed by p̂, ,zf = V̂, ,zf Ω̂, and g , ,zf = Û, ,zf , respectively. (p̂, ,zf denotes the precoding processing at the site of UÊ, and g , ,zf denotes the receiving processing at the site of UE , . This type of subscript contributes to distinguishing the two matrices.) Finally, we get the systematic ZF precoding scheme for users.
MMSE-Based
where Ĥ, ,mmse ∈ Ĉ, × is the submatrix of pseudoinverse matrix H † mmse and = . Note that the regularization factor approaches zero when the SNR is high, and thus we have HH † mmse ≈ I [12]. This means that those off-diagonal block matrices of HH † mmse converge to zeros with the increase in SNR. Hence, the matrix Ĥ, ,mmse is approximately in the null space of Ĥ, ; that is, Ĥ, ,mmse Ĥ, ≈ 0.
Then, the MMSE-BD algorithm also can be completed by applying the SVD operation to the effective channel matrix The user precoding and receive decoding beamforming matrices of MMSE are obtained as P̂, ,mmse = V̂, ,mmse Ω̂, , g , ,mmse = Û, ,mmse .
International Journal of Distributed Sensor Networks
LR-MMSE-Based Design.
In this subsection, LR is applied to reduce the dimension of the effective channel which replaces SVD in MMSE algorithm. For convenience, we term it as LR-MMSE (see Figure 2(b)). A powerful and famous reduction criterion for arbitrary lattice dimensions was introduced by Lenstra et al. in [20], and the algorithm they proposed is known as the LLL (or L 3 ) algorithm [21,22]. In order to reduce the complexity, a complex LLL (CLLL) algorithm was in [23], which reduces the overall complexity of the LLL algorithm by nearly half without sacrificing any performance. Through reducing the size of channels, CLLL can obtain better BER performance. In this paper, we employ the CLLL algorithm to implement the LR transform.
After the first MMSE precoding, we transform the MU-MIMO channel into parallel or approximately parallel SU-MIMO channels and the effective channel matrix for the UE , is H eff , ,mmse = H , Π̂, Q̂, ,mmse Ĥ, .
We perform the LR transformation on H eff , ,mmse in the precoding scenario [24]; that is, where T̂, is a unimodular matrix with det |T̂, | = 1 and all elements of T̂, are complex integers. The unimodular feature of T̂, guarantees that the energy will not change through the LR transform. The MMSE precoding is actually equivalent to the ZF precoding with respect to an extended system model [25]. The extended channel matrix H for the MMSE precoding scheme is defined as By introducing the MMSE method, a trade-off between the level of MUI and the noise is introduced (see also [12]). Then, the LR-M-D precoding filter is designed as where T̂, is the unimodular matrix for H eff , ,mmse . Then, the LR-MMSE relay forwarding filter is given by Since the lattice reduced precoding matrix P has near orthogonal columns, the required transmit power will be reduced compared to the MMSE precoding algorithms. Thus, a better BER performance than that of the BD precoding algorithms can be achieved by the proposed LR-MMSE precoding algorithm. The left work for receiver is to quantize the signalsd to the nearest vectors. The LR-MMSE algorithm is summarized in Algorithm 2.
Simulation and Performance Analysis
In this section, we analyze the proposed LR-MMSE algorithm in terms of computational complexity, architecture, BER, and achievable rates.
Computational Complexity Analysis.
In this section, we measure the computational complexity of the precoding algorithms we have introduced by the total number of floating point operations Per Second (FLOPS). Note that the LR algorithm has complex variable, and the average complexity of CLLL algorithm has been given in [26] by FLOPS. The number of FLOPS for the complex QR decomposition and the real SVD operation are given in [27]. Moreover, the FLOPS number in a × complex SVD operation is equivalent to its extended 2 × 2 real matrix. The total FLOPS number required by the matrix operations is summarized below: (i) multiplication of × and × complex matrices: 8 − 2 ; (ii) QR decomposition of a × ( ≤ ) complex matrix: (iii) SVD of a × ( ≤ ) complex matrix by only obtaining Λ and V: 32( 2 + 2 3 ); (iv) SVD of a × ( ≤ ) complex matrix by obtaining U and Λ and V: 8(4 2 + 8 2 + 9 3 ); (v) inversion of a × real matrix by Gauss-Jordan elimination: 4 3 /3; (vi) inversion of a × complex matrix by Gauss-Jordan elimination: The FLOPS of three types of precoding schemes are shown in Tables 1, 2, and 3, respectively. We assume the system is composed of 6 users ( = 3) equipped 2 antennas and one relay station equipped 12 antennas. From Table 1, apparently, we know that the complexity of SVD is very large, resulting in the BD's complexity reaching a level of million. On the contrary, the combined channel inversion needs only 22805.4 FLOPS which operates in merely one time. Meanwhile, the LR-MMSE can save 35% of FLOPS of MMSE, while MMSE has saved 89.8% of FLOPS of BD.
Next, we fix the antennas of users and reveal the low complexity of LR-MMSE's property by enlarging the number of users. Ruling = , we can obtain the performance comparison in Figure 3. Figure 3 shows that the complexity of BD-type precoding scheme grows rapidly with the increase in , while it grows gently with respect to MMSE and LR-MMSE. This phenomenon is due to the significant influence of dimensions with regards to SVD. Note that the QR decomposition also needs to be implemented in times, though there is only one time of the combined channel inversion. But, the QR decomposition is simpler than SVD.
Performance Analysis in Architecture.
Compared to the BD and MMSE scheme, the LR-MMSE has a simple architecture. At the site of each user, the decoding matrix can be omitted, since the precoding processing in formula (31) has directly decoded the effective channel. This outperformance can contribute to saving a significant power of mobile users, even the relay stations. Moreover, it is worth noting that we need not employ the self-interference eliminating technology, 114432 Total 206389 Table 3: Computational complexity of LR-MMSE in case (6,2,12).
Steps
Operations FLOPS Number due to the combined channel inversion having included the channel information of the receiver.
BER Performance Analysis.
Recall that, with the increase in SNR, we have Q̂, ,mmse Ĥ, ≈ 0 in the LR-MMSE algorithm. Thus, the MU-MIMO channel is approximately decoupled into equivalent SU-MIMO channel. In most schemes of considering the channel noise, RBD is a common method. A result is shown in [28], which is like formula (33), which is not converged to zeros. Comparing the above two characters, LR-MMSE has more accurate receiving. As we known, the condition number is a measurement of receiving error bits. In virtue of the defined condition number in [23], the channel matrix can be detected with respect to orthogonality. Taking an example in a system of linear equations, There are three cases for different kinds of matrices: (i) when the channel matrix is an orthogonal matrix, the condition number is 1; x will change a little with tiny changing b; (ii) when it is a singular matrix, the condition number is ∞; x will change significantly with tiny changing b; even b has not changed; (iii) when it is a nonsingular but not orthogonal matrix, the condition number will be large, and x changes between the above two cases.
As can be seen in Figure 4, the LR-MMSE has a smaller average cond(H) of the effective channel compared to MMSE and BD. Thus, a significant power reduction and better BER performance are obtained.
Simulation Results
Considering a specification of antenna with = 8, = 10, and , = 2, six users are divided into two symmetrical groups. The perfect channel scenario is applied in which = 1 and all channels' elements are fetched from the complex Gaussian process. For highlighting the performance of the proposed algorithm, we allocate the transmit power equally to each user. Figure 5 shows that the LR-MMSE algorithm has the best sum-rate performance when MMSE and BD have a similar achievable sum-rate. In particular, when the SNR is 30 dB, the proposed algorithm has a 3.208 bit/s/Hz gain compared to the BD algorithm. Since the low SNR will affect the diagonality after the precoding processing in MMSE scheme, there is an inevitable performance loss. However, by virtue of the size reduction of CLLL transform, the LR-MMSE has obtained a performance compensation.
International Journal of Distributed Sensor Networks Additionally, in the process of simulating, we discover that choosing the optimal matching matrix Π can achieve nearly 1 bit/s/Hz gain. In this work, we did not study the optimization of power and matching matrices.
Conclusion
In this work, we have proposed a low complexity precoding scheme of two-way MU-MIMO relay systems with QR decomposition and complex LLL transform instead of two times SVD. For simplicity, users are distributed in two symmetrical groups. The performance analysis show that the LR-MMSE algorithm has not only a low complexity, but also a better BER, a simple structure at the site of the receiving user, and a theoretical maximum. Finally, a higher achievable sum-rate of the system is confirmed by simulation. | 5,066.6 | 2015-06-01T00:00:00.000 | [
"Computer Science",
"Engineering"
] |
Phenotype of Transgenic Mice Carrying a Very Low Copy Number of the Mutant Human G93A Superoxide Dismutase-1 Gene Associated with Amyotrophic Lateral Sclerosis
Amyotrophic lateral sclerosis (ALS) is a progressive neurodegenerative disease of the motor neuron. While most cases of ALS are sporadic, 10% are familial (FALS) with 20% of FALS caused by a mutation in the gene that codes for the enzyme Cu/Zn superoxide dismutase (SOD1). There is variability in sporadic ALS as well as FALS where even within the same family some siblings with the same mutation do not manifest disease. A transgenic (Tg) mouse model of FALS containing 25 copies of the mutant human SOD1 gene demonstrates motor neuron pathology and progressive weakness similar to ALS patients, leading to death at approximately 130 days. The onset of symptoms and survival of these transgenic mice are directly related to the number of copies of the mutant gene. We report the phenotype of a very low expressing (VLE) G93A SOD1 Tg carrying only 4 copies of the mutant G93ASOD1 gene. While weakness can start at 9 months, only 74% of mice 18 months or older demonstrate disease. The VLE mice show decreased motor neurons compared to wild-type mice as well as increased cytoplasmic translocation of TDP-43. In contrast to the standard G93A SOD1 Tg mouse which always develops motor weakness leading to death, not all VLE animals manifested clinical disease or shortened life span. In fact, approximately 20% of mice older than 24 months had no motor symptoms and only 18% of VLE mice older than 22 months reached end stage. Given the variable penetrance of clinical phenotype, prolonged survival, and protracted loss of motor neurons the VLE mouse provides a new tool that closely mimics human ALS. This tool will allow the study of pathologic events over time as well as the study of genetic and environmental modifiers that may not be causative, but can exacerbate or accelerate motor neuron disease.
Introduction
Amyotrophic lateral sclerosis (ALS) is a degenerative disease affecting primarily the motor neurons in the spinal cord, brainstem, and motor cortex. Degeneration of the motor neurons leads to progressive paralysis, atrophy of denervated muscles, and ultimately death, with a median survival of less than 5 years. However, there is variability in ALS severity, with 20% of patients living longer than 5 years and 10% of patients living 10 years or more. Although most cases of ALS are sporadic (SALS), about 5-10% of ALS cases are familial (FALS), with approximately 20% of these resulting from mutations in the ubiquitously expressed Cu/ Zn superoxide dismutase (SOD1) gene [1,2]. Currently, 177 ALS causing mutations have been identified in the human SOD1 gene. There is significant heterogeneity of the disease reported in FALS subjects with SOD1 mutations [3]. Paradoxically, some siblings of FALS patients, possessing the same SOD1 mutation, do not show signs of the disease [4]. The role that SOD1 plays in the pathogenesis of ALS is not currently understood. The large variation in age of onset and severity in human ALS patients with specific SOD1 mutations lends support to the likelihood that there are other variables that determine disease expression. These variables likely include both environmental and genetic modifiers of disease [5,6].
Transgenic (Tg) mouse models of FALS containing mutant human SOD1 genes have led to an explosion of research into the causes of ALS [7][8][9]. The most utilized and best characterized Tg mice are the G93A mutant hSOD1 [Tg(hSOD1-G93A)1GUR], abbreviated G93A, which develop motor neuron pathology and clinical symptoms remarkably similar to those seen in ALS patients [9][10][11][12]. These mice show onset with weakness of the hind limbs and tremors at approximately 90 days of age followed by hind limb paralysis and death by 120-150 days of age. The timing of symptoms is directly dependent on the number of copies of the transgene [11,13].
The standard G93A mutant hSOD1 mouse model carries 25 copies of the mutant human SOD1 gene [14]. Loss of transgene copies can occur spontaneously during reproduction as a result of intra-locus recombination during meiosis. Our lab has developed a unique colony of G93A SOD1 mice that have only 4 copies of the G93ASOD1 gene with prolonged survival. These mice, designated as very low expressing (VLE) G93A SOD1 Tg, have a variable phenotype and do not always manifest clinical phenotype. Since the VLE mouse has a less penetrant phenotype with slower progression, it provides a model to study pathologic changes in sequence over a longer period of time so that the pathogenetic mechanisms can more easily be dissected. In addition, it also provides a model to study gene-environment interactions in disease expression as proposed in both familial and sporadic ALS [15]. The aim of this study is to characterize the clinical and pathologic phenotype of the VLE G93A SOD1 Tg mouse model of ALS which, like humans, has a variable penetrance.
Materials and Methods
Mice used in this study originated from a colony of C57BI/ 6JXSJL/J Fn mice hemizygous for the human G93A SOD1 gene. The VLE mice were derived from a spontaneously generated founder male that possessed only 4 copies of the G93ASOD1 gene. Transgene copy number was verified for every animal using QPCR as previously described [5,13].
Two hundred and twenty three mice, 125 Tg+ VLE and 98 Tgwild type controls, were evaluated for this study. Clinical status of each mouse was assessed by measuring weight and by ranking the splay reflex test on a 0-3 scale ( Figure 1). Thirty three of these mice were used for lumbar spinal cord motor neuron counts and TDP-43 immunohistochemistry. The mice were sacrificed at three time points, 9, 18, and 24 months of age. If prior to the scheduled time point the mouse demonstrated paralysis of one hind limb, or was unable to right itself in 30 seconds when placed on its side, the animal was sacrificed. Natural death was not an endpoint in this study. Mice were euthanized by intraperitoneal injection of Beuthanasia-D (solution of sodium pentobarbital and sodium phenytoin, Schering, Kenilworth, NJ) at a dose containing 150 mg/kg pentobarbital followed by transcardial perfusion with 4% formaldehyde in 0.1 M phosphate buffered saline (PBS). The lumbar spinal cord was dissected out, embedded horizontally in paraffin, sectioned in 10 um-thick sections and mounted on gelatin-coated slides.
For neuron counts, every third section was stained with 1% Thionin. The ventral horns of the spinal cord were identified in relation to the presence of the central spinal canal, and all motor neurons in Rexed layer 9 of the lumbosacral segments were counted. All images were collected and analyzed with an Olympus BX60 microscope equipped with a DP70 digital camera and DP software (Olympus America, Center Valley, PA, USA). Motor neurons counts were performed by two independent counters and averaged to limit subjectivity.
The motor neurons of the spinal cord were also qualitatively examined for the presence of TDP-43. Immunohistochemistry was performed using an avidin-biotin-peroxidase complex system (Vectastain Elite ABC Peroxidase Kit; Vector Laboratories, Burlingame, CA). Briefly, paraffin was removed by xylene treatment, and the sections rehydrated through descending grades of alcohol up to water. Non-enzymatic antigen retrieval was performed by heating the sections to 95uC in 0.01 M sodium citrate buffer (pH 6.0) for 40 min in a vacuum oven. After a 30min cooling period, the sections were rinsed, incubated in methanol/3% H 2 O 2 for 20 min to quench endogenous peroxidase, and blocked with 0.1% bovine serum albumin (BSA) with 5% normal goat serum. Primary antibodies against TARDBP/ TDP-43 (Proteintech, Chicago, IL; 1:500) were added and the sections incubated over night at room temperature in a humidifier chamber. After rinsing, sections were incubated with biotinylated anti-rabbit secondary antibodies for 1 h, rinsed, and then incubated with avidin-biotin-peroxidase complex. Sections were developed with a diaminobenzidine (DAB) substrate and counter stained with Hematoxylin.
Motor neurons were characterized for the location of TDP-43 staining from 0 (pale cytoplasm with distinct nuclear staining) to 2 (translocation of the TDP-43 staining from the nucleus to the cytoplasm). The resulting scores were compared by age and between mutant and control mice and most importantly, the percent of neurons with translocation of TDP-43 to the cytoplasm was noted. TDP-43 for each animal was characterized in duplicate by two independent observers and averaged to limit subjectivity.
Ethics
All animal procedures described in this study were carried out in accordance with and approved by the institutional animal care and use committee (IACUC) at Drexel University College of Medicine.
Genotyping
All of the VLE mice were positive for human SOD1 by QPCR with a delta CT versus Interleukin-2 (IL-2) consistent with four copies of the transgene [13]. Transgene negative litter mates were used as controls.
Clinical Status
Significant weight loss can be used as an indicator of disease progression in high copy-number transgenic mice. There were significant (p,0.05) differences in weight at all ages between male and female mice (males heavier than females) in both transgene positive and negative mice, therefore weight loss as an indicator of motor neuron loss was evaluated separately by gender. Although female wild-type mice were on average slightly heavier than their transgenic litter mates, the difference between their weight at any of the ages examined did not reach statistical significance (p. 0.05). For ages less than 18 months old, there was no statistically significant difference (p.0.05) in the weight of male transgenic mice as compared to their wild-type litter mates. However, for ages greater than 18 months old, the male transgene positive mice demonstrated significant (p,0.05) weight loss as compared to their wild-type litter mates ( Figure 2).
There was significantly (p,0.05) increased motor weakness (evaluated with the splay reflex) in the VLE mice as compared to control animals ( Figure 3). The transgenic mice demonstrated splay scores greater than 1 in 10% of the animals by 9 months. In mice between 9 and 18 months of age, 20% of animals showed splay .1 and 74% of mice older than 18 months showed splay . 1. Hind limb paralysis was observed in VLE mice as early as 22 months. However, only 18% of VLE mice older than 22 months demonstrated hind limb paralysis. The average survival for animals that developed hind limb paralysis was 625.0660.8 days. Not all VLE animals manifested clinically evident disease or shortened life span with approximately 20% of mice older than 24 months not manifesting any symptoms of motor neuron disease.
Motor Neurons
The alpha motor neurons, which are the largest cells (.30 mm diameter) in the spinal cord, are located in Rexed layer 9 and easily identified ( Figure 4). The lumbosacral spinal cord alpha motor neurons counts for VLE and control mice are illustrated in Figure 5. The wild-type control mice demonstrated a stable number of motor neurons (27696214, mean 6 se) with no statistically significant difference in numbers (p.0.05) between the 9, 18 and 24 month old animals (p.0.05). In contrast, as compared to controls, the VLE mice showed a progressive decrease in the number of lumbosacral spinal cord motor neurons that reached statistical significance (p,0.05) in the 18 and 24 month old mice ( Figure 5). Signs of individual degenerating motor neurons were observed in VLE mice as early as 9 month old with vacuolation even before significant neuronal loss.
TDP-43 Immunohistochemistry
The distribution of TDP-43 immunoreactivity in spinal cord motor neurons is illustrated in Figure 6. In young animals, most TDP-43 is located in the nucleus with light cytoplasmic staining ( Figure 6A). As the animals aged, the nuclear staining decreased and there was an increase in cytoplasmic staining ( Figure 6B and 6C). This shift in TDP-43 immunoreactivity was more pronounced in the VLE mice as compared to their wild-type litter mates (Figure 7). In the 9 month old mice ( Figure 7A) the percent of motor neurons in which the staining was nuclear only (score = 0) was 66.668.72% (mean 6 sd) in control mice and 56.164.95% in VLE mice. There was partial translocation to the cytoplasm in 29.567.43% of the motor neurons in the control mice vs. 37.463.71% in VLE mice. TDP-43 was translocated to the cytoplasm in 3.8261.69% of control motor neurons vs. 6.4861.76% in VLE mice. This was a statistically significant difference (p = 0.0321).
In the 18 month old mice ( Figure 7B), the percent of motor neurons with only nuclear staining (score = 0) in control mice was 67.264.62% compared to 53.065.33% in VLE. Partial translocation (score = 1) was observed in 28.963.66% of the motor neurons in the control mice compared to 36.563.98% in VLE mice. Cytoplasmic translocation (score = 2) was observed in 3.7761.94% of the motor neurons in the control mice compared with 10.6863.35% in VLE mice. These differences were all statistically significant (p,0.01).
In the 24 month old mice ( Figure 7C), the percent of motor neurons with only nuclear staining (score = 0) in control mice was 60.064.62% compared to only 51.065.64% in VLE mice. There was partial translocation of TDP 43 to the cytoplasm in 33.760.77% of the motor neurons in the control mice compared to 36.162.98% in VLE mice. Cytoplasmic translocation (score = 2) was observed in 6.1563.81% of the motor neurons in control mice compared to 12.8463.61% in VLE mice. These differences were statistically significant (p,0.05) for nuclear localization and cytoplasmic translocation. In contrast to control mice, there was a consistent trend to cytoplasmic localization of TDP-43 in the VLE animals at each age and the cytoplasmic translocation of TDP-43 increased over time ( Figure 7D).
Discussion
We have demonstrated that VLE mice carrying 4 copies of the mutant human G93A SOD1 gene on the mixed B6/SJL background have clinical motor and pathologic changes with neuronal loss and translocation of TDP-43 when compared to controls yet, not all of these animals demonstrate weakness or die at an earlier age. While a previous report described a prolonged survival in a B6 G93A SOD1 mouse colony with reduced copy numbers, these mice all showed onset at 12 months with survival to 15 months [16], suggesting a copy number of less than 8 and greater than 4 based on our previously published work in B6SJL G93ASOD1 mice [13]. In contrast to the VLE mice, these mice did not demonstrate variable penetrance. The variable penetrance despite neuronal loss and translocation of TDP 43 support the VLE model as an important tool to dissect pathologic changes over time and to examine the interaction of environmental and genetic factors in ALS onset and progression.
VLE mice demonstrate motor neuron loss at a slower rate with a protracted time course as compared to the 25 copy G93ASOD1 animals. At 9 months of age, even before significant motor neuron loss, there was vacuolation of the motor neurons observed similar to 25 copy G93ASOD1 mice at 30 days of age. Two year old VLE mice lost approximately 50% of their motor neurons; a loss comparable to a 70 day old 25 copy G93ASOD1 mouse. By two years of age 80% of the VLE mice had clinically evident disease with about one fifth of these animals at end stage. The remaining animals did not manifest any clinical signs of disease. The average survival of mice demonstrating paralysis was 625 days compared with 129 days in the 25 copy mouse. Thus there is clearly a slower time course of neuronal loss and a clinically variable phenotype in the VLE mice as compared to the 25 copy G93ASOD1 animals. In contrast to other G93A SOD1 mutant mouse models which develop an ALS phenotype that eventually leads to hind limb paralysis and death, not all VLE animals manifested clinically evident disease or shortened life span. Only 18% of VLE mice older than 22 months demonstrated hind limb paralysis with approximately 20% of mice older than 24 months not manifesting any symptoms of motor neuron disease.
This slow progression of disease is likely to be more reflective of human SOD1 mediated ALS and will allow the sequential study of pathologic changes much more easily than the rapid course of the 25 copy mouse. In fact, it has previously been noted that translocation of TDP-43 was not a prominent feature of the G93A SOD1 transgenic mouse model carrying 25 copies of the transgene. However, the slow progression and prolonged survival of the VLE mouse allowed us to observe significant translocation of TDP-43 into the cytoplasm from the nucleus. This would further support the VLE mouse as a more realistic model of human disease than the 25 copy G93ASOD1 mouse.
Finally, similar to FALS kindred's, not all VLE animals that carry the four copies of the transgene manifested clinically evident disease or shortened life span. In fact approximately 20% of these animals did not manifest any symptoms of motor neuron disease. This characteristic underscores the increased similarity of this model to human disease and will enable investigators to more easily study environmental and genetic factors that can provoke disease or alter phenotype by using this model. Simply, a comparison of the percentage of animals clinically affected, the timing of disease onset, and survival between groups of VLE mice who are exposed versus not exposed to a particular environmental factor will provide insight into the gene-environment interaction.
We recognize that variable phenotype could result from differential expression of mutant SOD1 by the VLE mice. However, we have previously reported on phenotype variability in transgenic mouse models of ALS [17]. Our study showed that in mice expressing the same number of G93A SOD1 transgene copies, the phenotype variability is mostly due to genetic background variability and not SOD1 expression. The authors feel that given the mixed genetic background (C57BL6/SJL) of the VLE mice most of the variability in ALS phenotype is likely due to genetic background variations and not differential expression of SOD1.
While we acknowledge that the slow progression and variable disease penetrance of the VLE model may demand longer time periods of study, we feel that using these animals in the appropriate way can actually accelerate our understanding of disease. In fact these animals would provide and ideal model to more easily study environmental and genetic factors that can provoke disease or alter phenotype. One example would be to examine the effects of pro-oxidant environmental risk factors and the genetic predisposition to the generation of free radicals in the VLE animals to determine the effects of exposure on the percentage of animals developing disease, the onset, the severity and the lifespan of affected animals. Furthermore, the slow tempo of neuronal loss and changes would enable a clearer picture of sequential events in the progression of motor neuron loss and the accompanying pathologic changes in spinal microglia and astrocytes.
In summary, we describe a new tool in the study of SOD1 mediated disease, the VLE mouse model with a variable penetrance of clinical phenotype, prolonged survival and protracted pathologic loss of motor neurons. These characteristics make the VLE mouse a valuable model for the study of genetic and environmental modifiers that may not be causative, but can exacerbate or accelerate motor neuron disease. | 4,425.4 | 2014-06-19T00:00:00.000 | [
"Biology",
"Medicine"
] |
Spacings Around An Order Statistic
We determine the joint limiting distribution of adjacent spacings around a central, intermediate, or an extreme order statistic $X_{k:n}$ of a random sample of size $n$ from a continuous distribution $F$. For central and intermediate cases, normalized spacings in the left and right neighborhoods are asymptotically i.i.d. exponential random variables. The associated independent Poisson arrival processes are independent of $X_{k:n}$. For an extreme $X_{k:n}$, the asymptotic independence property of spacings fails for $F$ in the domain of attraction of Fr\'{e}chet and Weibull ($\alpha \neq 1$) distributions. This work also provides additional insight into the limiting distribution for the number of observations around $X_{k:n}$ for all three cases.
Introduction
Let X 1:n ≤ X 2:n ≤ · · · ≤ X n:n be order statistics of a random sample X 1 , . . . , X n from a continuous cdf F. For 1 ≤ k ≤ n, we examine the clustering of data around the order statistic X k:n . This is done by an investigation into the limiting properties of the right and left neighborhoods formed by the adjacent spacings (X k+1:n − X k:n , . . . , X k+r :n − X k+r −1:n ) and (X k:n − X k−1:n , . . . , X k−s+1:n − X k−s:n ) for fixed r and s. We let n → ∞ and consider three scenarios: (i) Central case where k/n → p, 0 < p < 1; (ii) Intermediate case where k, n − k → ∞ and k/n → 0 or 1; (iii) Extreme case where k or n − k is held fixed. In the first two cases we show that, under some mild assumptions, these (r + s) spacings appropriately scaled with a common scale parameter converge weakly to a set of i.i.d. standard exponential random variables (rvs). In the extreme case, this conclusion holds only when F is in the domain of attraction of the Gumbel cdf G 3 , or the Weibull type cdf G 2;α with α = 1. A direct and useful consequence of such a result is that order statistics around a selected one arrive as in a homogeneous Poisson process.
Neighborhoods around a selected order statistic have been investigated by several authors in recent years. Almost all these results, starting with X n:n , have concentrated on the distribution of counts around it. We refer to a few, relevant to our results, from an exhaustive list: (Balakrishnan and Stepanov 2005;Dembińska et al. 2007;Pakes and Steutel 1997;Pakes 2009;Dembińska and Balakrishnan 2010). These authors typically consider neighborhoods of the form (X k:n − d, X k:n ) or (X k:n , X k:n + d) where the lengths of the intervals may or may not depend on n; in some papers, the d's are induced by the quantile function F −1 or are chosen to be random. While these approaches are beneficial from a technical perspective, it is more natural and practical to consider neighborhoods that are in the scale of the data collected. This is our motivation for considering the joint distribution of adjacent spacings. Our approach allows us to characterize the process governing the distribution of counts and provides additional insight into the asymptotic properties of the counts of cluster sizes around a specified order statistic.
Section 2 contains preliminaries that explore the properties of uniform and exponential order statistics; it introduces the von Mises conditions and the associated extreme value distributions. Section 3 is concerned with the joint distribution of a central order statistic and spacings adjacent to it on its right and left neighborhoods. The Poisson arrival process of adjacent order statistics is established there. Assuming von Mises conditions, Sect. 4 reaches a similar conclusion for the neighborhood of an intermediate order statistic. Section 5 displays the distributional structure of the extreme spacings assuming that F is in the domain of attraction of an extreme value distribution. Section 6 applies our results and describes the limiting distribution of the counts of observations around an order statistic. Section 7 discusses further applications of our results and contains concluding remarks.
Let f (x) denote the pdf and F −1 ( p), 0 ≤ p ≤ 1, be the quantile function associated with F(x), where F −1 ( p) = inf{x : F(x) ≥ p} for 0 < p ≤ 1, F −1 (0) = sup{x : F(x) = 0}. We interchangeably use x p and F −1 ( p) as the pth quantile. It is well known that if F is differentiable at x p with finite and positive pdf f (x p ), F −1 is differentiable at p with derivative 1/ f (x p ). Standard uniform and exponential rvs are, respectively, denoted by U and Z . An exponential rv with rate parameter λ will be denoted by Exp(λ), and Poi(λ) represents a Poisson rv with mean parameter λ. The sum of r i.i.d. standard exponentials is a Gamma rv, to be denoted as Gam(r ). A Weibull rv with shape parameter δ will be denoted by Wei(δ). Further, a standard normal rv will be denoted by N (0, 1) and its pdf by φ(·). The Z i 's and Z * i 's are i.i.d. Exp(1) rvs. The symbol ∼ indicates asymptotic equivalence.
The rv U i:n , 1 ≤ i ≤ n, is the ith order statistic from a random sample of size n from a standard uniform population. The distributional equivalence, X i:n d = F −1 (U i:n ), for any collection of order statistics from an arbitrary cdf F is helpful in our investigations.
Spacings near a uniform order statistic
The key to our approach is the following well-known exchangeable property of the uniform order statistics. Let U 0:n = 0 and U n+1:n = 1, and define the uniform spacing (1) Then, it is well known that the Δ i,n 's are exchangeable and for any fixed r , and for constants v i ≥ 0, i = 1, . . . r with r ≤ n, and r i=1 v i ≤ 1, the joint survival function of Δ 1,n , . . . , Δ r,n (and hence any collection of r Δ i,n 's) is given by (see, e.g., David and Nagaraja 2003, p. 135)
This means
That is, nΔ i,n forms an i.i.d. Exp(1) sequence Z i as n → ∞. The convergence is fast; Problem P.5.19 of Reiss (1989, p. 201) notes that there exists a constant C such that for every positive integer n and r ≤ n, where B denotes the family of all Borel sets. We record the implications of (2) and the exchangeability of the Δ i,n 's as a lemma given below; it uses the fact that the interarrival times being i.i.d. Exp(λ) rvs is a defining property of a homogeneous Poisson process with rate λ.
Lemma 1 Let U i:n denote the ith order statistic from a random sample of size n from a standard uniform distribution, and assume n → ∞. Then, for any k such that n − k → ∞, (n(U k+1:n − U k:n ), . . . n(U k+r :n − U k+r −1:n )) for any fixed r , and for any k → ∞ for any fixed s, where the Z i 's and Z * i 's are all mutually independent Exp(1) rvs. That is, inter-arrival times of successive order statistics in the right and left neighborhoods of kth uniform order statistic, upon scaling by n, produce asymptotically independent homogeneous Poisson processes if n, k, and n − k approach infinity. If k [resp. n − k] is bounded, the right [resp. left] neighborhood produces a Poisson process in the limit.
Spacings near an exponential order statistic
When F is standard exponential, it is well known that where the Z i 's are i.i.d. Exp(1) rvs. From this representation, it follows that . . , r for r ≤ n − k turn out to be i.i.d. Exp(1) rvs. Only in this scenario, we need finite and distinct scaling constants for the spacings to transform them into i.i.d. exponential rvs for any n, and hence asymptotically as well.
Extremes and von Mises conditions
Suppose there exist sequences of constants a n and b n > 0 such that P{(X n:n − a n )/b n ≤ x} converges to a nondegenerate cdf G(x) corresponding to a rv W . Then, we say F is in the domain of maximal attraction to G and we write F ∈ D(G). Then, it is known that G is necessarily of one of the three types given below.
The following are necessary and sufficient conditions on the right tail of F in order that F ∈ D(G). The first two are due to Gnedenko (1943) and the last one is due to de Haan (1970). (1)) is finite, and the following condition holds for every t > 0: (c) F ∈ D(G 3 ) iff the following hold: E(X |X > c) is finite for some c, and for all real t, where Our approach for the intermediate case assumes the following sufficient conditions that are applicable to absolutely continuous cdf's. The first two are due to von Mises (1936), and the last one is due to Falk (1989) and is weaker than the corresponding von Mises condition that assumes differentiability of the pdf f (see, e. g., David and Nagaraja 2003, p. 300).
(a) F ∈ D(G 1;α ) if f (x) > 0 for all large x and for some α > 0, where The family of limiting distribution for normalized X 1:n corresponds to that of −W where W has one of the above three types of cdfs; parallel necessary and sufficient, and sufficient conditions exist that impose conditions on the left tail of F.
Joint distribution of spacings
For 0 < p < 1, X k:n is a central order statistic if k n → p. For such an X k:n , Smirnov (1952; Theorem 3, p. 12) has shown (as pointed out by a reviewer) that if the condition F(x) = p has a unique solution x p . Since for any fixed j, n(X k+ j+1:n − X k+ j:n ) the limiting joint distribution of the spacings from an arbitrary cdf F can be linked to that of a collection of i.i.d. standard uniform rvs provided the first factor on the right in (9) above converges in probability to a nonzero constant. From (8), it follows that k+ j,n = U k+ j+1:n −U k+ j:n (defined in (1)) almost surely converges to 0. The first factor on the right in (9), if the following condition holds: f is positive, finite and continuous at x p .
This conclusion follows from the definition of the derivative of F −1 and its assumed continuity at p. Upon using (10), (9), Slutsky's Theorem, and Lemma 1, we conclude that jointly where the Z j 's are i.i.d. Exp(1) rvs if (11) holds. We can weaken the continuity assumption for f in (11) with the following condition: where 0 < p < 1. This assumption is similar to (17) in Dembińska et al. (2007) (given as (38) in Sect. 6 later). The condition (11) implies that (12) holds since the latter is satisfied upon dividing the numerator and denominator by h and taking the double limit; the converse is not true.
On the other hand, we can weaken the requirement for a finite nonzero f (x p ) by modifying a condition on F used by Chanda (1975) [see also Ghosh and Sukhatme 1981]. We assume that for some θ > 0. If f is indeed finite and nonzero at x p , then the above condition is satisfied with θ = 1. Whenever f (x p ) is finite and positive or (13) holds, there is a unique solution to F(x) = p and (8) holds. Based on the above discussion, we can now formally state the result for the central case.
Theorem 1 Let k/n → p ∈ (0, 1), and r and s be fixed positive integers.
(a) If condition (11) holds, or if (12) holds and f (x p ) is finite and positive, where the Z 's are i.i.d. Exp(1) rvs. Thus, the two counting processes defined by setting the jth event to occur, respectively, at times converge weakly to two independent homogeneous Poisson processes with unit intensity. (b) Assume (12) and (13) hold. Then, That is, the counting processes defined by setting the jth event of the process to occur at times n θ converge to i.i.d. renewal processes with Wei(1/θ ) renewal distribution. They reduce to homogeneous Poisson processes with unit intensity only when θ = 1 and f (x p ) is finite and positive.
Proof To prove part (a), we need to show that (10) holds whenever (11) holds, or if (12) holds and f (x p ) is finite and positive. Then, we would use (10), (9), Slutsky's Theorem, and Lemma 1. We have shown earlier that (10) holds whenever (11) is satisfied. If (12) holds and f (x p ) is finite and positive, the left side expression in (10) can be written as where the first factor converges to 1 and the second factor converges to 1/ f (x p ), both almost surely. Thus, (10) is established. For (b), the idea is similar. We note that n θ (F −1 (U k+ j:n + k+ j,n )− F −1 (U k+ j:n )) can be written as Assumption (13) coupled with (12) ensures that the first factor above converges almost surely to M( p, θ). Since P(n k+ j,n > w) → exp{−w} for all w > 0, has to be positive and finite, and the limiting arrival process would be Poisson.
Remark
The condition (13) does not imply (12); nor does it ensure that f (x p ) is finite and positive. Consider the pdf This is a corrected version of the pdf given in Chanda (1975), and discussed in Ghosh and Sukhatme (1981) (we thank a reviewer for noticing the error). The associated quantile function is given by This quantile function fails to satisfy the condition in (12) when p = 0.5 and η is a positive even integer, but satisfies (13) with θ = 1/(η + 1). Here,
Asymptotic independence of a central order statistic and spacings in its neighborhood
We now assume k/n = p + o(n −1/2 ) and establish the independence of X k:n and spacings around it.
The uniform parent
Using the (well-known) joint pdf of the consecutive standard uniform order statistics U k−s:n , . . . , U k:n , . . . , U k+r :n , we first obtain the joint pdf of appropriately normalized U k:n and the vector and thus determine the limiting form of the joint pdf. The joint pdf of U k−s:n , . . . , U k+r :n is given by Hence, and the Jacobian is ∂u ∂v = t n /n r +s . The joint pdf of for fixed r and s. Further, , using Stirling's approximations for the factorials and the expansion log The conclusion of the above discussion is summarized below.
Arbitrary parent
By establishing density convergence under the assumption that k/n = p + o(n −1/2 ), we have shown above that The conclusion in (16) also follows from Ghosh (1971) who has shown that if f (x p ) is positive and finite, We have shown in Sect. 3.1 that when k/n = p + o(1), if condition (11) holds or if (12) holds and f (x p ) is finite and positive, and if (12) and (13) hold, X k+ j+1:n − X k+ j:n Δ θ if (13) holds.
In view of Lemma 2, assuming k = np + o( √ n), we have established the asymptotic independence of the normalized spacings (X k+ j:n − X k+ j−1:n ) introduced in Theorem 1 and appropriately normalized X k:n under the conditions stated there. This discussion leads to the following result.
(a) If condition (11) holds, or if (12) holds and f (x p ) is finite and positive, (12) and (13) hold. Then, In both cases, the Z i 's and Z * i 's are Exp(1) rvs, and the r + s + 1 components in the limit vector are mutually independent.
3.3 Remarks-the central case Siddiqui (1960) considered higher order spacings around a central order statistic and showed that when F is continuously twice differentiable and f (x p ) is finite and positive, the rvs are asymptotically independent when k/n → p ∈ (0, 1) with r/n and s/n tending to zero; further, asymptotically the higher order spacings are Gam(r ) and Gam(s), respectively. We have proved a more refined result here with less assumptions on the properties of F, but have taken r and s to be fixed. Pyke's (1965) classic paper on spacings shows (Theorem 5.1) that n f (x p 1 )(X i:n − X i−1:n ) and n f (x p 2 )(X j:n − X j−1:n ) with i/n → p 1 and j/n → p 2 where 0 < p 1 = p 2 < 1 are asymptotically i.i.d. Exp(1) rvs. The key difference is that the spacings considered there are far apart, while our focus is on adjacent spacings around X i:n .
The asymptotic half-normal distribution of the normalized central order statistic under the conditions of part (b) of Theorem 2 is comparable to Chanda's (1975) conclusion; our condition (13) is on F −1 , whereas his comparable condition (given as (6) there) is on F.
The intermediate case
Here, we lean on the work of Falk (1989) and directly examine the convergence of the joint pdf of an intermediate order statistic X k:n and spacings around it. We assume k → ∞ but k/n → 1 as n → ∞ such that n − k → ∞ and assume one of the von Mises sufficient conditions stated in (5)-(7) holds. Theorem 2.1 of Falk (1989) shows that . (17) This is established by showing that the pdf of (X k:n − a n )/b n at x, for all real x. Consider the joint pdf of X k−s:n , . . . , X k:n , . . . , X k+r :n : and consider the transformation y 0 = (x k:n − a n )/b n , y 1 = (x k+1:n − x k:n )/c n , . . . , y r = (x k+r :n − x k+r −1:n )/c n , so that x k:n = a n + b n y 0 ; x k+ j:n = a n + b n y 0 + c n (y 1 + · · · + y j ), j = 1, . . . , r ; x k− j:n = a n + b n y 0 − c n (y * 1 + · · · + y * j ), j = 1, . . . , s.
Lemma 3 Suppose one of the von Mises conditions stated in
(a) For any real y 0 , 1−F(a n +b n y 0 ) 1−F(a n ) = n(1−F(a n +b n y 0 )) n−k → 1.
(b) If c n = b n / √ n − k, for any y 0 , y 1 real, f (a n +b n y 0 +c n y 1 ) f (a n ) → 1.
Proof (a) From Theorem 2.1 of Falk (1989), it follows that (17) holds under the conditions we have assumed, and the limit distribution is N (0, 1). From Theorem 1 of Smirnov (1967) [Remark (ii) of Falk (1989)], it then follows that [n − k + 1 + n(F(a n + b n y 0 ) − 1)]/ √ n − k + 1 → x for all real y 0 . Thus, √ n − k + 1 · 1 − 1 − F(a n + b n y 0 ) 1 − F(a n ) → y 0 since 1 − F(a n ) = (n − k)/n. This implies that (1 − F(a n + b n y 0 ))/(1 − F(a n )) → 1 for any real y 0 . (b) In the proof of his Theorem 2.1, Falk establishes that whenever one of the sufficient conditions stated in (5)-(7) holds, for any real y for which F(a n +b n y) → 1 (or equivalently a n + b n y → x 1 ) as n → ∞, f (a n + b n θ y)/ f (a n ) → 1 uniformly for θ ∈ (0, 1) where a n and b n are given in (17). Part (a) that we just proved implies that for any real y 0 , F(a n + b n y 0 ) → 1 as n → ∞. Thus, from (27), it follows that f (a n + b n θ y 0 )/ f (a n ) → 1 for all real y 0 . For large n − k and real y 1 , Using (27) with y = 2y 0 , we conclude that f (a n + 2y 0 θ b n ){ f (a n )} −1 → 1 uniformly for all 0 ≤ θ < 1. Since a n + b n y 0 + c n y 1 = a n + b n (y 0 + y 1 / √ n − k) is in (a n , a n + 2y 0 b n ) if y 0 > 0 and in (a n + 2y 0 b n , a n ) if y 0 < 0. Hence, f (a n + b n y 0 + c n y 1 )/ f (a n ) → 1 for all real y 0 = 0 and real y 1 . When y 0 = 0, for any real y 1 , f (a n + b n (1/ √ n − k)y 1 )/ f (a n ) → 1 since 1/ √ n − k ∈ (0, 1) and (27) holds. This completes the proof of the claim in (b).
With y = y * 1 + · · · + y * s > 0, consider the following component of τ 2 in (22): F(a n + b n y 0 − c n y) F(a n + b n y 0 ) = 1 − F(a n + b n y 0 ) − F(a n + b n y 0 − c n y) F(a n + b n y 0 ) = 1 − f (a n + b n y 0 − θ * c n y) F(a n + b n y 0 ) · c n y = 1 − yd k,n k where the second form above follows from the mean value theorem, θ * ∈ (0, 1), and d k,n = 1 F(a n + b n y 0 ) f (a n + b n y 0 + θ * c n (−y)) f (a n ) k n , where we have used the fact that c n = 1/n f (a n ). From part (a) of Lemma 3, the first factor of d k,n above converges to 1; from part (b), the second factor approaches 1, and from our assumptions about k and n made in the intermediate case, the third factor also approaches 1. Hence, d k,n → 1 as n, k → ∞. Thus, upon recalling (22), we obtain With y = y * 1 + · · · + y * j , the jth term in the product representing τ 3 in (23) is given by f (a n + b n y 0 + c n (−y)) F(a n + b n y 0 ) = k − j n f (a n + b n y 0 + c n (−y)) f (a n ) 1 F(a n + b n y 0 ) .
Using Lemma 3 as we did in proving d k,n → 1 as n and n − k → ∞, we conclude that the jth factor of τ 3 → 1 for all j and so does τ 3 . With y = y 1 + y 2 + · · · + y j , the jth term in the product representing τ 4 in (24) is given by f (a n + b n y 0 + c n y) 1 − F(a n + b n y 0 ) = n − k − j + 1 n(1 − F(a n )) · f (a n + b n y 0 + c n y) f (a n ) · 1 − F(a n ) 1 − F(a n + b n y 0 ) .
Since F(a n ) = k/n, the first factor above converges to 1, and Lemma 3 shows that the other two factors also approach 1 as n and n − k → ∞. Thus, τ 4 → 1. Finally, with y = y 1 + · · · + y r , consider the following component of τ 5 in (25): 1 − F(a n + b n y 0 + c n y) 1 − F(a n + b n y 0 ) = 1 − F(a n + b n y 0 + c n y) − F(a n + b n y 0 ) 1 − F(a n + b n y 0 ) = 1 − f (a n + b n y 0 + θ * c n y) 1 − F(a n + b n y 0 ) · c n y from the Mean Value Theorem = 1 − y f (a n + b n y 0 + θ * c n y) f (a n ) · 1 − F(a n ) 1 − F(a n + b n y 0 ) · 1 n(1 − F(a n )) where θ * ∈ (0, 1) and we have used the fact that c n = 1/n f (a n ). Lemma 3 implies that the first two factors of y converge to 1 as n, n − k → ∞, and the denominator of the last factor, n(1 − F(a n )), is n − k. Hence, τ 5 ∼ 1 − y 1 + · · · + y r n − k n−k−r → e −(y 1 +···+y r ) , y 1 , . . . , y r > 0, as n, n − k → ∞. Thus, we have formally proved the following theorem.
Remarks-the intermediate case
When F ∈ D(G 1;α ), a n f (a n ) 1 − F(a n ) = F −1 (k/n)c n n − k → α and hence α(n − k)/F −1 (k/n) can be chosen as c n . When F ∈ D(G 2;α ), the von Mises condition implies that α(n − k)/(x 1 − F −1 (k/n)) can be used as c n . When F ∈ D(G 3 ), [n f (a n )m(a n )/(n − k)] → 1 and we can use m(a n )/(n − k) as our c n . From Theorem 3, it follows that as in the central case, asymptotically, any two spacings, possibly of higher order, formed by nonoverlaping collections of order statistics around an intermediate order statistic X k:n are independent, and the collection is independent of X k:n .
Teugels (2001) has introduced a family C * of cdfs F with the following property: F has an ultimately positive pdf f and for all real y, whenever h(x) → 0 as x → x 1 . He states that the condition F ∈ C * 'slightly generalizes ' Falk's (1989) version of von Mises conditions (i.e., (5)- (7)). Assuming F ∈ C * , Teugels shows that upon normalization described above (i) X k:n is asymptotically normal and (ii) (X k:n − X k−s:n ) is asymptotically Gam(s). Their joint distribution and the asymptotic independence are not discussed there.
The upper extreme case
We now assume that k → ∞ such that n − k is fixed. It is well known that when F ∈ D(G) for G given in (4), ((X n:n − a n )/b n , . . . , (X k:n − a n )/b n ) where for any finite fixed j the vector (W 1 , . . . , W j ) has the same joint distribution as and if G = G 3 , where the Z i 's are i.i.d. Exp(1) rvs, and γ is the Euler's constant. The first three representations above are from Nagaraja (1982) who also shows that the joint limiting distribution (W 1 , . . . , W j ) is identical to the joint distribution of the first j lower record values from the cdf G. The representation in (32) is due to Hall (1978), and is more convenient when G = G 3 . Thus, whenever F ∈ D(G), the limiting form of the joint distribution of the normalized spacings and the concerned extreme order statistics can be described as follows: ((X n:n − X n−1:n )/b n , . . . , (X k+1:n − X k:n )/b n , (X k:n − a n )/b n , where the W j 's have one of the forms given in (29)-(32). We now specialize our results for each of the three domains.
The Fréchet domain
In this case, a n can be chosen to be 0 and b n to be F −1 (1 − 1/n) (= x 1−n −1 ). The representation in (33) for the limiting joint distribution along with (29) suggests that an extreme spacing is not asymptotically exponential, and the adjacent spacings are neither independent, nor identically distributed in the limit. The asymptotic independence of the spacings and the extreme order statistic also fail. Hence, when F ∈ D(G 1;α ), the asymptotic distributional structure for the extreme spacings differs from that for the central and intermediate cases.
From (28) and (29), we conclude that when F ∈ D(G 1;α ), the normalized higher order spacing, where the sum S j = Z 2 + · · · + Z j+1 is a Gam( j) rv that is independent of Z 1 . This distributional representation complements the work of Pakes and Steutel (1997) who have given an expression for the cdf of the limiting rv as (p. 192) They comment that this expression for the cdf 'does not seem susceptible to simplification for any choice of the parameter α'; for the other two domains, they provide explicit distributional representation that is equivalent to ours (see below).
The Weibull domain
Here, x 1 (= F −1 (1)) is finite and can be chosen to be our a n and the scaling constant b n can be chosen to be x 1 − x 1−1/n . From (33) and (30), we can conclude that the normalized adjacent spacings are asymptotically i.i.d. exponential iff α = 1 when F ∈ D(G 2;α ). Otherwise, they are all dependent. When α = 1 and k < n, the joint asymptotic distributional structure of ((X n:n − X n−1:n )/b n , . . . , (X k+1:n − X k:n )/b n , (X k:n − a n )/b n , (X k:n − X k−1: Thus, when α = 1, X k:n is asymptotically independent of the spacings in its left neighborhood, but is symmetrically dependent on the ones on its right. This conclusion is formalized in the following result. Theorem 4 When F ∈ D(G 2;α=1 ), for each fixed n − k and s, the asymptotic joint distribution of ((X n:n − X n−1:n )/b n , . . . , (X k+1:n − X k:n )/b n , (X k:n −a n )/b n , (X k:n − X k−1:n )/b n , . . . , (X k−s+1:n − X k−s:n )/b n ) has the representation given by (34) The standard uniform distribution is in D(G 2;α ) with α = 1 and hence has asymptotically i.i.d. extreme spacings. We had reached this conclusion earlier in Lemma 1 (recall x 1 − F −1 (1 − 1/n) = 1/n). But we have a more general result now that describes the symmetric dependence of the right neighborhood spacings on X k:n and is applicable to all F ∈ D(G 2;1 ). In fact with n − k fixed, given (X k:n − x 1 )/b n = u (< 0), (X n:n − X n−1:n )/b n , . . . , (X k+1:n − X k:n )/b n behave asymptotically like the spacings from a random sample of size n − k from a uniform distribution over (u, 0). For the form of the joint distribution, see, e.g., David and Nagaraja (2003;Sec. 6.4).
From (28) and (30), we conclude that when F ∈ D(G 2;α ), the normalized higher order spacing where the sum S j = Z 2 + · · · + Z j+1 is a Gam( j) rv that is independent of Z 1 . This is the conclusion of Theorem 7.2 in Pakes and Steutel (1997).
The Gumbel domain
Using (33) along with the representation for W j in (32), we conclude the following.
The representation in (35) shows that while X k:n is independent of spacings in its right neighborhood, it is correlated with the spacings in its left neighborhood, and this correlation decreases at the rate of 1/(n − k + j), j ≥ 1 as one moves away from it. This is in contrast with the situation when F ∈ D(G 2,1 ). Weissman (1978) has considered the limit distribution of ((X n:n − X n−1:n )/b n , . . . , (X k+1:n − X k:n )/b n , (X k:n − a n )/b n ), and has given the representation given by the first (n − k) components of the vector in (35). He has also noted the independence of these spacings and W n−k+1 (in his Theorem 2). The above theorem shows that using varying scaling sequences for the spacings, we can obtain i.i.d. standard exponential distributions in the limit. In particular, we have the following: ((n − j + 1)(X n− j+1:n − X n− j:n )/b n , j = 1, . . . , n − k + s) d → (Z 1 , . . . , Z n−k+s ).
From (28) and (32) or from the representation (35), we can conclude that when F ∈ D(G 3 ), the normalized higher order spacing, The last equality above follows from the representation for exponential order statistics given in (3). This is the conclusion reached in Theorem 7.1 of Pakes and Steutel (1997).
Cox processes
Whenever F ∈ D(G), we can say that the arrival process of order statistics in the left neighborhood of an upper extreme order statistic is asymptotically a Poisson process and is independent of its value only when F ∈ G 2 with α = 1. The arrival processes on both sides of the extreme order statistics are pure birth processes when F ∈ G 3 ; only the arrival process on the right side is independent of the order statistic. Harshova and Hüsler (2000) have shown that the arrival processes on the left neighborhood of the sample maximum are special Cox processes, when G is of Weibull (G 2;α ) or Gumbel (G 3 ) type cdf. Cox processes are mixed Poisson processes where the time-dependent intensity λ(t) is itself a stochastic process (Daley and Vere-Jones (2003;Sec. 6.2)). Harshova and Hüsler consider the counting process in the left neighborhood of X n:n , N n (·), defined by N n ([a, b) is a Cox process with stochastic intensity function λ(t) = e t−W in the Gumbel case where W has cdf G 3 ; and λ(t) = α(t − W ) α−1 in the Weibull case where W has cdf G 2;α . The cdfs G 2;α and G 3 are given in (4). Representations given in (30), and (31) or (32) provide another characterization of the resulting Cox processes in terms of the distribution of inter-arrival times of order statistics below the maximum.
Higher order extreme spacings
The representation for the special higher order extreme spacing involving the sample maximum (X n:n − X n− j:n , discussed above) can be expanded to other extremes. From (28)-(30), and (32), we conclude that as n → ∞, for fixed 1 ≤ i < j, (X n−i+1:n − X n− j+1:n )/b n converges in distribution to Here, the last distributional equality follows from (3). The above representations are extremely helpful in providing the asymptotic distribution theory for the number of order statistics around a specified extreme order statistic. This will be illustrated in the next section where all cases (central, extreme, and intermediate) will be considered.
Counts of observations around an order statistic
Consider the following count statistics that track the number of observations in the right and left neighborhoods of X k:n : K − (n, k, d) = #{ j : X j ∈ (X k:n − d, X k:n )}, K + (n, k, d) = #{ j : X j ∈ (X k:n , X k:n + d)}. Clearly, and thus the asymptotic distribution theory for spacings developed here can be directly applied to determine the limit distributions of the count statistics for appropriately chosen d that is dependent on n. Pakes and Steutel (1997) have used the link in (37) in the reverse direction in the extreme case where they derive the limit distribution of K − (n, n, d n ) first and use it to determine the limit distribution of the spacing X n:n − X n−k:n . As noted in the introduction, the literature on the investigation into the limit distribution of K − and K + is substantial. Poisson limits are generally obtained when d = d n is nonrandom but is dependent on the behavior of F around the concerned statistic. We now discuss implications of our results on spacings on the asymptotic distribution of counts and compare our results with only the most relevant results in the literature.
The central and intermediate cases-the Poisson counts
We have seen in Theorems 1 and 3 that the (X k+i:n − X k+i−1:n )/c n are asymptotically i.i.d. standard exponential for any fixed (positive or negative) integer i, where c n = 1/n f (x p n ) with p n ≡ p = lim(k/n) ∈ (0, 1) in the central case, and in the intermediate case, p n = k/n → 0 or 1 such that, respectively k or n − k → ∞. In other words, K − (n, k, λ 1 c n ) and K + (n, k, λ 2 c n ) are asymptotically independent, and Poi(λ 1 ) and Poi(λ 2 ) rvs, respectively. This conclusion matches with that of Pakes (2009) This is comparable to our condition (12), but there are differences in these conditions and their implications. While (12) specifies the behavior of the quantile function F −1 around p, (38) puts a similar condition on the property of the cdf F around x p . Dembińska et al.'s neighborhood is determined by d n = F −1 ( p+λ/n)−x p , a quantity dependent on the behavior of F −1 at ( p + λ/n). In contrast, our d n = λ/n( f (x p )) depends on the behavior of F −1 only at x p . When f is continuous around x p and f (x p ) is positive and finite, from L'Hospital's rule it follows that (38) is readily satisfied, and Under (38) and a similar condition on the left neighborhood, Dembińska and Balakrishnan (2010) have established the asymptotic independence of K − and K + . This independence readily follows from our Theorems 2 and 3. The technical conditions used in Dembińska et al. (2007) in the intermediate case (in their Theorem 6.1, for example) appear difficult to verify whereas the familiar von Mises conditions needed here are known to hold for many common distributions. In addition, our results show that counts in disjoint intervals are Poisson and independent, and also that these are independent of the location of X k:n . These finer conclusions on the limiting structure of the neighborhood cannot be reached using any of the currently available results in the literature on count statistics for the central and intermediate cases.
The upper extremes-non-Poisson and Poisson counts
Asymptotic distributions of K − (n, k, d) and K + (n, k, d) have been investigated by many authors when k or n −k are held fixed starting from the work of Pakes and Steutel (1997) who looked at K − (n, n, d). Assuming k is held fixed, Pakes and Li (1998) showed that K − (n, n − k, d) is asymptotically negative binomial, and Balakrishnan and Stepanov (2005) showed that K + (n, n − k, d) is asymptotically binomial. The success probability in these distributions is given by where x 1 is assumed to be infinite. Pakes (2009) has considered the limit distribution of K + (n, n − k, cb n ) with k fixed assuming that F is in the domain of attraction of either Fréchet or Gumbel distribution and the b n 's form the associated scaling sequence. When F ∈ D(G 1;α ), he shows that the limit distribution of K + (n, n − k, b n ) is mixed binomial with parameters k and random success probability that is a function of a Gam(k + 1) rv (his Theorem 5, part (a)). When F ∈ D(G 3 ), K + (n, n − k, λb n ) is shown to be asymptotically a Binomial rv with parameters k and success probability 1 − e −λ (his Theorem 4).
We now examine the consequences of the representations in (36) and the relations in (37). When k is fixed and F ∈ D(G 3 ), for any j, 1 ≤ j ≤ k, P(K + (n, n − k, λb n ) < j) = P(X n−k+ j:n − X n−k:n > λb n ) as n → ∞. Since the maximum value attainable is k, the limit distribution is binomial, a result noted above. Further, P(K − (n, n − k, λb n ) < j) = P(X n−k:n − X n−k− j:n > λb n ) resulting in a negative binomial distribution, a result shown by Pakes and Li (1998). They also derive the limit distribution of K − (n, n−k, λb n ) in other cases; the representations in (36) along with (37) yield us the same results. Of these, a commonly known distribution is obtained only when F ∈ D(G 2;1 ) in which case K + (n, n − k, λb n ) has a censored Poi(λ) distribution that is censored on the right at k; this conclusion was reached in Theorem 4.1 of Dembińska et al. (2007) under a set of technical conditions similar to the one given in (38). Further, K − (n, n − k, λb n ) will have a Poi(λ) distribution and these two statistics are asymptotically independent. Whenever F ∈ D(G 1;α or G 2;α =1 ), we can obtain the asymptotic distributions of K − and K + using (36) and (37) directly. For example, when F ∈ D(G 1;α ), we can use the corresponding representation in (36) to obtain the cdf of K + (n, n − k, b n ) in terms of Gamma rvs (in contrast with the mixed binomial representation of Pakes (2009) mentioned earlier). While closed form expression for the cdf may not be available, the needed probabilities can be evaluated using tractable univariate integrals that involve gamma type integrands that can be easily evaluated numerically. The link between Gamma and Poisson cdfs comes in handy in this simplification.
Discussion
We now provide further illustrations of applications of our results to distribution theory and inference.
Examples
Our examples thus far were the uniform and exponential populations, but our results are widely applicable since the conditions imposed here are satisfied by several common distributions. In the central case, we need positivity and continuity of the population pdf at x p to achieve independent Poisson arrival process in both right and left neighborhoods. von Mises conditions are satisfied by the common distributions that are in the domain of attraction of an extreme value cdf G (given in (4)) and thus the intermediate case also leads to independent Poisson arrival process for these distributions. The extreme case does not require the von Mises conditions, and provides interesting examples of situations where we do not get Poisson processes. For example, for F ∈ D(G 1;α ), a property satisfied by Pareto and loggamma distributions, the arrival process is no longer Poisson. Tables 3.4.2-3.4.4 of Embrechts et al. (1997) contain a good list of distributions in the domain of attraction of each of the three extreme value distributions along with the necessary norming (scaling) constants needed for the application of our results in the extreme case.
Our intermediate and extreme case discussions focused on the upper end of the sample. Parallel results hold for the lower end of the sample and upper-end and lowerend spacings can exhibit different types of clustering processes. For example, in the exponential parent case, upper extremes are in the Gumbel domain, and the lower extremes are in the Weibull domain with α = 1 (i.e., Exp (1)). Thus, for the lower extremes, we have a homogeneous Poisson arrival process in the right neighborhood, whereas for the upper extremes, we have a pure birth process in the left neighborhood of the concerned order statistic.
Inferential implications
Theorem 2 (a) can be used in the central case to provide (asymptotically) distributionfree estimates of x p and f (x p ) as noted by Siddiqui (1960) when he studied the joint distribution of X k:n , X k+r :n − X k:n , X k:n − X k−s:n [see Sect. 3.3]. It follows from Theorem 2 that n f (x p )(X k+r :n − X k−s:n ) is asymptotically Gam(r + s) and this fact can be used to provide estimates of f (x p ) and confidence intervals for the population pdf at the pth quantile. The asymptotic independence of n f (x p )(X k+r :n − X k−s:n ) and √ n f (x p )(X k:n − x p ) and their known familiar distributions can be used to find the distribution of the pivotal quantity (X k:n − x p ) √ n(X k+r :n − X k−s:n ) .
From Theorem 2, it follows that this rv behaves asymptotically as the ratio of a standard normal and an independent gamma rv (or a scaled Chi-square rv) and this distribution is free of f (x p ). It easily leads to an asymptotically distribution-free confidence interval for x p . A similar application of Theorem 3 would provide asymptotically distribution-free inference for the intermediate population quantile F −1 (k/n) and pdf f (F −1 (k/n)) when one of the von Mises conditions is assumed to hold.
In the extreme case, we have seen that the limit distributions of the top k order statistics are dependent on the domain of attraction. Weissman (1978) has discussed in detail inference on tail parameters (extreme quantiles and the tail index 1/α) based on these limit results.
Concluding remarks
It is interesting to note that norming/scaling constants for X k:n and the adjacent spacings are of the same order only for the extreme case (b n ); the limiting distributions are similar as well (functions of Exp(1) rvs). For the central and intermediate cases, the spacings and X k:n are scaled differently, and their limit distributions are different; the spacings are related to exponential, whereas X k:n relates to the normal. For the extreme and intermediate cases, our sufficient conditions that ensure the nondegenerate limit distributions for X k:n and for the adjacent spacings are the same. In the central case, asymptotic normality for X k:n requires k = np + o( √ n) (actually slightly less restriction on k would work), the asymptotic independence property of spacings holds whenever k = np + o(n).
We have focused here on neighborhoods of a single selected order statistic; this work can easily be extended to multiple neighborhoods. In the case of two or more central order statistics and their neighborhoods, we obtain a multivariate normal limit distribution for the selected order statistics and independent Poisson processes around them. Such a set up is considered in Theorem 3.1 of Dembińska and Balakrishnan (2010) where the independence of Poisson counts in right and left neighborhoods of multiple central order statistics is derived. In other cases (for example, one upper extreme, and another lower, considered in Theorem 2.1 of Dembińska and Balakrishnan (2010)), the resulting counting processes will turn out to be independent. | 11,631.2 | 2015-06-30T00:00:00.000 | [
"Mathematics"
] |
Genesis of magnetic anomalies and magnetic properties of archaeological sediments in floodplain wetlands of the Fossa Carolina
Floodplain wetlands are complex systems influenced by many natural and anthropogenic operators. Due to the influence of high and varying groundwater table and high organic contents, geophysical prospection in wetland floodplains quickly reaches the limits of its effectiveness. At the Early Medieval canal Fossa Carolina in southwest Germany, a study design employing magnetometry, drillings, sampling, and in situ rock magnetic measurements was used for environmental magnetic interpretation of magnetic anomalies in magnetograms and sediment layers. This approach offers reliable archaeological interpretation of magnetic anomalies and magnetic properties under the site specific sedimentological conditions of a floodplain wetland. It was also found that man‐made magnetic anomalies in the floodplain are due to the genesis of different remanent magnetizations – specifically, greigite (Fe3S4) can cause distinct magnetic anomalies in floodplains that can be recognized readily in surface magnetic data.
| INTRODUCTION
Alluvial and semi-terrestrial systems can contain remarkable archaeological record (Brown, 1997). Numerous archaeological and geoarchaeological studies have been carried out in alluvial systems during the last decades (Brown, 1997;Howard, 2003;Needham & Macklin, 1992). Every alluvial archaeological record represents a complex interaction between cultural and geomorphic operators (Needham & Macklin, 1992). To better understand the underlying material structure, the floodplain topography and its genesis must be considered, including all biological, chemical, and physical aspects of specific floodplain wetlands (Brown, 1997;Howard & Macklin, 1999).
Site-scale magnetic prospecting is one of the most powerful and effective tools in archaeological prospection (Aspinall, Gaffney, & Schmidt, 2008;Fassbinder, 2015;Kvamme, 2009). However, alluvial environments are often considered problematic for magnetometry.
Due to the limited depth range of magnetic prospecting the overlying sedimentary cover can offer difficult conditions for signal penetration (Clark, 1992). Also, the choice of instruments can produce different results (Linford, Linford, Martin, & Payne, 2007). Nevertheless, sitedependent variables are the primary factors for the detection and interpretation of archaeological structures by magnetic prospection (Cuenca-Garcia et al., 2018;English Heritage, 2008). In this case the most important site-dependent variables are rock, sediment, and soil magnetism (Dunlop & Özdemir, 1997;Evans & Heller, 2003;Jordanova, 2017;Liu et al., 2012;Thompson & Oldfield, 1986). A broad understanding of these site-specific variables is necessary for detailed interpretation of significant archaeological structures in magnetograms (Fassbinder, 2015).
Man-made magnetic anomalies have been found at several sites in river floodplains (Weston, 2001), including Fossa Carolina . So far, the genesis of these anomalies has only been discussed in general terms without considering the physical properties of floodplain sediments. Therefore, the aim of this article is to provide a detailed magnetic investigation of the genesis of magnetic anomalies in the floodplains of Fossa Carolina.
| STUDY AREA
In AD 792/793, Charlemagne started to build a navigable connection between the Rhine-Main catchment and the Danube catchment. This canal is called Fossa Carolina and was one of the major hydroengineering projects in Early Medieval Europe. Bridging of the Central European watershed was of high geostrategic relevance, due to the expanding fluvial communication and transportation network of the Franconian Empire at that time (McCormick, 2002;. Nevertheless, the canal was never finished and only remains of it are visible today (Schmidt et al., 2019;Schmidt, Werther, & Zielhofer, 2018). The post-Carolingian canal filling was reconstructed using excavation, drillings, and sediment analysis (Völlmer et al., 2018;.
The Fossa Carolina is located on a valley watershed between the Altmühl and Swabian Rezat Rivers in Franconia, Bavaria, southern Germany. The valley fills consist of Pleistocene sandy to loamy fluvial sediments. The canal has a length of 2.9 km, as proven by drillings , archaeological excavation (Werther et al., 2015), and direct-push sensing Völlmer et al., 2018). Furthermore, extensive multi-disciplinary research has been carried out at the Early Medieval canal over the last decade (Ettel, Daim, Berg-Hobohm, Werther, & Zielhofer, 2014;Hausmann et al., 2018;Kirchner et al., 2018;Linzen et al., 2017;Schmidt et al., 2018;Schmidt et al., 2019;Stele, 2017;. Therefore, the site offers an excellent case study. Two sections of the Fossa Carolina floodplain were selected for our study: the West-East Section and the Northern Section foothills of the Weissenburger Alb of the southern Franconian Jura (Jätzold, 1962) at an elevation of 413 to 420 m above sea level.
They are situated in water sensitive areas and covered by bogs, floodplain gley soils, and colluvial deposits (LfU-WMS, 2019; . The sections are divided by the current watershed. Due to construction of the canal the original watershed was relocated 850 m toward the northeast and the West-East Section . At present, the areas differ slightly in composition and elevation: few remains of the original excavation ramparts of Fossa Carolina have been detected in the Northern Section by LiDAR (light detection and ranging) digital elevation models (DEMs), and the entire structure is clearly visible in the East-West Section (Figure 1).
| METHODS
The study design is illustrated schematically in Figure 2. It is based on a sequence and combination of geophysical procedures and on-site methods. Esri geographic information system (GIS) ArcGIS Desktop was used for geodata handling, management, spatial analysis, and mapping (Esri, 2019). For geodetic surveying of spatial data (e.g. magnetic surveys and drillings) a Topcon HiPer II global navigation satellite system (GNSS receiver) was used (Topcon, 2019). The entire workflow comprises magnetometry, identification of drilling sites, drillings, sampling, and in situ rock magnetic measurements. Environmental magnetic interpretation was carried out based on the results of the magnetometry and in situ magnetic measurements ( Figure 2).
| Magnetometry
The magnetometry was carried out with a Bartington Instruments
| Drilling campaigns
Based on the results of the magnetometry, drilling locations were defined using a GNSS receiver. The magnetic anomalies that indicate canal structures were sampled with two core transects of three cores each using a closed 50 mm polyethylene inliner ( Figure 3).
| Magnetic susceptibility measurements
Volume magnetic susceptibility measurements were carried out with a Bartington Instruments MS3-Meter-MS2C sensor configuration to localize magnetically conspicuous layers in the core profile (Dearing, 1999). The closed inliners were passed through the loop of a MS2Csensor and measurements were carried out with a spatial resolution of 2 cm and a susceptibility accuracy of ±2 × 10 -6 SI (Bartington Instruments, 2019b). After mass anomalies (κ peaks) were identified in the cores, anomalies were sampled and their stratigraphic location determined by comparing the set of cores in the transects. Prior to their further analysis with a variable field translation balance (VTFB), the original samples were analysed for mass specific susceptibility (χ) and frequency dependent susceptibility (χ fd% ) with a Bartington Instruments MS3-Meter-MS2B sensor configuration (Bartington Instruments, 2019b;Dearing, 1999;Dearing et al., 1996).
| Variable field translation balance (VFTB)
Field dependent and temperature dependent magnetization measurements (Krása, Petersen, & Petersen, 2007) were carried out on all samples with apparent mass anomalies. Hysteresis loops and backfield curves with the key points saturation magnetization (M S ), saturation remanent magnetization (M RS ), coercivity force (B C ), and coercivity of remanence (B CR ) were estimated which allows a differentiation of F I G U R E 2 Work flow of the approach to magnetic anomaly interpretation presented in this study magnetic properties of mass anomaly producers (Day, Fuller, & Schmidt, 1977;Dunlop, 2002aDunlop, , 2002bEvans & Heller, 2003;Roberts, Tauxe, Heslop, Zhao, & Jiang, 2018). In addition, the shape of hysteresis loops and backfield curves allows a first magnetomineralogic typification (Fabian, 2003;Tauxe, Bertram, & Seberino, 2002;Tauxe, Mullender, & Pick, 1996). From temperature dependent magnetization measurements (M/T) Curie temperatures and crystallographic phase transitions of magnetic minerals can be inferred which allows identification of remanence carriers in the sample (Dunlop & Özdemir, 1997;Hanesch, Stanjek, & Petersen, 2006;Moskowitz, 1981). Visualization, analysis and interpretation of VFTB measurements were supported by the RockMagAnalyzer 1.0 software designed by Leonhardt (2006). 2.30 m below ground, this strongly ferrimagnetic layer is duplicated by strata with maximum κ values of 15 000 × 10 -6 SI. Interbedded are sandy alluvial sediments with moderate κ values between 80 and 150 × 10 -6 SI. The upper colluvial sediments also have moderate maximum volume magnetic susceptibility of 200 × 10 -6 SI. For further analysis of the mass anomaly remanence carriers, samples for VFTB measurements and mass specific and frequency dependent susceptibilities were taken from the layers with (extremely) high κ peaks in QP2 and QP1 (Figures 4 and 5).
| Magnetic properties of mass anomaly samples
Hysteresis loops for all mass anomaly samples in Figure 6 Table 1). The χ fd% value and its ratio with χ indicate superparamagnetic (SP) enhancement as a result of burning (Dearing, 1999). Hysteresis measurements point to the magnetite type with only sample 2 having a potbellied hysteresis loop shape parameter (Fabian, 2003). The values of B C and B CR are relatively low, which indicates soft magnetic material and, thus, high magnetic viscosity of the remanence carrier.
Hysteresis and coercivity ratios of both samples indicate vortex state, a mixture of different magnetic grains (Day et al., 1977;Dunlop, 2002a;Roberts et al., 2017). The thermomagnetic measurements in Figure 8(Q-X) and the calculated dominant magnetic phase transitions have marked differences (see also Table 1). The heating cycle of sample 2 (Figure 8(R)) has a magnetization increase between 150 and 250 C. This range demarcates the Curie temperature which could point to titanomagnetite (Fe 2.4 Ti 0.6 O 4 ) (Dunlop & Özdemir, 1997;McElhinny & McFadden, 2000;Soffel, 2002). The heating curve for sample 2 then decreases to zero magnetization at 600 C.
The heating curve for sample 4 (Figure 8(V)) has a convex shape until it rapidly reaches zero magnetization at 560-630 C (see second derivative in Figure 8(W)). The calculated dominant phase transition (Table 1) (Table 1; (Dearing et al., 1996)). They reach M S at 230 and 280 mT (Figure 6(A, C)). In contrast to samples from group 1, they have relatively low M S and much higher B C and B CR . The hysteresis loop shape for both samples is distinctly potbellied (Fabian, 2003). The ratios M RS /M S and B CR /B C indicate the dominance of high coercivity single domain (SD) particles. The heating cycles for both samples (Figure 8(B, J)) have three temperature dependent magnetic phase transitions. The first transition is marked by a bend in the heating curve with an increase of magnetization at 100 C, demarcation the Curie point of goethite (αFeOOH) (Dunlop & Özdemir, 1997). The then following steady magnetization increase ends abruptly at 350 C, which is indicative for a maximum unblocking or alteration temperature of greigite (Fe 3 S 4 ) (Skinner, Erd, & Grimaldi, 1964) as the dominant solid magnetic phase of samples 6 and 10 (Chang et al., 2014;Reynolds et al., 1994;Roberts, Chang, Rowan, Horng, & Florindo, 2011). After the general magnetization decrease, there is still a small peak at 580 C demarcating the partial transformation of greigite to magnetite.
Stratigraphically, these samples were situated below the samples from group 2 and are, thus, also part of the canal fillings (Figures 4 and 5). Although they have different hysteresis and backfield parameters, they have a similar thermomagnetic behaviour ( Figure 8(E-H, M-P)). Both heating curves decrease rapidly at 300 C to 400 C, which demarcates greigite as the dominant remanence carrier (Table 1). During the heating cycle greigite transforms entirely to magnetite. However, the greigites in samples 8 and 12 have different particle size distributions, as indicated by different hysteresis and coercivity ratios (Figure 6(B, D): Sample 8 is already saturated at 250 mT, while sample 12 reaches saturation at 300 mT). Also, B CR /B C for sample 8 is high (Figure 6(D) and Table 1), which indicates probably SP enrichment (Rowan & Roberts, 2006;Rowan, Roberts, & Broadbent, 2009).
| Magnetic anomaly interpretation: an environmental magnetic perspective
The magnetic properties of the various mass anomaly samples responsible for the linear magnetic anomalies in magnetograms at Fossa Carolina originate not only from different natural remanent magnetization (NRM), but also from different post-depositional genetic processes (Evans & Heller, 2003;Liu et al., 2012). The linear, intensive, and dipolar anomaly originates in a 40 cm thick layer of black, reddish-brown and red layers at a depth of 1.35 to 1.55 m. The highest κ values occur in a brown to reddish-brown, layer with a high allochthonous, ferrimagnetic titanomagnetite fraction. The high viscosity of the soft magnetic titanomagnetite and magnetite triggered the orientation of the remanence carriers in the soft water-soaked sediments toward Earth's magnetic field vector, thus causing intensive dipolar anomalies. In areas with simple positive magnetic anomalies, we assume that these layers are not under the influence of ground water (south of the Northern Section in Figure 3).
In natural environments titanomagnetite formation can occur in two forms: i. Primary formation of titanium rich magnetic iron compounds can only take place in magmatic rocks (Schmincke, 2010;Schön, 2011).
There are no basaltic rocks or their weathering products in the river basin of the Swabian Rezat (ArcTron 3D GmbH, 2009;Schmidt-Kaler, 1976). Therefore, the magnetic mass anomalies of the group 1 T A B L E 1 Selected environmental magnetic parameters and dominant magnetic phases (lower row) Figure 8).
samples must have been produced by an allochthonous/anthropogenic influence. This is supported by the fact that sediment colours and magnetic properties of these layers show clear signs of fire exposure. The findings of , that the titanomagnetite/magnetite layer, which produces the intensive dipolar anomaly, correlates stratigraphically with the construction of Fossa Carolina, and the fact that these intensive anomalies in the SQUID data of Linzen and Schneider (2014), which show that the same remanence carriers are found in the entire northern part of the canal, support the interpretation of the mass anomalies as anthropogenically introduced/influenced and burned materials. Additional (geo) archaeological research is necessary to determine the source and specific characteristics of this material.
Greigite-generated mass anomalies (groups 2 and 3), which partially follow the course of the canal and of the canal fills, is supported by the fact that samples 6 and 10 have similar magnetic behaviour as the SD greigites found by Faßbinder and Stanjek (1994), Hall, Cisowski, and King (1997), and Roberts et al. (1998). The characteristics of sample 12 are comparable to those reported by Jelinowska et al. (1998). The magnetic properties of sample 8 relate to genetic processes, which we interpret to be similar to those described by Rowan et al. (2009). All layers with greigite-generated mass anomalies may have undergone post-sedimentary remagnetization: These highly coercive SD greigites were generated as secondary/authigenic formations under highly organic, anaerobic conditions after deposition of the organic rich canal fillings. Thus, the organic canal fills offered the necessary conditions for producing mass anomalies due sulphidic remanence carriers and the sediments acquired their stable chemical remanence magnetization (CRM) (Snowball, 1997). These findings are similar to those of Stanjek, Fassbinder, Vali, Wägele, and Graf (1994) in gley soils, where the environmental conditions were similar to those in the canal fills. Also, the hysteresis and coercivity ratios of the greigite in the canal fills point at differences in the domain state of the dominant solid FeS magnetic phases. For a correct domain state diagnosis, further analyses such as FORC diagrams or scanning electron microscopy/transmission electron microscopy (SEM/TEM) observations are necessary (Chang et al., 2014;Roberts et al., 2019). Should these methods deliver uncertain results, methods such as X-ray diffraction could also be used (Linford, Linford, & Platzman, 2005).
In order to make such magnetic and non-magnetic analyses possible in the future, we recommend the use of closed polyethylene inliners when drilling/sampling, because they allow longer storage of the samples without any changes in the magnetic mineralogy during the storage (Stele, 2017). Samples for in situ magnetic measurements should be then taken from the inliner immediately before analysis.
This sample preparation approach prevents the oxidation of possible ferrimagnetic iron sulphides in the samples.
| CONCLUSIONS
At the Fossa Carolina in southern Germany a workflow comprising magnetometry, identification of drilling sites, drillings, sampling and rock magnetic measurements was used to interpret magnetic anomalies. The approach offers a reliable estimate of: • the depth of magnetic mass anomalies; • differentiation between natural and anthropogenically induced structures; and • genetic processes leading to the production of mass anomalies.
Our results indicate that secondary/post-sedimentary processes in the canal fillings produced magnetic anomalies in the Fossa Carolina. Greigite-generated mass anomalies in deeper sediment layers can cause detectable anomalies in near-surface magnetic surveys with a 1 m fluxgate at a depth of up to 3.5 m. There is also strong evidence that titanomagnetite/magnetite-generated mass anomalies at Fossa Carolina were anthropogenically introduced. Evidence of greigite in organic-rich canal fills in a European Watershed means that these processes can be expected in palaeochannels or active channels in middle European floodplains with similar characteristics. | 3,879.4 | 2019-12-16T00:00:00.000 | [
"Geology"
] |
The mitochondrial fission factor FIS1 promotes stemness of human lung cancer stem cells via mitophagy
Mitophagy, a form of autophagy, plays a role in cancer development, progression and recurrence. Cancer stem cells (CSCs) also play a key role in these processes, although it not known whether mitophagy can regulate the stemness of CSCs. Here, we employed the A549‐SD human non‐small cell lung adenocarcinoma CSC model that we have developed and characterized to investigate the effect of mitophagy on the stemness of CSCs. We observed a positive relationship between mitophagic activity and the stemness of lung CSCs. At the mechanistic level, our results suggest that augmentation of mitophagy in lung CSCs can be induced by FIS1 through mitochondrial fission. In addition, we assessed the clinical relevance of FIS1 in lung adenocarcinoma using The Cancer Genome Atlas database. An elevation in FIS1, when observed together with other prognostic markers for lung cancer progression, was found to correlate with shorter overall survival.
Mitophagy, a form of autophagy, plays a role in cancer development, progression and recurrence. Cancer stem cells (CSCs) also play a key role in these processes, although it not known whether mitophagy can regulate the stemness of CSCs. Here, we employed the A549-SD human non-small cell lung adenocarcinoma CSC model that we have developed and characterized to investigate the effect of mitophagy on the stemness of CSCs. We observed a positive relationship between mitophagic activity and the stemness of lung CSCs. At the mechanistic level, our results suggest that augmentation of mitophagy in lung CSCs can be induced by FIS1 through mitochondrial fission. In addition, we assessed the clinical relevance of FIS1 in lung adenocarcinoma using The Cancer Genome Atlas database. An elevation in FIS1, when observed together with other prognostic markers for lung cancer progression, was found to correlate with shorter overall survival.
Cancer stem cells (CSC) were first discovered in the hematopoietic system [1] and subsequently identified in a broad spectrum of solid tumors. CSCs comprise a group of cells with self-renewal and differentiation ability. They play a key role in tumorigenesis, metastasis and recurrence, as well as in the development of drug resistance [2][3][4][5][6][7][8][9].
Autophagy is an evolutionarily conserved biological process of energy metabolism. By degrading intracellular organelles and proteins, autophagy provides cells with biochemical reaction substrates for the maintenance of homeostasis under nutrient deprivation or other stressful conditions [10]. Autophagy is mainly divided into macroautophagy, small autophagy and selective autophagy, among which macroautophagy has been the focus of the field [11]. The key processes of autophagy are the formation of a bilayer membrane structure in the cytoplasm by autophagy-related proteins, bilayer membrane packaging of the substances that need to be degraded, the transportation of the contents of autophagic vesicles to the lysosome to form autophagosomes and, finally, the degradation of vesicle contents in the acidic environment of lysosomes [12]. Autophagy is an important metabolic form for normal cells. Researchers have reported that hematopoietic stem cells lost their ability to self-renew and regenerate after inhibition of autophagy [13]. Autophagy is essential for the maintenance of stemness of both stromal stem cells [14] and embryonic stem cells [15].
The role of autophagy in cancer is contextdependent. In the early stage of tumorigenesis, autophagy inhibits the development of tumors [16,17], with autophagy-related gene beclin1 acting as a tumor suppressor [18]. In the metastasis stage of cancer, autophagy augments tumor progression [19]. Emerging evidence supports a role of autophagy in the regulation and maintenance of the stemness of CSCs. The level of basal autophagic activity is higher in CSCs than that in normal tumor cells, and changing the autophagy level will affect the stemness of CSCs [20]. Liu et al. [21] found that mitophagy maintains the hepatic CSC population by removing mitochondriaassociated p53, which otherwise would be activated by PINK1 (i.e. phosphatase and tensin homolog-induced putative kinase 1) to suppress the expression of NANOG. However, a mechanistic understanding of the underlying autophagy regulation of the stemness of CSCs remains elusive.
Mitophagy is a form of selective autophagy that degrades damaged and aged mitochondria [22]. There is a strong correlation between mitochondrial division and the level of mitophagic activity. After mitochondrial division, senescent and damaged mitochondria are split. Thereafter, the damaged mitochondria are degraded by mitophagy and the excess reactive oxygen species (ROS) in the cell is metabolized at the same time [23]. Mitochondrial division is regulated by a number of mitochondrial fusion and fission genes, of which the most studied are the mitochondrial fission gene Drp1 and the mitochondrial fusion gene Opa1 [24]. Yamada et al. [25] showed that, in non-alcoholic fatty liver, mitophagy of hepatocytes augments the progression of the disease. The mitophagic activity of hepatocytes depends on the size of mitochondria, which appears to be regulated by Opa1 and Drp1. However, recent studies have shown that mitochondrial fission gene Fis1, encoding the FIS1 protein anchored on the outer membrane of the mitochondria, regulates mitochondrial division and the degradation of damaged mitochondria [26,27].
Mitochondria act as an energy supply of cells, and mitophagy is also important for the survival of tumor cells and CSCs [28,29]. A few recent studies have begun to show how mitophagy regulates the stemness of CSCs [30]. Pei et al. [31] showed that, in leukemia stem cells, FIS1 is highly expressed and FIS1-mediated mitophagy is required for the maintenance of the stemness of leukemia stem cells. However, whether FIS1 directly regulates mitophagy is not clear. During our characterization of mitophagic activities in lung adenocarcinoma A549-SD CSCs, we observed augmentation of stemness by mitophagy and the elevated expression of FIS1. The main finding derived from the present study is that FIS1 regulates mitophagy via mitochondrial division to promote the stemness of human lung CSCs.
Mitophagy promotes the stemness of cancer stem cells
For the present study, we employed the A549-SD human non-small cell lung adenocarcinoma CSC model that we have developed and characterized In a previous study conducted in our laboratory, the stemness of A549-SD cells was confirmed both in vitro and in vivo [32]. Before carrying out this research, we conducted some in vitro experiments to verify the stemness of A549-SD lung CSCs. Although the parental A549 cells grow adherently, A549-SD CSC cells grow in suspension (Fig. 1A). The results of the spheroid formation assay using 1000 cells in a six-well plate and the spheroid formation assay of single cells in a 96-well plate also showed that the spheroidization ability of A549-SD cells was stronger than that of the parental A549 cells (Fig. 1B,C). To verify whether mitophagy affects the stemness of CSCs, we changed the level of mitophagic activity by treating human lung CSC A549-SD with (a) carbonyl cyanide m-chlorophenyl hydrazone (CCCP) that induces mitophagy and (b) mitochondrial division inhibitor 1 (Mdivi-1) that inhibits mitophagy. These two drugs are classical drugs for mitophagy research [33,34]. After 24 h of treatment, mRNA expression of the stemness gene was examined (Fig. 1D). The expression of bmi1 was significantly decreased upon inhibition of mitophagy by Mdivi-1. To further verify the effect of mitophagy on the stemness, we performed the single-cell spheroid formation assay in a 96-well plate and the consecutive spheroid formation assay using 1000 cells in a six-well plate to measure spheroid formation efficiency. Inhibition of mitophagy upon treatment with Mdivi-1 significantly decreased spheroid formation (Fig. 1E,F). The lack of robust effect of mitophagy induction by CCCP on the stemness of A549-SD ( Fig. 1) could be ascribed to the significantly enhanced stemness in A549-SD cells as a result of CSC enrichment. Taken together, mitophagy activity is required for the maintenance of the stemness of A549-SD lung CSCs.
The mitochondria of non-CSC tumor cells differ from those of the CSCs
A study reported that the levels of autophagy and the distribution of organelles are different in tumor cells and CSCs [35]. To explore the mechanisms underlying the regulation of the stemness of CSCs by mitophagy, we compared the differences in mitochondrial function between A549 and A549-SD CSCs. We found that (a) A549-SD CSCs have lower levels of ROS ( Fig. 2A) and a higher membrane potential than A549 cells (Fig. 2Bi,ii). (b) Confocal immunofluorescence analysis revealed that the number of mitochondria in CSCs, visualized using TOM20 mito-tracker (Proteintech, Chicago, IL, USA), was much higher than that of A549 cells (Fig. 2C). (c) The size of individual mitochondria was counted using IMAGEJ (NIH, Bethesda, MD, USA). We found that the mitochondria of non-CSC tumor cells were larger than those of the CSCs (Fig. 2Di,ii). (d) The expression of the mitochondrial fission gene Fis1 was most significantly increased in A549-SD CSCs (Fig. 2E). This is consistent with the fact that mitochondrial content in A549-SD CSCs is higher than that in A549 cells. We next examined the relationship between mitochondrial division and mitophagy and the stemness of CSCs, respectively.
Mitochondrial division promotes mitophagy of cancer stem cells
To explore whether mitochondrial division promotes mitophagy in A549-SD CSCs, we used small hairpin RNA (shRNA)-mediated knockdown of the mitochondrial division gene Fis1 Fig. 3A,B) and examined the change in the mitochondria of A549-SD CSCs. Upon knocking down Fis1, although the size of the mitochondria increased, the number decreased (Fig. 3C,Di,ii), indicating the successful blockade of the mitochondrial division. To determine the effect of changes in mitochondrial division rate on mitophagy, we examined the autophagy marker proteins P62 and LC3. After knocking down Fis1, accumulation of P62 was enhanced, whereas the conversion of LC3I to LC3II was attenuated (Fig. 3E). These observations indicate that FIS1 may be a key factor in activating autophagy.
Mitochondrial fission promotes the stemness of CSCs
To examine the effects of mitochondrial division on the stemness of A549-SD CSCs, we performed a series of experiments in vitro. Knocking down of Fis1 lead to a significant reduction of: (a) the expression of two key stemness genes, bmi1 and aldh1 (Fig. 4A) and (b) the spheroid formation efficiency (Fig. 4B,C) in vitro. In vivo, we performed a subcutaneous tumor transplantation assay in nude mice to determine the effect of Fis1 knockdown on tumorigenesis efficiency. Accordingly, 100 000 tumor cells were transplanted subcutaneously on each side of the mouse and tumors were removed after 2 weeks of transplantation. Tumor formation was significantly reduced after Fis1 interference (Fig. 4Di,ii).
Collectively, we concluded that mitochondrial division augments the stemness of CSCs by promoting mitophagy.
The expression of FIS1 is lower in the early stage of lung cancer To determine the clinical relevance of our findings in lung adenocarcinoma, we analyzed clinical data from lung adenocarcinoma patients in the The Cancer Genome Atlas (TCGA) database. We found that the expression of FIS1 was lower in stage I lung adenocarcinoma compared to that of stage II/III/IV lung adenocarcinoma (Fig. 5A). Kaplan-Meier plotter analysis revealed that patients with low FIS1 expression had significantly better survival (Fig. 5B). These results indicated that FIS1 may be combined with other factors to provide a more accurate prognosis for lung cancer progression. The significance of this clinical finding requires further investigation and validation.
Discussion
In the present study, using the A549-SD non-small cell lung carcinoma stem cell model that we have developed and characterized [36], we have shown that activation of mitophagy by the Fis1 gene augments the stemness of lung CSC. Specifically, we have made several observations that have not been reported previously.
First, differences exist between the mitochondrial traits of lung CSCs and non-stem lung cancer cells. Lung CSCs have a higher mitochondrial content and a smaller mitochondrial size, as well as a denser perinuclear distribution, compared to non-stem counterparts. In addition, lung CSCs have lower ROS and higher mitochondrial membrane potential, consistent with prior reports in liver CSCs, leukemia CSCs and lung CSCs [37,38]. However, previous reports did not use stable and purified CSCs; thus, a mechanistic understanding of such differences with respect to the regulation of the stemness of CSCs has not been achieved. Second, we investigated the mechanism underlying the differences in the mitochondria features between lung A549-SD CSCs and parental A549 cells. We found that mitochondrial fission gene Fis1 is expressed at a higher level in A549-SD CSCs compared to parental A549 cells. We then made the hypothesis that a mitochondrial fission gene such as Fis1 regulates the stemness of lung CSCs. Pei et al. [31] showed that in leukemia CSCs, the highly expressed FIS1 could regulate the stemness of leukemia CSCs by mitophagy. In the present study, using the stable A549-SD CSC model, we showed that, upon Fis1 silencing, mitochondrial division was inhibited and the stemness of A549-SD was also attenuated. This set of observations indicates that FIS1 augments the stemness of CSCs by promoting mitochondrial division. To our knowledge, our present study, for the first time, has demonstrated the direct regulation of the stemness of CSCs by FIS1. However, the mechanisms regulating FIS1 expression remain need to be elucidated. Furthermore, the specific mechanism underlying the regulation of mitophagy by FIS1 has not been investigated.
Third, we determined the clinical relevance of FIS1 in lung adenocarcinoma by analyzing the relationship between the expression of FIS1 in lung cancer patients and the clinical stage of lung cancer in the TCGA database. We found that there is an inverse relationship between FIS1 expression and the clinical stage of lung cancer. Furthermore, high FIS1 expression confers a shorter overall survival. Clinical validation studies are required to confirm the significance of FIS1 expression.
Materials and methods
Animals BALB/c nude mice were purchased from the Experimental Animal Centre of Chongqing Medical University (Chongqing, China). Animal studies were conducted in accordance with an approved protocol and with the institutional animal welfare guidelines of Chongqing Medical University.
Cell culture
A549 cells was purchased from the cell bank of the Shanghai Institute of Life Sciences, Chinese Academy of Sciences (Shanghai, China). A549 cells were cultured with RPMI 1640 culture medium that contained 1% amphotericin B, 1% penicillin-streptomycin and 10% fetal bovines serum. The derivative CSC A549-oncosphere cells (A549-SD) were isolated and characterized as described previously [36] and the stemness of A549-SD cells had also been verified [32]. Cells were cultured in Dulbecco's modified Eagle's medium/ F12 1 : 1 culture medium with 2% B27, 1% amphotericin B and 1% penicillin-streptomycin. All cells were kept in the incubator with 5% CO 2 and constant humidity at 37°C.
Six-well plate serial spheroid formation assay
In this assay, single-cell suspensions were plated at 1000 cells/well in a six-well plate with 2 mL of culture medium. After 1 week in culture, clonogenic spheroids containing more than 50 cells were counted under microscopy. Spheroid cultures were then collected and single-cell suspensions were prepared for setting up the second round of the assay. The assay was repeated for three consecutive rounds.
Ninety-six-well plate single-cell cloning assay Single cell suspension was prepared and one cell was added into a well with 100 lL of culture medium. Single-cell seeding in each well was confirmed by examination under a microscope and wells containing a single cell were marked. After culturing at 37°C in 5% CO 2 for 1 week, colonies exceeding 50 cells were counted.
Reverse transcriptase-PCR (RT-PCR)
Total RNAs were isolated with Trizol agent (Thermo Fisher). In brief, 1 mL of Trizol agent were added into the Eppendorf tube containing cells, then incubated on the ice for 10 min, followed by the addition of 200 lL of chloroform. The mixture was centrifuged at 4°C for 15 min followed by the addition of isoamyl alcohol and then incubation on the ice for 10 min. Finally, the mixture was centrifuged at 4°C for 10 min followed by the addition of 70% ice-cold ethanol and then another centrifugation in at 4°C for 10 min. The pellet was dissolved using 20-30 lL of diethyl pyrocarbonate water. RT-PCR was conducted using Prime Script RT Master Mix (Takara, Kyoto, Japan) in accordance with the manufacturer's instructions.
Packaging of shRNA lentivirus
The small interfering RNA (siRNA) of FIS1 was purchased from GenePharma (Suzhou, China). Plasmid construction and lentivirus packaging were also prepared and supplied by GenePharma. Lentivirus-mediated short hairpin RNA was obtained using H1/GFP&Puro (GenePharma).
Real-time quantitative PCR
The reaction volume for the quantitative PCR was 10 µL, which comprised 5 µL of Master Mix agent (Takara), 0.8 µL of cDNA, 0.8 µL of primer and 3.4 µL of diethyl pyrocarbonate water. Triplicates were set up for every sample. The PCR was conducted in accordance with the manufacturer's instructions. The sequences of PCR primers are provided in Table 1.
Western blotting
The present study used 12% resolving gel and 6% condensing gel. The voltage of electrophoresis was 120 V, whereas the current for member transfer was 350 mA. After transfer, the membrane was blocked with 5% BSA (BD, Franklin Lakes, NJ, USA) and then incubated with the primary antibody at 4°C for 12 h, followed by incubation with the secondary antibody at room temperature for 1 h. The results were acquired using electrophoresis gel imaging (Bio-Rad, Hercules, CA, USA).
Subcutaneous tumor transplantation in nude mice 1 9 10 5 cells were injected to each side of the hinder leg of each nude mouse. After 2 weeks, mice were killed and tumors were obtained and weighed.
Immunofluorescence
Cells were collected, fixed with ice-cold 70% ethanol for 30 min on the slide and treated with hydrochloric acid for 1 h. After washing with phoshate-buffered saline three times, the slides were blocked with 10% BSA (BD) solution at 37°C for 1 h. Thereafter, cells were incubated with the primary antibody at 4°C for 12 h, followed by incubation with the secondary antibody at 37°C for 1 h. Images were taken using a confocal microscope (Leica, Wetzlar, Germany).
Clinical data analysis mRNA expression of FIS1 in lung adenocarcinoma was evaluated by TCGA Research Network (http://cancerge nome.nih.gov). To analyze the survival of patients with lung adenocarcinoma, patient samples were analyzed using Kaplan-Meier plotter analysis (http://kmplot.com).
Statistical analysis
Data were analyzed by one-way analysis of variance using PRISM, version 8 (GraphPad Software Inc., San Diego, CA, USA) and are presented as the mean AE SD. P < 0.05 was considered statistically significant. The size of the mitochondria was measured using IMAGE J. | 4,198.6 | 2021-05-29T00:00:00.000 | [
"Biology"
] |
Application of Taguchi Method for Optimization of Friction Stir Welding Process Parameters to Joining of Al Alloy
In this study, the joining of 6061-T4 Al alloy plates are carried out using friction stir welding (FSW) process and the process parameters are optimized using Taguchi method. The rotational speed, welding speed and axial force are the process parameters taken into consideration. The optimum process parameters are determined with reference to tensile strength of the joint. The results indicate that the rotational speed is highest significant parameter to deciding the tensile strength of the joint. The result shows that optimal values of process parameters are to get a maximum tensile strength of friction stir welded AA 6061 is 162 MPa. Friction stir welding (FSW) is a solid state joining process that invented at The Welding Institute (TWI) United Kingdom in 1991, is a viable technique for joining aluminium alloys that are difficult to fusion welding [1]. No defects are observed in FSW like porosity, alloy segregation and hot cracking, and welds are produced with good surface quality and thus no post weld cleaning is required [2]. There have been a lot of efforts to understand the effect of process parameters on material flow behavior, microstructure formation and mechanical properties of friction stir welded joints. The effect of some important process parameters on weld properties is major area for researchers [3-5]. In order to study the effect of FSW process parameters, most of follow the traditional experimental techniques, i.e. varying one parameter at a time while other parameters are constant, this conventional parametric design of experiment approach is time consuming. Taguchi statistical design is a powerful tool to identify significant factor from many factors by conducting relatively less number of experiments. Though research work applying Taguchi method on various processes have been reported in literatures [6-11], it appears that the optimization of FSW process parameters of 6061-T4 aluminium alloy using Taguchi method has not been reported yet. Considering the above facts, the Taguchi method is adopted to analyse the effect of each processing parameters (i.e. rotational speed, welding speed and axial force) for optimum tensile strength of friction stir welded joints of 6061-T4 aluminium alloy. Taguchi method Taguchi, a Japanese quality engineer widely recognized as the father of quality engineering [12], addresses quality in two main areas: off-line and on-line quality control. Both of these areas are very cost sensitive in the decisions that are made with respect to the activities in each. Off-line quality control refers to the improvement in quality in the product and process development stages. On-line quality control refers to the monitoring of current manufacturing processes to verify the quality levels produced [13]. The most important difference between a classical experimentaldesign and a Taguchi methodbased robust design technique is that the former tends to focus solely on the mean of the quality characteristic, while the later considers the minimization of the variance of the characteristic of interest. Although the Taguchi method has drawn much criticism due to several major limitations, it has been able to solve single response problems effectively.
A B S T R A C T
In this study, the joining of 6061-T4 Al alloy plates are carried out using friction stir welding (FSW) process and the process parameters are optimized using Taguchi method.The rotational speed, welding speed and axial force are the process parameters taken into consideration.The optimum process parameters are determined with reference to tensile strength of the joint.The results indicate that the rotational speed is highest significant parameter to deciding the tensile strength of the joint.The result shows that optimal values of process parameters are to get a maximum tensile strength of friction stir welded AA 6061 is 162 MPa.
Friction stir welding (FSW) is a solid state joining process that invented at The Welding Institute (TWI) United Kingdom in 1991, is a viable technique for joining aluminium alloys that are difficult to fusion welding [1].No defects are observed in FSW like porosity, alloy segregation and hot cracking, and welds are produced with good surface quality and thus no post weld cleaning is required [2].There have been a lot of efforts to understand the effect of process parameters on material flow behavior, microstructure formation and mechanical properties of friction stir welded joints.The effect of some important process parameters on weld properties is major area for researchers [3][4][5].In order to study the effect of FSW process parameters, most of follow the traditional experimental techniques, i.e. varying one parameter at a time while other parameters are constant, this conventional parametric design of experiment approach is time consuming.Taguchi statistical design is a powerful tool to identify significant factor from many factors by conducting relatively less number of experiments.
Though research work applying Taguchi method on various processes have been reported in literatures [6][7][8][9][10][11], it appears that the optimization of FSW process parameters of 6061-T4 aluminium alloy using Taguchi method has not been reported yet.Considering the above facts, the Taguchi method is adopted to analyse the effect of each processing parameters (i.e.rotational speed, welding speed and axial force) for optimum tensile strength of friction stir welded joints of 6061-T4 aluminium alloy.
Taguchi method
Taguchi, a Japanese quality engineer widely recognized as the father of quality engineering [12], addresses quality in two main areas: off-line and on-line quality control.Both of these areas are very cost sensitive in the decisions that are made with respect to the activities in each.Off-line quality control refers to the improvement in quality in the product and process development stages.On-line quality control refers to the monitoring of current manufacturing processes to verify the quality levels produced [13].The most important difference between a classical experimentaldesign and a Taguchi methodbased robust design technique is that the former tends to focus solely on the mean of the quality characteristic, while the later considers the minimization of the variance of the characteristic of interest.Although the Taguchi method has drawn much criticism due to several major limitations, it has been able to solve single response problems effectively.
The Taguchi method attempts to optimize a process or product design and is based upon three stages, as follows: 1. Concept design or system design 2. Parameter design 3. Tolerance design The following are the steps to be followed for process parameter optimization [14]: Step 1: Determine the quality characteristic to be optimized.
Step 2: Identify the noise factors and test conditions.
Step 3: Identify the control factors and their alternative levels.
Step4: Design the matrix experiment and define thedata analysis procedure.
Step 5: Conduct the matrix experiment.
Step 6: Analyze the data and determine optimum levels for control factors.
Step 7: Predict the performance at these levels
FSW process parameters
It has been clearly shown in the literature [15][16][17][18] that FSW process parameters such as tool geometry rotational speed, welding speed and axial force significantly influence the process and play a major role in deciding the quality of the weld.
The detailed list of FSW process parameters are listed below: 1. Rotational speed of the tool (rpm) 2. Welding speed (mm/min) 3. Axial load (KN) 4. Tool geometry (i) D/d ratio of tool (ii) Pin length (iii) Tool shoulder diameter, D (mm) (iv) Pin diameter, d (mm) (v) Tool inclined angle (•) In the present investigation, three process parameters, i.e. rotational speed, welding speed and axial force are considered.Trail experiments are carried out using thick rolled plates of 6061 AA to fix the working range of FSW process parameters.When the rotational speed is lower than 800 rpm, low frictional heat is generated which results in poor plastic flow of the material during welding and contain defects like pinhole or tunnel in weld zone; when the rotational speed is higher than 1000 rpm causes excessive release of stirred material to the upper surface, which resultants left voids in the weld zone and poor surface quality.Similarly, when the welding speed is lower than 60 mm/min, pin holes type of defects are observed due to excessive heat input per unit length of the weld; when the welding speed is higher than 100 mm/min associated with low heat input, poor plastic flow of the material which causes some defects are observed at the weld zone.When the axial force applied on the tool by machine head is lower than 6 KN, sufficient heat is not generated which causes tunnel and crack like defect at the weld zone are observed; when the axial force is higher than 8 KN, large mass of flash and excessive thinning are observed due to higher heat input.Hence, the range of process parameters such as tool rotational speed is selected as 800-1000 rpm, the welding speed is selected as 60-100 mm/min and axial force is selected as 6-8 KN.The FSW process parameters along with their range and values are given in Table 1.
Selection of orthogonal array (oa)
Before selection of particular OA following points must be considered.
The number of factors and interactions of interest 2. The number of levels and interactions of interest
As three levels and three factors are taken into consideration, L9 OA is used in this investigation.Only the main factor effects are taken into consideration and not the interactions.The degrees of freedom (DOF) for each factor is 2 (number of levels − 1, i.e. 3 − 1 = 2) and therefore, the total DOF will be 6(= 3 × (3-1)).As per Taguchi method, the total DOF of selected OA must be greater than or equal to the total DOF required for the experiment.So an L9 OA having 8 (=9-1) degrees of freedom are selected for the present analysis..
Experimental procedures
The material used in this study is 5 mm thick sheets of 6061-T4 aluminium alloy.Chemical composition of base metal is given in Table 2.The rolled plates are cut into required dimension (300 mm long and 150 mm wide) for friction stir welding.Welding is carried out in butt joint configuration using friction stir welding machine.The butt joints are fabricated normal to the rolling direction.The experiments are conducted using parameters of the designed L9 OA table 3.
The American Society for Testing and Materials (ASTM E8) standard is used for preparing the tensile test specimens.The wire cut electro discharge machine (EDM) is used for prepared the smooth profile tensile test specimens.To minimize the machining error (noise), three specimens are prepared at each set of parameters in the designed matrix.The 27 prepared tensile specimens are subjected to tensile testing and ultimate tensile strength of each specimen is evaluated.
Signal to Noise Ratio
The signal to noise S/N ratio is calculated based on the quality of characteristics intended.The objective function described in this investigation is maximization of the tensile strength, so the larger the best S/N ratio is calculated.The formula for S/N ratio is given below.
( ) ∑ ( )
Where n is number of experiments and y is observed response value.
In this study, the tensile strength value of the FSW joints is analyzed to study the effects of the FSW process parameters.The experimental results are then transformed into means and signalto-noise (S/N) ratio.In this work, 9 means and 9 S/N ratios are calculated and the estimated tensile strength, means and signalto-noise (S/N) ratio are given in Table 4.The main effects of average mean and S/N ratio values of all levels are calculated and listed in Table 5 and 6.It is clear that a larger S/N ratio corresponds to better quality characteristics.Therefore, the optimal level of process parameter is the level of highest S/N ratio.Based on both mean and S/N ratio, indicated that the tensile strength at maximum when rotational speed, welding speed and axial force are at level 2. The main effects for mean and S/N ratio are plotted in Fig. 1 and 2.
Analysis of Variance (ANOVA)
ANOVA test is performed to find out the significant factor statistically.The purpose of ANOVA is to find out the significant process parameters which affect the tensile strength of FSW joints.The ANOVA result for both mean and S/N ratio is calculated and given in table 7 and 8 respectively.The F-test is being carried out to study the significances of process parameters.The high F value shows that the, factor is highly significant to affecting the response of process.In this study, results of ANOVA show that the rotational speed is highly significant factor and plays an important role to affecting the tensile strength of FSW joints.
Predicted Value of Tensile Strength
Based on the experiments, the optimum level setting is A2B2C2.The additive model to evaluate the predicted tensile strength is taken from the literature [10].The average values of parameters are taken from table 6 and predicted the value of tensile strength.
Confirmation Run
The confirmation experiments are carried out by setting the process parameters at optimum levels.The rotational speed, welding speed and axial force are set at 900 RPM, 80 mm/min and 7KN respectively.Three tensile specimens are subjected to tensile test and the average value is 162 MPa of the friction stir welded AA 6061.
Conclusion
1.The L9 Taguchi orthogonal designed experiments of FSW on aluminium alloy AA 6061 are successfully conducted.2. The FSW process parameters are optimized to maximize the tensile strength of joint.The optimum level levels of the rotational speed, welding speed and axial force are found to be 900 RPM, 80 mm/min and 7 KN respectively.3. The rotational speed plays an important role and contribution 67 % of the overall response, welding speed and axial force contribute 28% and 1.4 % respectively of the overall response.
Fig.1 Main Effects Plot for MeansFig.2Main Effects Plot for S/N ratio of tensile strength at second level of rotational speed B2= average value tensile strength at second level of welding speed C2= average value of tensile strength at second level of axial force T= overall mean of tensile strength
Table 2 :
Chemical composition of Al alloy
Table 4 :
Mean value and S/N Ratio
Table 5 :
Main effects of the process parameters
Table 6 :
Response table for Means and S/N Ratio
Table 8 :
ANOVA for S/N Ratio | 3,256.4 | 2013-03-13T00:00:00.000 | [
"Materials Science"
] |
Multicarving for high-dimensional post-selection inference
We consider post-selection inference for high-dimensional (generalized) linear models. Data carving (Fithian et al., 2014) is a promising technique to perform this task. However, it suffers from the instability of the model selector and hence may lead to poor replicability, especially in high-dimensional settings. We propose the multicarve method inspired by multisplitting, to improve upon stability and replicability. Furthermore, we extend existing concepts to group inference and illustrate the applicability of the methodology also for generalized linear models.
Introduction
We consider post-selection inference in high-dimensional (generalized) linear models. Statistical inference in high-dimensional models is challenging: in a frequentist setting, the main methods use some bias-corrected estimators of the Lasso [37,35,14] or of Ridge regression [5], and Cai and Guo [8] provide refined optimality results for such techniques. On the other hand, post-selection inference provides a very different approach for constructing confidence statements in high-dimensional models. Post-selection inference is attractive as it is closer in some vague sense to what practitioners like to do, namely to apply first some model selection in order to restrict the set of covariates and make the problem feasible. Post-selection inference has long been viewed as rather ill-posed [17] until Berk et al. [3] provided a conservative approach to improve its image. More recent work by Fithian, Sun and Taylor [10], Tian and Taylor [30], Taylor and Tibshirani [28] and others lead to interesting new inferential tools. The current work is building on those contributions.
The instability of post-selection inference. Post-selection inference deals with the problem of inference statements, after having selected a set of covariates using a data-driven algorithm or method. For post-selection inference in high-dimensional (generalized) linear models, a very popular model selection method is the Lasso [31]; and in fact, in this work, we only focus on the Lasso as model selector. Among the main concerns when using the Lasso or any other variable selection method is its instability. The selected model, say, by the Lasso, has low degree of replicability due to its instability arising from correlated covariates and/or high noise scenarios. Thus, the inference after model selection might be very non-replicable if the model selector leads to different results for small perturbations of the data. Take getting new realizations from the same data generating process as an example. Our new multicarving proposal is a possible remedy to make post-selection inference more reproducible.
A variety of approaches to get valid tests and confidence intervals after model selection have been developed. In order to put our proposal in some context, we discuss briefly the ones most relevant to our work in the following.
A simple approach for valid inference is to split the data into two parts and use the first half for selection and the second half for inference [36]. Thus, the idea is very similar to any validation scheme using data splitting.
This simple single data splitting method has certain drawbacks. Since splitting the data is a random process, the inference statements change if a different split is chosen. If we repeat this process multiple times, we observe that the obtained p-values per predictor change a lot: Meinshausen, Meier and Bühlmann [21] call this phenomenon the "p-value lottery". For the Lasso selector, this is especially accentuated as it is highly non-stable and potentially selects quite different models depending on the split. Therefore, results obtained through this method are not replicable at all unless one fixes the split. In order to receive more stable and replicable p-values, Meinshausen, Meier and Bühlmann [21] suggest splitting the data multiple times, say, B = 50 times leading to p-values P (b) j for each split b = 1, . . . , B and each predictor j = 1, . . . , p. The p-values per predictor are aggregated using quantile functions and adequate correction terms. Although there is still randomness involved, the results should become more stable with increasing B in the spirit of the law of large numbers. This technique is referred to as multisplitting.
To avoid confusion, we save the term post-selection inference for techniques that perform inference on the same data as used for selection and refer to the methods from [36] and [21] as (multi)splitting. Post-selection inference for a (generalized) linear model can be achieved by calculating or simulating a constrained null distribution, where the constraints reflect the selected model. Lee et al. [16] analyze the case of Lasso selection in a linear model. They show that the Karush-Kuhn-Tucker (KKT) criteria, which are necessary conditions for the Lasso solution, lead to a polyhedral constraint on the observed response vector. Using this constraint, they derive a truncated normal distribution which allows for valid inference. A drawback of this method is a loss in power introduced by those polyhedral constraints. Similar constraints have been derived in [32] for sequential regression problems: compared to Lasso selection for fixed value of λ, those constraints increase in dimensionality rather quickly, since every step of the procedure results in additional constraints.
Somewhere in between data splitting and post-selection inference is a technique called data carving [10]. In order to distinguish data carving from methods as in [16], we refer to the latter as pure post-selection inference in the following. Due to the loss in power introduced by pure post-selection inference, Fithian, Sun and Taylor [10] prefer not to use all observations for the selection process. Further, they prove that completely discarding the fraction of data used for model selection in the inference stage leads to inadmissible tests. Instead, one should use as much information of the selection data as is still usable and should only discard the information that was actually needed to obtain the given selection. This means that one "carves" the data. One can reuse the selection constraints introduced for pure post-selection inference but imposes them on the selection data only. This method outperforms pure post-selection inference and simple sample splitting with respect to power. Though, it is computationally much more involved under certain model assumptions. Naturally, pure postselection inference can be seen as a special case of data carving, and Fithian, Sun and Taylor [10] refer to it as Carve 100 .
Barber et al. [1] introduce the concept of knockoff filters for model selection and inference. Their main idea is to compare the measurable effect of the regressor covariates to the corresponding effect of their "knockoff copies" which should behave statistically equivalent for covariates with no true underlying effect. Barber et al. [2] adapt this methodology to the high-dimensional setting and post-selection inference. The data is split into two parts for that purpose, one for selection and one for inference only. The authors also suggest a method which can "recycle" some of the information from the selection data in the inference stage, which resembles the data carving idea. However, they condition not only on the selection event but on the full observation of the selection data. This has the advantage that the selection process on the first part of the data can be arbitrarily and is not restricted to methods for which one can sample from the data conditional on the selection event.
Berk et al. [3] provide an inference technique that is valid given any preceding model selection procedure, potentially, inspecting all of the data. This is possible by using the so-called PoSI (post-selection inference) constant K. This constant is defined as the minimal value such that the maximal absolute t-statistic maximized overall possible predictor variables and submodels is at most equal to K with probability at least 1 − α. The advantage of this method is that it leaves all freedom to the practitioner for the selection process without losing validity. For example, visual inspection of the data through a human, which is done quite often in practice, is allowed. On the other hand, this method is quite conservative by construction. Furthermore, calculating the constant K gets computationally involved such that the authors only suggest to use their method for up to p ≈ 20. Despite the nice theoretical framework, the method is not suited for high-dimensional statistics, which is our focus.
Recent developments by Kuchibhotla et al. [15] lead to computationally efficient procedures with similar guarantees. They derive a method to construct confidence regions such that they contain the true parameter in any submodel simultaneously with probability at least 1 − α. Due to this simultaneous cov-erage any possible model selection can be applied and the true parameter is still contained in the constructed region. Naturally, this method is also rather conservative. Especially, it cannot gain power from a sparsity assumption due to the simultaneous coverage in all submodels.
Relation to other work and contribution
Meinshausen, Meier and Bühlmann [21] as well as Fithian, Sun and Taylor [10] emphasize different drawbacks of the simple idea of data-splitting for inference in high-dimensional statistics and show how to improve on them. Therefore, we focus on how to optimally combine those improvements leading to our "multicarving" method. Since we work with the Lasso as model selector, we also build on the results of Lee et al. [16].
We further elaborate two more extensions of data carving in a linear model that can be combined with multisplitting in the same fashion. The first one concerns group testing. There are many developments in high-dimensional statistics for testing groups of covariates for significance instead of single covariates, see for example [34], [23], and [12]. Group tests are of particular use as with many (highly correlated) covariates, it might be overly ambitious to correctly detect the individual active variables, whereas groups of variables might be more realistic to detect. Hierarchical testing schemes are particularly attractive for this task; see for example [18] and [26]. Secondly, we provide extensions of multicarving to generalized linear models. Pure post-selection inference in logistic linear regression is discussed in Taylor and Tibshirani [28] who rely on asymptotic Gaussianity. As for the linear model, pure post-selection inference is suboptimal regarding power, thus, we extend their argument to the data carving approach. We only provide a detailed discussion for the case of logistic linear regression. Though, similar adjustments could be done for other generalized linear models.
Methodology for high-dimensional post-selection inference
We first consider the methodological framework for linear models and summarize multisplitting (Section 2.2.1) as well as data carving (Section 2.2.2). This serves as a basis to develop our novel multicarving procedure for single covariates (Section 2.3) and an extension to group inference (Section 2.5) and logistic regression or other generalized linear models (Section 2.6). While those developments focus on hypothesis testing, we discuss confidence intervals in Section 2.4.
High-dimensional linear model and inference for single variables
We assume to have a response vector Y = (Y 1 , . . . , Y n ) and a (fixed) design matrix X ∈ R n×p , where p n. This yields a linear model of the form where = ( 1 , . . . , n ) consists of i.i.d. N 0, σ 2 entries with known or unknown variance σ 2 and β ∈ R p is the unknown parameter of interest. We represent vectors in boldface, whereas scalars and matrices are written in usual letters. We write y for a given realization of the random vector Y. We use index 1 (X 1 , Y 1 and y 1 ) and index 2 (X 2 , Y 2 and y 2 ) to denote selection data and data used for inference only, respectively. Further, we assume that the active set S = {j; β j = 0} is sparse, i.e., s = |S| n such that inference using ordinary least squares would be possible on the data if the true active set was known.
After data-driven model selection, we deal with a subsetS of sizes = S . We aim to perform inference based on this subsetS. We write XS for the matrix X restricted to the selected columns. Likewise, X 1,S and X 2,S denote selection and inference data restricted to the selected columns. Generally, a distinction has to be made whether we test for the entries of the full β ∈ R p or if the test is made with respect to Here, βS ∈ Rs corresponds to the selected submodel and is defined as the best linear predictor in the given model. We write X + S for X S XS −1 X S , i.e., the generalized inverse of XS. We introduce corresponding null hypotheses for groups of variables in Section 2.5. Typically, an inference statement for (2) would be more favorable, since we are interested in the true underlying model. Though, tests for (3) are valid under weaker assumptions.
Of particular interest is the screening property. Screening is defined asS ⊇ S or in words, screening asks for all active variables being part of the selected model. If this holds, we have βS j = β j ∀j ∈S. Thus, tests valid for (3) are also unbiased for (2) assuming screening. Importantly, screening is a requirement on the initial model selection process and not on the following inference calculation. We focus on model selection using the Lasso. The screening property for the Lasso is rather delicate to achieve in the finite sample case. Though, it can be guaranteed with probability 1 for n → ∞ under adequate conditions. Such conditions are discussed in [20], [22] and [4], see also the book by Bühlmann and van de Geer [7].
Previously proposed methods
We first review some earlier work which serves as a basis for our new proposal in Section 2.3.
Multisplitting for inference
In this section, we briefly summarize the multisplitting method introduced in [21]. Multisplitting works as follows: For each b = 1, . . . , B: 1. Randomly split the data into two disjoint groups of sizes n 1 and n 2 . 2. FindS (b) using X 1 and y 1 .
using X 2,S (b) and y 2 with ordinary least-squares; for j / , 1 to correct for multiplicity using Bonferroni adjustment.
The fourth step is designed to control the family-wise error rate (FWER). Throughout this work, we use lower case letters (p) for raw p-values that result from a test and upper case letters (P, Q) for p-values resulting from any correction or aggregation. The default value for splitting is n 1 = n 2 . It remains to aggregate the B p-values for covariate j. Valid aggregation is possible by using a quantile of fixed fraction γ ∈ (0, 1] as with q γ being the empirical quantile function. Since a good choice of γ might not be known a priori, one can also optimize γ over a range [γ min , 1] where γ min ∈ (0, 1]. This yields a different p-value The additional factor (1 − log (γ min )) corrects for optimizing over all possible quantiles. A typical choice is γ min = 0.05, yielding a correction factor of (1 − log (0.05)) ≈ 3.996. Without any screening assumption, those p-values actually test the following null hypothesis for some given covariate j HS (1) ,...,S (B) 0,j : βS (7) Given two conditions, Meinshausen, Meier and Bühlmann [21] derive asymptotic (for n → ∞) FWER control with respect to null hypothesis (2). The conditions are: The screening condition, as argued before, leads to βS (b) and makes the inference statement valid for the true underlying parameter vector.
The sparsity condition enables us to do least-squares inference, implicitly assuming that X 2,S (b) has full column rank for all b.
If screening held in the finite sample case as well, the error control could be formulated in a non-asymptotic sense. Although this is usually not the case, the simulations in [21] as well as ours show that multisplitting controls the type-I error with respect to (2) clearly better than single-splitting when screening cannot be guaranteed. This can be explained by the "p-value lottery": Every split results in different p-values for the selected variables. There are chances that some true non-active variables are significant for some splits. After aggregation, only variables that are significant in a decent number of splits remain significant overall. Due to the variability of these p-values over different splits, chances are that fewer non-active variables get rejected after aggregation than in the average single split. Thus, multisplitting leads to better error control.
Data carving
In this section, we discuss the idea of data carving introduced in [10]. We focus on the special case of the linear model (1) with Lasso selection, which we will later extend to logistic regression and other generalized linear models. We emphasize that they provide a theoretical framework that could be applied to a much broader spectrum of problems.
The main conceptual idea of data carving is summarized in the following statement [10]: "The answer must be valid, given that the question was asked." Thus, one should control the selective type-I error rate The hypothesis HS 0 is a general notation for a hypothesis as, e.g., in (3). Define the event M (Y 1 ) as S , HS 0 selected , the selection event using data {X 1 , Y 1 }. We write M (Y 1 ) since X 1 is assumed to be fixed. Then, the requirement (8) can be equivalently stated as Simple data splitting on the other hand controls the following error Thus, more conditioning is done than would theoretically be needed, since M (Y 1 ) does not contain all information about Y 1 but only guarantees that it results in the observed selection event.
To perform inference controlling the error in (9), one needs to understand the distribution of Y M (Y 1 ). The first step is to understand the selection event M (Y 1 ). We focus on our case of interest, inference in the linear model (1) using Lasso selection. More precisely, let Lasso selection be defined as follows There exist different definitions of the Lasso that are equivalent after rescaling. We use definition (10) following [16] where this selection event is fully characterized. The set of Y 1 that would lead to the sameS forms a union of polyhedra in R n1 . If we additionally condition on the signs of the parameters' Lasso estimates, sign β j ∀j ∈S, this union is shrunk to a single polyhedron. Dealing with a single polyhedron is easier both computationally as well as from a theoretical perspective. Hereafter, we additionally condition on the signs at the price of a small loss in power. This single polyhedron can easily be described by linear inequality constraints, e.g., AY ≤ b. Those constraints can be split into "active" (A 1 Y ≤ b 1 ) and "inactive" (A 0 Y ≤ b 0 ) constraints which define statistically independent events. Further, X + S Y is independent of the inactive constraints such that it is also independent while conditioning on the active constraints, i.e., X + . Therefore, we can ignore the inactive constraints for inference purposes which are based on X + S Y. For simplicity, we refer to AY ≤ b as being the active constraints only.
Fithian, Sun and Taylor [10] b in a given model. As βS is unknown, the conditional distribution is not tractable yet. To deal with this problem, one can treat the unknown parameters as nuisance parameters in an exponential family which one can get rid of by conditioning accordingly. Generally, one has to decide between the "saturated model" and the "selected model": has n degrees of freedom and βS = X + S μ is the best linear predictor based on the selected model (cf. Equation (4)). If we consider the saturated model, which includes more parameters than the selected model, more conditioning has to be done. This leads to a drop in power but with the advantage that tests are valid for (3) without any screening assumption. The selected model view is generally more powerful since less conditioning is done but it needs stronger assumptions to hold. The existence of βS such that E [Y] = XSβS is exactly the screening condition. If screening holds, either approach is valid to test (2). Since we are mainly interested in this null hypothesis, we focus on the selected model leading to more powerful tests under screening. In Section 2.4, we elaborate further on the saturated method and its advantages. Consider the selected model. To perform inference for covariate j, one has to condition onto XS \j Y. After applying this conditioning, the random vector of interest X + S j Y is independent from the unknown parameters βS −j . This leads to a degenerate truncated multivariate Gaussian distribution with no more unknown nuisance parameters. The truncation is defined by the selection event.
To test the null hypothesis, one further assumes βS j = 0.
Thus, one is interested in Note that we can use one-sided tests, since we implicitly condition on the sign of β j by restricting ourselves to the single polyhedron AY ≤ b. If we have selected a correct model such that the selected model view is applicable, σ is known, and j / ∈ S is not a true active variable, then we have p j (Y) ∼ Unif [0, 1]. This null distribution is not easily tractable and thus the probability is hard to calculate. Though, it can be sampled from using MCMC. This means that data carving achieves higher power compared to sample splitting at the price of a substantially higher computational cost. We present an applicable MCMC sampling scheme in Appendix B.
In the saturated viewpoint, only one degree of freedom remains after conditioning (cf. Section 2.4). Therefore, one can deal with a univariate truncated normal such that the exact probability, the analogue of (11), can be calculated efficiently using the CDF of a Gaussian. Thus, the trade-off between the selected and the saturated model also involves a computational component.
So far, we assumed σ to be known. If this is not the case, σ 2 could be handled as further nuisance parameter, which is resolved by additionally conditioning on Y 2 . However, this nonlinear constraint disables some of the computational shortcuts which all linear constraints allow for. In our simulations, we use some estimate σ wherever the variance is assumed to be unknown and proceed as if it was known initially. For completeness, we mention that the distribution when additionally conditioning on Y 2 is not Gaussian anymore. The corresponding null distribution can still be approximated using a different MCMC sampling technique. Note that this is only possible for the selected model. In the saturated model, one would end up imposing one quadratic and n − 1 linear equality constraints onto an n-dimensional vector. This would only leave two points to sample from such that no inference is possible.
Novel multicarving for valid inference
Meinshausen, Meier and Bühlmann [21] have theoretically argued and empirically shown that splitting several times and aggregating is to be preferred over a single-split approach. On the other hand, Fithian, Sun and Taylor [10] have shown that discarding all selection data in a splitting set-up is mathematically inadmissible and typically less efficient. To overcome this problem, they introduce the idea of data carving. Nevertheless, their approach potentially suffers from a similar p-value lottery as discussed in [21] since it is initiated by randomly splitting the data into two disjoint groups of given sizes; one for selection and inference, and the other one for inference only. Therefore, it is often difficult to replicate. Thus, we advocate the idea of applying data carving multiple times in order to a) overcome the p-value lottery and b) avoid the proven inadmissibility of any splitting procedure. We use the following procedure: For b = 1, . . . , B: 1. Randomly split the data into two disjoint groups of sizes n 1 and n 2 .
2. FindS (b) using X 1 and y 1 with Lasso selection.
j for the given split and selected model according to (11), for j / , 1 to correct for multiplicity using Bonferroni adjustment.
As in multisplitting, we include the fourth step in order to control the FWER. Different corrections could be applied to obtain some less restrictive error control such as the false discovery rate (FDR) as discussed in [21]. There is a trade-off involved in choosing n 1 and n 2 . The higher we set n 1 , the higher the probability of screening gets, which is required for valid tests. On the other hand, more power remains for the second stage, namely, the inference calculation, for higher values of n 2 . We empirically analyze this trade-off in our simulations in Section 4. To get one p-value per predictor, we use the same aggregation techniques as presented in Section 2.2.1, resulting in a single p-value Q j (γ) or P j . In our simulations, we focus on optimizing over the quantiles as described in (6) instead of using a fixed predefined quantile γ. To distinguish the different methods, we call this procedure multicarving and the method described in Section 2.2.2 single-carving.
Saturated view and confidence intervals
Naturally, one wants to perform inference without the screening assumption. As mentioned in Section 2.2.2, we can use the saturated model from [10] for this purpose. In the saturated view, we do not assume the selected submodel to completely define the mean parameter μ but only to approximate it as in (4). In order to get rid of the unknown parameters and create a tractable distribution, we have to condition on to P ⊥ η Y = P ⊥ η y. Here, we define η ≡ X + S j , leading to η μ = βS j . As P ⊥ η has rank n − 1, there remains only one degree of freedom after conditioning, namely, in the direction of η. Therefore, one deals with a univariate truncated Gaussian where the truncation comes from invoking the selection event AY ≤ b. Inference statements can be calculated efficiently using the CDF of a univariate Gaussian. A detailed explanation of this procedure can be found in [16].
This can be done regardless of the quality of the selected submodel. Therefore, the saturated viewpoint leads to valid tests for null hypotheses (3) (singlecarving) and (7) (multicarving) without any screening assumption. However, if screening fails, the best linear predictor in the submodel is generally non-sparse. This means that there is no j ∈S s.t. βS j = 0 and there cannot be any false positives with respect to those null hypotheses. Therefore, such tests for null hypotheses without any screening assumption are not of particular interest.
Nevertheless, those tests can be used to determine confidence intervals. As for any test, confidence intervals for multicarving can be found by inverting it. Dezeure et al. [9] give a detailed explanation of how to compute confidence intervals for multisplitting. We refrain from giving a full theoretical result for our derived method, but remark that their construction does not require the individual p-values to origin from a sample splitting procedure as long as they are valid. Therefore, this approach can be directly adopted to multicarving by calculating carving p-values but keeping the remaining scheme the same. We focus on the construction without multiplicity correction. For covariate j, this leads to a (1 − α)-confidence interval (CI) such that P βS where βS (b) j are defined through (4). This is of particular interest when βS Therefore, it appears natural to omit the screening assumption and to adopt the saturated model for our confidence intervals. Further, the use of the saturated model leads to more efficient computation. We focus on two-sided confidence intervals for two reasons. First, having both a lower and an upper bound might be more informative for a practitioner. Second, sign βS is not necessarily the same for all splits b in which covariate j is selected such that combining different splits to a one-sided confidence interval is not appropriate. Thus, the confidence intervals in this case are not the exact inversion of the hypothesis tests.
Notably, if one were to apply simultaneous tests for different null hypotheses in the selected model, this could be done by just calculating a single MCMC chain and relying on the idea of importance sampling afterwards. However, to get a precise enough statement for such simultaneous tests, more MCMC samples might be required than for just calculating a p-value such that this extra statement is not for free.
Extension to group testing
In a high-dimensional set-up with potentially correlated predictors, finding individual active variables is often too ambitious. Especially, the Lasso selector struggles with distinguishing between two or more highly correlated variables. Therefore, one might prefer to test several variables as a group. We define the null hypothesis for a given group G as for the full model coefficients. LetG =S ∩ G be the variables in our group that have been selected then we define the null hypothesis in the selected model as The practitioner often wants to test multiple groups or test groups in a hierarchical fashion, say, in a data-driven way. Of course, a multiplicity correction has to be applied which is possible for any valid group test which controls the type I error. We refer to [19] for a detailed explanation of a hierarchical testing procedure and corresponding multiple-testing correction.
(Multi)splitting for group inference
Groups of variables can be tested for significance in the same way as single variables by splitting the data. The extension to groups follows naturally as in the low-dimensional case by applying partial F-tests instead of t-tests. This can be done either with a single split or multiple splits using the previously mentioned aggregation techniques (5) and (6).
(Multi)carving for group inference
The above mentioned (multi)splitting techniques for group inference suffer from the same inadmissibility issue as in the single variable case as more conditioning than necessary is applied. Therefore, we suggest a slight transformation of the data carving idea which makes it applicable to testing for group significance. We focus on the selected viewpoint meaning that our derivation will actually only be valid if a correct model has been found. We emphasize that the saturated model could be extended to inference for groups with very similar adjustments.
Inference for a group follows the single variable case closely. Firstly, note that the selection event is completely unchanged by the idea of testing group significance afterwards as we still apply Lasso for model selection. Thus, we can still invoke the selection event by conditioning on AY ≤ b. Based on [10], does not depend on βS −G such that there are no more unknown parameters in our model under the null hypothesis (14). Due to this independence from the nuisance parameters, we can base the inference on X + S G Y or functions thereof. We advocate the use of the following test statistic In words, it is a directed sum of projections in to all directions corresponding to the group's variables. Including sign β j in our test statistic is valid, since we additionally condition on having observed the parameters' signs for the sampling procedure. This additional conditioning is not mandatory for valid inference but simplifies the computation (cf. Section 2.2.2). The success of the sum can be intuitively justified as potentially no variable has a significant effect by itself, but the group as a whole could have.
There are two main reasons to perform such a group test instead of aggregating p-values of single variables in the group. First, since we are interested in the null hypothesis for the group, it seems more appropriate to conduct a test that treats all the variables within in the group in the same way instead of applying and aggregating multiple tests each of which focussing on a different variable. Second, as the calculation of any null distribution requires to sample a MCMC chain, fewer chains have to be created when looking at a group simultaneously. Though, this comes at the price of a higher dimensionality to sample from compared to treating covariates individually.
We need to sample from the (approximate) null distribution to perform tests. As in the single variable case, the carving procedure leads to a Gaussian distribution subject to linear equality and inequality constraints, which can be sampled from as presented in Appendix B with few adjustments.
In contrast to testing of single variables, the group problem remains multidimensional in the saturated view (for G > 1) as one conditions on all but the group's variables. To sample from this saturated model, some more changes would be needed, especially the conditioning in B.1 has to be adjusted, while B.2 has to be omitted.
In Section 3.2, we establish the validity of our group test on a single split. This validity is enough to enable multicarving using standard aggregation techniques (5) or (6). When testing for several groups, the fourth step of the multicarving procedure given in Section 2.3 must be adapted to a suitable multiplicity correction factor. The factor which enlarges the p-values naturally depends on the construction of the different groups. Some possible choices for disjoint groups are p/ G and S [b] / S [b] ∩ G , where the latter can be different for every split. For a more elaborate description of this procedure as well as an extension to hierarchical testing, see [18] and [26].
Extension to logistic regression
Not all data can be described and approximated well by the linear model given in (1). We extend the inference method to be applicable to generalized linear models and focus on logistic regression only in the following. Many of the ideas could be carry over to different generalized linear models too, after applying the right transformations.
In logistic regression, we have a binary response vector Y ∈ {0, 1} n and some matrix of predictor variables X ∈ R n×p . For every entry Y i of Y, the probability of being 1 is modelled as for some unknown parameter vector β ∈ R p , the target of our inference. We denote the i-th row of X by X i .
In a classical low-dimensional setting with p < n, this would be fitted using the MLE or equivalently by minimizing the negative of the log-likelihood for an observation y. The log-likelihood l (β) is defined as The negative of the above formula can be minimized, for example, by using a Newton algorithm, which leads to solving an iteratively reweighted least squares (IRLS) problem as derived in [13]. Starting with some initial estimate β 0 , one iterates Thus, in every step a weighted least-squares problem with weight matrix W , which iteratively changes, is solved. This explains the name of the procedure. By further defining this can be reformulated as a usual least-squares problem [9] β t+1 = arg min In the low-dimensional case, Dezeure et al. [9] suggest to perform the inference as if the final iterate follows Y w ∼ N (X w β, I). This approach is asymptotically valid because if this was the case, one would have which is the limiting distribution of the MLE. This can be seen by noting that the covariance matrix is the plug-in estimate of the inverse Fisher information.
As for the linear model (1), the MLE cannot be uniquely found for p > n since X W X is not invertible anymore. Therefore, one also depends on some sort of shrinkage. One can use the Lasso, i.e., an 1 -penalty, in the same fashion as for the linear model and solve the following minimization This minimizer can be found similarly as in the non-penalized case by adding the penalty term in every update [11] β t+1 = arg min Thus, the final Lasso estimate will (approximately) fulfil where X w and y w are functions of the estimate β itself. As this is exactly a Lasso fit as in (10), the estimate β will also fulfil the KKT criteria defined by X w and y w . Therefore, we can formulate the constraint AY w ≤ b, which the observed adjusted response is required to fulfil.
In the high-dimensional case with Lasso selection, it is an obvious approach to calculate inference statements as if Y w ∼ N (X w β, I) AY w ≤ b inspired by the inference techniques in the low-dimensional setting. Or in other words, proceed as in the usual Gaussian case using our new transformed data X w and Y w . This can be done likewise for either pure post-selection inference or data carving.
Taylor and Tibshirani [28] provide an argument for the first case. Their main assumption is √ n-consistency of the Lasso estimator β. This condition is discussed, for example, in [7]. Under this assumption, the "one-step estimator" β ≡ X + w,S Y w would have the same limiting Gaussian distribution as the usual low-dimensional MLE if no selection was applied. After some technicalities, which we do not want to recite here, they are able to derive the corresponding constrained limiting distribution from this non-selective CLT.
Importantly, this theory was derived for the fixed-p case. Especially, √ nconsistency of the Lasso estimator typically only holds for fixed p. An argument for the high-dimensional case p n → ∞, if any exists, is yet to be found. Recent developments by Sur and Candès [27] and Zhao, Sur and Candes [38] regarding the limiting distribution of the MLE suggest that one has to additionally assume at least s = O (n) in order to derive such an argument.
We empirically test the adaption of pure post-selection inference for logistic regression to data carving in our simulations without giving a full theoretical argument. Presumably, such an argument, at least for the fixed-p case, could follow using similar concepts as in [28].
Other types of generalized linear models are often fitted in the same fashion using (penalized) IRLS. Whenever this is the case, one can apply our carving method to the transformed data, i.e., X w and y w , which behave asymptotically Gaussian.
Multicarving and aggregation. As in Section 2.3, we apply this method of calculating p-values to various splits and aggregate as described in Section 2.2.1. Those aggregation techniques are proven to be unbiased given screening. Obviously, assuming that aggregation is performed over p-values that are all valid themselves given screening.
Here, the p-values are only asymptotically valid even under screening. Asymptotic validity of the aggregation over asymptotically valid p-values has not yet been theoretically studied in depth. Therefore, we cannot restate the same theoretical results for logistic regression as were derived in [21] for multisplitting and which we adapt in Section 3.1 for multicarving in a linear model. Nevertheless, applying multicarving to logistic regression does not result in any problem with type-I error control in our simulations so that we can advocate its use.
Theoretical properties
We elaborate here the theoretical properties of multicarving and the extension to group testing for (multi)carving in the selected view, requiring the screening assumption in (A1). Without the screening assumption, (multi)carving is still valid controlling the type I error in great generality when taking the saturated view. Then, at the price to be often overly conservative, confidence intervals with guaranteed coverage should be preferred over tests, see also Sections 2.2.2 and 2.4. Throughout this section, we assume that the data follow the linear model (1) with Gaussian errors.
Multicarving for the linear model
Validity of our multicarve method follows naturally from validity of singlecarving and multisplitting. Assuming screening in split b and known variance, we know from the theory of data carving that p (b) j as defined in (11) follows Basically, this uniformity of the p-value is the only thing needed to construct the proofs of Theorems 3.1 and 3.2 in [21]. Therefore, we can restate their theoretical result for the aggregation methods. Though, we slightly alter the assumptions on the model selection procedure. We assume Asymptotic screening: lim n→∞ P S ⊇ S = 1 (as in Section 2.2.1). (A1) The difference in the second condition yields from the fact that one has to invert X 2 X 2 to perform inference using splitting, while X 1 X 1 has to be inverted for data carving. Actually, the condition is rank X 1,S =s and we implicitly assume this to follow from the sparsity condition. Our simulations suggest to use n 1 > n 2 , thus this altered sparsity assumption is less restrictive. Using those two conditions, we establish FWER control for our multicarve procedure.
where the probability is with respect to the data sample. The statement holds regardless of the B random sample splits.
The analogue result holds when aggregation is not done with a fixed quantile γ but with the optimized quantile and the adequate correction term.
where the probability is as in Theorem 1.
For proofs, we refer to the appendix of [21] invoking the fact that p (b) j is stochastically larger than Unif [0, 1] under our assumptions.
Some more technicalities have to be considered for error control in a practical set-up. First, in order for the uniformity assumption to hold, we depend on a good convergence of the MCMC approximation. Second, since we refrain from conditioning on Y 2 , we need to know the variance, which is often rather unrealistic. Though, we emphasize that the same theoretical result would hold in the unknown variance case when actually using the conditioning trick. Further, when using an overestimate of σ, tests become likely more conservative such that type-I error control is given at least as good as with the true variance parameter. However, this cannot be guaranteed in all cases. A discussion on this issue can for example be found in the supplemental materials of [33]. Third, since we work with finite data, there is no way to guarantee the screening assumption in general. For analogous reasons as argued in Section 2.2.1, chances are that multicarving corrects the type-I error better than single-carving in such setups. However, if screening becomes too unlikely, breaches in the error rate are likely to happen for multicarving as well. This is especially an issue for highly correlated covariates which make the Lasso selection very difficult. We analyze this effect in our simulations in Section 4.1 and Appendix C.1.
Data carving for group testing
In this section, we focus on the theoretical properties of our group test applied to a single group using a single split. Using Theorem 3, results for multicarving then follow from standard arguments.
At the base of our group test is the following lemma, which is proven in Appendix A. Lemma 1. Let Y be generated by the linear model (1) with Gaussian errors. Let G be some group with G > 0, whereG = G∩S. Assume that the screening property (S ⊇ S) and (Ã2) hold, and σ is known. Then, the probability law of is completely defined by our parameter of interest βS G .
Using this lemma, we can base our inference statement on the conditional distribution of X + S G Y. Let y be some observation, then we define our selected group p-value as This probability can be calculated since we additionally condition on the only remaining unknowns in the model. Notably, this exactly defines the "probability of observing a value at least as extreme as the observed statistic" under null hypothesis (14). Thus, it fulfils the desired property of a p-value, which leads to the following theorem.
Theorem 3. Let Y be generated by the linear model (1) with Gaussian errors. Assume that the screening property (S ⊇ S) and (Ã2) hold, and σ is known. Let y be a realization of Y and pG (y) for some group G with G > 0 be calculated as in (16). Then, under null hypothesis (14), it holds Now further define a general group p-value for group G as Then, we can establish error control of our procedure. Let y be a realization of Y and p G (y) for some group G be calculated as in (17). Then, under null hypothesis (13) and for any α ∈ (0, 1], it holds The proof is available in the Appendix A. The technicalities mentioned at the end of Section 3.1 apply in the same fashion for our group test.
Numerical results
In this section, we provide detailed results of the performance of our proposed methods in simulation studies. All results were obtained using the programming language R [29]. As an overall summary, we find that multicarving exhibits often an advantage, sometimes being substantial, over multisplitting or single-carving methods.
Multicarving for the linear model
We tested our multicarve method testing for single variables in the linear model on several simulation set-ups and we present here the results for two of them. In the Appendix C.1, we add further results for variations of these set-ups where we also show the limitations of (multi)carving.
We do not restrict ourselves to successful screening, we assume the variance to be unknown and estimate it, and lastly, we select our model through crossvalidated Lasso with regularization parameter λ 1se . All these choices are (in part only slightly) deviating from our theoretical assumptions. In particular, by choosing λ through cross-validation, more information of Y is used than invoked in the selection event, making the inference biased. There are first approaches to correct for this additional bias, for example, in [30]. However, we refrain from applying any of these, since they will get computationally more involved and because our empirical results do not show any significant violation of the selective type-I error rate (8) using cross-validation. Perhaps though, this should be done with a certain precaution as, e.g., Taylor and Tibshirani [28] report bad error control using a cross-validated λ for post-selection inference in a Cox model.
We For aggregation over the different splits, we optimize over quantiles as in (6). Starting with the default value in the multisplitting literature, γ min = 0.05, we noticed that this makes the procedure sometimes overly optimistic leading to poor error control. For some intuition of this effect, assume that there is a true active predictor X j and a decently correlated predictor X k for which the null hypothesis holds true. In order to falsely reject this null hypothesis, X k must be selected as a proxy for X j in at least γ min B of the random sample splits. Of course, this is more likely the lower we set γ min . Therefore, we additionally consider γ min = 0.3 to have a comparison. Using a larger γ min is also favorable for computational reasons since less MCMC samples are required to be able to find a significant aggregated p-value for the smallest possible quantile, namely the γ min -quantile; see Appendix C.3 for more details.
Toeplitz design
In a first scenario, we sample X once from a multivariate Gaussian distribution with mean zero and a Toeplitz covariance matrix Σ with Σ ij = ρ |i−j| with ρ = 0.6, and we then treat it as fixed design. The coefficient vector β is 5sparse, and the active predictors are {1, 5, 10, 15, 20}, each of which having a coefficient equal to 1. The standard deviation is set to σ = 2, leading to a signal-to-noise ratio (SNR) of approximately 1.71. For each simulation run, the variance estimate σ 2 is calculated through cross-validated Lasso on the entire data set and is used globally for all splits and inference methods.
In Figure 1, we present the outcome of the simulations for the Toeplitz design. Each performance measure represents 200 simulation runs. Although screening cannot always be guaranteed, FWER and power are calculated with respect to (2) with rejection level α = 0.05. Carving using the entire data for selection, i.e., f = 1, is performed using a different algorithm, namely, the exact calculation from [16]. We emphasis this using a cross in the figures.
The left-hand side of Figure 1 illustrates that neither single-splitting nor single-carving controls the error at 5% for f = 0.5 and f = 0.75. Though, this is not a violation of our theoretical result, error control would hold when only looking at successful screening. For carving, the power initially increases in f and decreases in the larger values of f . This can be explained by the trade-off between more successful screening of the true active set and losing power for the inference stage as more constraints are imposed. The same holds for splitting and multicarving when looking at lower values of f as eventually too few active variables are selected in the first stage such that no decent power remains. As indicated by the inadmissibility statement in [10], carving outperforms splitting with respect to power. The important question is now whether multicarving introduces some improvement over single-carving. The single-carve method has the highest power starting from f = 0.75, where f = 0.5 can be basically ignored as error control is not given at all. The multicarve method with γ min = 0.3 performs best among all carving methods regarding FWER for all values of f . Multicarving with γ min = 0.05 seems to be inferior in this scenario. Thus, there is a trade-off between higher power and better error control. The highest power with FWER ≤ 5% is obtained at f = 0.9 for all carving methods with a value of 0.59 (single-carving), 0.51 (γ min = 0.3) and 0.50 (γ min = 0.05). So, the single-carve method is favorable in this situation. However, this comparison is not quite fair since the methods have different FWER. Therefore, we additionally look at an adjusted power, i.e., the rejection level of the underlying hypothesis tests is adjusted such that each method has an FWER of exactly 5% for each value of f ; see Figure 2. Carving is still superior to splitting although the multisplit method with γ min = 0.3 is now competitive for lower values of f . All three carving methods reach their optimum at f = 0.9, with an adjusted power of 0.67 (single-carving), 0.73 (γ min = 0.3) and 0.61 (γ min = 0.05).
In Appendix C.1.1, we present further results for Toeplitz designs where ρ is changed to 0.3 and 0.9 respectively. Our assumption that the correlation level highly affects the performance is confirmed. Especially, none of the methods in scope performs particularly well for the scenario with ρ = 0.9 since the initial selection using the Lasso is very unlikely to screen successfully.
Saturated viewpoint.
As discussed in Section 2.4, testing for null hypotheses (3) (single-carving) and (7) (multicarving) while omitting the screening assumption is not particularly meaningful as βS (b) is fully dense. Therefore, the saturated viewpoint without the screening assumption has no advantage for testing null hypotheses. However, in order to assess the power drop mentioned in Section 2.2.2, we test for null hypothesis (2) using inference in the saturated model. For the set-up discussed above, this leads to the following performance measures. The highest power with FWER ≤ 5% is 0.44 for singlecarving (f = 1), 0.44 for multicarving with γ min = 0.05 (f = 0.9) and 0.41 for multicarving with γ min = 0.3 (f = 0.95). The corresponding highest adjusted power is 0.50 (single-carving), 0.61 (γ min = 0.05) and 0.71 (γ min = 0.3), all of which obtained at f = 0.9. Thus, the saturated approach leads to lower power and adjusted power as anticipated. Though, this drop is less distinct for the adjusted power as the additional conservatism also leads to better type-I error control. Furthermore, we see that for multicarving the differences are less pronounced than for single-carving. For computational reasons, the saturated viewpoint might, therefore, be an interesting alternative for our multicarve procedure.
In Section 2.4, we further introduced the idea of multicarving confidence intervals, where omitting the screening assumption and using the saturated method appears to be more natural. We present a corresponding analysis in Section 4.2.
PoSI. In Section 1, we mentioned the work by Kuchibhotla et al. [15] which generally provides stronger guarantees at the price of increased conservatism. In order to assess this conservatism, we executed a small simulation study applying their method to this Toeplitz design. For this, we use the software available in the GitHub repository cited in [15].
For the model selection, we use cross-validated Lasso on all data. Thus, the models on which we perform inference are the same as for pure post-selection inference used above. After calculating the 95% confidence regions for βS in all s dimensions, we reject the null-hypothesis for covariates j for which 0 is not within the region. With this technique, we did not receive a single rejection over 1000 simulation runs. Thus, the expectation that the inference method is very conservative is confirmed.
Since their method is not restricted to Lasso selection but allows for any possible method in the selection step, we tried a different approach. Namely, we applied an "oracle" selection that always selects the correct submodel, i.e., S = S. However, not a single rejection was observed even using this best possible selection. This further confirms the assumption that guaranteeing simultaneous coverage in all submodels is too restrictive for this simulation set-up.
Semi-synthetic Riboflavin data
Since simulated data sometimes behaves somewhat more nicely than real data, we also test the methods on "semi-synthetic" set-ups, meaning that the X matrix comes from some real data set. We simulate the response Y from (1) with known β.
We use the Riboflavin data set with n = 71 and p = 4088, which was made publicly available by Bühlmann, Kalisch and Meier [6]. The original response measures the Riboflavin production rate for 71 samples of strains of Bacillus subtilis and gives the data its name. The X matrix contains the log-expression level of 4088 genes for each of these strains.
For our simulations, we set β to be 2-sparse and use an SNR of 16. The active variables are chosen at random for every simulation run and their respective coefficient is set to 1. Since this can result in very different signal strength depending on the correlation between the 2 variables, we fix the SNR on a per run basis by always adjusting σ such that Var(Xβ) σ 2 = 16. Here, Var (Xβ) denotes the empirical variance of the true underlying signal. We choose this rather sparse set-up with high SNR since otherwise Lasso selection works very poorly in this high-dimensional set-up and none of the inference methods has good performance. To illustrate this, we repeat the same simulation with 4 active predictors; compare with Appendix C.1.2.
For the selection, we again perform cross-validation on the given split. To be more realistic, we stick to the unknown σ assumption. With the estimation technique described before, we realized that P [ σ ≥ σ] is empirically quite low in this scenario. Therefore, we choose the more conservative approach of calculating a new σ for every split as where β b is calculated on the selection data only but y and X are the full data. The results obtained for the Riboflavin data with a sparsity of 2 are shown in Figures 3 (FWER and power) and 4 (adjusted power). This set-up is now highly in favor of our multicarve method. Especially, the highest power obtained for FWER ≤ 5% is 0.42 (single-carving), 0.60 (γ min = 0.3) and 0.69 (γ min = 0.05); see Figure 3. The multicarve methods reach this maximum at f = 0.9, while single-carving only obtains error control starting from f = 0.95 and higher. There is a power versus FWER trade-off between the two different values of γ min .
The adjusted power is slightly in favor of the lower value γ min = 0.05 as illustrated in Figure 4. More precisely, the highest adjusted power is 0.75 (γ min = 0.05) and 0.71 (γ min = 0.3) for the multicarve method. Both these values are obtained for f = 0.95. Single-carving reaches its maximum of 0.46 at f = 0.9. Thus, the adjusted power clearly prefers multicarving as well.
We note that although we increase both SNR and sparsity, the adjusted power is not (much) better than in the previous set-up. This can be intuitively explained by the following two reasons: First, p n ≈ 58 in the Riboflavin design is much larger than p n = 2 in our Toeplitz design. Second, there are variables with a very high empirical correlation of up to around 99%, making them hardly distinguishable in the selection stage.
Confidence intervals
We apply our method for confidence intervals to the same set-up with X simulated from a multivariate normal distribution with Toeplitz ρ = 0.6 covariance matrix as in Section 4.1.1. As we explicitly omit the screening assumption, we use a different estimate σ for every split as in (18). The target parameters βS (12) are defined including an intercept. Naturally, whenever screening works, this intercept term vanishes. We use B = 50 splits and aggregate according to (6) with γ min = 0.05. The obtained intervals are targeted to be 95%-confidence intervals (95%-CI).
In Tables 1 and 2, we compare the performance of our carving confidence intervals to the ones obtained using multisplitting implemented in [9]. Those results are based on 200 simulation runs.
The obtained intervals are generally rather conservative as the false coverage rate is always far below the theoretical bound of 5%. Notably, for f = 0.5, the intervals obtained through carving are not actually shorter than those from splitting. The advantage of carving is that the intervals get shorter in a first phase when increasing f . By increasing f , the selected models become more stable and likewise, βS fraction of f = 0.9 which outperforms every other configuration with respect to at least three interval lengths. Further, it also performs comparably well with respect to the false coverage rate as every configuration with lower false coverage rate suffers from substantially longer intervals.
In a further analysis, we look at the length of the confidence intervals of all covariates that were selected at least once within the B = 50 splits. For the other variables, there is no real interpretation of the coverage in Equation (12). Further, not selecting a covariate at all in 50 splits is a rather strong indication for the variable generally being inactive such that treating it as if it has an infinite confidence interval length does not seem appropriate. However, there are still many variables obtaining an infinite interval length, namely, those that are selected at least once but less than γ min B times, i.e., once or twice in this set-up.
In Table 2, we report the median over the 200 simulation runs over several quantiles of the interval lengths among the selected variables. Due to the possibility of infinite interval lengths, we focus on quantiles instead of averages.
Again, we note that for f = 0.5, the intervals obtained through multicarving are longer than those from multisplitting. However, the power of multicarving comes from the ability to raise the selection fraction without losing all information for the inference stage. The 50 selected models become more stable for larger values of f and fewer covariates are selected in total over the B splits. The total number of distinct variables selected over all the splits is 96 and 20 on average for f = 0.5 respectively f = 0.99. With fewer features under consideration, a higher fraction of those is selected sufficiently often such that powerful inference is possible. Those effects are visible in Table 2 as the quantiles of the intervals using multicarving mostly become shorter when increasing f . For carving, there is also a natural countereffect as information for the inference stage is lost, thus the quantiles of interval lengths are not strictly decreasing.
In summary, our confidence intervals obtain the desired coverage stated in Equation (12). Further, multicarving brings an advantage compared to mul- Table 2 Results of length of confidence intervals. Median is taken over simulation runs of several quantiles over lengths of 95%-CI of variables that were selected at least once in B = 50 splits. tisplitting because of the possibility to perform well using a higher selection fraction f .
Data carving for group testing
In order to see how well our group test performs, we compare it with results presented in [12]. The authors consider two scenarios testing either a small or large group based on data simulated using different covariance structures. Testing a large group in a dense scenario is described below. Results of group testing for a small group in a sparse and high correlation scenario are illustrated in the Appendix C.2. The dense alternative with many small non-zero coefficients is a set-up where testing single variables is difficult. More precisely, p is 500 and n is varied in {250, 350, 500, 800}. The feature matrix X is generated from normally distributed features having a Toeplitz covariance matrix with ρ = 0.6. The parameter vector is defined as β j = δ for 25 ≤ j ≤ 50 and β j = 0 otherwise. We
Single-carving for group testing
We perform inference using our group test introduced in Section 2.5. As in Section 4.1, we vary the fraction of data used for selection f in {0.5, 0.75, 0.9, 0.95, 0.99, 1}. We start with just using a single split, i.e., B = 1, for inference. Notably, for the group test, inference using f = 1 is obtained with MCMC sampling as well. Since we condition on all but the covariates of interest, we generally have more than 1 degree of freedom such that an easy calculation as in [16] is not possible. The only exception to that is if G = 1, which is algorithmically equivalent to single variable testing.
For the selection, we perform cross-validation. Based on the assumption that Lasso might eliminate many of the covariates with weak signal, we use λ min instead of λ 1se . To assess the variance parameter σ, we use a global estimate obtained with cross-validation and λ min on all data.
In Table 3, we show the results for the dense alternative. For each combination of δ, n, and f , we report the empirical rejection rate (ERR), i.e., the fraction out of 200 simulation runs in which the null hypothesis is rejected at level α = 5%. For δ = 0, this measures the type-I error, for δ > 0, this measures the power.
For fixed δ > 0 and f , the power increases in the number of observations n, and for fixed n and f , it increases in the signal strength δ. This conclusion is to be expected.
The FWER is controlled for all combinations of f and n, for most combinations even conservatively. The fraction f = 0.5 has always the lowest power because selection works not overly well. In many settings, f = 1 is also suboptimal with respect to power as too little power remains for the inference stage. Fractions f = 0.9 to f = 0.99 are all competitive and perform similar. This is in good accordance with our results testing for single variables in Section 4.1. Table 3 can be compared to [12, Table 1] for δ in {0, 0.04, 0.06} and n in {250, 300, 500}, where six different methods are evaluated in this scenario. Our method with fractions between f = 0.75 and f = 0.99 is amongst the best with respect to power in each set-up. Especially, it has clearly higher power than their method φ Σ (1) for δ = 0.04, whereas the power is similar for δ = 0.06. The power of the method φ Σ (0.5) is comparable to the power of our group test but their method attains slightly lower values. Though, their method φ Σ controls the error more conservatively such that a clear statement in favor of either method is not possible.
If we summarize the results from the dense scenario in this section and the results from the sparse scenario in Appendix C.2, we can state that our method does not have the best performance in all possible set-ups. Though, it is competitive in all of them, while all competitors have some set-ups where they do not work well at all. Thus, our group test, which results from a very simple adjustment of the data carving idea, offers some valuable results.
Multicarving for group testing
In Section 4.1, we see that the multicarve method usually has better error control than single-carving. Based on this observation, it is to be expected that multicarving could further improve on group inference in scenarios where the error is not controlled conservatively (cf. Table 3). Therefore, we test multicarving for group testing as well. Indeed, with multicarving, no ERR above the target level 5% occurs for δ = 0 in either alternative. However, the ERR for δ > 0, i.e., the power, is sensitive to the choice of the tuning parameters f , γ or γ min , and B. Especially, in the two scenarios under consideration, aggregation using a fixed quantile clearly outperforms the use of an optimized quantile according to Equation (6).
In the following, we present results obtained using B = 20 splits and a fixed quantile for aggregation of γ = 0.05 in order to show the possibilities of multicarving. We emphasize that these choices work comparably well such that in general, when no such comparison is possible, one could expect slightly lower power using multicarving for group testing. Those results are shown in Table 4.
We consider the dense alternative. For multicarving, the highest ERR for δ = 0 is 5%, whereas it is 8% for single-carving. Naturally, there is some fluctuation involved in those empirical values. Nevertheless, this difference indicates an improvement of multicarving over single-carving. For most scenarios with δ > 0, a selection fraction of f = 0.5 is favorable. The intuitive explanation is that although S [b] ∩ G might on average be smaller than with higher fractions f , it is still "big enough" in a decent number of splits. In these splits, the lower f allows for a more powerful inference statement making the method more powerful overall after aggregation. Notably, using B = 20 and γ = 0.05 (fixed quantile for aggregation) is equivalent to a Bonferroni corrected minimum pvalue (cf. Equation (5)). Thus, only the most significant split is of importance. We now compare the power in Table 4 to that for single-carving in Table 3. Using a selection fraction of 0.5, multicarving outperforms any single-carving configuration in all scenarios unless δ = 0.02 and n = 250. Thus, using multiple splits and aggregating can bring a clear improvement. Though, this is rather sensitive to the choice of the tuning parameters as mentioned above.
In summary, the natural extension of our group test using multiple splits leads to a performance boost. Especially, the error can be controlled on a more conservative level using multiple splits. A drawback of the method is its sensi- Table 4 Empirical rejection rate at level 5% for the dense alternative using multicarving. tivity to tuning parameters. If those happen to be chosen poorly, power might be lower compared to single-carving.
Multicarving for logistic regression
We conduct a similar simulation study as in Section 4.1 for the logistic model (15). We reuse the matrix X coming from a Toeplitz covariance design from Section 4.1 with dimensions n = 100 and p = 200. The active variables are {1, 5, 10, 15, 20}, each of which having a coefficient of 2. After having noticed that cross-validated Lasso tends to select overly sparse models in logistic regression, at least in this set-up, we alter the selection technique. Namely, we select a Lasso model with a given number of selected variables or if there is no such model, the largest model with fewer variables. Inspired by [21], we choose this number to be n 6 = 16. Just as for cross-validated Lasso, this introduces a slight bias to our test as λ is determined in a data-dependent fashion and is not predefined. We stick to our usual tuning parameters, i.e., B is varied in All methods are rather conservative in this set-up. Especially, no value of the FWER above the 5% level occurs. Furthermore, no significant findings are observed for splitting which results in a power of 0. There exist probably better algorithms for calculating low-dimensional p-values in logistic regression than the ones used for splitting here. Though, as it is not of primary interest to our work, we did not investigate this further. Single-carving has clearly higher power than multicarving, whereas the latter controls the error on a more conservative level. The highest power obtained is 0.28 (single-carving), 0.16 (γ min = 0.3) and 0.14 (γ min = 0.05). All these maxima are reached at f = 0.75. Pure postselection inference has a power of 0.088. Thus, the conjecture that the constraints might be too restrictive is confirmed.
For the trade-off between power and error control, we consider the adjusted power as defined in Section 4.1. Interestingly, multisplitting is now quite competitive. The interpretation is that although p-values are generally larger than 5%, there is still a distinction between active and non-active variables. The best adjusted power of the multisplit method is 0.54. As the curve seems to increase towards lower values of f , we further tested f = 0.3 and f = 0.4. Neither leads to an increase in the adjusted power for multisplitting such that we can assume that the optimum is reached around f = 0.5. Multicarving clearly outperforms single-carving with the respective maxima being at 0.67 (γ min = 0.3), 0.64 (γ min = 0.05) and 0.49 (single-carving). Pure post-selection obtains an adjusted power of 0. 16. In summary, we can state for this data that either of the carving methods improves on pure post-selection inference. The choice between multicarving and single-carving is a trade-off between power and FWER. Our definition of adjusted power, which makes the different methods have equal FWER, is in favor of multicarving.
Runtime considerations
Our method is computationally quite involved while performing empirically well. Details are discussed in the Appendix C.4. The computational bottleneck is the MCMC sampling required to calculate p-values and therefore, we ignore the other steps for our considerations. An approximate bound is O BE s 4 for multicarving, where the expectation is due to the fact thats is non-constant over splits.
Another popular inference technique for high-dimensional statistics is the de-biased Lasso [35]. A total of p + 1 Lasso fits have to be calculated on the entire data. Thus, it scales as O p 2 . For high-dimensional data with p n and the standard assumptions ≤ n 1 ≤ n on the Lasso, we have accordinglỹ s p. Then, our multicarve method is more efficient than the de-biased Lasso for p → ∞ if n = O p 1/2 .
Discussion and conclusions
We provide new developments based on the idea of data carving [10]. Particularly for high-dimensional scenarios, we improve upon standard data carving.
First, we introduce multicarving in the spirit of multisplitting. Our simulation study shows that multicarving generally leads to better error control and its adjusted power is better than for the single-carve method. Furthermore, multisplitting and multicarving not only aim to reduce the FWER but also to make results more replicable. It is very plausible that our multicarve method clearly increases replicability compared to single-carving, due to the instability of the Lasso model selector.
Second, we present group inference, a natural extension of single variable testing. Such a group test can be applied using single-carving or using the advocated multicarving. In simulation examples, either variant appears to be competitive to several methods discussed in [12].
Last, we adapt data carving to make it applicable to logistic linear regression and other generalized linear models. Those adjustments are based on the central limit theorem and follow from similar ideas as already introduced for low-dimensional data and for pure post-selection inference. Our simulation study leads to the same conclusions as for the linear model. In particular, data (multi)carving in the logistic case leads as well to a performance increase compared to pure post-selection inference.
User-friendly R-software for all of the described (multi)carving methods is available on GitHub, see https://github.com/cschultheiss/Multicarving.
Appendix A: Proofs
Proof of Lemma 1 We require Assumption (Ã2) such that X 1,S X 1,S −1 is defined. As in Section 3.1, we implicitly assume rank X 1,S =s to follow from the sparsity condition. This inverse is implicitly included in A and b. Using the screening assumption, we know E [Y] = XSβS. Thus, we can write the unconditional distribution of Y as follows where c XSβS, σ 2 denotes the normalizing constant of the Gaussian distribution. We see that X S Y is the sufficient statistic, while βS is the natural parameter as σ is assumed to be known. Conditioning on the selection event Y AY ≤ b leads to a different exponential family with the same sufficient statistic X S Y and natural parameter βS but different normalizing constant, say, c , compare with [10,Section 3].
Here, we split into the parameter that we want to perform inference for βS G and the nuisance parameter in the model βS −G . From the theory of exponential families, we know that the conditional law X G Y XS \G Y, AY ≤ b does not depend on βS −G . We now want to establish the same result for For simplicity, we assume XS = XS \G XG such that it can be separated into variables being part of the group and the others. The result holds w.l.o.g., since permutations of the matrix' columns do not change our inference statement. Then, we get , making it independent from βS −G as well. Naturally, the subset X + S G Y is conditionally independent too. Based on our two assumptions, the only parameters in the model are βS −G and βS G . Thus, after establishing independence from the former, the only parameter left in the model is the latter, which is exactly the lemma's statement.
Proof of Theorem 4
For a group G, we either have G > 0 or G = 0. Assume the former case first. Due to screening, we know βS j = β j ∀j ∈S, which leads to βS j = β j ∀j ∈G asG ⊆S. Null hypothesis (13) then directly implies which corresponds to null hypothesis (14). Therefore, all assumptions of Theorem 3 are fulfilled, leading to the uniform distribution of the p-value. Error control can thus be stated as In the other case ( G = 0) we have Thus, we obtain error control in either case, which closes the proof.
Appendix B: Sampling from a linearly constrained Gaussian
The algorithm presented in this section is strongly based on the GitHub repository cited in [10] for their simulations. However, since there seems to be no written documentation of the algorithm itself and the theory behind, we provide it for the interested reader.
For simplicity, we will suppress indexS, since we implicitly assume to work in a selected submodel throughout this section.
In order to do inference for variable j, the goal is to sample from Y ∼ N Xβ, σ 2 I n subject to AY ≤ b, (X −j ) Y = (X −j ) y ≡ d and β j = 0. The first condition leads to boundaries on the sampling region, the second one changes both the mean parameter and the covariance matrix, and the last one further changes the mean and creates a null distribution.
B.1. Change of mean and covariance
Let Z be a Gaussian random vector with mean μ and covariance Σ. We are interested in E Z CZ = d ≡μ and Cov Z CZ = d ≡Σ. To find those, split Z into One can see (e.g., by calculating the covariance) that the second term is independent of CZ, thus unchanged by the conditioning, while the first part is completely defined by the conditioning. Thus, we havẽ In our problem of interest, we have μ = X −j β −j (after setting β j = 0), Σ = σ 2 I n , and C = (X −j ) . This yields Most importantly, the mean term does not have any dependence on β −j such that we can calculate an inference statement without knowing the other coefficients.
B.2. Computational shortcuts: linear transformations
Since all constraints are linear, they can also be guaranteed for linear transformations of Y if not too much dimensionality reduction is applied.
Define the least squares solution on all data as β = X X −1 X Y and the one on the selection data only as Then, two vectors which are well suited to fulfil all constraints after transformation are Since those are linear transformations, they will still be Gaussian with mean and covariance that can be easily derived from those of Y. Further, the constraints transform to We use the bracket notation for the indices to indicate that row j of the resulting matrix has to be omitted. And, by using the active constraints from [16], we have where ξ denotes the signs of the parameters' Lasso estimates. This can be transformed to Thus, we have transformed the linear equality and inequality constraints and can proceed as if we were to sample from Y by firstly adjusting the mean and the covariance matrix as described in Section B.1. The choice of whether to sample from U or V is rather simple: just use whichever has lower dimensionality in order to increase efficiency. As stated in Section 2.2.2, one would further condition on Y 2 in the unknown variance case. Though, this constraint is not transformable to U or V, thus the dimensionality could not be reduced. Therefore, we use an estimate of the variance instead of the (theoretically beautiful) conditioning idea for our simulations.
B.3. Whitening
In order to make the MCMC algorithm simpler, we would like to always sample from zero mean unit variance independent Gaussians (i.e., white Gaussians). This can be achieved by applying a further linear transformation. We need a forward map transforming the initial point and an inverse map transforming back the MCMC sample.
Assume that we sample from Y ∼ N (μ, Σ) AY ≤ b which is achieved by applying the transformations from the previous two sections. Here, Σ ∈ R n×n has rank r = n + 1 −s, i.e., Σ is not full-ranked whenevers > 1. This is as we lose some degrees of freedom after conditioning (cf. Section B.1). Further, define matrices Σ 1 2 ∈ R n×r and Σ − 1 2 ∈ R r×n such that These can be found, e.g., by using the eigenvalue decomposition of Σ. Then, our forward map is and accordingly, the inverse map is Note that W W −1 Y = Y ∀Y and further W −1 W Y for all Y fulfilling the equality constraints, thus all Y we are interested in.
Importantly, the boundary constraint AY ≤ b has to be transformed as well. This is possible by i.e., the constraints in the whitened space. With these whitened constraints at hand, the only thing left is to sample from a white Gaussian subject to linear inequality constraints. Notably, since Σ − 1 2 is a wide matrix (r < n unlesss = 1), we transform into a lower-dimensional space. Therefore, the transformation into the withened space leads to a further dimensionality reduction, which makes the sampling more efficient.
B.4. Sampling from a linearly constrained white Gaussian
The MCMC algorithm presented in this section is as well based on the mentioned GitHub repository. Though, we emphasize that any algorithm approximating a white Gaussian with linear inequality constraints could be invoked in this place using the same preprocessing steps (cf. Sections B.1-B.3).
For simplicity, reuse all initial names, thus we want to sample from Y ∼ N (0, I) subject to AY ≤ b and let y 0 be a point fulfilling the constraints. More precisely, y 0 is the preprocessed version of the observed vector.
The idea is to move in every step t in a given random direction η t , while keeping the projections into its orthogonal complement fixed, i.e., Or in other words, we want to sample from This is in exact analogy to the set-up in [16] for pure post-selection inference using η t as direction of interest and y t−1 as observation to base the inference on. Thus, the boundary derived for pure post-selection inference can be reused, making η t Y t a univariate truncated Gaussian with known mean and variance. One can easily sample from this leading to a new point y t . For every Y t , this can be repeated for a new random direction η t such that the whole constrained space should be explored. After enough steps, the samples should approximate the null distribution sufficiently well. An alternative algorithm that could be used for the actual MCMC sampling is the Hamiltonian Monte Carlo algorithm described in [25]. An implementation thereof is available in the R-package tmg [24].
Appendix C: Additional numerical results
This section contains additional numerical results and details about runtime considerations.
C.1. Multicarving for the linear model
We consider slight variations of the simulation set-ups in Section 4.1. Especially, we look at scenarios where the selection stage is rather hard leading to low probability of screening. This can have a negative impact on the performance of the inference methods for multiple reasons. First, without screening the theoretical validity for the error control is not given anymore. Second, selecting less true active predictors leads to less potential for true rejections such that the power drops.
C.1.1. Toeplitz design with different correlation parameter
As we mention in Section 3.1, the correlation between predictors has a high impact on the success of screening in the finite data set-up and accordingly, on the performance of our procedure. To analyze this effect, we redo our simulation for the Toeplitz design in Section 4.1.1 with different correlation parameter ρ. We test the values ρ = 0.3 and ρ = 0.9 and otherwise proceed as before. We sample the predictor matrix X once for each value of ρ. To make things as comparable as possible, we fix the noise level such that Var(Xβ) σ 2 = 1.71 as it was in the set-up in Section 4.1.1.
We first consider ρ = 0.3 for which we show the obtained FWER and power in Figure 7 and the obtained adjusted power in Figure 8. Comparing this to our base case in Section 4.1.1, we see that the curves for power and adjusted power are much higher while as the FWER are at a lower level. Thus, this problem is a lot easier to handle by all the inference methods at hand as one would expect. Our conclusions are similar to the set-up with ρ = 0.6 comparing the different methods. Single-carving obtains the highest power with FWER ≤ 5% with a maximum of 0.79 whereas the multicarving methods reach 0.74 (γ min = 0.05) and 0.75 (γ min = 0.3). Though, multicarving controls the error more conservatively leading to better adjusted power with respective maxima of 0.90 for either carving method and 0.84 for single-carving. Two things shall be noted: First, f = 0.75 is now competitive with higher selection fractions which can be explained by the empirical success rate of screening that is already rather high (76.5%) for f = 0.75. Second and related, multisplitting is also more competitive since screening works reasonably well for selection fractions for which there is still some power left using only the second part of the data for inference.
For the high-correlation case with ρ = 0.9, the results are displayed in Figures 9 and 10. Note that the plotting range is restricted to a more representative area and that some FWER symbols above the level 40% are thus missing. As expected, those results now look much worse. Especially, neither carving method is able to control the FWER at 5% for any selection fraction. Of course, this relates to the low probability of screening which is only at 7.9% even when using all the data for the selection stage. Nevertheless, we still see that multicarving leads to better error control than single-carving except for f = 0.5, where the FWER for single-carving is 43% and for γ min = 0.05 it is even 71%. Accordingly, the best adjusted power is also better for multicarving. Though, all the values are on a very low level with respective maxima of 0.052 for either value of γ min and 0.045 for single-carving. Further, we note that multisplitting with γ min = 0.3 performs roughly as well as multicarving with respect to the adjusted power for f = 0.5 and f = 0.75. We think that this is because the performance of each method is mainly driven by the selection quality in this scenario such that the blessings of multiplicity are more pronounced than those of carving.
In summary, our assumption that lower correlation leads to better performance and vice-versa is confirmed in this analysis. Especially, none of the inference techniques in scope works well in a scenario where the selection stage is very difficult and screening is very unlikely. Nevertheless, we can still see some positive effect of using multiple splits in this scenario.
C.1.2. Semi-synthetic Riboflavin data for sparsity 4
We redo the simulation as in Section 4.1.2 setting the sparsity to 4 without changing anything else. The respective results are presented in Figures 11 (FWER and power) and 12 (adjusted power).
Note that we restrict the plotting area of the y-axis to a maximum of 0.2 such that some values of the FWER are non-visible. At first glance, one sees that the power is generally quite low for all methods while as the error is above the 5% level for many set-ups leading also to low adjusted power. As in Appendix C.1.1, this relates to the difficulty for the selection stage. Screening only worked in 9.8% of the simulation runs using all data for selection and naturally even less for any subset. For comparison, screening worked in 81.3% of the instances in the sparser alternative, which makes the problem much easier.
In this set-up, multicarving with γ min = 0.05 has the highest power for all f , while γ min = 0.3 has the lowest FWER amongst the three carving methods. The highest power obtained controlling the FWER at 5% is in favor of using γ min = 0.3 with a value of 0.065. The other two methods obtain respective maxima of 0.055 (γ min = 0.05) and 0.026 (single-carving). Especially, singlecarving only reaches error control at f = 1 which is pure post-selection inference. The adjusted power is slightly higher for γ min = 0.05 than for γ min = 0.3 with maximal values of 0.078 and 0.069. For single-carving, the maximal value is 0.048. In summary, multicarving is to be preferred over single-carving in this difficult set-up.
As one of the main difficulties in this scenario is the bad screening property, a natural adaption is the use of λ min instead of λ 1se for selection with crossvalidation. This leads to larger selected models and could potentially increase the probability of screening. Our simulation confirms that this leads to a performance boost with the highest adjusted power for multicarving now being 0.148. For simplicity, we refrain from showing the results in detail. It has to be mentioned though that the use of λ min leads to a substantial increase in runtime as more variables are selected. We elaborate this effect further in Section C.4.
C.2. Data carving for group testing: sparse scenario
We refer to Section 4.3 for more details about the implementation and further discussion.
For the sparse scenario, we choose β to be sparse and the active covariates are strongly correlated with other covariates. The number of covariates p is as well 500, and X is simulated using the following covariance structure Σ jl = 0.8 if1≤ j = l ≤ 5 0.6 |j−l| otherwise.
Thus, Σ is the same Toeplitz matrix as in the dense alternative described in Section 4.3 unless for the first five variables. The parameter vector is defined as β 1 = β 3 = δ and β j = 0 otherwise, meaning that the active variables are within the highly correlated set. This time δ is varied over {0, 0.1, 0.2, 0.3, 0.4, 0.5} and n over {250, 350, 500}. The response Y is generated as before, leading to SNR in {0, 0.036, 0.144, 0.324, 0.576, 0.9}. In this scenario, we are interested in the null hypothesis (13) for the group G = {1, 2, . . . , 5}.
C.2.1. Single-carving for group testing: sparse scenario
In Table 5, we report the empirical rejection rate for the scenario with very sparse β and highly correlated features. This scenario seems to be easier to handle than the dense scenario. Especially, the error is controlled at a more conservative level, with the highest error being 1.5%. For the power, the tendencies are similar as before. For δ ∈ [0.1, 0.2], f = 0.5 and f = 1 have generally the lowest power, while the highest power is obtained with f ∈ [0.75, 0.95]. Starting from δ = 0.3, f = 1 leads to the lowest power, while the other ERR are mostly exactly 1. These results are to be compared to [12, Table 3] for δ in {0, 0.2, 0.3}, where they test the six methods in the sparse scenario. Their proposed methods φ Σ (0.5) and φ Σ (1) have lower power than our method for δ = 0.2, while error control works very reliably for all three methods. The power is (almost) at 1 for all methods (except for their method φ I ) for δ = 0.3. The methods φ hdi and φ FD obtain values of power comparable to our method at a price of clearly higher error.
C.2.2. Multicarving for group testing: sparse scenario
The results for multicarving are illustrated in Table 6. As for single-carving, error control in the highly correlated sparser alternative is no issue with the multicarve method. Namely, no ERR above 1.5% occurs for δ = 0 for multicarving either. Again, using a selection fraction of f = 0.5 seems to be favorable for multicarving.
Looking at Table 5, one sees that none of the single-carving configurations outperforms multicarving with f = 0.5 in any scenario with δ > 0. Therefore, Table 5 Empirical rejection rate at level 5% for the sparse alternative using single-carving. we can state that multicarving brings an improvement in this alternative as well when choosing the tuning parameters properly
C.3. Effect of the aggregation parameter on the runtime
Using a larger γ min for the aggregation in (6) is favorable for computational reasons. First, only variables present in at least γ min B models have to be tested for. The higher this threshold is, the more variables can be omitted directly, reducing computing time. Second, if we account for the multiplicity correction that we impose through considering multiple variables and aggregating over multiple splits, raw p-values of αγ miñ s (1 − log (γ min )) or smaller should be possible. Otherwise, one can never observe a significant effect occurring from P j = (1 − log (γ min )) Q j (γ min ) (cf. Section 2.2.1). Accordingly, we need at least s (1 − log (γ min )) αγ min (19) MCMC samples to use the method to full capacity. This requirement decreases in γ min and is about 11 times higher for γ min = 0.05 than for γ min = 0.3.
C.4. Details for runtime considerations
We discuss what influences the runtime of multicarving and how to further speed it up. Especially, we want to assess how the runtime behaves as p n → ∞. We first review the structure of our method. For a total of B times, the data is split into two parts, a model is selected on the first part, and p-values are calculated using the carving idea. For those p-values, a separate calculation for all of thes selected variables is necessary. Lastly, the B p-values of the different splits are aggregated per covariate. We ignore splitting the data, the initial selection stage, and the aggregation for our considerations since the computational bottleneck is the MCMC sampling required to calculate p-values.
Naturally, the runtime scales linearly in B. For every split,s MCMC chains have to be sampled and one needs O (s/α) samples in order to have the possibility to observe a significant result. Multicarving takes B (1−log(γmin)) γmin times as long as single-carving due to using multiple splits and the aggregation over the different splits (cf. Equation (19)). Though, in practice convergence of the chain is another issue such that for single-carving more than the minimally required samples are likely to be generated and the difference between the two methods is slightly reduced. For single-carving and multicarving, there is a factor of s 2 involved as one needss chains of size O (s/α). Lastly, sampling happens in a min (s + 1, n 2 + 1)-dimensional space subject tos inequality constraints (cf. Appendix B). We discuss two algorithms in the Appendix B.4 and the choice of the MCMC algorithm influences the runtime.
Pakman and Paninski [25] state that for their algorithm the exact run time also depends on the shape of the constraint such that a general statement cannot be made. There are steps of complexity O min (s + 1, n 2 + 1) 2 and O min (s + 1, n 2 + 1)s involved, which can be bounded by O (s) 2 . However, the number of such calculations needed depends on the selection event's geometry.
For the hit-and-run algorithm adapted from the GitHub repository cited in [10], every step involves solving a problem of the complexity of pure postselection inference as in [16]. Due to the matrix equation involved in calculating the bounds, this leads to a complexity of O (min (s + 1, n 2 + 1)s) ≤ O s 2 .
For both algorithms, we come up with an approximate bound of O BE s 4 for multicarving where the expectation is due to the fact thats is non-constant over splits.
In comparison, if we use the saturated viewpoint instead, p-values for every variable are determined by calculating bounds once taking at most O (n 1s ) steps. Assumings = O (n 1 ), the inference process can be bounded by O BE s 3 such that a factor ofs is saved. Though, it might be less appropriate to ignore the initial Lasso selection for runtime considerations in the saturated model.
Notably, there are several ways to speed up multicarving algorithmically. We want to state the two most obvious. As mentioned in Section C.3, not all covariates have to be tested for but only the ones selected in at least γ min B of the splits. This means that the algorithm described in Section 2.3 has to be adjusted to selecting B models first and performing inference afterwards, while the final outcome is not altered by this change. This improvement is more pronounced for higher values of γ min . The exact same adjustment could also be applied to multisplitting. Second, not every MCMC chain has to be run to the full extent as in Equation (19). If it is already clear with fewer iterates that a covariate cannot be shown to be significant, the chain can be aborted in an earlier stage as for p-values clearly above the significance level the precision is less important. | 22,460.2 | 2020-06-08T00:00:00.000 | [
"Mathematics",
"Computer Science"
] |
Electromagnetic Spectrum Allocation Method for Multi-Service Irregular Frequency-Using Devices in the Space–Air–Ground Integrated Network
The management and allocation of electromagnetic spectrum resources is the inner driving force of the construction of the space–air–ground integrated network. Existing spectrum allocation methods are difficult to adapt to the scenario where the working bandwidth of multi-service frequency-using devices is irregular and the working priorities are different. In this paper, an orthogonal genetic algorithm based on the idea of mixed niches is proposed to transform the problem of frequency allocation into the optimization problem of minimizing the electromagnetic interference between frequency-using devices in the integrated network. At the same time, a system model is constructed that takes the minimum interference effect of low-priority-to-high-priority devices as the objective function and takes the protection frequency and natural frequency as the constraint conditions. In this paper, we not only introduce the thought of niches to improve the diversity of the population but also use an orthogonal uniform crossover operator to improve the search efficiency. At the same time, we use a standard genetic algorithm and a micro genetic algorithm to optimize the model. The global searchability and local search precision of the proposed algorithm are all improved. Simulation results show that compared with the existing methods, the proposed algorithm has the advantages of fast convergence, strong stability and good optimization effect.
Introduction
The deployment and application of 5th-generation mobile communication (5G) has opened the era of the Internet of everything and promoted the integration of the information industry and other traditional industries [1]. With the commercialization of 5G technology, many countries and organizations have made forward-looking layouts for 6th-generation communication (6G) [2,3]. In order to support full coverage of the network and the highspeed mobility of users, the breadth and depth of the existing communication range of the ground network are extended. The 6G technology will cover network infrastructure such as airspace and space and integrate ground and non-ground networks to provide full coverage of space, sky and land [4,5].
The space-air-ground integrated network integrates space-based networks of orbiting satellites, high-attitude platforms (HAPs) and HAPs consisting of unmanned aerial vehicles (UAVs) [6,7]. The ground-based network, composed of traditional wireless communication systems, can achieve seamless coverage of the whole domain and provide users with ubiquitous communication services, which is the core development direction of 6G technology in the future [8]. At present, the research on the space-air-ground integrated network mainly includes joint beamforming, power allocation and integration with the Internet of Things (IoT). In [9], the authors proposed a joint optimization design for a non-orthogonal multiple access (NOMA)-based satellite-terrestrial integrated network (STIN), where a satellite multicast communication network shares the millimeter wave spectrum with a cellular network employing NOMA technology. In [10], the authors investigated the multicast communication of a satellite and aerial integrated network (SAIN) with rate-splitting multiple access (RSMA) to satisfy the explosive access demand of IoT devices. However, due to the differences of services among space-based networks and ground-based networks, the heterogeneity of frequency-using devices and the diversity of management, the space-air-ground integrated network is facing great challenges in the management and allocation of electromagnetic spectrum resources [11,12].
Spectrum allocation technology has been of great interest to researchers at home and abroad. In general, spectrum allocation technologies can be divided into the following three categories: Firstly, there are spectrum allocation technologies with regions as allocation objects [13,14]. This technology takes "cell" as the frequency object and the interference between cells as the constraint condition. Genetic algorithms, particle swarm optimization algorithms and other heuristic algorithms are used to achieve the frequency demand of each cell. Such methods solve the spectrum allocation problem to a certain extent but do not specifically consider the resource requirements of specific devices. Thus, they cannot be extended to the actual problem of different types of devices having different resource requirements and different frequency priorities. Secondly, there is spectrum allocation technology based on the frequency-using device as the allocation object [15,16]. This technology takes specific frequency-using device as the spectrum allocation objects, takes protection frequencies and prohibition frequencies as constraints, and takes same frequency interference and adjacent frequency interference between devices as optimization objectives to model. At the same time, an improved heuristic algorithm is used to solve the problem. Such algorithms achieve fine spectrum allocation at the device level, but the modeling of the spectrum allocation problem does not meet the actual business needs and does not take into account the difference in frequency priority. Thirdly, there is spectrum allocation or access technology based on cognitive radio or dynamic spectrum access [17,18]. This technology divides frequency-using device into two categories-primary users and secondary usersand allocates resources to secondary users on the premise of ensuring the performance requirements of primary users. The constraints are the maximum interference threshold in the same channel and the maximum transmission power of each secondary user. The optimization goal is to maximize the spectral efficiency of each user or the number of users successfully allocated. Reinforcement learning or heuristic algorithms are usually used to solve the problem. These algorithms can adapt to the dynamic electromagnetic environment and distributed decision-making scenarios and have good flexibility and robustness. However, the problem of differentiation of frequency demands of users at all levels has not been solved, and the method of dividing users at all levels cannot adapt to scenarios with different priorities.
Generally speaking, the basic thought behind existing technology is similar: abstracting spectrum resources into several resource blocks with different properties while taking the electromagnetic interference between devices, device transmission power or various frequency restrictions as constraints. The purpose is to minimize the interference to the devices and meet the frequency requirements of each device. At the same time, various algorithms are used to solve the electromagnetic spectral distribution problem to a certain extent. However, the space-air-ground integrated network is a huge and complex system; it is a highly integrated system of multi-dimensional networks. The frequency requirements of different business types of devices vary greatly, and the existing technology cannot meet the irregular frequency requirements of various business types of devices in complex networks. In addition, in the actual business scenario, each frequency-using device performs different types of business, and different business types have different contributions and impacts to the network. Therefore, frequency requirements for devices providing high-priority services should be preferred. Existing techniques do not consider the practical problem of different frequency-using priorities for each device. For the above problems, this paper firstly conducts mathematical modeling on the spectrum allocation problem of frequency-using devices in the multi-service irregular scenario and proposes a multi-service irregular frequency allocation method based on mixed niche orthogonal genetic algorithms. The proposed algorithm introduces niche thoughts based on the clearing mechanism to improve the diversity of the population and proposes orthogonal uniform crossing operators to improve search efficiency. Moreover, the proposed algorithm uses a standard genetic algorithm (SGA) and a micro genetic algorithm (MGA) to improve the global search capability and local search accuracy.
System Model
We show the spectrum allocation system model of frequency-using devices with different operating frequencies and priorities in the space-air-ground integrated network in Figure 1. The system model consists of a scene description, constraint conditions and objective functions. When a large number of frequency-using devices apply for working frequency bands from the spectrum management system, the spectrum management system will allocate spectrum according to the natural frequency band of each device, its priority, the system's rejection frequency band, and the interference between devices. Assuming that there are N devices in the system, the differences between devices of different business types are mainly reflected in the parameter numerical level, and device of different business types can be described in k unified ways. This paper specifically describes a device k using the parameters shown in Table 1, where p k is the value of device priority. The higher the priority, the smaller the value of p k . The variables x k , y k and z k , respectively, represent the longitude, latitude and height of the device deployment point; f kgs and f kge represent the start and end frequencies of the device's natural frequency band, respectively; f ks and f ke represent the start and end frequencies of the device's working frequency band, respectively; the device's working frequency band must be within the range of the natural frequency band.
Control link
According to the scenario description, the spectrum allocation scheme must meet the following three constraints: • The operating frequency band of each device shall be within its inherent frequency band range, as shown in Equation (1) : • The operating frequency band of each device should not overlap with the protection frequency band of its protection area, as shown in Equation (2) : • The operating frequency band of each device shall not overlap with any rejection frequency band, as shown in Equation (3) :
Objective Function
The objective of the model is to avoid interference with or reduce interference to high-priority devices. Interference can be considered from two aspects: Firstly, frequency domain interference: if the working frequency bands do not overlap, then there is no frequency domain interference between the two devices. If there is overlap, the severity of interference is described by the frequency spectrum overlap degree. Based on this, the frequency domain interference coefficient is defined as shown in Equation (4): where f ic = 0.5 · ( f is + f ie ) is the center frequency of the operating frequency band of the i-th device, and i, j ∈ [1, N]. Secondly, airspace conflict: assuming that the transmitter device sends signals with power P i and the receiver device sensitivity is S j , in the process of signal propagation, path loss L ij is generated, as shown in Equation (5) [19]: where D ij represents the straight-line distance between the i-th device and the j-th device, and i, j ∈ [1, N], i = j. The spatial interference coefficient is defined to quantify spatial interference, as shown in Equation (6): When I( f ic ) = 1, there is spatial interference between the originating device and the receiving device. Otherwise, there is no airspace interference.
The device interference coefficient a i (F c ) is defined to quantify the degree that each device does not interfere with other devices, as shown in Equation (7): where . . , f Nc },p max is the priority value with the largest absolute value, that is, the lowest priority value. The fitness function S(F c ) is defined as the objective function of the spectrum allocation model, as shown in Equation (8): The optimization objective of the model is to find the appropriate working frequency band for each device so that the interference of low-priority-to-high-priority devices is minimized.
In conclusion, the spectrum allocation model as shown in Equation (9) is established:
Description of the Algorithm
In this paper, a method is proposed to solve the multi-service irregular spectrum allocation problem. Firstly, several spectrum allocation schemes that meet the constraints are randomly generated. The content of the spectrum allocation scheme is the operating frequency band information of each device. Then, each spectrum allocation scheme is encoded with real numbers, and the scheme is mapped into a chromosome array. Each chromosome array is an individual, and the fitness function to evaluate the performance of each individual is the objective function of the spectrum allocation model. Randomly generated individuals form the initial population, which is used as the starting point for the genetic algorithm iteration. In terms of solving the algorithm, this paper proposes a new Hybrid Niche Orthogonal Genetic Algorithm (HNOGA). The steps of the HNOGA algorithm are shown in the pseudocode of Algorithm 1 .
Algorithm 1 HNOGA pseudocode 1: Input: population size M, number of elite individuals q, subpopulation neighborhood search probability P m1 , optimal individual neighborhood search probability P m2 , and maximum iteration number T; 2: Output: the optimal individual I b ; 3: Start: 4: Step 1: The initial population with population size M was generated by random coding, the fitness of each individual in the initial population was calculated, and the top q individuals with the highest fitness were copied and saved (q < M). Studies show that when q = 0.3M, the performance of the algorithm is optimal. 5: Step 2: Genetic algorithm optimization 6: for iteration t = 1 : T do 7: Main genetic algorithm operation: random league selection, orthogonal uniform crossover, uniform mutation; 8: All the individuals whose fitness is not 0 in the population and whose subpopulation contains greater than or equal to 2 individuals are optimized by the micro genetic algorithm with probability P m1 , and then the niche operation is performed; 9: The individual with the highest fitness in the whole population was selected to perform micro genetic algorithm optimization with probability P m2 ; 10: Perform niche operations; 11: The top q individuals with the highest fitness in the population are copied and retained, and the top M individuals with the highest fitness are used to form the population of the next iteration. 12: Step 3: Output the optimal individual I b of the last generation population; 13: End: Obtain the optimal individual I b .
In each iteration, a standard genetic algorithm (SGA) is first used for global optimization (this step is referred to as the "main genetic algorithm" in this paper), and then the optimized population is niche operated: Firstly, the normalized Euclidean distance between individuals in the population is calculated. If the normalized Euclidean distance of some individuals is less than a certain threshold, they are clustered into a subpopulation, and all individuals in each subpopulation have similar genetic characteristics. Secondly, the individual with the highest fitness in the subpopulation retains its original fitness value, and the fitness of the other individuals is set to 0. This operation makes the probability that the genetic characteristics of each subpopulation will be eliminated in the selection process of the next iteration very low to ensure the genetic diversity of the population. After the niche operation, the original population is divided into several subgroups. Then, the individual with the highest fitness is selected from each subpopulation with population size greater than 1, and the micro genetic algorithm is used for a neighborhood search. Thus, the full utilization of neighborhood information is realized, and the search accuracy and efficiency of the algorithm are improved.
The flowchart of the HNOGA algorithm is shown in Figure 2 .
Coding and Initial Population Generation
The HNOGA algorithm uses real coding to map from the spectrum allocation scheme to the chromosome array. Assuming that there are N devices in the region, the chromosome is an array of real numbers G s = {g 1 , g 2 , · · · , g N } of length N. Each gene locus g k of the chromosome corresponds to a device, and the value of g k is the central frequency of the operating band of the corresponding device. Assuming that the working bandwidth of the k-th device is B k , the start frequency of the working band of the device is f ks = g k − 0.5B k and the end frequency is f ks = g k + 0.5B k , k ∈ [1, N].
The HNOGA algorithm uses random coding to generate the initial population: for each device, the center frequency is randomly generated within its natural frequency band range, and the value is taken as the gene value of the corresponding position of the device on the chromosome. A complete individual can be generated by performing the above operations for all device. It should be noted that if the randomly generated center frequency value does not meet the three constraints of the model, it should be generated again until the location meets the constraints. Performing the above group size operation times generates the initial group.
Function of Fitness
The fitness function of the HNOGA algorithm is the same as the objective function of the system model. The fitness f (G s ) of an individual G s = {g 1 , g 2 ,...,g N } is defined as shown in Equation (10):
Selection Operator
The choice operator of the HNOGA algorithm is stochastic tournament selection. Compared with the commonly used roulette selection, the tournament selection strategy has the advantages of high solution accuracy and fast solution speed, so it is widely used in the construction of genetic algorithms [20,21]. The basic steps of the tournament selection strategy are as follows: (1) total of t individuals are randomly selected from the parental population; (2) Among the t individuals, the individuals with the highest fitness are selected and retained in the middle group; (3) assuming that the parental population size is M, the above steps are repeated M − 1 times to form a complete intermediate population.
The variable t is an artificially set parameter whose value greatly affects the actual effect of tournament selection. In this algorithm, the value t is set to 2.
Crossover Operator
In this paper, a uniform crossover operator based on an orthogonal experiment is proposed; it is called an orthogonal uniform crossover operator.
Orthogonal experimental design is an efficient and economical experimental design method that mainly studies the influence mechanism of several specific factors in a system at different levels on the overall state or performance of the system. It designs a variety of the most-representative test schemes based on the principle of orthogonality to evenly disperse and match the levels of factors with the factors. The effect of a few tests is equivalent to that of a full test. The main tool for orthogonal experimental design is the orthogonal table, which is a matrix arranged by rows and columns and is usually represented as L x (q y ), where L represents the orthogonal table, x represents the number of trials, q represents the number of factor levels, and y represents the number of factors. This paper uses a two-level orthogonal table.
The orthogonal uniform crossover operator is an improvement and supplement to the uniform crossover operator, and it uses a similar operation mode as the uniform crossover operator. Each gene of the progeny chromosome is derived from one of the alleles at the corresponding location of the two parental chromosomes that have been paired [22,23]. The difference is that the uniform crossover operator randomly selects crossed genes to obtain two individuals, while the orthogonal uniform crossover operator quickly selects the best combination of alleles of parental individuals through orthogonal experiments to obtain only one individual. Specifically, the process of finding the best combination of alleles can be regarded as a two-level orthogonal experiment in which the factors are the genes on the chromosome and the two-level of each factor is the specific values of the corresponding alleles of the two parental chromosomes. According to the chromosome length (i.e., the number of factors), a two-level orthogonal table can be established for the experiment. Each experiment gets a new individual. The fitness of all individuals obtained by the orthogonal experiment is calculated, and the individual with the highest fitness is taken as the result of the orthogonal uniform crossover.
Mutation Operator
The mutation operator of the HNOGA algorithm uses a uniform mutation operator, and the specific operation steps are as follows: (1) Each locus of an individual is designated as the point of variation.
(2) For each gene locus, a uniform mutation operation is performed with mutation probability p m .
(3) It is determined whether the working frequency band meets the three constraints of the model. If it does, continue; otherwise, go back to the second step.
(4) All individuals perform the above operations. Compared with the variation strategy, the uniform variation strategy has greater variation intensity and better searchability. It can not only improve the diversity of the population, but also avoid the prematurity of the algorithm, which is suitable for solving complex problems.
Niche Operation
The HNOGA algorithm uses a niche technology based on a Clearing Procedure. Its basic idea is to classify all individuals based on the Euclidean distance between individuals, divide the whole population into multiple subgroups, and select the individual with the highest fitness within each subgroup to participate in further optimization [24,25]. The steps are as follows: Step 1: If the fitness of two individuals is not 0, the normalized Euclidean distance between the two individuals is calculated. For any two individuals G i = {g i1 , g i2 , · · · , g iN }, G j = g j1 , g j2 , · · · , g jN , their normalized Euclidean distance is defined by Equation (11): Step 2: For any two individuals, if their normalized Euclidean distance is less than or equal to the niche radius D, the fitness of the two individuals is compared. The fitness of the individuals with small fitness is set to 0, while the fitness of the individuals with large fitness remains unchanged and they are placed in a subgroup. If their normalized Euclidean distance is greater than the niche radius D, no operation is performed.
Step 3: All individuals perform the above operations. These non-zero fitness individuals are essentially representatives of a class of individuals in the original population. Niche operations keep them in place in the iteration process, retaining the genetic characteristics of such individuals in the population and thus maintaining the diversity of the population. At the same time, this also avoids the phenomenon of a high-performance individual quickly replacing other individuals in the iterative process, resulting in the algorithm falling into a local optimum.
From the perspective of mathematical optimization, if the objective function of the problem solved by the genetic algorithm is a complex multi-peak function, the clustering phenomenon of some individuals in the population may be caused by the fact that these individuals search the neighborhood of the same local optimal solution. In the real genetic algorithm population, a class of individuals classified by Euclidean distance may be located near the same extreme point. The non-zero-fitness individual is the one closest to the local optimal solution after the niche operation. Then, a local search with higher precision can be further approached or even found with this individual as the starting point. Every extreme point of a multimodal function may be the global optimal solution, so the idea of a local search based on the individuals near the extreme point is significant for determining the global optimal solution. Local search methods are often referred to as "hill-climbing operators".
Micro Genetic Algorithm
In this paper, MGA is extended and a real micro genetic algorithm is proposed [26,27]. The length of the chromosome G m = g p1 , g p2 , · · · , g pN is equal to the number of devices N, and each gene location of the chromosome g pk corresponds to one device. The value g pk is the perturbation momentum of the center frequency of the operating frequency band of the corresponding device. The perturbation momentum should meet g pk ∈ −g pb , g pb , where g pb is the perturbation amplitude, namely the maximum perturbation momentum allowed by the algorithm k ∈ [1, N]. The micro genetic algorithm randomly generates the initial population in the same way as the main genetic algorithm.
Assuming that the optimization result of the main genetic algorithm is G s = {g s1 , g s2 , · · · , g sN } and an individual of MGA is G p = g p1 , g p2 , · · · , g pN , then G s will generate a new individual G under the action of G p . The action mechanism is shown in Equation (12): ..,g N } = g s1 + g p1 , g s2 + g p2 , · · · , g sN + g pN (12) The objective of the micro genetic algorithm is to find an appropriate perturbation g s to obtain a new individual G with the highest performance possible. Therefore, the objective function of the micro genetic algorithm is shown in Equation (13): MGA used in this paper adopts three genetic operators, namely random league selection, orthogonal uniform crossover, and uniform mutation. It should be noted that MGA may produce illegal individuals in the process of coding and mutation, and G s will generate new individuals that do not meet the constraints of the system model under the action of the individual. In this paper, the method to deal with this problem is to re-encode or mutate until a legitimate individual is generated.
Through the combination of the main genetic algorithm and MGA, the HNOGA algorithm forms an iterative process including an outer loop and inner loop. When the number of outer loops is too many and the number of inner loops is too few, the local searchability of the algorithm decreases. When the number of outer loops is too small and the number of inner loops is too large, the global optimization ability of the algorithm is limited. Therefore, the ratio of internal and external circulation times should be reasonably allocated. The literature [28] puts forward that the optimal ratio of internal to external circulation is 0.3 to 0.5. In addition, an elite strategy is also used in the MGA, in which the worst individual in each iteration is replaced by the best individual in the previous iteration.
Experimental Data Setting
Assume that there are 10 devices in the system, and the working bandwidth of each device is different. The data for each device are shown in Table 2 . In the actual scenario, there is only one protected zone. The start frequency and end frequency of the protected band in this zone are 150 MHz and 175 MHz, respectively. The protected area is a cuboid area with a latitude between 20 and 60 degrees, a longitude between 1 and 4 degrees, and a height between 20 and 50 m. There is only one denial spectrum in actual service scenarios. The rejection frequency band starts at 25 MHz and ends at 50 MHz. It can be seen that the total bandwidth of the commonly available frequency band is 200 MHz, while the sum of the working bandwidth of each device is 224 MHz. Since this paper does not consider the case of allocation failure, interference between devices cannot be avoided.
Simulation of Orthogonal Uniform Crossover Operator
To verify the performance of the orthogonal uniform crossover operator, this paper has carried out simulation experiments on SGA using the single-point crossover operator [29], two-point crossover operator [30], uniform crossover operator [31] and orthogonal uniform crossover operator. In the above four experiments, the selection operator used by the GA is a random league selection operator, the mutation operator is a uniform mutation operator, and the elite strategy is used in all of them. The population size M is 100, the largest number of iterations T is 200 times. The mutation probability p m is 10 %. The single-point crossover and multipoint crossover probability p c is 100%. The uniform crossover probability p uc is 70%. The content of the elite strategy of each algorithm is to replace the worst individual of the offspring with the best individual of the parent. Each algorithm is run 100 times respectively, and the average fitness of the optimal individual in each iteration of each experiment is taken to draw the simulation curve shown in Figure 3 . Figure 3 shows that the convergence rate of the uniform orthogonal crossover operator and the optimization effect are better than those of the other three kinds of common crossover operators. Compared with the other three strong crossover operators, the randomness of the uniform orthogonal crossover operator search efficiency is higher. At the same time, the orthogonal homogeneity operator can select the approximate optimal cross result of the two paired individuals through the orthogonal experiment.
Based on 100 executions of each of the above four experiments, the optimal fitness, worst fitness, average fitness, and variance of each experiment are shown in Table 3: Table 3. Experimental results of crossover operator performance verification. It can be seen from Table 3 that the optimal fitness, the worst fitness and the average fitness of the orthogonal uniform crossover operator in the results of 100 runs are better than those of the other three crossover operators, and it has the advantage of good stability.
Performance Simulation of a Mountain Climbing Operator Based on the Micro Genetic Algorithm
In this paper, MGA, simulated annealing algorithm (SA) [32] and Tabu search algorithm (TS) [33] are respectively applied to the standard genetic algorithm as mountain climbing operators. At the same time, the three algorithms are simulated 100 times, respectively. The three algorithms all use random league selection, uniform crossover, uniform mutation and elite strategy. The population size M is 100, the maximum iteration number T is 200, the mutation probability p m is 10%, and the crossover probability p uc is 70%. The population size of the micro genetic algorithm is 10, the maximum number of iterations is 20, the perturbation amplitude g pb is 10, and the mutation probability is 1%. The Tabu length of TS is 5, the maximum number of iterations is 20, and the neighborhood search method adopts uniform mutation with a mutation probability of 0.5%. SGA adopts the strategy described in the literature. The maximum number of iterations is 20, and the annealing rate is 0.01 in SA. The content of the elite strategy of each algorithm is to replace the worst individual of the offspring with the best individual of the parent.
Each algorithm is run 100 times respectively, the average fitness of the optimal individual in each iteration of each experiment is taken, and the simulation curve is drawn as shown in Figure 4. As can be seen from Figure 4, the convergence speed and optimization effect of MGA are significantly better than for other two hill-climbing operators because MGA can search multiple directions simultaneously from the starting point, which is more efficient.
Based on executing each of the above three algorithms 100 times, the optimal fitness, worst fitness, average fitness, and variance of the optimization results of each algorithm are shown in Table 4 . In the statistical results of 100 experiments, the optimal fitness, the worst fitness and the average fitness of MGA are better than for the other two mountain climbing operators. However, for any algorithm, the performance improvement in some aspects is often offset by a performance reduction in other aspects. When MGA is applied to the standard genetic algorithm, its stability is between that of the other two algorithms. This is caused by a large number of random factors in the process of MGA.
HNOGA Algorithm Performance Simulation Experiment
In this paper, simulation experiments are conducted on the HNOGA algorithm, SGA, niche genetic algorithm based on scavenging mechanism (NGAC)) [34], improved ant colony optimization (IACO) [35] algorithm and greedy algorithm, and the above algorithms are run 100 times. In terms of simulation parameters, the population size M of the HNOGA algorithm is 100, the maximum number of iterations T is 200, the mutation probability p m is 1%, the niche neighborhood search probability p m1 is 80%, the optimal individual neighborhood search probability p m2 is 30%, and the number of elite individuals q is 30. The population size of the micro genetic algorithm is 5. The maximum number of iterations is 10. The perturbation amplitude g pb is 10, and the mutation probability is 1%. SGA uses random league selection, uniform crossover, uniform mutation and elite strategy. The population size M is 100, the maximum number of iterations T is 200, the mutation probability p m is 10%, and the crossover probability p uc is 70%. The content of the elite strategy is to replace the worst individual in the offspring with the best individual in the parent. NGAC uses random league selection, uniform crossover and uniform mutation and adopted the same elite strategy as the HNOGA algorithm. The population size M is 100, the maximum number of iterations T is 200, the mutation probability p m1 is 10%, the crossover probability q is 70%, and the number of elite individuals is 15. In IACO, the number of ants m is 15. Pheromone constant Q is 20, pheromone factor α is 2, pheromone volatile factor ρ is 0.3, and the maximum number of iterations T is 200 times. The greedy algorithm is only used as the control group in this experiment, and its results are directly shown in the figure without reflecting the iterative process. Each algorithm is run 100 times, the average fitness of the optimal individual in each iteration of each experiment is taken, and the simulation curve is drawn as shown in Figure 5.
As can be seen from Figure 5, although the greedy algorithm has the advantage of a simple process, it easily falls into local optima, and the optimization effect is weaker than that of ICAO, NGAC and HNOGA. Due to the complexity of the multi-service irregular spectrum allocation problem, the optimization effect of the standard genetic algorithm is poor and the convergence speed is slow. The result of 200 iterations is only slightly better than that of the greedy algorithm. Compared with the standard genetic algorithm, the optimization effect and convergence speed of the niche genetic algorithm based on the scavenging mechanism is significantly improved due to its better population diversity. Compared with the other two genetic algorithms, the convergence speed of the HNOGA algorithm, which combines orthogonal uniform crossover, niche technology and neighborhood search of the micro genetic algorithm, is greatly improved, and the final average optimization effect has obvious advantages. Based on executing each of the above three genetic algorithms 100 times, the optimal fitness, worst fitness, average fitness, and standard deviation of the optimization results of each algorithm are shown in Table 5 . It can be seen from Table 5 that the optimal fitness, the worst fitness and the average fitness of the HNOGA algorithm in the results of 100 runs are better than those of the other three, and it has the advantage of relatively good stability.
HNOGA Algorithm Priority Function Verification Simulation Experiment
In this paper, the function of the HNOGA algorithm to meet the frequency demand of high-priority devices is experimentally verified. Firstly, the verification index-the non-interference coefficient U i -is proposed, which is defined as Equation (14): As can be seen from the above equation, the non-interference coefficient U i represents the degree to which device i is not interfered with by other devices.
The HNOGA algorithm is repeated for 100 experiments, the non-interference coefficient and average value are calculated according to the above steps, and the red regular bandwidth curve in Figure 6 is drawn. It can be seen that the degree of interference of high-priority devices is lower than that of the low-priority device, and the curve does not have any jitter, showing a strictly monotonically decreasing trend. This shows that the HNOGA algorithm meets the functional requirements of priority to meet the frequency demand of high-priority devices.
Conclusions
In the space-air-ground integrated network, the existing spectrum allocation technology cannot be applied to the multi-service scenario where the frequency devices have irregular working bandwidths and different priorities. This paper proposes a hybrid niche orthogonal genetic algorithm to solve the problem. The algorithm improves population diversity by introducing niche technology based on a clearing mechanism, proposes a new orthogonal uniform crossover operator to improve the search efficiency of the algorithm, and uses the joint optimization of the standard genetic algorithm and micro genetic algorithm to significantly improve the global searchability and local search accuracy of the algorithm. The simulation results show that the method in this paper effectively solves the problem of electromagnetic spectrum allocation of multi-service-type devices under irregular frequency demand and has the function of preferentially meeting the frequency demand of high-priority devices. In addition, compared with various existing improved genetic algorithms, the hybrid niche orthogonal genetic algorithm proposed in this paper has the performance advantages of good optimization effect, fast convergence speed and strong stability.
Conflicts of Interest:
The authors declare no conflict of interest. | 8,474.2 | 2022-11-27T00:00:00.000 | [
"Computer Science"
] |
Sparseness bounds on local operators in holographic $CFT_d$
We use the thermodynamics of anti-de Sitter gravity to derive sparseness bounds on the spectrum of local operators in holographic conformal field theories. The simplest such bound is $\rho(\Delta) \lesssim \exp\left(\frac{2\pi\Delta}{d-1}\right)$ for CFT$_d$. Unlike the case of $d=2$, this bound is strong enough to rule out weakly coupled holographic theories. We generalize the bound to include spins $J_i$ and $U(1)$ charge $Q$, obtaining bounds on $\rho(\Delta, J_i, Q)$ in $d=3$ through $6$. All bounds are saturated by black holes at the Hawking-Page transition and vanish beyond the corresponding BPS bound.
Introduction
There has been a recent surge of interest in precisely characterizing conformal field theories with a weakly coupled Einstein gravity dual, with equations now accompanying folkore from the past. The most quantitative work has focused on conformal field theories in two dimensions, though there has also been progress on higher-dimensional theories. The difficulties brought on by higher dimensions are clear: the constraining infinite-dimensional Virasoro symmetry is absent and modular invariance of the torus partition function does not immediately provide constraints on the space of local operators.
In this paper we will use the familiar thermodynamics of gravity in asymptotically anti-de Sitter spacetimes to provide quantitative sparseness bounds on the spectrum of local operators of holographic conformal field theories. This approach began with [1], which showed that the thermodynamics of gravity in AdS 3 is reproduced if and only if the spectrum of operators with scaling dimension ∆ < c/6 and ∆ ∼ O(c) obeys ρ(∆) exp(2π∆). This methodology was subsequently generalized to supersymmetric theories [2], correlation functions [3], and higher-dimensional theories on tori [4]. The universality of the thermodynamics for holographic CFTs on tori can also be derived from the special center symmetry structure of such theories through the Eguchi-Kawai mechanism [5].
In two dimensions, the low-temperature and high-temperature thermodynamics are related to one another by modular invariance. This is what allows one to capture the entire thermodynamic phase structure by constraining only the low-lying (∆ < c/6) operators. Unfortunately, in higherdimensional theories on S d−1 , there is no obvious high-temperature/low-temperature duality. But there is still a universal feature of the gravitational phase structure that we can aim to reproduce from the CFT: the Hawking-Page phase structure [6], where, as a function of some external chemical potentials, the vacuum-subtracted free energy (or the entropy) jumps from O(1) to O(N k ) for k some positive number. (For notational simplicity we will ignore the possibility of intermediate scalings O(N m ) for 0 < m < k.) More specifically, we will reproduce the fact that the theory is confined (O(1) scaling in the entropy) below the Hawking-Page transition temperature T HP .
To illustrate the basic idea, consider the finite-temperature canonical ensemble with normalization E vac = 0 and a deconfining phase transition at β c ∼ O(1). Then, Since Z(β) = e −β∆ ρ(∆)d∆, the O(1) behavior of log Z may be ruined if the density of states ρ(∆) grows too quickly for states with ∆ O(N k ). More precisely, we have log Z(β > β c ) ∼ O (1) if and only if ρ(∆) e βc∆ for ∆ O(N k ) .
In the worst-case scenario where the bound is saturated for all states, we have Hence, log Z is O(1) for all β > β c + for 1 as long as is not exponentially small in N . While a deconfinement transition is generically expected for large-N adjoint CFTs on compact spaces [7,8], it is the precise temperature at which the transition occurs which gives us mileage. In particular, applying the above argument to the well known Hawking-Page transition at inverse temperature β HP = 2π d−1 gives us a bound on the spectrum of local operators of holographic CFTs: This bound applies to the entire spectrum, but above the transition temperature, bulk thermodynamics tells us that the large-N density of states is given by the degeneracy of the black hole dominating the ensemble, which is generically smaller than our bound (see figure 1). Interestingly, our bound must be saturated at the transition point, since at leading order in N we can write where we are assuming that immediately above the transition we have equivalence of canonical and microcanonical ensembles, i.e. E c ≡ E βc is a well-defined energy level stable to fluctuations. Applied to AdS/CFT, this argument means that our bound will be saturated by the black hole at the Hawking-Page transition. In appendix A, we invert the logic behind this fact to provide a field-theoretic density of states interpretation for the Bekenstein-Hawking entropy.
For the remainder of this paper, we generalize eq. (4) using known classical black hole solutions to bound the density of operators of the dual CFT with given scaling dimension ∆, spins J i and U (1) charge Q. For d = 2 the bounds will reduce to those of [1]. Importantly, these bounds are more constraining in d > 2 than in d = 2, because for d = 2 modular invariance implies that, if a single deconfining phase transition occurs, it must occur at β = 2π independent of coupling. Indeed, free symmetric orbifolds (which are not dual to weakly coupled Einstein gravity theories) have a transition at β = 2π just like AdS 3 gravity, and α perturbation theory around AdS 3 gravity leaves the Hawking-Page temperature unchanged [9]. On the other hand, in higher dimensions the deconfining temperature tends to increase as interactions are turned on. For example, in both ABJM theory and N = 4 super Yang-Mills, it can be checked that β HP (λ = 0) > β HP (λ = ∞) for 't Hooft coupling λ [8,10,11], with further calculations suggesting monotonic behavior between the free and strongly coupled theories [12][13][14]. This means that log Z ∼ O(1) for a smaller range of temperatures as the interaction strength is decreased. By the argument above, this means that weakly coupled CFTs must be less sparse-they must have ρ(∆) e 2π∆/(d−1) somewhere in their spectrum. The fact that strong interactions are necessary to reproduce the precise low-temperature phase structure of AdS gravity in higher dimensions has been translated into a simple bound on the density of local operators. The violation of our bound is a sharp diagnostic of "how much" interactions have to sparsify a spectrum. There is another interesting aspect to these bounds that we will discuss in section 4: they imply an O(1) density of states beyond corresponding BPS/unitarity bounds. For example, taking ∆ < 0 implies log ρ(∆) ∼ O(1), which looks like a coarse unitarity bound. The layout of the rest of the paper is as follows. In section 2, we provide the methodology behind obtaining our bounds more carefully. In section 3, we provide calculational details for deriving our various bounds. Analytic bounds are possible for three parameters, either mass and two spins or mass, one spin and one U (1) charge, but for four or more parameters, we must resort to numerics. Two-parameter analytic results are summarized in table 1. In section 4, we discuss the connection of our bounds to BPS/cosmic censorship bounds. In section 5, we speculate on the connection between the high-lying spectrum or high-temperature thermodynamics and our bounds on the lowlying spectrum. We will begin with an analysis of the Cardy-Verlinde formula, which correctly gives the entropy above the Hawking-Page temperature T HP for holographic CFT d on S d−1 [15]. After discussing the many limitations of this formula, we instead focus on a more robust feature of the high-temperature thermodynamics: the extended range of validity of a high-temperature effective field theory. In appendix A, we provide a field-theoretic density of states interpretation for the Bekenstein-Hawking entropy of black holes at Hawking-Page phase transitions. In appendix B, we provide details for calculations in 4 ≤ d ≤ 6.
Method for obtaining bounds
In this section, we explain more carefully our method for obtaining bounds on the allowed density of states of operators with U (1) charge and spin for holographic theories with a confining phase transition. We consider a grand canonical ensemble at finite temperature β, with m angular velocities Ω i and a single chemical potential for U (1) charge Φ for CFT d : where the integral goes over the spectrum of the theory and we sum over repeated indices in the exponential. Except when otherwise noted, we will always normalize the ground state energy (even for d = 2) to zero. The extension to additional chemical potentials is trivial. A confining phase transition means that log Z[β > β c (Ω i , Φ)] ∼ O(1), i.e. the free energy does not scale with N for temperatures below some critical temperature β −1 c (Ω i , Φ). The chemical potentials Ω i , Φ and β span an (m + 2)-dimensional space, and the confinement-deconfinement phase transition happens on a co-dimension one critical surface β = β c (Ω i , Φ). The O(1) scaling of the free energy requires that the density of states be bounded from above, It is simple to minimize the right-hand-side with respect to the potentials β, Ω i , and Φ to provide the tightest bound. In the case of Ω i = Φ = 0 the minimization gives β = β c for E > 0 and the bound becomes ρ(E) e βcE , while for E < 0 gives β → ∞ and our bound vanishes. This behavior is generic: the minimum of eq. (7) always lies either on the critical surface or at β → ∞ which gives vanishing degeneracy. The set of values for charges which separates the two behaviors corresponds to a unitarity/BPS bound. To see the two behaviors in general, we first impose parity symmetry under Ω i → −Ω i such that the critical surface is an even function of the chemical potentials Ω i and Φ. Since eq. (7) is invariant under In the former case we then minimize along the critical surface, while in the latter case the bound simply vanishes. The minimization along the critical surface is for a given set of charges {E, J i , Q}, and ∇ = (∂/∂Ω 1 , . . . , ∂/∂Ω m , ∂/∂Φ). Until this point, the discussion applies to states with general U (1) charge Q and momenta J i in large-N gauge theories with a confining phase transition. Focusing on local operators in holographic CFTs with a semiclassical Einstein gravitational dual, we restrict to dimensions 2 ≤ d ≤ 6 and the spatial manifold S d−1 . The Hawking-Page temperature in the bulk will serve as the deconfinement temperature in the CFT. To find the Hawking-Page transition, we compare the on-shell action of the relevant black hole solution to that of vacuum AdS. The vacuum AdS solution will have topological identifications and constant gauge field to match the inverse temperature, angular velocities, and chemical potential for U (1) charge of the black hole. When the black hole has charge and spin, the deconfinement temperature will depend on the chemical potential Φ and angular velocities Ω j . Below this temperature, the dual CFT is in a confined phase (dual to the AdS vacuum) and above this temperature the dual CFT is deconfined (dual to a black hole).
We consider the most general black holes in d + 1 dimensions for the cases d = 2 through d = 6 with a single U (1) charge and d 2 spins. These black holes are asymptotic to a (spinning) Einstein static universe (ESU) which, in the Lorentzian case, has topology R × S d−1 . Classical solutions for the generically spinning charged black hole in dimensions d = 5 and d = 6 depend on choice of supergravity truncation and so our results in those cases should be considered in that context. Nevertheless, bounds obtained from these solutions are similar to their lower dimensional Analytic expressions are only applicable when they are real and positive; when they become complex or negative it means the charges admit a set of chemical potentials for which E − Ω i J i − ΦQ < 0 and the minimization procedure lands at β → ∞ instead of the Hawking-Page surface. This leads to an O(1) density of states. Notable in this table is the absence of a bound for operators with U (1) charge in 2d CFTs. Electrically charged static black holes in three dimensions have interesting but somewhat peculiar thermodynamic properties-see [16,17]. Among these properties is the fact that if one wants to include a bulk Maxwell field, the black hole mass is not bounded from below [18]. If one wants to consider only a Chern-Simons term -which is necessary to describe a U (1) current on the boundary -there are new difficulties in finding the dominant saddle. It is unclear how to match asymptotics as any non-zero holonomy of the gauge field remains constant along the radial direction. A holonomy in the spatial direction would lead to a singularity at the origin for the vacuum AdS phase, while a holonomy in the thermal direction would lead to a singularity at the horizon for the black hole phase. If one includes both Maxwell and Chern-Simons terms for the same U (1) gauge field, the spacetimes include closed timelike curves in the asymptotic region [19]. Thus we cannot consistently analyze this situation in Einstein gravity coupled to U (1) Chern-Simons and/or Maxwell gauge fields.
Bounds on operators
In this section, we derive our bounds for electrically charged operators with spin in CFT dimension d = 2 through d = 6. We begin with d = 3 in section 3.1, giving all details of the derivation of the bound. For general d we state our analytic results, without derivation, for single-charge spinless operators in section 3.2, single-spin uncharged operators in section 3.3, double-spin uncharged operators in section 3.4, and single-spin single-charged operators in section 3.5.
In the case with four or more parameters, we do not have an analytic bound but present numerical results in 3.6. Figures for our numerical results will be presented together at the end of this section to emphasize the similarities between dimensions. The bound on the density of states decreases when charge or spin is added, to the point that no states are allowed beyond a curve that exactly coincides with the BPS bound. As we will see, when the parameters satisfy the BPS condition and admit a BPS black hole, our bound is saturated by the entropy of the BPS black hole, This is a special case of the fact that generic black holes at the Hawking-Page transition have an entropy which saturates our bound.
Example
In d = 3, the AdS-Kerr-Newman black hole is the generic electrically charged, spinning black hole with AdS 4 asymptotics. Its thermodynamics were first studied in [20]. In the limit of zero spin, the thermodynamics reproduces [21,22], and in the limit of zero charge reproduces [23][24][25]. The metric may be written where the metric functions and Maxwell field, A, are The mass M , angular momentum J, and electric charge Q-calculated via boundary integrals-are Note that we follow the convention of [26] for the normalization of Killing vectors as the associated conserved charges generate the SO(d, 2) algebra. To find the on-shell Euclidean action, we evaluate The second term is the Gibbons-Hawking-York boundary term and the last term is a local boundary counterterm that regularizes the action [27]. The horizon angular velocity and inverse Hawking temperature of these black holes are .
The appropriate thermodynamic potential for spin, however, is the difference between Ω h and Ω ∞ , the angular velocity of the boundary ESU. One way to find this Ω ∞ is to boost the boundary metric to a static frame through a coordinate change We then obtain, The parameter Φ is chosen so that the gauge potential vanishes on the outer horizon, defined by ∆ r (r + ) = 0. Notably, this is the potential difference between the horizon and the conformal boundary, and serves as a chemical potential for U (1) charged operators in the CFT.
where k = ∂ t + Ω H ∂ φ is the null generator of the horizon. Subtracting the vacuum AdS result from the AdS-Kerr-Newman result gives We can replace {r + , a, q} with {β, Ω, Φ} using eq. (14) and (16). At fixed {β, Ω, Φ}, there are two competing stable phases-a large AdS-Kerr-Newman black hole and vacuum AdS. The bulk undergoes a Hawking-Page phase transition when the two saddle point solutions exchange dominance, in other words when ∆I E = 0. In the limit of zero charge, the Hawking-Page transition occurs at r + = 1.
In the limit of zero angular momentum, the Hawking-Page transition occurs at r + = √ 1 − Φ 2 . For non-zero charge and angular momentum, it is simplest to extremize with respect to Ω and r + . Obtaining the critical values for Ω and r + , we find that whereĴ = J/∆,Q = Q/∆. Note that ifĴ +Q > 1, eq. (19) breaks down and the correct minimization gives an O(1) density of states. This limit corresponds to the BPS bound ∆ = |J| + |Q| for the lightest charged, spinning state. Notably, at ∆ = |J| + |Q|, the upper bound on our density of states exactly matches the degeneracy of the corresponding BPS black hole with those charges, Again we see that the upper bound on the density of states is saturated by the degeneracy of the bulk black hole at the Hawking-Page transition and is greater for all other black holes (see figure 1). For ∆ = |J| + |Q|, in d = 3, the black hole at the phase transition is a BPS black hole.
Charged, spinless operators
To bound the density of states of charged, spinless operators, we examine the thermodynamics of (d + 1)-dimensional AdS-Reissner-Nordström black holes. Using the conventions of [22], the mass, global U (1) charge, U (1) potential, and inverse temperature for this black hole are where ω d−1 is the area of the unit (d−1) sphere, and c = 2(d − 2)/(d − 1). The vacuum subtracted Euclidean action is As before, there are two competing stable phases at fixed Φ, β. The first is the AdS vacuum with m = q = 0 and constant gauge potential and the second is a large black hole, both at inverse temperature β. Solving for ∆I E = 0, it is clear that for r + > √ 1 − c 2 Φ 2 , black holes dominate the grand canonical ensemble while the vacuum dominates below. The corresponding Hawking-Page temperature is Interestingly, for Φ = 1/c, the Hawking-Page temperature 1/β HP vanishes and an extremal black hole dominates the grand canonical ensemble. To find our density of states, we extremize β HP (Φ)(∆−ΦQ) and find the bound for charged operators is for ∆ ≥ |Q|/c. The lower limit on the energies is the BPS bound for these black holes. Supersymmetry appears through considering Einstein-Maxwell as a consistent truncation of some supergravity theory. The fact that there cannot exist states lighter than the BPS bound ∆ > Q/c, can be seen from our bound eq. (23), which vanishes (more precisely, is O(1)) in the BPS limit ∆ = Q/c. Unlike the previous subsection, the bound on the density of states at the BPS limit vanishes. This is consistent with the nakedly singular nature of these BPS states.
Single spin, uncharged operators
For uncharged operators with a single spin, the dual bulk black hole is the (d + 1)-dimensional Kerr black hole, analyzed first in [23] for d > 2. For d = 2, we work with the spinning BTZ black hole [28]. The relevant thermodynamic parameters for these black holes are the uncharged single spin limit of sections 3.
so black holes dominate for r + > 1. The inverse temperature for the Hawking-Page transition is Figure 2: The bound for operators with spins J a and J b in d = 4 (left), d = 5 (middle), d = 6 (right). Curves To find the density of states, we extremize β HP (Ω)(∆ − ΩJ) with respect to β HP and find where againĴ = J/∆. The d = 3 case is obtained by taking the limit. The unitarity bound is ∆ ≥ |J|, which can also be understood as a BPS bound by taking the limit of zero U (1) charge. The result for d = 2 agrees with the HKS bound [1]. It is notable that in this case, the bound from cosmic censorship agrees with the BPS bound, ∆ − c/12 ≥ |J| [23], where we have normalized E vac = −c/12. However, the HKS bound allows states down to ∆ = |J|, which is the saturation point of the unitarity bound ∆ ≥ |J|. This only occurs in d = 2: all higher-dimensional bounds obtained by our method will coincide with BPS bounds. Because of similarities with multiple spin operators derived in the next sections, we also note that the single spin bound may be written as log ρ(∆, J) π∆s (27) where s is the smallest non-negative solution to
Multiple spin and zero charge operators
Analytic expressions are possible for two spins and zero U (1) charge. Here, the bulk black holes are spinning AdS-Myers-Perry black holes in dimension d > 3, whose metrics can be obtained from the zero charge limit of the gauged supergravity solutions [29,30] in d = 4, 5 respectively and from the zero charge, two spin limit of [31] We find that extremizing β HP (Ω a , is equivalent to finding the smallest nonnegative solution to whereĴ i = J i /∆ and our bound is For completeness, we will solve eq. (30) explicitly. First, define In d = 4, the bound is where A 4 = 3x 3 (3x − 4)y 2 + 6((x − 6)x + 6)y 4 + y 6 + x 3 + 3xy 2 − 6y 2 1/3 , In d = 5, the bound is In d = 6, the bound is where A 6 = 9(5x + 6)y 2 − x 3 + 3 −3x 3 (5x + 4)y 2 + 6(5x(5x + 18) + 54)y 4 − 375y 6 One must be careful with these expressions to always take the principal root, which is generally complex, though the bound is always real for |J a | + |J b | ≤ ∆. For instance, in the no spin limit, A 6 → exp(iπ/3) and B 6 → 3. Like in the previous section, there is a unitarity bound |J a | + |J b | = ∆ which can be understood as a BPS bound by taking the limit of zero U (1) charge. It can be shown |J a | + |J b | → ∆, only when |J i |, ∆ → 0 or they both diverge. In the first case, our bound vanishes and is consistent with the bulk, while in the latter case the bound diverges and is saturated by the divergent entropy of the corresponding black hole. We close this section with a remark on the triply spinning case. Though it must be solved numerically, the bound on triply spinning operators can be obtained from the simple expression The smallest non-negative solution to this expression gives our bound,
Single spin and single charge operators
Bounds for single spin and single charge operators exist in d > 2. We already derived the bound for d = 3 in section 3.1. In d = 4, we take the single spin limit of the black hole in [29]. In d = 5 and d = 6, we choose the single spin and single charge black hole from [30] and [31], respectively. It is worth noting that the generically spinning, charged black holes with AdS 6 and AdS 7 asymptotics are not pure Einstein-Maxwell, whose generically spinning solutions are not known in these dimensions, but are rather truncations of minimally gauged supergravity. Their zero-spin limit is not AdS-Reissner-Nordström and so this limit will not agree with section 3.2. Relevant thermodynamic quantities and vacuum subtracted Euclidean actions are listed in appendix B, in the single spin and single charge limit. As in d = 3, it is easiest to find Φ(r + , Ω) at the Hawking-Page transition and then minimize over β(r + , Ω)(∆ − ΩJ − Φ(r + , Ω)Q). In d = 4, we have the odd feature (see section B.1) that b = 0 does not imply Ω b = 0 or J b = 0. However, this choice gives a nice analytic bound which can be written purely in terms of ∆, J a , Q. Defining we have ρ(∆, J a , J b , Q) π∆ Notable in this bound is the BPS limit, J a + J b + √ 3Q = ∆, which does not vanish but, as in d = 3, reproduces the entropy of the corresponding BPS black hole, In d = 5, we get the bound log ρ(∆, J, Q) π∆ 4 Here, the density vanishes in the BPS limit |J| + Q = ∆. Finally, in d = 6, our bound is log ρ(∆, J, Q) 2π∆ 15 Here too, the density vanishes in the BPS limit |J|+Q = ∆. The vanishing at |J|+Q = ∆ in d = 5, 6 is a consequence of the fact that BPS black holes only exist for J a , J b , Q non-vanishing [30,31].
Numerical Results
In the previous sections, we calculated analytic bounds for operators with up to three parameters. To obtain bounds for operators with four or more parameters, we must resort to numerics. With d/2 angular potentials and one chemical potential, the Hawking-Page temperature is a d/2 + 1 dimensional hypersurface. For a given set of charges, {J 1 , J 2 , ..., J d/2 , Q}, we then numerically find the minimum value of where for simplicity we have scaled out an overall factor of ∆ so that all charges fall in a finite range.
In the full ensemble, the BPS bound is ∆ = 1 For energies below this bound, the density of states vanishes at leading order in N . Because the equations we need to solve are algebraic, no sophisticated numerical techniques are necessary. We discretize the thermodynamic potentials and (hatted) parameters which have finite range. Angular potentials are bounded from above by the speed of light of the boundary ESU, Ω i = 1 and the electric potential is bounded from above by cosmic censorship. The spins and electric charges, scaled by the energy, also have finite range, typically {Ĵ i ,Q} ∈ [0, 1] but this depends on the normalization of A µ . The exact limits can be found in the appendix using the BPS bounds. We divide these intervals into equally spaced grids of N = 100 points. For each grid point labeled by the potentials' ( d/2 + 1) coordinates, we used the built-in "NSolve" function in Mathematica to obtain the black hole radius at the Hawking-Page transition giving us the critical surface defined in section 2. Once obtained, we calculate eq. (45) for each grid point in the spins' and charge's ( d/2 + 1) coordinates. Then, for each point {Ĵ i ,Q} we searched for the minimum value of eq. (45) over the potentials, imposing the lower bound of zero. Because eq. (45) is exponentiated for the density of states, the lower bound determines where a single state is allowed-this is the BPS/unitarity bound of the CFT. Beyond this point (or curve), our procedure allows no states.
As checks on the numerics, we verified that our curves did not vary appreciably as a function of the grid sizes and that they agreed with the analytic results in the previous subsections. In figures 5 through 8, we plot the bound on the density of states in each dimension. Notable in these plots is the entropy of BPS black holes, plotted as a gray surface. Our bounds end on this surface, giving the entropy of these black holes, and then vanish, marking the BPS bound of the CFT. Furthermore, as we pointed out in section 3.5, with only one spin and charge, there are no BPS black holes and hence the gray entropy surface vanishes.
BPS, cosmic censorship, and sparseness bounds
In previous sections, we saw that our bound on the density of states vanishes at leading order in N for states that violate the BPS bound in d > 2. This is intriguing since we generically considered nonsupersymmetric (Einstein-Maxwell) theories, without using any embedding into supergravity. The appearance of a coarse BPS condition suggests that bulk thermodynamics knows about the consistent supergravity extension. Its appearance is due to the upper bounds on the chemical potentials in the confined phase of strongly coupled holographic theories. To see this, consider the case of finite temperature and a single angular potential. The confined phase always satisfies Ω ≤ 1, which means minimizing exp (β(E − ΩJ)) in the confined phase will give zero for J > E, since then we can pick Ω = 1 and β → ∞. Had the confined phase admitted Ω > 1, then our bound would rule out states with J > J c where J c < E.
The bulk gravitational theory also has an additional bound -the cosmic censorship (CC) bound, that arises by demanding that there are no naked singularities. In general these two bounds are different: for ∆ BP S the lower bound implied by the BPS bound and ∆ CC the lower bound implied by cosmic censorship, we have ∆ BP S < ∆ CC for fixed U (1) charge or fixed spins, i.e. BPS states violate cosmic censorship. In the case with both U (1) charge and spin, there is a line J(Q) along which ∆ BP S = ∆ CC if there is at least one spin in d = 3, 4 and at least two spins in d = 5, 6 (see figure 9). We find that in the cases where ∆ BP S is strictly smaller than ∆ CC , our bound vanishes at the BPS bound, while in the case where the BPS bound coincides with the CC bound, the maximum of our bound reduces to the entropy of the extremal black hole. Masses between cosmic censorship and BPS must have superextremal bulks but our bound allows for an O(N k ) number of states in this range. In this section, we quickly review the BPS and CC bounds to compare to the sparseness bounds we obtain from the Hawking-Page transition. For singly spinning black holes, as mentioned in previous sections, there is a unitarity bound that can also be understood as a Q → 0 limit of a BPS bound, Thus ∆ BP S = J becomes the lower bound on the allowed energy levels. This energy is also found to be strictly less than the cosmic censorship bound. In the limit ∆ → ∆ BP S , we find that our bound gives vanishing degeneracy of states at leading order in N . This is consistent with the fact that the only uncharged spinning BPS states are superextremal, and hence have O(1) entropy. There are no extremal black holes with only one spin in d ≥ 5, which is easily seen from the emblackening factor of the Kerr metric in Boyer-Lindquist coordinates [23] ∆ r = (r 2 + a 2 )(1 + r 2 ) − 2mr 4−d .
However, there is still a "speed limit," a → 1, required for stable bulk black holes.
For singly charged black holes, the BPS bound is given by At fixed charge, this energy is strictly less than the CC bound. Again we find that as ∆ → ∆ BP S our bound gives vanishing degeneracy of states at leading order in N . For non-spinning black holes in Einstein-Maxwell theory, the BPS bound is only rigorously known in d = 3, 4 where embeddings into supergravity theories have been found. The same qualitative results are true for charged spinning black holes -at fixed charges the BPS energy is less than or equal to the cosmic censorship bound on energy. In the case of single charge in AdS d+1 , single spin single charge or double spin in AdS 6 and AdS 7 , BPS states are always superextremal, and we find that our bound vanishes in the BPS limit. However superextremal states that lie between the BPS bound and the cosmic censorship bound, are nakedly singular and have O(1) entropy, but our bound allows for O(N k ) states.
In the case of single spin single charge in AdS 4 , single or double spin single charge in AdS 5 , double spin single charge in AdS 6 , and double spin or triple spin single charge AdS 7 , an extremal black hole that saturates the BPS bound exists for specific values of {∆, J i , Q}. In such cases, the maxima of our bound reduces to the entropy of the extremal black hole. For generic values of {∆, J i , Q} between the BPS bound and the CC bound, the black hole is superextremal and has O(1) entropy, but our bound still allows for O(N k ) states. These features are shown for the four-dimensional Einstein-Maxwell-AdS theory in figure 9. For fixed Q and ∆, it is clear that there exist states with J CC < J < J BP S which are allowed by our bound but must be superextremal. That such states can be allowed is not surprising considering the stability of AdS black holes. Charged rotating black holes can often be obtained via dimensional reduction of spinning supergravity black holes in higher dimensions (see sections B.2 and B.3). Spinning black holes have superradiant instabilities by which the black hole should decay to the most stable spinning charged states (i.e. BPS). This instability is reflected in the lower dimension because the extremal black hole is not supersymmetric and hence unstable. Our bound allows for a finite number of superextremal states to which the extremal black hole can decay. Recent work relating BPS and cosmic censorship bounds can be found in [32,33].
Comments on the high-lying spectrum
Our bounds imply a range of vacuum dominance (β > β HP ) that matches the phase structure of Einstein gravity in the bulk. It is interesting to ask if the high-temperature phase structure (β < β HP ) can be reproduced without additional assumptions. This is what was done in [1,4] by using modular invariance of the torus partition function. Since we are considering theories on S d−1 , where ordinary modular invariance is absent, we need other tools.
We begin with an analysis of the Cardy-Verlinde formula [15], which was proposed as a higher-dimensional analog to the Cardy formula on S d−1 : R is the radius of S d−1 of the CFT, E = E ext + E subext , and E cas ≡ 2E subext . E ext and E subext are the extensive and subextensive pieces of the thermodynamic energy. This formula reproduces the entropy of AdS-Schwarzschild black holes above the Hawking-Page transition but is known to fail for generic theories [34].
A very important aspect of this formula is that, unlike the ordinary Cardy formula, it is canonical in nature. E cas is in no sense the ground state energy of the theory-as stated above, it is calculated by extracting the subextensive piece from E ≡ E . That E cas cannot be a single energy level is clear by matching to the high-energy scaling S ∼ E (d−1)/d , which shows that E cas ∼ E (d−2)/d at leading order; in particular it has to scale with E. Furthermore, to compute E cas , one has to have knowledge of log Z since E = −∂ β log Z. But this means one already has knowledge of S = (1 − β∂ β ) log Z. So, the Cardy-Verlinde formula should be understood as a repackaging of thermodynamic quantities into a suggestive form. If not for the similarity to the ordinary Cardy formula it would be essentially meaningless. The parameters appearing in the Cardy formula, on the other hand, do not require knowledge of the thermodynamics. To have the Cardy-Verlinde formula reduce to the Cardy formula for d = 2, as is often stated, one has to shift the definition of E cas by the vacuum energy −c/12.
Nevertheless, the fact that the thermodynamic quantities can be repackaged in this way for holographic theories is nontrivial. It is then natural to ask how general it is-does it depend on the field theory manifold? Can one add potentials for electric charge and angular momentum? It turns out the formula fails for a holographic theory on flat slices, like a torus. This is because E cas = 0 for such theories, making the formula meaningless. A constant shift only works for d = 2, since for d > 2 it would give incorrect asymptotic scaling S ∼ √ E. For this case, one has to instead use the higher-dimensional Cardy formula, which can be derived from modular invariance and is true for generic conformal theories [35,36]. On hyperbolic slices, it was shown that the formula fails but can be fixed by defining E sub = Ecas 2k [37], where spherical slices have k = +1 and hyperbolic have k = −1. With this definition, E sub is strictly positive. While no explanation was given for this substitution, we will use a high-temperature effective field theory to explain this result at the end of this section.
The formula fails generally when chemical potentials are added, although it can be fixed by making appropriate modifications in some cases. It has been shown that the entropy of Reissner-Nordström is reproduced by the Cardy-Verlinde formula on substituting E ext for E ext − ΦQ 2 , where Φ and Q are the U (1) potential and the electric charge respectively [38]. While for Kerr-Newman black holes, thermodynamic quantities defined with respect to an asymptotically rotating frame can be shown to satisfy the Cardy-Verlinde formula [39]. However in these modified definitions, E ext loses its meaning as being the extensive part of the energy. For more complicated solutions like multi-charged or multiply-spinning black holes in gauged supergravity models, one can still fix the Cardy-Verlinde formula by making changes to E cas and E ext [37], however these changes are quite complicated in terms of the CFT thermodynamic quantities [40]. Thus, there does not seem to exist a universal modification that works for every case. While it is tempting to think the form of the Cardy-Verlinde formula implies a connection between high-lying and low-lying states, the difficulties outlined above, coupled with the fact that E cas is not a fixed low-lying energy, suggest otherwise.
Two approaches, which we will point out but not pursue, are to investigate the notions of "emergent circles" [41] and "detachable circles" [42]. In this context, the notion of emergent circles says: The quotient is performed on the Hopf fiber for the odd-dimensional sphere represented as a circle fibered over CP n . In this highly lensed limit, there is an emergent modular invariance that appears, since a highly lensed sphere behaves like S 1 × CP n for the purpose of leading-order thermodynamics.
Coupling this with the special pattern of center symmetry breaking of strongly coupled holographic CFTs [5] may give an avenue to relating the theory on S 2n−1 /Z p→∞ back to the theory on S 2n−1 . For n = 3 there is even a nontrivial Hawking-Page phase structure in the bulk with calculable β(p) HP that can be used to provide a bound on the density of states ρ(E) e β(p) HP E on S 3 /Z p , connecting the round sphere p = 1 to the case with an emergent modular invariance p → ∞. The notion of "detachable circles" in this context relates a finite-temperature conformal theory on S d−1 to the theory on H d−1 /Z at some inversely related temperature: By restricting our theory to be gapped at finite temperature (which is the generic situation), we can use the effective theory approach introduced in [43]. This approach allows us to write down the following effective action for the theory dimensionally reduced over the thermal circle: This is to be understood as a perturbative expansion around β → 0. Powers R (n) are to be understood as all possible combinations of contractions of the Riemann tensor, with e.g. different coefficients between R µν R µν and R 2 which are suppressed for simplicity. This effective theory makes clear that the high-temperature theory on a hyperboloid is related to the high-temperature theory on the sphere by sign flips in the terms of the effective theory with odd powers of curvatures. Certain large-N theories may have a sufficiently extended range of validity for this effective theory such that we can relate the theory on H d−1 /Z back to the theory on S d−1 . This effective theory also explains why the Cardy-Verlinde formula works for hyperbolic slices with the definition E sub = Ecas 2k : this is a simple way to achieve the sign flips implied by the effective theory.
Conclusion
In this paper, we derived quantitative sparseness conditions on holographic CFTs with a semiclassical Einstein dual. To arrive at these conditions, we used the fact that there generically exists a Hawking-Page transition between vacuum AdS and a large asymptotically AdS black hole at a particular temperature and set of thermodynamic potentials. Such a phase transition implies a discontinuous jump in the free energy from O(1) to O(N k ) and hence the CFT can only support a finite number of states before it deconfines. The difficulty in satisfying such bounds comes from the fact that interactions tend to sparsify a spectrum, so generic weakly interacting theories have dense spectra which violate our bounds. An interesting aspect of these bounds is that that log ρ = O(1) for masses below the BPS bound. In situations where a bulk BPS black hole exists at the bound, its entropy saturates our bound, which then discontinuously drops to O(1) consistent with the bulk. It is interesting to see the appearance of the BPS bound in the cases with U (1) charge without inputting supersymmetry.
Sparseness assumptions figure prominently into simplifying limits of conformal bootstrap techniques. The usual style of argument is that a sufficiently sparse spectrum allows you to pick up only the contribution of the vacuum in a particular OPE expansion. This was most recently utilized in the bootstrap approach [44] to the "large charge" expansion [45,46]. It would be interesting to explore the connection of our quantitative sparseness bounds to these bootstrap techniques.
A sparse low-lying density of local operators is often invoked as a requirement for a CFT to have a semiclassical Einstein dual, but for d > 2, a precise definition of "sparseness" was lacking. In this work we have provided a quantitative sparseness bound on the allowed density of local operators in the CFT. This bound enforces vacuum dominance of the gravitational path integral at low temperatures. It is a sharp diagnostic for how much interactions have to "sparsify" a spectrum, since it is violated by weakly coupled holographic theories. It would be interesting to connect this sparseness condition to a different sparseness condition, the gap to the higher-spin operators [47], both of which need to be satisfied for a weakly coupled Einstein gravity dual, and both of which are violated for weakly interacting holographic CFTs in d > 2.
A Black hole entropy from deconfining phase transitions As we saw throughout this paper, our bounds on the density of states are saturated by the black hole at the deconfining phase transition. We can invert this logic to produce a derivation of black hole entropy from field-theoretic considerations. Since we would have to input a deconfinement temperature (and more assumptions) in the general case, let us focus on d = 2 where we can get by with minimal assumptions.
Assume a large-c CFT in d = 2 has a single first-order deconfining phase transition. By modular invariance it must occur at β = 2π. We use a normalization consistent with modular invariance, E vac = −c/12. We also know from the modular bootstrap that E β=2π = 0 [48]. By the generic description of first-order phase transitions as an exchange of saddles, we can use the vacuum energy and E β=2π = 0 to deduce that E β=2π− = c/12 up to corrections in . Since ∆ c − S c /β c = O(1) =⇒ S c = β c ∆ c = β c (E c + c/12) at leading order in c, this gives us a prediction for the thermal entropy, where we have deduced E c = c/12 and β c = 2π purely from field-theoretic considerations. Notice that "c" is doing double duty here. For β = 2π − we are in the deconfined phase of a large-c theory, so we can coarse grain to translate into a density of states as in [1]. Altogether we have the formula log ρ(E = c/12) = πc/3 .
This agrees precisely with the bulk, where the ensemble is dominated by a BTZ black hole below β = 2π, and so by continuity the density of states at β = 2π is given by the Bekenstein-Hawking entropy of the BTZ black hole at the Hawking-Page transition. One could also dispense of the assumption that the transition is first order and so described by an exchange of saddles to obtain a formula of the sort log ρ(E) = 2πE where E ≡ E 2π− . Notice that the Cardy formula log ρ(E) = 2π cE/3, which is true for E → ∞, matches onto the formula given above. This is completely expected, since in the case of d = 2 the bound ρ(∆) e 2π∆ implied by the phase transition assumed here can be used to prove the validity of the Cardy formula down to ∆ ∼ c/6 [1]. So this result is weaker, but the different route taken is illuminating and can potentially be applied in other cases where arguments like that of [1] are absent.
B AdS 5 , AdS 6 , and AdS 7 details
In section 3, we exhibited results for bounds on the density of states with d = 4, 5, and 6 dimensional boundaries. The metrics and derivation of thermodynamic quantities, including the Euclidean actions are straightforward and follow the same steps as in d = 2, 3 but the expressions are longer and not directly illuminating. Below, we expound on the steps that lead to the bounds above. In particular, we collect results for AdS 6 and AdS 7 whose derivation is distributed over multiple papers in the literature. First, we discuss the derivation for d = 4. We set G = 1 everywhere below.
B.2 AdS 6
The bounds on the density of states are meant to serve all holographic CFTs in their respective dimensions. However, there are no bottom-up solutions for Einstein-Maxwell gravity in d = 5 and d = 6. This may not be surprising as higher form fields and Chern-Simons terms seem natural in higher dimensions, especially in consistent supergravity trunctations. Instead, one must search for the most generic supergravity solution with AdS asymptotics and fewest bulk fields. The most generic choice we could find in the literature is the black hole in [30]. This comes from a dimensional reduction of massive type IIA supergravity on a hemisphere of S 4 [49]. This supergravity theory should arise as the near horizon limit of a D4-D8 brane configuration and is dual to a d = 5, N = 2 superconformal field theory. The bosonic field content of six dimensional N = 4, SU (2) gauged supergravity is a graviton, a two-form potential, a one-form potential, the gauge potentials of SU (2) Yang-Mills and a scalar. We can truncate to the sector where only one U (1) of the SU(2) is excited. Then, the bosonic Lagrangian is with F (2) = dA (1) . We now set g = 1.
The Gibbs free energy, defined by is equivalent to the background subtracted on-shell Euclidean action divided by −β, ∆I E = −βG.
In the limit of zero charge, this agrees with the generically spinning black holes in six dimensions with no charge [24]. However, it turns out the a = b = 0 solution is not the AdS-Reissner-Nordström black hole, but rather the black hole in [49]. In table 1, for the charged static case, we instead presented the result from [22], where the action is the one calculated in section 3.2.
B.3 AdS 7
The d = 6 case follows [31]. These solutions come from reducing eleven-dimensional supergravity on S 4 leading to seven dimensional N = 4, SO(5) gauged supergravity. Note that this can be thought of as coming from the near horizon limit of a stack of M 5 branes and is dual to the sixdimensional, N = (2, 0) SCFT. For singly charged black holes, we choose to truncate to the U (1) 3 Cartan subgroup. The bosonic fields are a graviton, a self dual 3-form potential, two U (1) gauge fields and two scalars. Turning off one of the scalars in the gauged theory sets the two U (1) fields equal and the Lagrangian is where X = e −φ 1 / √ 10 . The self-duality condition reads For this work, we set g = −1 (this must be negative for BPS states). As in six dimensions, it is more straightforward to calculate the Gibbs free energy. The relevant thermodynamic quantities are Ω i = a i [(1 + r 2 + ) j =i (r 2 + + a 2 j ) + qr 2 + ] + q j =i a j i (r 2 + + a 2 i ) + q(r 2 + + abc) , Φ = q(2m + q)r 2 + i (r 2 + + a 2 i ) + q(r 2 + + abc) , For brevity, we used a i ∈ {a, b, c}. The parameter m is 2m = 1 + r 2 r 2 (r 2 + a 2 )(r 2 + b 2 )(r 2 + c 2 ) + q(2r 2 + a 2 + b 2 + c 2 ) + 2qabc r 2 + q 2 r 2 .
The limit q → 0 agrees with the Myers-Perry-AdS 7 solutions, but like d = 5, the non-spinning limit does not match the Reissner-Nordström result of Chamblin et al. | 11,943.4 | 2017-11-08T00:00:00.000 | [
"Physics"
] |
Microbiomes of Two Pest Fly Species of Pennsylvania Mushroom Houses
Simple Summary Flies inhabit mushrooms or consume them incidentally throughout their lifecycle. In wild conditions, these flies are part of the recycling process and are not considered pests. However, in commercial mushroom houses, these flies can become problematic, as compost pests, through physical damage of mushroom production through consumption, or as vectors of pathogens or nematodes. Here, we describe the bacterial associates of two fly pest species of Pennsylvania mushroom houses. Abstract Mushroom cultivation vastly improves the yield of mushrooms under optimized, controlled conditions, but may be susceptible to opportunistic colonization by pest species that can establish themselves, as well as the pathogens and pests they may transmit. Here, we describe our investigation into the bacterial communities of adult Lycoriella ingenua (Diptera: Sciaridae) and Megaselia halterata (Diptera: Phoridae) collected from button mushroom (Agaricus bisporus) production houses in Pennsylvania. We collected adult flies and sequenced the hypervariable v4 region of the bacterial 16S rRNA using the Illumina MiSeq. The most abundant bacterial genus detected in both species was Wolbachia, but phylogenetic analysis revealed that the infections are from different clades. Future studies include the characterization of Wolbachia infections on fly behavior and biology, comparison of microbial diversity of fly species colonizing wild mushrooms, and other microbiota that may contribute to the success of certain pest fly species.
Introduction
Fungi are integral components of many ecosystems.Many fungi produce sporocarps, macroscopic structures in which sexual spores develop and from which spores are released.Fungi play many roles in nature: predators, parasites, mutualists, and/or recyclers.A given fungal species may be a mycorrhizal companion of nearby plants, food and shelter for developing invertebrates, and may itself be parasitized by microorganisms while also consuming bacteria and benefiting from bacterial breakdown products.Mycetophagous flies in turn may consume fungal material (mycelia or fruiting bodies) and utilize volatiles from fungal pathogens or bacterial breakdown products to select optimal oviposition sites [1].
Wild sporocarp-forming fungi are populated with a rich diversity of flies [2].In the northeastern United States, the fruiting structures of the fungal genus Agaricus are predominantly colonized by members of the families Drosophilidae, Phoridae, and Tipulidae [3].In contrast, only a handful of fly species are considered economic pests of mushroom production worldwide.Fly pests are predominantly from the families Phoridae, Sciaridae, and Cecidae [4,5].In Pennsylvania mushroom farms, the two major pest species are Lycoriella ingenua (Dufour 1839, Family Sciaridae) and Megaselia halterata (Santos Abreu, 1921, Family Phoridae) [6].
Agricultural monocultures are optimized for maximum crop yield but are also susceptible to opportunistic colonization by pest species that can establish themselves, as well as the pathogens and pests they may transmit.Mushroom houses serve as an ideal experimental environment in which to study the dynamics of microbe-fly-cultivated crop interactions when conditions are optimized for mushroom crop production.Here, we describe our investigation into the bacterial communities of adult L. ingenua and Megaselia halterata collected from button mushroom (Agaricus bisporus) production houses in Pennsylvania.
These two pest species have distinct but overlapping biologies.L. ingenua consumes the compost material (and any associated microbes) prior to the addition of mushroom spawn [7].Once the compost is fully colonized by Agaricus bisporus, the populations of L. ingenua decline [8].In contrast, populations of M. halterata thrive on mycelial growth, gradually building up from spring until fall and then declining when mushrooms are harvested and beds replaced [9].We predicted that there would be some overlap in bacterial community composition between the two fly species, but we suspected there would be differences that might be unique to each fly species.We collected specimens of each species, extracted their nucleic acids, and sequenced the bacterial 16S rRNA gene sequences using the Illumina MiSeq sequencing platform.We detected some overlap in bacterial diversity and identified two phylogenetically distinct Wolbachia sequences.
Collection Sites and Sampling
We collected Lycoriella ingenua and Megaselia halterata adults by aspiration from two mushroom production houses in Kennett Square, Chester County, PA, at two times (May and October of 2014; Table 1).Flies were collected in separate vials by species (L.ingenua or M. halterata) and transported alive on ice to University Park, PA, to prevent damage to the nucleic acids.Flies were then placed at −80 • C until processed.
Sample Processing for Bacterial 16S rRNA Sequencing
To describe the bacterial community composition, we extracted genomic DNA from 40 individual flies (Table 1).DNA extractions were conducted by macerating each adult fly in individual 1.5 mL sterile polypropylene centrifuge tubes using sterile polypropylene pellet pestles in tissue lysing (TL) solution from the Omega E.Z.N.A. Tissue DNA kit (SKU#: D3396, Norcross, GA, USA), following the provided manufacturer's protocol for extraction from tissue.Extracted nucleic acid samples were submitted for Illumina MiSeq paired-end sequencing of the v4 region of the bacterial 16S rRNA gene.The sequencing facility used the original Caporaso 515F/806R primers [10] (since the 2016 updated EMP primers were not yet available at the time of sequencing in 2015).Sequences (~290 bp) returned from the facility were demultiplexed with primers and adaptors, and barcodes were removed.
Analysis of Bacterial Communities of Flies
Demultiplexed sequences were quality checked, dereplicated, merged (trimmed to 220 bp), filtered to remove chimeric sequences, aligned, and analyzed in RStudio using Dada2 version 1.33.0 and Phyloseq version 1.48.0 and other data visualization tools in several R packages (microbiome 1.26.0,mia 1.12.0,microViz 0.12.3)[11][12][13][14][15]. Reads were assigned taxonomic identity using the Dada2 taxonomy assigner and Silva (v138) reference database of eubacterial 16S ribosomal RNA sequences [11,16,17].The Amplicon Sequence Variant (ASV) table is provided in the Supplementary Materials.After an initial analysis, we detected taxa that matched mitochondrial sequences or were suspected to be contaminants from other samples sequenced in the same run (see "Post-Illumina sequence confirmation and phylogenetic placement of Wolbachia sequences").Thereafter, we filtered out reads matching "mitochondria", "Rickettsia", and "Rickettsiella" before continuing with diversity analyses.Shannon and inverse Simpson indices were used for measuring richness and evenness.Statistical comparisons within species were performed using the Kruskal-Wallis test.Community dissimilarity (Bray-Curtis index) was evaluated between groups.Principal Coordinate Analyses (PCoA) were performed and plotted to visualize bacterial community structure between groups using Phyloseq.Statistical comparison between groups was performed to run permutational multivariate analyses of variance (PERMANOVA using 999 permutations).The significance level was set to 0.05.
Post-Illumina Sequence Confirmation and Phylogenetic Placement of Wolbachia Sequences
Our samples were sequenced with samples from other arthropod studies (two different mosquito species and a tick species).Because of this, it was important to confirm the presence of the three bacterial genera that were also detected in one or more of those arthropod hosts.We used PCR primers to test for Rickettsia, Rickettsiella, and Wolbachia [18,19].
We did not detect the presence of Rickettsia or Rickettsiella, both of which were taxa in high abundance in tick samples sequenced on the same run.However, we did detect fragments of Wolbachia 16S rRNA using primers W-Specf (5 ′ -CATACCTATTCGAAGGGATAG-3 ′ ) and W-Specr (5 ′ -AGCTTCGAGTGAAACCAATTC-3 ′ ) to amplify a 438 bp fragment [18].Amplicons from four fly samples (two of each fly species) were gel-separated, purified, and submitted for Sanger sequencing.Amplicons were aligned to known GenBank deposited sequences of Wolbachia, trimmed to 330 bp to eliminate indels, and phylogenies estimated using Maximum Likelihood with MEGA X using the best-estimated model of evolution selected by jmodeltest [20,21].
Relative abundance revealed a marked difference between the two taxa.Wolbachia (Class Alpha-Proteobacteria: Family Anaplasmataceae) was the dominant taxon in all individuals of L. ingenua, but it was not always the most abundant taxon in M. halterata "Mh", while those of Lycoriella ingenua (Sciaridae) are denoted by "Li".Top horizontal bar denotes fly species: lighter blue represents specimens of M. halterata; darker blue represents L. ingenua.Heatmap generated using microViz, using the viridis color palette option "rocket" [15].
Relative abundance revealed a marked difference between the two taxa.Wolbachia (Class Alpha-Proteobacteria: Family Anaplasmataceae) was the dominant taxon in all individuals of L. ingenua, but it was not always the most abundant taxon in M. halterata (Figure 2).Gamma-bacteria Klebsiella (Enterobacteriaceae) and Pseudomonas (Pseudomonadaceae) were often more abundant in M. halterata than in L. ingenua (Figure 2).Rare reads with a prevalence of less than 50% and detection below 0.1% were aggregated into "Other".Barplot generated using the R package "microbiome" [13].
Bacterial Diversity
The L. ingenua microbiota was not evenly distributed, and fewer taxa (lower richness) were identified compared to M. halterata.We detected significant differences in diversity between collection times in L. ingenua, but not in M. halterata (Figure 3).While the alpha diversity measures of L. ingenua individuals differed between May and September collections, this was not the case for M. halterata (most M. halterata bacterial taxa clustered together regardless of dates).Rare reads with a prevalence of less than 50% and detection below 0.1% were aggregated into "Other".Barplot generated using the R package "microbiome" [13].
Bacterial Diversity
The L. ingenua microbiota was not evenly distributed, and fewer taxa (lower richness) were identified compared to M. halterata.We detected significant differences in diversity between collection times in L. ingenua, but not in M. halterata (Figure 3).While the alpha diversity measures of L. ingenua individuals differed between May and September collections, this was not the case for M. halterata (most M. halterata bacterial taxa clustered together regardless of dates).
Figure 3. Comparison of alpha diversity by date within each fly species using inverse Simpson and
Shannon indices.There was a significant difference in diversity in L. ingenua samples collected in May versus September.This difference in diversity was not observed in M. halterata between collection dates.Barplot generated using the R package "microbiome" [13].
When we examined the beta diversity, we observed that the diversity measures of the two species were distinct from each other, although there was some clustering between fly species that corresponded to the fall collection (Figure 4).We confirmed that there was a significant interaction between fly species and collection date using a PER-MANOVA analysis (Species p = 0.001; Date p = 0.022; Species*Date p = 0.008).We therefore analyzed the two species separately to confirm the effect of the collection date.In both species, there was a significant effect of collection date (Li, R2 = 0.36726, p = 0.005; Mh, R2 = 0.13511, p = 0.003).There was a significant difference in diversity in L. ingenua samples collected in May versus September.This difference in diversity was not observed in M. halterata between collection dates.Barplot generated using the R package "microbiome" [13].
When we examined the beta diversity, we observed that the diversity measures of the two species were distinct from each other, although there was some clustering between fly species that corresponded to the fall collection (Figure 4).We confirmed that there was a significant interaction between fly species and collection date using a PERMANOVA analysis (Species p = 0.001; Date p = 0.022; Species*Date p = 0.008).We therefore analyzed the two species separately to confirm the effect of the collection date.In both species, there was a significant effect of collection date (Li, R2 = 0.36726, p = 0.005; Mh, R2 = 0.13511, p = 0.003).
Figure 4.
Two-dimensional density plot for both fly species.Data are plotted using a Principal Coordinate Analysis (PCoA) with the Bray-Curtis dissimilarity measure.Samples closer together are more similar in diversity of samples than those that are farther away.Microbial community diversity of L. ingenua and M. halterata specimens clustered separately in spring-collected specimens, but were more diffuse and overlapped between species in the fall-collected specimens.PCoA plot generated using the R package "microbiome" [13].Shapes refer to collection dates; colors represent fly species (Circles = 16May2014, Triangles = 18Sept2024, Blue = "L.ingenua", Yellow = "M.halterata").
Sequence Confirmation and Phylogenetic Placement of Wolbachia Sequences
Fragments of Wolbachia 16S rRNA sequences from four samples (two from each fly species) were amplified, gel-purified, and submitted for Sanger sequencing.We confirmed that the Wolbachia sequences identified in the dataset were not due to sequence contamination and that the isolates from each fly species were from phylogenetically distinct clades (Supergroup E for L. ingenua and Supergroup B for M. halterata) (Figure 5).Coordinate Analysis (PCoA) with the Bray-Curtis dissimilarity measure.Samples closer together are more similar in diversity of samples than those that are farther away.Microbial community diversity of L. ingenua and M. halterata specimens clustered separately in spring-collected specimens, but were more diffuse and overlapped between species in the fall-collected specimens.PCoA plot generated using the R package "microbiome" [13].Shapes refer to collection dates; colors represent fly species (Circles = 16May2014, Triangles = 18Sept2024, Blue = "L.ingenua", Yellow = "M.halterata").
Sequence Confirmation and Phylogenetic Placement of Wolbachia Sequences
Fragments of Wolbachia 16S rRNA sequences from four samples (two from each fly species) were amplified, gel-purified, and submitted for Sanger sequencing.We confirmed that the Wolbachia sequences identified in the dataset were not due to sequence contamination and that the isolates from each fly species were from phylogenetically distinct clades (Supergroup E for L. ingenua and Supergroup B for M. halterata) (Figure 5).
Figure 5.
Maximum Likelihood phylogenetic tree of Wolbachia 16SrRNA sequences from each fly species.Analyses were conducted using MEGA X [21].The evolutionary history was inferred using the Maximum Likelihood method and the Tamura-Nei model.The bootstrap consensus tree was inferred from 1000 replicates.Branches with less than 50% bootstrap support are collapsed.Initial trees are obtained by Neighbor-joining and BioNJ algorithms.Evolutionary rate differences among sites were modeled using gamma distribution with the inclusion of some evolutionary invariable sites (+G, +I).All positions with less than 95% coverage were excluded.The trimmed and aligned fragment length was 330 bp.Anaplasma marginale and Rickettsia bellii were outgroups.Lm07 and Lm15 (stars) = Lycoriella ingenua; Mh06 and Mh17 (circles) = Megaselia halterata.Supergroups E and B represent Wolbachia Supergroups into which the fly Wolbachia sequences clustered.
Discussion
While there have been other microbiome studies in mushroom cultivation settings, they largely focus on the mushrooms and the substrate, on associated fungi, or on viruses [9,[22][23][24][25].In this study, we address the presence of bacteria in two mushroom fly pest species.We observed distinct bacterial community compositions and also observed an effect of collection date, particularly in L. ingenua.
We investigated the two collection dates in order to assess whether microbial communities changed between spring and fall populations.The bacterial community composition is highly dynamic and dependent on the ecology of the mushroom house.After the first flush (=crop) of mushrooms, some taxa (e.g., Proteobacteria) decline, but are replaced in abundance by Actinobacteria and Firmicutes [25].It is therefore conceivable that flies intimately associated with mushroom compost and casing might acquire some of their microbiota, but the extent to which they both harbored the same taxa was not known.
The microbial diversity between the two fly species overlapped somewhat in the fall.Maximum Likelihood phylogenetic tree of Wolbachia 16SrRNA sequences from each fly species.Analyses were conducted using MEGA X [21].The evolutionary history was inferred using the Maximum Likelihood method and the Tamura-Nei model.The bootstrap consensus tree was inferred from 1000 replicates.Branches with less than 50% bootstrap support are collapsed.Initial trees are obtained by Neighbor-joining and BioNJ algorithms.Evolutionary rate differences among sites were modeled using gamma distribution with the inclusion of some evolutionary invariable sites (+G, +I).All positions with less than 95% coverage were excluded.The trimmed and aligned fragment length was 330 bp.Anaplasma marginale and Rickettsia bellii were outgroups.Lm07 and Lm15 (stars) = Lycoriella ingenua; Mh06 and Mh17 (circles) = Megaselia halterata.Supergroups E and B represent Wolbachia Supergroups into which the fly Wolbachia sequences clustered.
Discussion
While there have been other microbiome studies in mushroom cultivation settings, they largely focus on the mushrooms and the substrate, on associated fungi, or on viruses [9,[22][23][24][25].In this study, we address the presence of bacteria in two mushroom fly pest species.We observed distinct bacterial community compositions and also observed an effect of collection date, particularly in L. ingenua.
We investigated the two collection dates in order to assess whether microbial communities changed between spring and fall populations.The bacterial community composition is highly dynamic and dependent on the ecology of the mushroom house.After the first flush (=crop) of mushrooms, some taxa (e.g., Proteobacteria) decline, but are replaced in abundance by Actinobacteria and Firmicutes [25].It is therefore conceivable that flies intimately associated with mushroom compost and casing might acquire some of their microbiota, but the extent to which they both harbored the same taxa was not known.
The microbial diversity between the two fly species overlapped somewhat in the fall.This could be explained by their respective biologies.L. ingenua is a generalist mycophagous insect, readily consuming the mycelium, colonized compost, mushroom primorida, and all parts of the fully developed sporocarps.In contrast, M. halterata is more selective ("oligophagous"), feeding only on actively growing hyphal tips [26].As the two populations increase over the season from spring to fall, they experience increased competition for resources that become depleted over time, and consequently may acquire or share microbes present in the substrate.
Because bacterial read counts were much higher from L. ingenua than from M. halterata (Figure 1), it was important to consider this when interpreting differences in microbial communities between fly species.For instance, Serratia was found in higher absolute abundance in L. ingenua, but this only accounted for proportionally less than 10% of the total taxonomic abundance.The absolute and relative abundances of Klebsiella and Pseudomonas were higher in M. halterata than in L. ingenua; the dominant bacterial taxon detected in both fly species was confirmed to be Wolbachia.
Wolbachia occurred at higher relative (and absolute) abundance in L. ingenua than in M. halterata.It is not unusual to find Wolbachia in fly species.However, phylogenetic analysis suggests that the Wolbachia found in both fly species may have been acquired independently.Wolbachia sequences contained in the fly species were determined to be from different clades.The presence of Wolbachia in both fly species was confirmed (post-Illumina sequencing) because of a concern that the sequences represented contamination from mosquito samples that were also sequenced in the same run.However, while the M. halterata Wolbachia was from a similar clade to Culex pipiens Wolbachia, it was distinct from the Wolbachia sequenced from the mosquito samples.Further, the Wolbachia identified in L. ingenua was from a completely different cluster (within the Supergroup E clade), a clade that has been previously associated primarily with springtails and several mite species [27,28].One predatory mite (Hypoaspis aculeifer), known to be an effective biocontrol agent against both species of flies, was not observed or known to be in these mushroom houses, but even if it had been present, it is not a species known to harbor Wolbachia [7,29].Since the sequencing was conducted on whole flies (flies were too small to dissect for sufficient DNA for sequencing), we cannot exclude the possibility that the Wolbachia detected came from infected springtails or mites that may have been consumed by fly larvae in the mushroom mats.Whether or not the Wolbachia found in L. ingenua existed in the flies as a co-evolved associate or acquired through horizontal transfer through interactions with other organisms in mushroom beds (e.g., springtails or mites) needs further research.
Pseudomonas is a ubiquitous bacterial taxon, and several species of Pseudomonas have been described from mushroom farms.Its presence was therefore not a surprise in our sequence data.While some species of Pseudomonas are important enhancers of mushroom development (metabolizing compost compounds that might otherwise inhibit A. bisporus primordial development), other species (e.g., P. tolaasii and P. reactans) are known pathogens [22,24].Although we detected Pseudomonas in both fly species, we did not isolate or characterize them and cannot ascribe their nature as pathogenic, beneficial, or commensal within the mushroom house microbiome.
Limitations of the Study
Our study had some limitations that could be addressed in future studies.In some flies (e.g., Bactrocera tryoni, the Queensland fruit fly), the microbial communities can shift from immature to adult stages [30].We are unable to speculate as to the effect the bacterial communities have on larval development or adult behavior because our samples were collected in the same year, albeit from different seasons.
We cannot speculate whether the Wolbachia detected in this study caused sex ratio distortion or reproductive effects because we did not separate the males or females, nor did we examine immature life stages.However, since Wolbachia infections can be cleared through antibiotic treatment, we could potentially examine the behavior and interactions of Wol-bachia-free flies with mushroom substrates, other invertebrates, and associated microbiota in future studies.
We do not know the extent to which the microbial communities of flies and mushrooms are shared, how the microbiota influences fly behavior, or the secondary ecological impacts of fly microbiota on parasitoids of mushroom flies, springtails, predatory mites, or nematodes.These are areas that could be explored further in later studies.
Future Studies
Mushroom cultivation began in the 1600s, but structures or caves were not used until the early 1900s [31].Modern mushroom houses were started in the early 1900s but were quickly plagued by sciarid pests [32].While earlier attempts at control included chemical applications, the quick development of resistance necessitated changes in cultivation practices and biocontrol agents as integrated pest management strategies [33,34].Cultivation practices such as compost pasteurization and the use of chemicals and biocontrol agents (e.g., predatory mites, entomopathogenic nematodes and fungi) can help control fly pests, but exclusion is preferred [7,35,36].It should be noted that pasteurization alone may not be sufficient, as adult female sciarids (L.ingenua) are attracted to compost and the volatiles released by pathogenic molds [37].
Cultivated button mushroom farming represents a rich microbial ecosystem under fairly homogeneous environmental conditions.The current study connects one more piece of the multitrophic ecological puzzle, but many questions are still unanswered.For instance, can the interactions and dynamics of mushroom flies with other microbial (e.g., viruses or nematodes) or invertebrate associates be used to facilitate the biocontrol of flies, bacterial pathogens, or fungal pathogens of mushroom houses?
One such study we are currently exploring is the identification of viral communities and the potential for both of these fly species to serve as vectors of mushroom pathogens.In a preliminary viral study of mushroom flies, we identified a putative fungal hypovirus in the spawn (unpublished data).While we did not detect that hypovirus in flies, it is conceivable that (given the polyphagous nature of L. ingenua) the fly might serve as a vector and/or reservoir of viral pathogens of fungi.In a study by Liu et al. (2016), a mycovirus of the plant pathogen Sclerotinia sclerotiorum (named Sclerotinia sclerotiorum hypovirulence-associated DNA virus 1) was shown to infect and replicate in L. ingenua, and to be transovarially transmitted [38].
One of the biological differences between the two fly species is that, while L. ingenua is found throughout the growing season, M. halterata populations build up from spring until fall, and then decline in winter.M. halterata adults leave the mushroom houses in fall to mate and may maintain their populations outside of the houses.However, evidence of phorid presence was not detected in adjacent residential properties [7,9].One future goal is to identify alternative hosts that may support M. halterata or serve as refugia for overwintering.Another is to compare the microbial dynamics between years, since, if M. halterata overwinters outside of mushroom houses, it is likely exposed to different microbial pressures and could bring external pathogens back into the mushroom houses the following year.
While the purpose of the exploration of mushroom fly microbial dynamics was to identify biocontrol options in a cultivated setting, a broader ecological question we could not ignore was the following: What are the factors that dictate which fly species becomes a pest?Wild mycophagous flies are dependent on a resource whose abundance is heavily affected by rainfall and other variables.Thus, resource unpredictability would likely favor polyphagy, not host specialization, in mycophagous flies [1].The diversity of mycophagous fly taxa encountered in wild mushrooms reflects this and represents an arena for resource competition.Wild mushrooms (Agaricus spp. in particular) in the northeastern United States are largely colonized by mycophagous drosophilids (Drosophila and Leucophenga spp.), wood gnats, mushroom flies, and crane flies [1,3].Although a limited food source (e.g., single basidiocarp) might result in inter-and intraspecies competition and subsequent reduction in size, an effectively inexhaustible food source (such as a mushroom house) would likely favor mushroom flies.
In theory, any mycophagous fly in the vicinity should benefit from such an abundance of resources.In other regions of the USA and worldwide, other fly species are also pestiferous: cecids Mycophila speyeri and Heteropeza pygmaea can damage mushroom production, while house and stable flies are nuisance pests of compost heaps [5].Cultivation and control practices have been successful in excluding or eliminating former pests such as mites and springtails in commercial production houses [6,33].One future study we are interested in is an in-depth investigation of the multitrophic dynamics of fly colonization of wild mushrooms in adjacent wooded areas to identify possible explanations for the exclusion or establishment of other fly species in cultivated settings.
What role Wolbachia plays in the lifecycle of either of these fly species is unknown.Further studies would include attempts to cure colonized flies of Wolbachia infections and determine whether/how this affects biology, behavior, or pathogen vector competence.We can also determine the population genetics of the Wolbachia isolates in mushroom houses and in wild mushroom populations.
Figure 3 .
Figure3.Comparison of alpha diversity by date within each fly species using inverse Simpson and Shannon indices.There was a significant difference in diversity in L. ingenua samples collected in May versus September.This difference in diversity was not observed in M. halterata between collection dates.Barplot generated using the R package "microbiome"[13].
Figure 4 .
Figure 4. Two-dimensional density plot for both fly species.Data are plotted using a Principal Coordinate Analysis (PCoA) with the Bray-Curtis dissimilarity measure.Samples closer together are more similar in diversity of samples than those that are farther away.Microbial community diversity of L. ingenua and M. halterata specimens clustered separately in spring-collected specimens, but were more diffuse and overlapped between species in the fall-collected specimens.PCoA plot generated using the R package "microbiome"[13].Shapes refer to collection dates; colors represent fly species (Circles = 16May2014, Triangles = 18Sept2024, Blue = "L.ingenua", Yellow = "M.halterata").
Figure 5 .
Figure5.Maximum Likelihood phylogenetic tree of Wolbachia 16SrRNA sequences from each fly species.Analyses were conducted using MEGA X[21].The evolutionary history was inferred using the Maximum Likelihood method and the Tamura-Nei model.The bootstrap consensus tree was inferred from 1000 replicates.Branches with less than 50% bootstrap support are collapsed.Initial trees are obtained by Neighbor-joining and BioNJ algorithms.Evolutionary rate differences among sites were modeled using gamma distribution with the inclusion of some evolutionary invariable sites (+G, +I).All positions with less than 95% coverage were excluded.The trimmed and aligned fragment length was 330 bp.Anaplasma marginale and Rickettsia bellii were outgroups.Lm07 and Lm15 (stars) = Lycoriella ingenua; Mh06 and Mh17 (circles) = Megaselia halterata.Supergroups E and B represent Wolbachia Supergroups into which the fly Wolbachia sequences clustered. | 6,150.2 | 2024-07-01T00:00:00.000 | [
"Environmental Science",
"Biology"
] |
A New Implementation of Genome Rearrangement Problem
Unsigned reverse genome rearrangement is an important part of bioinformatics research, which is widely used in biological similarity and homology analysis, revealing biological inheritance, variation, and evolution. Branch and bound, simulated annealing, and other algorithms in unsigned reverse genome rearrangement algorithm are rare in practical application because of their huge time and space consumption, and greedy algorithms are mostly used at present. By deeply analyzing the domain of unsigned reverse genome rearrangement algorithm based on greedy strategy (unsigned reverse genome rearrangement algorithm (URGRA) based on greedy strategy), the domain features are modeled, and the URGRA algorithm components are interactively designed according to the production programming method. With the support of the PAR platform, the algorithm component library of the URGRA is formally realized, and the concrete algorithm is generated by assembly, which improves the reliability of the assembly algorithm.
Introduction
With the development of biotechnology, biological information data is growing explosively. At the same time, the improvement of computer computing ability and the development of Internet make it possible to store and process large-scale data. How to use computer technology to extract useful information from these data is imminent. erefore, bioinformatics emerges as the times require. Bioinformatics covers the comprehensive application of biology, computer science, and mathematics.
rough the collection, processing, storage, dissemination, analysis, and interpretation of biological information, the biological significance of a large amount of data is clarified and understood.
In organisms, a chromosome is composed of a gene sequence, and the genome is a collection of chromosomes [1]. In order to determine the similarity or homology between the two organisms and reveal the problems of biological heredity, variation, and evolution, we often arrange and compare the two DNA sequences of two organisms according to certain rules. en, through a series of basic operations such as character editing (insert, delete, and replace), one sequence is transformed into another. e minimum number of editing operations required to complete this conversion is the edit distance of two sequences. At the level of a single gene, genetic sequences evolve by editing these characters, so edit distance is a useful measure of evolutionary distance. However, at the chromosomal level, genetic sequences are mainly evolved by global genome rearrangements. ere are five basic rearrangements, namely, translocation, transposition, duplication, deletion, and reversal. Translocation refers to the partial exchange of two nonhomologous chromosomes in a genome. Translocation refers to the exchange of two contiguous gene subsequences in a chromosome. Replication refers to the replication of a continuous gene subsequence on a chromosome. Deletion refers to the deletion of a continuous gene on a chromosome, and inversion refers to the sequence reversal of a continuous gene on a chromosome. Inversion is the most common of these five forms, especially in organisms with only one chromosome. For example, the only difference between the gene sequences of the two most famous bacteria, Escherichia coli and Salmonella typhimurium, is the inversion of a subsequence of the chromosome sequence [2]. In the fruit fly, genus Drosophila, inversion can reflect the differences between species and within species more frequently than translocation or other processes [3]. In these examples, the importance of inversion shows that the research on the algorithm of genome rearrangement only by inversion (we call it reverse genome rearrangement) is a valuable step to study the evolutionary distance at the chromosome level.
In this paper, we design an abstract generic algorithm component library of unsigned reverse genome rearrangement algorithm (URGRA) based on greedy strategy, which improves the reliability and reusability of algorithms in this field. e second section introduces the genome rearrangement problem and related research and briefly describes the domain modeling technology and formal methods. e third section analyzes the domain of reverse genome rearrangement algorithm domain, establishes the domain feature model of URGRA, identifies the common features and variable features, establishes the relationship between features, and designs algorithm components and component interaction model. In Section 4, we show the process of developing a reverse sorting algorithm based on the First Descending Strip Reversal (FDSR) based on the component library and give the experimental results of the algorithm. Finally, we summarize and prospect the full text.
Reverse Genome Rearrangement
Problem. Given two chromosomes, they are represented by τ and σ, respectively, τ � τ 1 , τ 2 , τ 3 , . . . , τ n , σ � σ 1 , σ 2 , σ 3 , . . . , σ n , where σ i and τ i represent a gene on the chromosome, and let ρ � [i, j] (1 ≤ i ≤ j ≤ n) denote the inversion interval acting on the chromosome. σ·ρ denotes that the subsequence What we want to seek is the minimum value of inversion operation; that is, to find a series of inversion interval ρ 1 , ρ 2 , ρ 3 , . . ., ρ r makes σ·ρ 1 ·ρ 2 ·ρ 3 . . .ρ r � τ and r is the smallest. We call r the reverse distance or inversion distance of σ and τ and record it as d (σ,π) (although it cannot guarantee that the inversion process represents an actual evolutionary sequence, it can give us a lower bound of the number of rearrangements that have occurred and indicate the similarity between two species. erefore, scientists are interested in the minimum number of reversals [4]).
In the extension arrangement, if π i and π i+1 are adjacent numbers (0 ≤ i ≤ n), then π i and π i+1 are adjacency; otherwise, π i and π i+1 are breakpoints. In this paper, the interval between two adjacent breakpoints in π is defined as a strip, that is, the largest fragment without a breakpoint. e further differentiation of strip can be divided into ascending strip and descending strip. e strip with only one element can be defined as either an ascending strip or descending strip, but it is conventionally defined as a descending strip (0 element and n + 1 element are always defined as ascending strip).
It has been proved that if the gene has direction, the reverse genome rearrangement problem is polynomial time solvable, and a very effective algorithm has been found. However, if the gene is unsigned, the reverse genome rearrangement problem is NP hard [5,6]. In this paper, we mainly study the problem of unsigned reverse genome rearrangement. If there is no special explanation below, genome rearrangement will mean the unsigned reverse genome rearrangement. In 1995, David Sankoff and Kececioglu began to study the inversion distance problem and discussed the greedy approximation algorithm based on breakpoint elimination and an accurate algorithm of the branch and bound algorithm for the first time [7]. en, Afna and Pevzner designed a 1.75-fold approximation algorithm for genome inversion sequencing without symbol, with a time complexity of O (n 2 ). In 1998, Christie gave a polynomial approximation algorithm with an approximation degree of 1.5 and a time complexity of O (n 4 ) [8]. Berman et al. designed a polynomial approximation algorithm with an approximation degree of 1.375 by using the signed reverse sorting algorithm and the cycle decomposition approximation algorithm [9]. In the twentieth century, Professor Mo Zhongxi's team of Wuhan University designed a greedy algorithm based on breakpoint graph, and Professor Zhu Daming of Shandong University designed a greedy algorithm based on the First Descending Strip Reversal (FDSR) [10]. Many researchers have developed various algorithms since then.
At present, the diversity and complexity of the reverse genome rearrangement algorithm make it impossible for many users to choose the algorithm suitable for different DNA sequence characteristics, which leads to unnecessary errors in the research process. On the other hand, it is difficult to understand the structure of the reverse genome rearrangement algorithm, which will affect the correct use of the algorithm in the actual situation. Because of the low abstraction of the reverse genome rearrangement algorithm, the reusability and reliability of the reverse genome rearrangement software are affected. erefore, it is necessary to study the reverse genome algorithm at the domain level. e research on the algorithm family is helpful to extract the commonness and variability of the algorithm and provide support for the formal development of the reverse genome rearrangement algorithm.
Feature Modeling Technology.
Domain modeling needs to determine key concepts and feature modeling of key concepts [11]. Feature engineering [12] holds that feature is a first-order entity that runs through the software life cycle, spans the problem space and solution space, and reduces the difference in demand awareness between users and software developers through features. In FODA [13], features are regarded as aspects, qualities, characteristics, and so forth that are visible, obvious, or characteristic to users in software systems. Features are domain knowledge accumulated by users' experts after long-term use or research in a domain.
Feature modeling is the activity of modeling the commonality and variability of features and the relationship between them. Literature [14] puts forward a feature-oriented domain modeling method (FODM), which considers the characteristics of service, function, and behavior of the domain and obtains the feature model through service analysis activities, function analysis activities, behavior analysis activities, domain terminology analysis activities, common variability analysis activities, interactive process analysis activities, and quality demand analysis activities.
Formal Method PAR.
e Par [15][16][17][18][19] (partition and recur) method is a unified algorithm programming method based on partition and recur. It makes full use of mature programming technologies such as data abstraction, function abstraction, software reuse, and class genus, to realize the formal development of complex algorithm problems.
at is, through a series of formal transformations to the problem specification, a fast and correct algorithm is obtained, and then an executable language program is obtained through a series of formal equivalent transformations or software conversion tools. It consists of the following elements: SNL (structured natural language), Radl (recurrence based algorithm design language), APLA (Abstract Programming Language), a set of model transformation rules, and a set of automatic conversion tools and executable programs among requirements model, algorithm model, and abstract program model. e APLA language fully embodies modern programming ideas such as function abstraction and data abstraction, which make it very suitable for describing abstract algorithm programs. In APLA, all the combined data types and their related operations adopt the generic mechanism. e generics are mainly divided into two categories: ① type parameterization, the introduction of the keyword sometype, which can be used to define type variables, and the basic type of the combined data type can be directly described in the form of parameters in the type declaration and ② subroutine parameterization: the func and proc keywords are provided in APLA to declare the process parameters and function parameters. When declaring these parameters, you only need to define the operation contains several variables and the type of each variable, and it can be instantiated by taking a subroutine implementation as argument. Apla is not only the target language of Radl-Apla program converter but also the source language of Apla to Ada, Java, C++, Python, and other executable language program converters, which is beneficial to the formal development of reusable components.
Domain Analysis.
Here, we deeply analyze the core ideas of three typical greedy algorithms. ① Input a gene permutation π to judge the validity of gene arrangement. If it is not legal, the output is wrong. ② e extended π is an extension arrangement, π � (0, π 1 , π 2 , π 3 , . . . , π n,n+1 ), and the inversion distance d is set to 0. ③ Judge whether π is the identity permutation; if not, enter into ④. If it is, enter ⑨. ④ If the first strip other than strip 0 is ascending, it is reversed to a descending strip. ⑤ Find the position i of the smallest term (i.e., the last term) in the first descending strip and the position j of its inverse adjacent element. ⑥ If the inverse adjacent element is the first element of a descending strip, the strip where the inverse adjacent element is located is first flipped. ⑦ If the position of j is on the right side of i, the inversion interval is ρ � [i + 1, j] and if the position of j is on the left side of i, the inversion interval is ρ � [i, j + 1]. ⑧ Let π � πρ, increase the inversion distance d by 1 and return to ③. ⑨ Output reverse distance d.
Algorithm Based on Breakpoint Graph.
e algorithm based on breakpoint graph is a greedy algorithm with time complexity O (max{b3 (π), nb (π)}) and space complexity O (n), which is developed by Mo Zhongxi's team of Wuhan University. Its main process is as follows: ① Input the extension permutation π and judge whether π is legal; if not, end. en, judge whether it is the identity permutation. If it is, end, otherwise, determine the inverse permutation π −1 of π and the breakpoint set B π of π, π′s breakpoint location table T. Let i � 1.
② Based on the breakpoint location table, find out all the reverse intervals in π that can eliminate two breakpoints and one breakpoint, and store them in S2 and S1, respectively.
Algorithm Based on Eliminating Breakpoints.
e algorithm based on eliminating breakpoints is a greedy algorithm with an approximate degree of 2 designed by David Sankoff and Kececioglu. e main process can be summarized as follows: ① Input the extension permutation π, judge the correctness of permutation, and calculate π-1, array down, and array up (array down and up can judge whether there are ascending and descending strip in a certain interval within O(1)). ② Find an inversion interval that can eliminate two breakpoints. ③ If there is no reverse interval ρ that can eliminate two breakpoints, find a reverse interval ρ that can eliminate one breakpoint and make the new arrangement after πρ have a descending strip. ④ If the above does not exist, then find an inversion interval that can eliminate a breakpoint. ⑤ If the above does not exist, find the inversion interval [i, πi−1], if πi≠i (i is the minimum position of subscript of element πi≠i). e above five steps until π are the identity permutation. According to the above analysis, we can use a unified flow chart to express the idea of each algorithm, as shown in Figure 1.
Domain Modeling.
In the following, we will use the feature modeling method proposed by academician Mei Hong's team to conduct feature modeling on the URGRA domain and construct the feature model based on the characteristics of service, function, and behavior in the URGRA domain. e reverse manipulation service (rever-se_mani) is the core service in this domain, and sequence validity check (seq_check), gene permutation storage table manipulation (perm_store_mani), greedy algorithm mode option (greedy_op), auxiliary permutation storage table manipulation (auxiliary_permu_mani), judging whether it is identity arrangement (is_sorted), and output are the main functions in this field. Where seq_check, perm_store_mani, greedy_op, output are required functions, auxiliar-y_permu_mani, is_sorted are optional functions, for greedy_op, Breakpoint_diagram_op, FDSR_op, and break-point_op are its behavior characteristics. For output, out-put_mode is its significant behavior characteristic, and there are three main behavior characteristics: inversion process output (procedure_op), inversion distance output (dis-tance_op), and inversion interval output (interval_op). For auxiliary_permu_mani, break_store, π −1 and break_-pos_store, and array of π −1 , up and down, are behavior characteristics. Based on the above analysis, a feature model for this domain is constructed, as shown in Figure 2.
Different features in that feature model realize a complete domain feature model through interaction, and the interaction between the features in the feature model needs to be reflected by the constraints and dependency between the included features. erefore, aiming at the feature model established above, we design the feature interaction model in the URGRA domain.
rough the establishment of the URGRA feature model, it is analyzed that the algorithm mainly includes three characteristics of the change process: permuta-tion_mani, greedy_op, and output. In addition, the input of the algorithm in this field is gene sequence, and the legitimacy of sequence information needs to be checked before algorithm execution. So, the major artifacts in this domain are the seq_check artifact, the perm_store_mani artifact, the greedy_op artifact, and the output artifact. Other features in the feature model are used as auxiliary components, and the interaction model of components is established according to the dependency between components, as shown in Figure 3.
Wherein, the nodes connected by solid lines represent the basic features that must be included in the URGRA domain, and the direction represented by the arrow represents the execution priority of the four features from high to low. e dotted arrows represent the associated operations required during algorithm assembly, such as the use of auxiliary storage table operations for greedy mode selection. e dotted line indicates the interaction between two features during the execution of the algorithm; for example, when using the inversion output feature, when selecting the permutation process, the distance, or the inversion interval output, the gene permutation storage table operation is required.
Type and Algorithm Component Design.
Here, we further analyze the abovementioned interaction design model of the URGRA domain feature model and algorithm component and encapsulate them into two abstract data type (ADT) components and a reverse rearrangement algorithm component. By virtue of the high abstraction of the Apla program, good support for ADT, and easy formal derivation and correctness verification, we carry out the formal design and implementation of the URGRA model based on Apla code. is generic ADT name is perm_store, which contains a type parameter elem, which can accept either integer or character types. type perm_store � private is the storage space specification, which specifies that the storage space used by this self-defined ADT is private. init(var p: perm_store; permutation: array[0 ... n, elem]) is used to dynamically allocate storage space and initialize it. check (p: perm_store; permutation: array[0...n, elem]) is to verify whether the gene sequence is correct. isSorted(p:perm_store; permutation:array[0...n,elem]) is to determine whether the permutation is the identity arrangement. setValue(p: perm_store; i:integer; permutation:array[0...n,elem]) and Journal of Healthcare Engineering getValue(p:perm_store; permutation:array[0...n,elem]; i:integer) function is to set the element value and get the element value. output(p:perm_store; permutation:array[0...n,elem] � NULL; distance:integer) indicates the inversion distance, inversion process, and inversion interval of the output permutation, and only the inversion distance is output by default. reversal(p:perm_store; permutation:array [0...n,elem]; i:integer; j:integer; aux:auxiliary_permu) indicates the inversion of the permutation. distance_mani(p: perm_store; distance: integer) indicates the operation to reverse the distance. e operation of this self-defined ADT type specified in Apla should pass this self-defined ADT type as an argument to a function or procedure as an operation object, so the above operations have a variable p of type perm_store. function get_value(a:auxiliary_permu; i:integer): elem; enddef e ADT contains a procedure generic parameter someproc initialization_ auxiliary(sometype:elem) and an integer parameter n, so that the generic program can support instantiating different greedy algorithm modes. type per-mission_mani � private is a storage space description, which is used to describe that the storage space used by the selfdefined ADT is private. procedure set_value(a:auxiliar-y_permu; i:integer) and function get_value(a:auxiliar-y_permu; i:integer) are to set and obtain element values.
Results and Discussion
e computer is configured with AMD A10-7300 Radeon R6,10 Compute Cores 4C+6 G 1.90 GHz, 12 GB memory, and Window 7 operating system.
We used real data to carry out the inversion test. Both human and mouse chromosomes have the same gene fragments, totaling 193 genes. ese genes are described as follows [20]: Read the mouse gene arrangement from sourcedata.txt, then output the inversion result to targedata.txt, as shown in Figure 4. e running time is shown in Figure 5. e FDSR algorithm developed by the formal PAR method runs for 3 ms.
With the formal method PAR, we first accurately describe the functional specification of reverse genome rearrangement in Radl language and then develop loop invariants based on the new strategy of developing loop invariants. en, we develop an Apla algorithm program based on the obtained algorithm specification and loop invariants, thus formally implementing perm_store, auxiliary_permu type components, and reverse rearrangement algorithm component. Finally, we use these three components and assemble the FDSR algorithm with the support of the PAR platform. Compared with some existing algorithms, our formal developed algorithm ensures the reliability and robustness of the algorithm program and improves the assembly flexibility of the assembly algorithm by means of component assembly, which is convenient for researchers to maintain and optimize.
Conclusions
Reverse genome rearrangement is a hot topic in bioinformatics research, and its implementation algorithm has been widely studied. Because of the flexibility of its algorithm design strategy, this kind of algorithm presents more diversity and complexity. In this paper, the generative programming technology is used to deeply analyze the field of reverse genome rearrangement algorithm based on greedy strategy, find out the common features and variable features, design highly abstract program components based on Apla language by using formal method PAR, and generate FDSR algorithm by automatically assembling components supported by PAR platform, thus improving the reliability and reusability of algorithm components and reducing the development cost. Our team has formally implemented the pairwise sequence alignment algorithm component library and the multiple sequence alignment algorithm component library [21,22]. In the next step, the insertion, deletion, and replacement of single characters in sequence alignment will be taken into account, so as to develop corresponding components and expand the scope of assemblable algorithms, to better analyze biological similarity and homology.
Data Availability
e data used to support the findings of this study are available in [20].
Conflicts of Interest
e authors declare that there are no conflicts of interest regarding the publication of this paper. Journal of Healthcare Engineering | 5,070.8 | 2021-01-23T00:00:00.000 | [
"Computer Science"
] |
3D Fractals as SERS Active Platforms: Preparation and Evaluation for Gas Phase Detection of G-Nerve Agents
One of the main limitations of the technique surface-enhanced Raman scattering (SERS) for chemical detection relies on the homogeneity, reproducibility and reusability of the substrates. In this work, SERS active platforms based on 3D-fractal microstructures is developed by combining corner lithography and anisotropic wet etching of silicon, to extend the SERS-active area into 3D, with electrostatically driven Au@citrate nanoparticles (NPs) assembly, to ensure homogeneous coating of SERS active NPs over the entire microstructured platforms. Strong SERS intensities are achieved using 3D-fractal structures compared to 2D-planar structures; leading to SERS enhancement factors for R6G superior than those merely predicted by the enlarged area effect. The SERS performance of Au monolayer-over-mirror configuration is demonstrated for the label-free real-time gas phase detection of 1.2 ppmV of dimethyl methylphosphonate (DMMP), a common surrogate of G-nerve agents. Thanks to the hot spot accumulation on the corners and tips of the 3D-fractal microstructures, the main vibrational modes of DMMP are clearly identified underlying the spectral selectivity of the SERS technique. The Raman acquisition conditions for SERS detection in gas phase have to be carefully chosen to avoid photo-thermal effects on the irradiated area.
Introduction
Surface-enhanced Raman scattering (SERS) spectroscopy is based on the enormous enhancement of Raman scattering of molecules adsorbed on suitable metallic nanostructures. The amplification of signals in SERS only occurs in very close vicinity (ca. < 10 nm) to the metal substrate and relies on the electromagnetic interaction of light with metals. There are two widely accepted mechanisms for SERS, i.e., the chemical mechanism (CM) and the electromagnetic mechanism (EM). CM is based on a charge transfer between the molecule and the substrate. Because of this transfer, the polarizability of the molecule increases and consequently, the Raman cross-section of the molecule. This enhancement is usually 10-10 2 . However, the EM is based on the enhancement of the local electromagnetic field that Micromachines 2018, 9, 60; doi:10.3390/mi9020060 www.mdpi.com/journal/micromachines results in a significant increase in Raman signal, proportional to |E| 4 . This increase can become 10 8 or more. The interaction of the light with the metal substrate produces large amplifications of the laser field through excitations generally known as plasmon resonances [1]; either localized surface plasmon resonances (LSPRs) for 0D and 1D nanostructures or propagating surface plasmon polaritons (SPP) for 2D tunable nanostructures [2]. Nowadays, Surface Enhanced Raman Scattering (SERS) outstands as one of the leading techniques for label-free ultrasensitive vibrational fingerprinting of a variety of molecular compounds [3][4][5][6][7]. In a recent review paper on explosives and chemical threats detection [8], SERS has been identified as key technique because it combines several attractive features such as ultrasensitivity, high speed, simple sample manipulation, comparatively low cost, multiplexing detection capability (narrow bands of vibrational Raman scattering) and portability. Despite its exceptional advantages, additional efforts on the fabrication of SERS substrates are still required to circumvent the following limitations: deficient target-substrate contact, poor homogeneity and reproducibility of the signal within a substrate and limited re-usability of the substrates. In particular, the degree of control and reproducibility of commercial as well as research-based SERS substrates are still questionable and significant improvements in terms of performance, process standardization and sample-to-sample reproducibility need to be done.
The development of robust, reproducible and cost-effective SERS substrates containing a large number of hot spots that can give rise to high enhancement factors; and capable to co-localize the targeted molecules in the hot spots is therefore a key requirement towards the widespread use of SERS as practical analytical technique [9]. Most of the reported SERS substrates take advantage of the hot spots limited to a single cartesian plane. Thus, the scattered volume calculated from the laser spot and the laser-probe interaction depth is widely infra-utilized. Accordingly, 3D SERS active substrates with considerable extension in the z-direction are becoming an active focus of research [9][10][11][12][13][14]. Hsu et al. presented a Sierspinki carpet fractal structure by a solid-state electrochemical patterning technique on 40 nm thick Ag film [15] as 3D platform for SERS EM field enhancement. The fractal structure includes features of sizes ranging from 150 nm to 4 µm. The structures were protected with an Al 2 O 3 layer of l.6 nm to isolate the CM effect. The highest electromagnetic enhancement was observed for the size-range 150/750 nm, but also the intensity of the photon counts increased for the 1.2 µm and even for 4 µm structures, by a factor of 100 and 16, respectively. The EM field supported by the fractals further extends spatially compared to continuous 2D silver surfaces.
The combination of self-assembly of NPs with top-down fabrication of periodic surfaces has demonstrated clearly benefits in terms of large-scale and highly reproducible preparation [8][9][10]. Theoretically, the number of hot spots in the 3D nanostructures is greater than that of the 2D counterparts. Furthermore, the extension of a SERS substrate from 2D to 3D leads to about a larger overall surface area, and this promotes adsorption and detection of more target molecules. Finally, the tolerance in focus (mis)alignment along the z-direction could lead to more reproducible SERS signals.
This work explores the use of 3D-fractals with metallic coatings as SERS active substrates for label free detection of G-nerve agents in gas phase. It is well known that Chemical Warfare Agents (CWAs) are poor Raman scatterers with cross-sections in the range of 10 −29 cm 2 sr −1 molecule −1 . This characteristic feature precludes any possibility of analyte detection at low concentration levels without special enhancement processes. Available detection and identification systems for airborne chemical threats remain sorely lacking in terms of the sensitivity and selectivity achievable in the short response time that a real incident would require. CWAs are especially insidious in this respect, given the low concentrations needed to cause potentially lethal effects (for instance, with Sarin gas the value stands at 0.064 ppm for 10 mins) [16]. Indeed, field experiences in simulated scenarios reveal numerous false alarms and error rates so high that it becomes nearly impossible to identify in real time unknown or not explicitly searched-for agents.
The goals of this paper are to characterize the Raman signal intensity enhancements of these 3D-fractal structures as a function of generation of bifurcations (denoted as 1G and 3G fractals) and their performance is compared with non-patterned colloidal gold films. Moreover, the applicability of 3D-fractal microstructures for the spectral identification of Sarin surrogate, i.e., dimethyl methylphosphonate, in gas phase at ppm concentration level is demonstrated.
3D-Fractal Fabrication
The method of engineering of a 3D-fractal structure is based on a combination of anisotropic etching of silicon and corner lithography [17,18]. Basically, a silicon (Si) wafer with thermally grown SiO 2 was patterned in buffered hydrofluoric acid (BHF) using a resist mask with a regular pattern of holes (25 µm diameter and 25 µm periodicity). The silicon which was unprotected, was anisotropically etched in potassium hydroxide (KOH) in order to create inverted pyramidal-shaped pits. The remaining oxide mask was stripped. Next, the wafer was uniformly coated with 160 nm of low-pressure chemical vapor deposited silicon nitride (Si 3 N x ) (see Figure 1A). The next step, called corner lithography, was used to isotropically etch the silicon nitride in hot phosphoric acid (H 3 PO 4 ) (see Figure 1B) and leaves only a dot in the corner of the inverted pyramid (see Figure 1C). The following step was a LOCOS process (LOCal Oxidation of Silicon). In this stage, bare silicon was locally oxidized at 1100 • C (45 mins, yielding 77 nm SiO 2 ) using the silicon nitride dots as mask. The silicon nitride in the corner of each inverted pyramid was stripped with H 3 PO 4 . Next, the unprotected silicon in the pyramidal apex was etched anisotropically using tetra methyl ammonium hydroxide (TMAH; 125 mins) which formed a single octahedral shape feature at the vertex of the pyramid (see Figure 1D), denoted as 1 G (1st generation 3D-fractal). The first level of processing was finished by stripping of SiO 2 and depositing around 88 nm of nitride. The entire process was repeated in order to create the second (2G; see Figure 1F), and third generation (3G; see Figure 1H) of fractal microstructures. To facilitate handling of the hollow 3D fractal surfaces, anodic bonding of the processed silicon wafer to a Mempax glass wafer (500 µm thick) was performed, followed by dissolvation (back-etching) of the silicon. Thus, all 1G and 3G silicon nitride fractals are positioned on a glass substrate. their performance is compared with non-patterned colloidal gold films. Moreover, the applicability of 3D-fractal microstructures for the spectral identification of Sarin surrogate, i.e., dimethyl methylphosphonate, in gas phase at ppm concentration level is demonstrated.
3D-Fractal Fabrication
The method of engineering of a 3D-fractal structure is based on a combination of anisotropic etching of silicon and corner lithography [17,18]. Basically, a silicon (Si) wafer with thermally grown SiO2 was patterned in buffered hydrofluoric acid (BHF) using a resist mask with a regular pattern of holes (25 μm diameter and 25 μm periodicity). The silicon which was unprotected, was anisotropically etched in potassium hydroxide (KOH) in order to create inverted pyramidal-shaped pits. The remaining oxide mask was stripped. Next, the wafer was uniformly coated with 160 nm of low-pressure chemical vapor deposited silicon nitride (Si3Nx) (see Figure 1A). The next step, called corner lithography, was used to isotropically etch the silicon nitride in hot phosphoric acid (H3PO4) (see Figure 1B) and leaves only a dot in the corner of the inverted pyramid (see Figure 1C). The following step was a LOCOS process (LOCal Oxidation of Silicon). In this stage, bare silicon was locally oxidized at 1100 °C (45 mins, yielding 77 nm SiO2) using the silicon nitride dots as mask. The silicon nitride in the corner of each inverted pyramid was stripped with H3PO4. Next, the unprotected silicon in the pyramidal apex was etched anisotropically using tetra methyl ammonium hydroxide (TMAH; 125 mins) which formed a single octahedral shape feature at the vertex of the pyramid (see Figure 1D), denoted as 1 G (1st generation 3D-fractal). The first level of processing was finished by stripping of SiO2 and depositing around 88 nm of nitride. The entire process was repeated in order to create the second (2G; see Figure 1F), and third generation (3G; see Figure 1H) of fractal microstructures. To facilitate handling of the hollow 3D fractal surfaces, anodic bonding of the processed silicon wafer to a Mempax glass wafer (500 μm thick) was performed, followed by dissolvation (back-etching) of the silicon. Thus, all 1G and 3G silicon nitride fractals are positioned on a glass substrate.
Preparation of 3D-Fractals Active SERS Substrates
The hierarchical 3D-fractal platforms are fabricated by combining corner lithography and anisotropic Si-etching with electrostatically driven Au NPs assembly. The top-down fabrication sequence creates microstructured platforms required to extend the SERS-active area into 3D, and the self-assembly of Au NPs ensures homogeneous coating of SERS active Au NPs over the entire microstructured platforms (see Figure 2). Furthermore the fractals could extend the EM field enhancement.
Preparation of 3D-Fractals Active SERS Substrates
The hierarchical 3D-fractal platforms are fabricated by combining corner lithography and anisotropic Si-etching with electrostatically driven Au NPs assembly. The top-down fabrication sequence creates microstructured platforms required to extend the SERS-active area into 3D, and the self-assembly of Au NPs ensures homogeneous coating of SERS active Au NPs over the entire microstructured platforms (see Figure 2). Furthermore the fractals could extend the EM field enhancement. Spherical gold nanoparticles 22 nm in size, Au@citrate NP, were synthesized via a modified version of the Turkevich-Frens method [19]. Briefly, 50 mL of aqueous solution (1.1 mM) of HAuCl4 (50% Au basis) was heated to 70 °C under stirring, and then 5 mL of preheated sodium citrate solution (3.8 mM) was added. The solution was kept at 70 °C until a red-wine colour appeared, circa 10 mins. Then, the liquid was allowed to cool to room temperature. The synthesis experiments have been performed by the platform of Production of Biomaterials and Nanoparticles of the NANBIOSIS ICTS, more specifically by the Nanoparticle Synthesis Unit of the CIBER in BioEngineering, Biomaterials & Nanomedicine (CIBER-BBN).
Following the fabrication of 3D fractals, they were metallized with silver (99.99% silver pellets from Kurt J. Lesker Company, Jefferson Hills, PA, USA) via electron beam evaporation (Edwards auto-500, 3•10 −7 mbar, 32 mA, 5.3 KV). In a second step, gold nanoparticles (Au@citrate NPs) were assembled on the substrates by electrostatic interactions to generate a 3D monolayer-over-mirror configuration. For such purposes, the SERS substrates were incubated in a poly(diallyldimethilammonium) chloride aqueous solution (PDDA), 0.2% wt for 4 h; followed by rinsing with deionized water. Afterwards, the substrates were immersed in Au@citrate nanoparticles solution (0.19 mg/mL) for 16 h at 4 °C. Then the SERS substrates were rinsed with deionized water and dried at room temperature. Table 1 shows the main characteristics of the SERS active substrates studied in this work, where reference samples used for comparison purposes are also included. Spherical gold nanoparticles 22 nm in size, Au@citrate NP, were synthesized via a modified version of the Turkevich-Frens method [19]. Briefly, 50 mL of aqueous solution (1.1 mM) of HAuCl 4 (50% Au basis) was heated to 70 • C under stirring, and then 5 mL of preheated sodium citrate solution (3.8 mM) was added. The solution was kept at 70 • C until a red-wine colour appeared, circa 10 mins. Then, the liquid was allowed to cool to room temperature. The synthesis experiments have been performed by the platform of Production of Biomaterials and Nanoparticles of the NANBIOSIS ICTS, more specifically by the Nanoparticle Synthesis Unit of the CIBER in BioEngineering, Biomaterials & Nanomedicine (CIBER-BBN).
Following the fabrication of 3D fractals, they were metallized with silver (99.99% silver pellets from Kurt J. Lesker Company, Jefferson Hills, PA, USA) via electron beam evaporation (Edwards auto-500, 3·10 −7 mbar, 32 mA, 5.3 KV). In a second step, gold nanoparticles (Au@citrate NPs) were assembled on the substrates by electrostatic interactions to generate a 3D monolayer-over-mirror configuration.
For such purposes, the SERS substrates were incubated in a poly(diallyldimethilammonium) chloride aqueous solution (PDDA), 0.2% wt for 4 h; followed by rinsing with deionized water. Afterwards, the substrates were immersed in Au@citrate nanoparticles solution (0.19 mg/mL) for 16 h at 4 • C. Then the SERS substrates were rinsed with deionized water and dried at room temperature. Table 1 shows the main characteristics of the SERS active substrates studied in this work, where reference samples used for comparison purposes are also included.
SERS-Raman Measurements
An alpha300 R-confocal Raman Imaging®spectrometer of WITec (Wissenschaftliche Instrumente und Technologie GmbH, Ulm, Germany) was used (480 nm as lateral spatial resolution). Raman-SERS spectra were collected in backscattering geometry. Excitation of the samples was carried out with laser 785 nm at room temperature. Although this wavelength does not correspond to the maximum absorption of the SERS substrates (see Figure S1 of Supplementary Materials), this laser has been selected to avoid the photodegradation of the dimethyl methylphosphonate (DMMP) molecule that we observe when 633 nm laser was used.
In this work, rhodamine 6G (R6G) was chosen as probe molecule and the characteristic band for C-C stretching exhibited at 1512 cm −1 was selected for Enhancement Factor (EF) and SERS Gain quantification.
This EF quantifies how much the Raman signal is amplified with respect to normal conditions, giving information on the field enhancement provided by the structures. It is calculated to assess the SERS activity of the prepared substrates and it is given by: where I SERS is the Raman-band intensity corresponding to the number of molecules analyzed on the SERS substrate, N SERS , and I Raman and N Raman are the intensity and number of molecules without the presence of the SERS substrate, respectively.
To calculate N Raman a Raman spectrum was acquired on a liquid droplet of R6G 1 mM aqueous solution at 50×, 15 mW, 100 s acquisition conditions: The number of R6G molecules within the interaction volume of the laser, 2.1 × 10 −12 cm 3 , was calculated as follows [20].
To establish the values for I SERS a 2 µL droplet of 1 µM R6G (aqueous solution) was deposited on the different SERS substrates (see Table 1). The droplet was allowed to evaporate. The area-size covered by the dried droplet on the different substrates was measured and the total 3D surface area, including the fractal structures, was calculated. It is assumed that the R6G molecules are distributed evenly across the dried spot. The spectra were recorded for R6G at 20×, 1.6 mW, 0.1 s acquisition conditions. The R6G spectrum of the SERS substrate was measured in the center of the droplet once dried. The number of R6G molecules within the effective laser spot, N SERS was estimated by the following equation [21]: On the other hand, the SERS Gain provides quantitative information on the signal gain that one has to expect from a specific SERS sensor with respect to a reference Raman experiment. The SERS Gain is calculated as the ratio between the SERS and Raman Intensities, I SERS and I Raman respectively, normalized to the different powers (P SERS and P RAMAN ), integration times (t SERS ; t Raman ) and molecular concentration (c SERS ; c Raman ) used in the experiment. Therefore, the SERS Gain was calculated by the following equation [22]: SERS experiments for detection of traces of CWA vapors were carried out with DMMP, often used as a Sarin gas simulant thanks to its chemical structure similarities ( Figure 3) and its much lower toxicity. For measuring DMMP in gas phase, the SERS substrate was mounted on a home-made gas chamber (2.7 × 10 −2 cm 3 ) placed at an angle of 90 • with respect to the 785 nm laser excitation beam (Figure 4). A nitrogen stream (10 STP cm 3 min −1 ) was passed through a bubbler containing liquid DMMP, placed in a thermal bath at room temperature, and fed to the microfluidic chamber. The SERS measurements were immediately performed without any stabilization period. The value of the concentration of DMMP in the feed stream was calibrated with a permeation tube (MT-PD-Experimental, 107-100-7845-HE3-C50) of known concentration and corresponds to 1.2 ppmV. The 785 nm excitation laser was coupled through 20× objective (Numerical Aperture, N.A., 0.5; spot diameter 1.9 µm) or 50× objective (N.A., 0.8; spot diameter 1.2 µm) respectively. The SERS spectra are shown after baseline subtraction using WITec Control 1.60.
Characterization Techniques
Scanning electron microscopy (SEM) images were obtained using a FEI INSPECT 50 system equipped with a FEG source of electrons. ImageJ analysis was used to obtain the Au@citrate NP density onto the SERS substrates from three different SEM images, randomly selected of each one. Atomic force microscopy (AFM) measurements (Multimode 8 from Veeco-Bruker; tip, OMCL-AC240TN-W2 from OLYMPUS, XY resolution < 20 nm) were conducted in tapping mode in air to investigate the topography of the metallic coatings. Roughness was estimated by Gwyddion 2.45 analysis of topography images. The UV measurements were performed in a Jasco V-670 spectrometer equipped with a DRIFT chamber for the measurement of solid surfaces.
Morphological Characterization of the 3D Fractals SERS Active Substrates
Recently, plasmonic systems consisting of metal nanoparticles separated from metal films by
Characterization Techniques
Scanning electron microscopy (SEM) images were obtained using a FEI INSPECT 50 system equipped with a FEG source of electrons. ImageJ analysis was used to obtain the Au@citrate NP density onto the SERS substrates from three different SEM images, randomly selected of each one. Atomic force microscopy (AFM) measurements (Multimode 8 from Veeco-Bruker; tip, OMCL-AC240TN-W2 from OLYMPUS, XY resolution < 20 nm) were conducted in tapping mode in air to investigate the topography of the metallic coatings. Roughness was estimated by Gwyddion 2.45 analysis of topography images. The UV measurements were performed in a Jasco V-670 spectrometer equipped with a DRIFT chamber for the measurement of solid surfaces.
Characterization Techniques
Scanning electron microscopy (SEM) images were obtained using a FEI INSPECT 50 system equipped with a FEG source of electrons. ImageJ analysis was used to obtain the Au@citrate NP density onto the SERS substrates from three different SEM images, randomly selected of each one. Atomic force microscopy (AFM) measurements (Multimode 8 from Veeco-Bruker; tip, OMCL-AC240TN-W2 from OLYMPUS, XY resolution < 20 nm) were conducted in tapping mode in air to investigate the topography of the metallic coatings. Roughness was estimated by Gwyddion 2.45 analysis of topography images. The UV measurements were performed in a Jasco V-670 spectrometer equipped with a DRIFT chamber for the measurement of solid surfaces.
Morphological Characterization of the 3D Fractals SERS Active Substrates
Recently, plasmonic systems consisting of metal nanoparticles separated from metal films by nanometer scale gaps have attracted great attention as SERS substrates, because the LSPRs of metal nanoparticles can couple with the propagating SPPs at the surface of metal films when precise gap regions between metal nanoparticles and the films are provided [23]. In this work, a preliminary evaluation of the minimal thickness for the silver coating was performed on flat, oxidized silicon wafers (Siltronix Double side polished Si (100) with ca. 1 µm of SiO 2 ) previous to metallization of 3D-fractal structures. Two different silver evaporation times were selected, i.e., 90 and 500 s, which result in film thicknesses, measured by AFM, of 12 ± 1 and 77 ± 3 nm respectively. As we can observe in Figure 5, when the amount of silver deposits increases, the formation of a continuous layer is enabled. In addition, the average roughness decreases slightly with increasing film thickness, i.e., 2.8 ± 0.1 nm and 2.6 ± 0.2 nm. Accordingly, 3D-fractal structures were metallized with 77 nm thickness of silver. nanoparticles can couple with the propagating SPPs at the surface of metal films when precise gap regions between metal nanoparticles and the films are provided [23]. In this work, a preliminary evaluation of the minimal thickness for the silver coating was performed on flat, oxidized silicon wafers (Siltronix Double side polished Si (100) with ca. 1 μm of SiO2) previous to metallization of 3Dfractal structures. Two different silver evaporation times were selected, i.e., 90 and 500 s, which result in film thicknesses, measured by AFM, of 12 ± 1 and 77 ± 3 nm respectively. As we can observe in Figure 5, when the amount of silver deposits increases, the formation of a continuous layer is enabled. In addition, the average roughness decreases slightly with increasing film thickness, i.e., 2.8 ± 0.1 nm and 2.6 ± 0.2 nm. Accordingly, 3D-fractal structures were metallized with 77 nm thickness of silver. The assembly of Au@citrate NPs was performed on the different SERS substrates using the electrostatic interaction between the negatively charged sodium citrate groups anchored on the Au@citrate NPs and the positively charged PDDA layer on the SERS substrate. Figure 6 shows the results of Au@citrate NPs deposition over 2D and 3D active SERS substrates. In general, the experimental procedure yielded a homogeneous monolayer of Au@citrate NPs with a surface density above 600 NPs/μm 2 , i.e., a coverage degree of around 25%. As it was expected, no differences are observed for samples without Ag due to the effectiveness of the intermediate PDDA coating. The Au@citrate NPs density on the flat surfaces is around a 6% higher compared to the 3D-fractal surfaces. These minor deviations are attributed to the higher surface roughness for the Ag coatings when evaporated on flat substrates compared to fractals (see Figure 6B,E,H). The assembly of Au@citrate NPs was performed on the different SERS substrates using the electrostatic interaction between the negatively charged sodium citrate groups anchored on the Au@citrate NPs and the positively charged PDDA layer on the SERS substrate. Figure 6 shows the results of Au@citrate NPs deposition over 2D and 3D active SERS substrates. In general, the experimental procedure yielded a homogeneous monolayer of Au@citrate NPs with a surface density above 600 NPs/µm 2 , i.e., a coverage degree of around 25%. As it was expected, no differences are observed for samples without Ag due to the effectiveness of the intermediate PDDA coating.
The Au@citrate NPs density on the flat surfaces is around a 6% higher compared to the 3D-fractal surfaces. These minor deviations are attributed to the higher surface roughness for the Ag coatings when evaporated on flat substrates compared to fractals (see Figure 6B,E,H).
SERS Performance for Au Monolayer over Mirror Configuration
The SERS performance of the Au monolayer-over-mirror configuration, i.e., of the Glass_Ag_AuNPs sample, is highlighted by comparing it with Au@citrate NPs coated onto non-metallized glass substrates, i.e., the Glass_AuNPs sample. Figure 7 shows the SERS response of both samples under identical conditions upon exposure to DMMP vapor, at different integration times. On the silver coated sample, i.e., Glass_Ag_AuNPs, the characteristic molecular fingerprint of DMMP molecules adsorbed on the SERS substrate is clearly distinguished [24]. As the integration time increases, more signal is collected by the detector, and for an integration time of 20 s (or larger) is possible to obtain a signal with a value of around 200 cts for the more intense peak at 710 cm −1 . As was explained in the introduction section, due to the toxicity of Sarin gas a fast detection is pursued. The main vibrational modes of DMMP: 710 cm −1 (ν (P-CH3)), 782 cm −1 (νas (O-P-O)), 980 cm −1 and 1280 cm −1 (ν (P = O) can be perfectly identified when using Au monolayer-over-mirror configuration. On the other hand, the non-metallized Glass_AuNPs sample gives a higher SERS intensity but the spectral fingerprint of the target is hindered by the Raman lines for citrate coating, i.e., 887 cm −1 -950 cm −1 (ν (C-COO)) and 1167 cm −1 (δ (COO)). In addition, a broad band in the region 1450 to 1700 cm −1 , attributed to amorphous carbon, is observed. The carbon formation agrees with the photodecomposition of the citrate molecules covering the Au NPs. Our hypothesis is that in the case of a non-metallized surface, the local temperature on the substrate increases up to levels that lead to photodecomposition due to the poor thermal conductivity of the underlying glass (1 W/K•m) compared to Ag films (410 W/K•m). Such heat-transfer related effects are especially noteworthy when dealing with gas phase compounds. Based on these control experiments, we consider the monolayer-over-mirror to be the optimal configuration for our subsequent 3D SERS experiments.
SERS Performance for Au Monolayer over Mirror Configuration
The SERS performance of the Au monolayer-over-mirror configuration, i.e., of the Glass_Ag_AuNPs sample, is highlighted by comparing it with Au@citrate NPs coated onto non-metallized glass substrates, i.e., the Glass_AuNPs sample. Figure 7 shows the SERS response of both samples under identical conditions upon exposure to DMMP vapor, at different integration times. On the silver coated sample, i.e., Glass_Ag_AuNPs, the characteristic molecular fingerprint of DMMP molecules adsorbed on the SERS substrate is clearly distinguished [24]. As the integration time increases, more signal is collected by the detector, and for an integration time of 20 s (or larger) is possible to obtain a signal with a value of around 200 cts for the more intense peak at 710 cm −1 . As was explained in the introduction section, due to the toxicity of Sarin gas a fast detection is pursued. The main vibrational modes of DMMP: 710 cm −1 (ν (P-CH 3 )), 782 cm −1 (ν as (O-P-O)), 980 cm −1 and 1280 cm −1 (ν (P = O) can be perfectly identified when using Au monolayer-over-mirror configuration. On the other hand, the non-metallized Glass_AuNPs sample gives a higher SERS intensity but the spectral fingerprint of the target is hindered by the Raman lines for citrate coating, i.e., 887 cm −1 -950 cm −1 (ν (C-COO)) and 1167 cm −1 (δ (COO)). In addition, a broad band in the region 1450 to 1700 cm −1 , attributed to amorphous carbon, is observed. The carbon formation agrees with the photodecomposition of the citrate molecules covering the Au NPs. Our hypothesis is that in the case of a non-metallized surface, the local temperature on the substrate increases up to levels that lead to photodecomposition due to the poor thermal conductivity of the underlying glass (1 W/K·m) compared to Ag films (410 W/K·m). Such heat-transfer related effects are especially noteworthy when dealing with gas phase compounds. Based on these control experiments, we consider the monolayer-over-mirror to be the optimal configuration for our subsequent 3D SERS experiments.
Evaluation of the SERS Enhancement Factor and SERS Gain for 3D Fractals Active SERS Substrates
Typical SERS spectra of R6G (1 mM) are presented in Figure 8 for the 2D and 3D SERS active substrates prepared in this work, as well as the normal Raman spectrum of R6G solution (1 mM) used as a reference for the calculation of EF-values (see Section 2.3). The corresponding EF-values, calculated following Equations (1)-(3), are summarized in Table 2.
The importance of the 3D-fractal architectures is demonstrated by comparing them with the planar Glass_Ag_AuNPs, where the latter gives rise to an enhancement factor of 1.9 × 10 4 , 6-fold and 18-fold lower than the 1G and 3G fractals covered with Ag_AuNPs, which have EF-values of 1.18 × 10 5 and 3.51 × 10 5 , respectively. The enhanced surface area of 3D-fractals, due to their extension in z-direction, is already considered for the estimation of NSERS. Thus, the registered differences can only be attributed, due to the fractal size, to surface plasmon modes generated by the edges and corners, as observed previously for Sierspinki fractals with feature sizes of 750 nm, 1.2 μm and 4 μm [15]. Table 2. Enhancement factor and SERS gain for the 2D and 3D SERS substrates studied in this work.
Evaluation of the SERS Enhancement Factor and SERS Gain for 3D Fractals Active SERS Substrates
Typical SERS spectra of R6G (1 mM) are presented in Figure 8 for the 2D and 3D SERS active substrates prepared in this work, as well as the normal Raman spectrum of R6G solution (1 mM) used as a reference for the calculation of EF-values (see Section 2.3). The corresponding EF-values, calculated following Equations (1)-(3), are summarized in Table 2.
The importance of the 3D-fractal architectures is demonstrated by comparing them with the planar Glass_Ag_AuNPs, where the latter gives rise to an enhancement factor of 1.9 × 10 4 , 6-fold and 18-fold lower than the 1G and 3G fractals covered with Ag_AuNPs, which have EF-values of 1.18 × 10 5 and 3.51 × 10 5 , respectively. The enhanced surface area of 3D-fractals, due to their extension in z-direction, is already considered for the estimation of N SERS . Thus, the registered differences can only be attributed, due to the fractal size, to surface plasmon modes generated by the edges and corners, as observed previously for Sierspinki fractals with feature sizes of 750 nm, 1.2 µm and 4 µm [15]. To obtain an estimation of the EM enhancement on the fractal structures, the Raman signal was collected at selected z-positions in a mapping area of 20 × 20 μm (1 spectrum/μm), see Figure 9. Bright areas represent higher Raman intensity and comparing the xy-images at different z-positions, i.e., 5, 8, 10 and 12 μm from the ground level, it is evident the higher electromagnetic field on the corners, apices or tips of the microstructures. This enhancement is modest, just 1.7 folds higher, when compared with those exhibited by Sierspinki fractals with feature sizes of 150 nm [15]. However, due to the higher extension of the 3D fractal structures in the z direction, 12 μm vs. 40 nm, larger interfacial area for gas-solid contact and improved limit of detection for gas sensing could be expected. Figure 10 comparatively shows the recorded SERS spectra on a 1G_Ag_AuNPs sample upon exposure to DMMP in vapor phase (1.2 ppmV) with the focus plane at two z-positions: 0 μm (bottom part of the fractal) and 5 μm (top part of the fractal). The most intense signal is appearing at 706 cm −1 for both heights, tentatively assigned to the P-C stretching mode and used as Raman reference line. To obtain an estimation of the EM enhancement on the fractal structures, the Raman signal was collected at selected z-positions in a mapping area of 20 × 20 µm (1 spectrum/µm), see Figure 9. Bright areas represent higher Raman intensity and comparing the xy-images at different z-positions, i.e., 5, 8, 10 and 12 µm from the ground level, it is evident the higher electromagnetic field on the corners, apices or tips of the microstructures. This enhancement is modest, just 1.7 folds higher, when compared with those exhibited by Sierspinki fractals with feature sizes of 150 nm [15]. However, due to the higher extension of the 3D fractal structures in the z direction, 12 µm vs. 40 nm, larger interfacial area for gas-solid contact and improved limit of detection for gas sensing could be expected. To obtain an estimation of the EM enhancement on the fractal structures, the Raman signal was collected at selected z-positions in a mapping area of 20 × 20 μm (1 spectrum/μm), see Figure 9. Bright areas represent higher Raman intensity and comparing the xy-images at different z-positions, i.e., 5, 8, 10 and 12 μm from the ground level, it is evident the higher electromagnetic field on the corners, apices or tips of the microstructures. This enhancement is modest, just 1.7 folds higher, when compared with those exhibited by Sierspinki fractals with feature sizes of 150 nm [15]. However, due to the higher extension of the 3D fractal structures in the z direction, 12 μm vs. 40 nm, larger interfacial area for gas-solid contact and improved limit of detection for gas sensing could be expected. Figure 10 comparatively shows the recorded SERS spectra on a 1G_Ag_AuNPs sample upon exposure to DMMP in vapor phase (1.2 ppmV) with the focus plane at two z-positions: 0 μm (bottom part of the fractal) and 5 μm (top part of the fractal). The most intense signal is appearing at 706 cm −1 for both heights, tentatively assigned to the P-C stretching mode and used as Raman reference line. for both heights, tentatively assigned to the P-C stretching mode and used as Raman reference line. However, when focused at z = 0 the signal-to-noise ratio hinders its proper identification. In addition, the characteristic peak of DMMP at 780 cm −1 , attributed to PO 2 bending, is clearly observed on the top of the fractals. Both peaks are shifted with respect to normal Raman of the DMMP in liquid phase. This observation agrees with the hydrogen bonding type interactions between citrate from Au NPs and DMMP molecule, as described in the literature [21]. The citrate coating acts as an effective trap for the target molecules in the immediate vicinity of the metallic surface where the maximum electromagnetic enhancement is achieved. It is important to note that the spectra in Figure 10 are taken at 1s integration time, under these conditions a value of around 2000 cts is obtained, indicating a large enhancement of the electromagnetic field in the vicinity of the molecule. When the spectra were taken at longer acquisition times, the signal around 1600 cm −1 corresponding to amorphous carbon appeared, indicating that the high energy concentrated on the molecule was able to burn it. A similar result was obtained when 3G fractals were used for measurements. Thus, we can conclude that in the case of the DMMP molecule the energy should be modulated to avoid decomposition of the molecule and a trade-off between electromagnetic field enhancement and stability of the molecule should be achieved.
SERS Detection in Gas Phase of Sarin Surrogate
However, when focused at z = 0 the signal-to-noise ratio hinders its proper identification. In addition, the characteristic peak of DMMP at 780 cm −1 , attributed to PO2 bending, is clearly observed on the top of the fractals. Both peaks are shifted with respect to normal Raman of the DMMP in liquid phase. This observation agrees with the hydrogen bonding type interactions between citrate from Au NPs and DMMP molecule, as described in the literature [21]. The citrate coating acts as an effective trap for the target molecules in the immediate vicinity of the metallic surface where the maximum electromagnetic enhancement is achieved. It is important to note that the spectra in Figure 10 are taken at 1s integration time, under these conditions a value of around 2000 cts is obtained, indicating a large enhancement of the electromagnetic field in the vicinity of the molecule. When the spectra were taken at longer acquisition times, the signal around 1600 cm −1 corresponding to amorphous carbon appeared, indicating that the high energy concentrated on the molecule was able to burn it. A similar result was obtained when 3G fractals were used for measurements. Thus, we can conclude that in the case of the DMMP molecule the energy should be modulated to avoid decomposition of the molecule and a trade-off between electromagnetic field enhancement and stability of the molecule should be achieved. To evaluate the homogeneity of the SERS substrate and the reproducibility of the present fabrication method, two different 1G fractal samples were prepared with the same protocol, and 6 different spots were randomly selected on the top and on the bottom to record the SERS signal of 1.2 ppmV DMMP for each substrate ( Figures 11A and S2 of the Supplementary Materials). Sample 1 and 2 result in similar intensities with a relatively low standard deviation for the 6 spots measured, which is an evidence of the good reproducibility of the present fabrication method. To evaluate the homogeneity of the SERS substrate and the reproducibility of the present fabrication method, two different 1G fractal samples were prepared with the same protocol, and 6 different spots were randomly selected on the top and on the bottom to record the SERS signal of 1.2 ppmV DMMP for each substrate ( Figure 11A and Figure S2 of the Supplementary Materials). Sample 1 and 2 result in similar intensities with a relatively low standard deviation for the 6 spots measured, which is an evidence of the good reproducibility of the present fabrication method. | 8,905.4 | 2018-01-31T00:00:00.000 | [
"Chemistry",
"Materials Science"
] |
Design and Implementation of Welding Mobile Robot Using a Proposed Control Scheme Based On Its Developed Dynamic Modeling for Tracking Desired Welding Trajectory
This paper presents a proposed control scheme that makes the combination of a kinematic controller (KC) and an integral sliding mode controller (ISMC) for a welding mobile robot (WMR) to track a desired welding path. First, a posture tracking error vector is defined and a kinematic controller is designed based on kinematic modeling to make the tracking error vector go to zero asymptotically. Second, a sliding surface vector is defined based on the velocity tracking error vector and its integral term. And then, an integral sliding mode dynamic controller is designed based on developed dynamic modeling to make velocity tracking error vector also go to zero asymptotically. The above controllers are obtained by backstepping method. The stability of system is proved based on the Lyapunov stability theory. To implement the designed tracking controller, a control system is developed based on DSP F28355 and ATmega328. A scheme for measuring the posture tracking error vector using torch sensor is presented. The simulation and experiment results are shown to illustrate effectiveness and the applicability to the welding industry field of the proposed controller.
To solve the problem of trajectory tracking of welding mobile robot, this research presents a proposed control scheme that makes the combination of a kinematic controller (KC) based on the kinematic modeling and an integral sliding mode dynamic controller (ISMC) based on the developed dynamic modeling considering at voltage level for the WMR to track a desired welding trajectory at a desired velocity. The above controllers are obtained by backstepping method. The system stability is proved using the Lyapunov stability theory. To implement the designed tracking controller, a control system is developed based on DSP F28355 and ATmega328. A scheme for measuring the posture tracking error vector using torch sensor is presented. The simulation and experiment results are shown to illustrate effectiveness of the proposed nonlinear controller.
II.
SYSTEM DESCRIPTION AND MODELING In this section, the system description, dc servo motor modeling, the kinematic and dynamic models of a welding mobile robot (WMR) are presented. Fig. 1 shows the 3D configuration of a welding mobile robot (WMR) used in this research in three sides. It consists of platform, two wheels, DC servo motors and encoders, welding torch, power supply and a electronic control system, etc. Fig. 2 shows the 2D configuration for geometric model of the WMR. For simplifying the modeling of WMR, the assumptions are given as follows [10][11][12][13][14]:
System description
(1) Kinematic's parameters such as wheel's radius r and distance b are known exactly.
(2) Moment of inertia of WMR is constant during welding process. (3) A disturbance vector exerted on the WMR consists of surface friction and slip phenomenon bewteen wheel and the ground. (4) Motion surface is a smooth horizontal plane. (5) The WMR has two driving wheels for flatform motion, and those are positioned on an axis passed through the WMR geometric center, (6) Two passive wheel which have zero constraint are installed in front and rear of the flatform at the bottom for balance of WMR. So their motion can be ignored in the kinematic and dynamic models. A welding torch is located to coincide the axis through the center of the two driving wheels. The radius of welding curve is sufficiently larger than turning radius of the WMR.
DC Servo Motor Modeling
This section presents the modelingof DC servo motor [9]. Schematic of the the DC servo motor plus wheel is shown in Fig. 3.
The relation between m Torque of DC servo motor is given by ki ta The kinematic equations for the center point of the WMR are set up as the following: is the actual velocity vector.
The relationship between , v and the angular velocities of right wheel rw and left wheel lw is given by In where l is assumed to be constant.
Developed Dynamic Modeling
In Fig. 2, using the references from [10] to [14], the developed dynamic equations of the WMR considering at DC Servo motor voltage level is rewrited as follows: Because torch length l is controllable based on the torch slider. The first derivative of e yields First, the kinematic controller is designed as follows ( ) cos 3 3 3 1 1 33 15) and the length of torch satisfies ,, C C C is a positive values. Second, the developed dynamic controller with voltage control input vector for DC servo motors is designed as follows: The velocity error vector v e is defined as The sliding surface vector S v is defined as Where v K is a positive diagonal matrix and is an integral sliding surface vector.
Third, the auxilary control law The Lyapunov function candidate is defined as follows: where the conponents of the V function are chosen as: With the velocity control input Eq. (15), the 1 V becomes The derivative of S v in Eq. (19) is as the following In other hand, from Eqs. (11) and (17) Subtituting Eq. (28) into the first derivative of 2 V in Eq.
Mesurement of tracking error using tourch
sensor In order to measure the tracking errors, a mechanical measurement scheme using potentiometers is shown in Fig. 5 [9]. Two rollers are placed at points O2 and O3. Two sensors for measuring the errors are needed. That is, they are one linear potentiometer sensor for measuring ds and one rotating potentiometer sensor for measuring the angle between the torch and the tangent line of the wall at the welding point.
International Journal of Advanced Engineering Research and Science (IJAERS)
[ Vol-4, Issue-10, Oct-2017] https://dx.doi.org/10.22161/ijaers.4.10.13 ISSN: 2349-6495(P) | 2456-1908(O) www.ijaers.com Page | 77 where O2 and O3 are the center points of roller O2 and O3 respectively, O1 is the center point of O2O3, W is the point on torch holder, rs is the radius of the roller, ds is the length measured by the linear potentiometer, and e3 is the angle measured by the rotating potentiometer. In Figure 5, the reference welding path is a line. When the reference welding path is a curve, Eq. (30) is also valid if the distance O2O3 is sufficiently small and the radius of the welding path is enough large. Fig. 6 shows the control system configuration of welding mobile robot. The control system is based on the integration of microcontroller DSP F28335 and ATmega328. The microcontroller ATmega328 are used for two DC servo motor control signal of left wheel and right wheel and torch slider controller. The microcontroller DSP F28335 is used for main controller. The three servo controllers are controlled by main controller. The main controller functionalized as master links to the three servo controllers via I2C communication.
Slave unit 3 ATmega328
Servo controller of wheel 1 Servo controller of wheel 2 Servo controller of torch slider The two A/D ports of the DSP F28355 are connected with the two potentiometers for sensing the tracking errors. The master unit send the commands to the slave controllers via I2C communication,respectively. The master unit can be used to interface other devices such as display and keypad devices for manual control. The sampling time of control system is 10ms. The slave unit integrates ATmega328 with motor drivers for the DC servo motor control. This slave controller can perform a complete servo operation with a closed loop feedback control using an encoder for velocity control of two wheels and position of welding torch. The experimental welding mobile robot is shown in Fig. 7 and its dimensions are shown in Table 3. Fig. 8 The simulation results for tracking welding path are shown in Figs. 9 to 17. Fig.s 9 and 10 show the movement of the WMR along the desired welding trajectory for the time beginning and full time 4000 seconds. The simulation results for error tracking vector during 5 seconds at beginning and 4000 secconds of full time are shown in Fig.s 11 and 12. The errors go to zero from 3 seconds. The linear velocity of welding point is shown in Fig. 13; its is shown that the linear velocity at the welding point W of the WMR has quick change at the first time and converges to the constant velocity in the vicinity of 7.5 / mm s as desired after 1 seccond. Fig. 16 shows the linear velocity and torch length of WMR. It goes from 145 mm at initial time to 163.3mm after about 3 seconds and keeps the that value for remain time. Fig. 17 shows that the control input voltage vector u changes rapidly at the start time converges to small values from 0.3 second for the full time. Figs. 18 and 19 show the welding process of welding mobile robot and experimental welding line result respectively. The simulation and experimental results are shown that the WMR has good welding path tracking performance. It is so that the welding mobile robot can be applied in the practical welding industry field. VI. CONCLUSION This paper presents the proposed control scheme that makes the combination of a kinematic controller (KC) and an integral sliding mode dynamic controller (ISMC) based on the developed dynamic modeling for the welding mobile robot to track a desired welding trajectory at a desired velocity under external disturbances. The control laws are obtained by backstepping method. The system stability is proved based on Lyapunov theory. The control law stabilizes the sliding surface vector and makes the tracking error vector go to zero asymptotically. To implement the designed tracking controller, the control system is developed based on DSP F28355 and ATmega328. A scheme for measuring the posture tracking error vector using torch sensor is presented. The simulation and experiment results are shown to illustrate effectiveness and the applicability to the welding industry field of the proposed controller. | 2,301.6 | 2017-10-01T00:00:00.000 | [
"Materials Science"
] |
Dynamic model of store/rack interface reactions
This research demonstrated the ability to predict the reaction loads transmitted to an aircraft bomb rack due to the inertial forces acting on an external store relying only on store mass properties, accelerometer data and geometry. Once theoretical equations were developed, a full-scale static ground test was conducted to provide data for model, has traditionally been immune to the unpredictability of the defense budget, but the pressure to streamline must now be considered in all aspects of the flight test process.[ 11 To compensate for the Shrinking budgets and increased requirements, new and more versatile test techniques and data processing systems must be developed. Standard flight verification and refinement. Flight test data for final validation were accumulated during a carrier suitability flight test program conducted at the Naval Air Warfare Center, Patuxent River, Maryland on an F-14 aircraft with instrumented BRU-32/A bomb ejector rack and a GBU24B/B 2,000-lb. bomb. In the 300 milliseconds following arrestment, forces and moments up to 15,000 Ibs. and 150,000 in-lbs., respectively, were calculated at the store CG. Compared to the measured data, agreement was found in form and magnitude for the calculated interface reactions. Critical lug and swaybrace rod reactions averaged less than 9% absolute error.
INTRODUCTION
The demands on aircraft flight-testing have increased exponentially during the past few decades. Significant reductions in the budgets for weapons systems testing require collection of more test information for less money. In addition to the effects of cost reductions and downsizing environments, the desire to make rapid production decisions to retain program funding is critical. Aircraft flight-testing test processes must be reexamined to maximize a test asset's availability. Highly complex, expensive aircraft can no longer afford to sit idle while recorded data are being reduced. Each test must be optimized to produce the information which allows the most effective use of flight time while contemporaneously processing and analyzing the collected data more quickly and at a lower cost. [ 13 Concurrently, military aircraft are becoming more versatile in design while the externally carried ordnance has become larger and heavier. With the entrance of the F/A-lSE/F as the United States Navy's (USNs) premier fighter and attack aircraft of today and the future, the majority of this burden will lie on its wings. Though the aircraft is new, the store/aircraft interface is approximately the same as originally designed in 1976. The BRU-32/A Bomb Ejector Rack (BRU-32), figure 1-1, was created for the USN and, through a number of new features, was better than the stateof-the-art suspension systems of the time. [2] The BRU-32 is still the primary ejector rack for carriage of single, external ordnance on the F/A-18 Hornet and F-14 Tomcat. Increases in aircraft capability and ordnance weight and size create larger reaction forces at the BRU-32 interface points. The aerodynamic and inertial loading of the store is transferred to the BRU-32 through six points of contact: two lughook connections that react only tensile forces and four swaybrace rods only able to react in compression. The magnitudes of these reactions have been rising steadily over the past 15 years due to increases in aircraft capability as well as in the mass of the external ordnance. The ability of the BRU-32 to withstand the required forces is often considered marginal based on analytical computations. Accurate predictions of reactions at the storehack interface (SRI) points are difficult and, in the past, generally shown to be unsuccessful during flight-testing.
Without accurate and readily available prediction models, all ordnance with the possibility of creating large reactions in the BRU-32 must go through time-consuming and expensive structural flight test programs with fully instrumented store and rack combinations. Strain gages, either machined within or mounted on the surface of the lugs, provide data for the lughook reactions; instrumented swaybraces have externally mounted gages for data collection. This instrumentation is costly in the application of the gages as well as in the data collection and reduction. Furthermore, the gages are very fragile and, especially during the typical installation and removal procedures of the larger stores, are easily damaged resulting in wasted effort in the delay or repetition of test flights. added costs and impact of critical program milestones.
Although an aircraft's envelope of flight operations contains a plethora of combinations of airspeed, altitude, etc.. many of the largest reactions experienced at the SRI are found during carrier suitability flight-testing. With maximum sink rates for a typical fighter aircraft over 20 feet per second (fps), the oscillating reactions due to inertial forces immediately after the impact of an arrested landing will almost completely eclipse all aerodynamic load contributions at that same time. Structural flight test programs typically begin with carrier suitability flight tests.
Objective
The objective of this research was to demonstrate the ability to accurately predict the reaction loads at the SRI via a BRU-32 on an aircraft during an arrested landing. The reaction forces were found without employing specialized SRI instrumentation, thereby avoiding the time and expense of installing and calibrating the delicate strain gages as well as installation of strain gage signal conditioning. Preloads accumulated at the interface during ordnance loading and handling were also examined. The final equations are based only on the combination of external loads that subsequently result in the maximum forces generated at each swaybrace and hook and are generalized for multiple aircraft and rack combinations. The final objectives of this research are summarized in figure 1-2. The objectives have been completed in conjunction and cooperation with the USN. Their interest in the successful completion of this project has allowed the use of previous flight test data as well as funding for the additional ground testing and data reduction required for model development and verification.
Approach
The approach for this research was to create BRU-32 reaction equations based on inertial loads and preloads only. This was accomplished by theoretically modeling the inertial reactions and comparing the expected results with data from an instrumented ground test as well as an actual flight test. For the ground test, a 2,000-lb. store was hung fiom an instrumented SRI in a laboratory and subjected to eight predetermined load conditions comprised of single and multiple forces and moments. These data were collected as truth data and used to improve the theoretical models by reducing the number of assumptions necessary due to the static indeterminacy of the reactions and parallel nature of the loads. The newly developed equations were then compared with actual flight test data from carrier suitability flight-testing where, immediately after an arrestment, aerodynamic loads were shown to be negligible. Any differences between the actual data and the predictions were then resolved at a finer resolution by accounting for the small variations in individual store geometry and loading conditions. The predictions for in-flight SRI reactions due to aerodynamic loads can be found through a similar process as the inertial loads, with modeling and comparisons to actual flight data. The inertial predictions for arrested landings are also adequate for inertial-only predictions during normal flight. Full aerodynamic considerations were beyond the scope of this research.
The carrier suitability flight test data were collected during a test program conducted at the Naval Air Warfare Center Aircraft Division (NAWCAD), Patuxent River, MD with an F-14 aircraft and a fully instrumented BRU-32. The final results provided an accurate prediction model for calculating the loads at the BRU-32 SRI given the store mass properties, store accelerometer data, geometry and flight test conditions. Further work may address real-time data collection, lighter stores, additional aircraft types, or various bomb ejector racks carrying single or multiple stores.
BRU-3YA Bomb Ejector Rack
In the 1970s, the USN created a design specification for a bomb ejector rack for the F/A-18 aircraft that included several new features that were non-existent in the current The BRU-32 combines two sets of hooks (including 14-inch and 30-inch spacings) designed for carriage of stores with two suspension lugs. [3] Stores with 10-inch to 28-inch diameters are automatically swaybraced when the hooks become latched and secondary adjustments are not required. [2] The rack assembly weighs approximately 76lbs. and has a pitching moment of 1.52 dugs/f,2. [4] The static ejection acceleration for the 2,370-lb. GBU-24 from the BRU-32 is approximately 11 Gs. [5]
ANALYTICAL MODELING
This paper addresses research related to calculating the reactions at SRI points without the need for strain gages. The reaction equations were derived via a classical analysis and the results are primarily a function of directly measured or computed accelerometer data. Once developed, a FORTRAN program was created to calculate the SRI reactions based on geometry, inertial forces, known preloads, etc. The program determined the translational and rotational accelerations at the store's center of gravity (CG) using data from accelerometers placed around the surface of the store. The inertial loads at the store CG (three forces and three moments) were then found from the accelerations using Newton's Laws, kinetics of rigid bodies in three dimensions and Euler's Equations of Motion as an extension of d'Alembert's Principle. The store CG loads were distributed to the six SRI interface points and the reactions calculated. The most accurate analytical tool to determine store carriage loads for all conditions is a solution to the full Navier-Stokes equations. [6] Solutions to the complete Euler equations have been solved for various shapes by generating a grid to fit the configuration then extending the grid into the flowfeld. [7,8] Both processes, however, are too difficult and time consuming for most configurations. A simpler method was required to predict real-time SRI reactions during a flight-test program.
The store and rack assembly forms a slightly flexible (in the rack), statically indeterminate structure due to the number and parallel nature of the loads. Rigidities in future designs were not known and the load paths could not be analytically or experimentally determined for an exact solution; therefore, rigid body assumptions were incorporated for all components. The interface points at the SRI consist of four swaybrace rods and two lug/hook combinations: swaybrace rods only take compressive loads in the y-z (lateral and vertical) plane; hooks can react in the x-y plane (longitudinal and lateral) and exert tensile force along the zaxis.
To verify the structural integrity of a store and rack combination, the first consideration was how to interpret the test data. Failure can be defined many ways, from almost imperceptible yielding to complete separation. [9,10] The USN definition of a failure with regard to store carriage is constituted by unintended separation of the store from the suspension equipment, separation of any part of the store or suspension equipment at ultimate or lower loads, or material fracture or yielding of the store or suspension equipment. Limit loads are defined as the maximum expected loads in normal operation of the store and suspension equipment; yield and ultimate loads are defined as 115% and 150% of the limit loads, respectively. [3] A dynamic analysis was used to create the equations required for inertial load calculations. A number of procedures for complete dynamic analyses have been considered in academia for dynamic loading, including dividing a system into substructures and performing a discrete element idealization on each substructure to obtain the necessary stiffness matrices and mass matrices.[l 11 A variety of procedures have also been utilized to obtain the substructure mass matrices, including lumped mass formulations in which a displacement method (considering geometric compatibility along each substructure boundary) is employed in the final substructure coupling. [ 12,131 Substructure methods relying on the displacement method have been employed in analysis of both static [11,14] and dynamic [ 15,161 behavior of structures. Dynamic behavior generally requires a Rayleigh-Ritz procedure to reduce the order of system matrices.
Additionally, Gladwell introduced the Branch-Modes technique with the advantage of a diagonal system stifhess matrix that is formed along with the normal-mode analysis of the substructure.
[l2] All of these methods, however, are very time-intensive and require many input parameters. It is also difficult, if not impossible. to solve a parallel set of equations.
To more quickly solve for the store's translational and rotational accelerations, accelerations were calculated using the individual linear accelerometer data and their location on the store. First principles were incorporated to derive the equations of acceleration fiom raw accelerometer data.
Accelerations
The general motion of a rigid body is equivalent, at any given instant, to the sum of a translation in which all of the particles of a body have the same velocity and acceleration of a reference particle A, and of a motion in which particle A is assumed to be fixed. The angular velocity and angular acceleration of a rigid body at a given instant are therefore independent of the choice of the reference point.
[ 171 From Euler's theorem that the general displacement of a rigid body with a fvred point A is equivalent to a rotation of the body about an axis through A, it can be shown that Considering a 2,000-lb. bomb as a rigid body, the acceleration can be rewritten in terms of the acceleration of the store CG, zcc, and the acceleration measured by an Vol. 6-2728 4 accelerometer on the store surface, a,,. The term ri,, represents the relative acceleration felt by the surface accelerometer with respect to the store CG due to the store's angular acceleration or rotational velocity about the CG.
To find the acceleration at the store CG, equation (3) is rearranged to The measured accelerations at each accelerometer zA4 are known while the relative accelerations require calculation.
Allowing zi, to coincide with another point on the surface and point A to represent the CG, equations (2) and (4) were combined to find the equation of acceleration about the CG based on the measured accelerometer data.
From equations (4) and (5), the acceleration of a store accelerometer relative to the store CG zi,,, is defined as Since W , and sf are all three-dimensional vectors, they can be expressed as the combination of their 7 , J and k directional components: By expanding equations (6) through (9), the three final equations for relative acceleration broken down by component along the X -, Y'-and Z'-axes are found as felt by the accelerometer with respect to the store CG.
The final equations to calculate the components of translational acceleration at the store CG based on the measured accelerations, geometry, angular velocities and angular accelerations were found by combining equation (5) with equations (lo), (1 1) and (12):
For accelerometer data, the measured accelerations were all tangential allowing the calculation of the angular accelerations when combined with the store and instrumentation geometry. Accurate physical locations of the accelerometers on the store were critical to avoid coupling effects of the angular, centripetal and translational accelerations that could not analytically be separated. Additionally, these equations contain scalars instead of vector expressions; therefore, correct sign convention is critical. The values of r,, r,, and rz physically relate the store CG to the locations of the accelerometers and must follow global sign conventions vice simply being positive distance values. Accelerometers are fastened to an orthogonal riser block and not necessarily fastened directly on the surface of the store. The measured accelerations were labeled and defined as shown in table 2-1. Z-axis accelerometer used for roll measurement.
Located on port side of store, preferably at store CG across from uzs,roll.
Z-axis accelerometer used for pitch measurement. Redundant use of accelerometer data allows the necessary calculations to be completed with a minimum of six accelerometers (i.e. sharing a =-accelerometer for pitch, roll and vertical acceleration calculations, etc.). All coupled accelerometers (i.e. ayF,,, and ayAJzaw, etc.) should be placed at the necessary and identical coordinates to minimize error due to acceleration coupling (i.e. a roll acceleration producing false pitch acceleration readings due to uneven lateral placement, etc.).
\
The magnitude of the store roll acceleration ax was computed using two linear z-axis accelerometers placed on opposite sides of the store's longitudinal centerline although not necessarily in the same longitudinal coordinate. Equation (16) provides the average reading of roll acceleration of the store CG. The second term in equation (16) removes an effective roll acceleration component due to pitch acceleration for the case when the two accelerometers are at different longitudinal locations. Yaw acceleration does not affect this parameter and translational acceleration along the z-axis is canceled out via subtraction within both terms in equation (16). The value of Ly,roll corresponds to the positive lateral distance between the accelerometers used for the roll acceleration calculations, which are typically located just above the surface of the store on orthogonal blocks. Lxron and LXpirch correspond to the positive longitudinal distances between the z-axis accelerometers measuring roll and pitch, respectively. With the store rigid body assumptions imposed to this point, the roll accelerometers do not have to be equidistant laterally or vertically from the store CG or center of rotation.
The magnitude of the store pitch acceleration ay was computed using two linear z-axis accelerometers placed on the same side of the store with one forward and one aft of the longitudinal store CG. Equation (17) provides the average reading of pitch acceleration about the store CG, with the second term removing an effective pitch acceleration component due to roll acceleration when the two accelerometers are at different lateral locations. Yaw acceleration does not affect this parameter and translational acceleration along the z-axis is canceled out in equation (17). Ly,roll and denote the positive lateral distances between the z-axis accelerometers measuring roll and pitch, respectively. Due to the store rigid body, the pitch accelerometers do not have to be equidistant longitudinally or vertically from the store CG or center of rotation.
The magnitude of the store yaw acceleration Q, was computed using two linear y-axis accelerometers placed on the same side of the store's longitudinal centerline with one positioned forward of the store CG and the other aft of the store CG. Equation (18) provides the average reading of yaw acceleration of the store CG. The second term in equation (1 8) removes an effective yaw acceleration component due to roll acceleration for the case when the two accelerometers are at different vertical locations. Pitch acceleration does not affect this parameter and translational acceleration along the y-axis is canceled out in equation (18). The values of Lx,ymI. and Lz,ym, correspond to the positive longitudinal and vertical distances, respectively, between the accelerometers used for the yaw acceleration . calculations. The yaw accelerometers do not have to be equidistant laterally or longitudinally from the store CG or center of rotation due to the rigid body.
Forces and moments at the store CG require accurate acceleration data. Data for equations (16) through (1 8) are from accelerometers placed on the store in such a way to eliminate interference fiom other accelerometers. Translational accelerations (equations 13 through 1 5 ) are already corrected for interference effects.
Inertial Loads
If coordinate axes are selected such that they are the principal axes with an origin at the body mass center or at a point fvred in inertial space, and if the angular velocity (r3 of the coordinate system is the same as that of the body, the rotational motion of a rigid body can be described as: Although the mass moments of inertia are found from mass property measurements, the translational velocities ( v, , v, and v, ) and angular velocities ( u ,~, U, and wz ) cannot be found directly and must be approximated numerically by integrating over any time step (t2 -t , ) as The forces resisting the inertial loads, as calculated with equations (20) and (21), are equal in magnitude but opposite in direction to the inertial forces as found with the accelerations experienced by the store during any particular maneuver.[3[ The external reactions resisting a particle undergoing a maneuver must then be in the opposite direction of the resisting force, or in the same direction as the acceleration of the particle. Therefore, equations (20) and (21) not only represent the forces and moments felt by the store due to the measured and calculated accelerations, these equations also represent the reactions experienced externally by the bomb rack unit that are resisting the stores inertial loads. The coordinate system used for the store geometry as well as all inertial loads and reactions is defined by the right-hand convention shown in figure 2-1. Y ti.
Reactions at SRI
The lateral distance between two swaybrace contact points The accurate distribution of the three forces F , F,, Fz and three moments A4, M,,, M, from the store CG to the SRI was critical in calculating the reactions in the BRU-32. Once the loads at the store CG were dynamically calculated for each time step as described in equations (20) and (21), the resulting reactions at the individual interface points were found through an assumption of quasi-static force and moment equilibrium. Using the principle of superposition, the six load cases were considered independently and the calculated interface reactions were combined to form the overall interface reactions at each time step.
There are six points of contact between a store and a typical aircraft bomb rack: four swaybrace rods and two lughook combinations. The swaybrace rods can only be loaded via compressive forces and unloaded due to a reduction of compressive forces. The swaybrace rods cannot support tensile loads and the ball-and-socket design prohibits any moment reactions; physical separation between the store and swaybrace pad is possible. The hook reactions can only support tensile loading from the lugs; separation between the hook and lug is possible.
The BRU-32 is a steel bomb rack and is assumed to have rigid body characteristics. Figure 2-2 shows the sign convention and layout geometry of a typical SRI assuming the requisite 30-inch lug spacing of a larger store (versus smaller stores requiring 14-inch lug spacing).
Lugs (Aft lugvisible, within a single assembly is typically 4.24 inches. Reloads in the interface points will typically be found in the hooks and swaybrace rods during loading of a store onto a bomb rack. The magnitude of the preloads are likely to change or redistribute during taxiing and flight maneuvers but will generally return to the original values in steady state flight. These loads can be predicted based on historical averages and consistent loading techniques, but must be accounted for in the reaction equations. Although the swaybrace rods cannot react in tension (nor can the lugs react in compression), the swaybrace rods are not restricted from unloading due to the presence of a tensile reaction and thereby having the compressive reaction reduced as if the rods were responding in tension. Furthermore, the lugs could effectively react in a compressive fashion if there were already a sufficient tensile reaction present as to not allow the lugs to unload into an overall compressive state. Therefore, the magnitudes of the reactions at the interface points were changed to maintain equilibrium with the unloading reactions. To account for unloading, the reactions calculated for each interface point were split equally between the reacting component and the unloading component. Throughout the equations outlined below, the interaction locations experiencing an unload during either a positive or negative load will have the magnitude of that reaction reduced until the reaction has reached zero (fully unloaded) at which point the entire load in that direction will only be reacted by the interface point that is increasing in magnitude. The reactions at the interface points due to each load type represent the total magnitude of each reaction expected in the SRI during a state of zero preload. However, some amount of preload is present during all flight-testing and will generally be of sufficient magnitude to keep the interface points fkom ever reaching zero during typical flight maneuvers. As the store experiences external forces, opposing reactions will simultaneously load and unload equally to maintain equilibrium. In all six load cases, the four swaybrace rods will have a loading and unloading portion. For example, a positive pitching moment K,, will proportionally add to the forward swaybrace reactions while decreasing the aft swaybrace reactions. Likewise, a positive lateral load Py will increase the reactions on the two starboard swaybrace rods while proportionally decreasing the port swaybrace rod reactions. The reaction due to a vertical load P, must be split between the swaybrace rods and the hooks. Similar to the swaybrace rods, the forward and aft lugs will split the reactions between loading and unloading portions due to a pitching moment My as well as the vertical reaction due to a longitudinal force P,. The
Program to Calculate Reactions at the SRI
A FORTRAN program was created to calculate the SRI reactions at each time step for ground or flight-testing, as all necessary data must be entered for each test, and requires the following information: the physical parameters of the store, including weight, radius, accelerometer layout, CG, moment of inertias, etc.; the store and aircraft aerodynamic parameters; the raw accelerometer data; and the preflight preloads at the SRI. The model was modified and verified with ground and flight test data. Once all of the necessary data was entered into the program, the translational and rotational accelerations at the store CG were found with equations (13) through (18). Inertial forces and moments at the store CG were then calculated with equations (20) and (21), respectively. The reactions at the SRI were found using superposition and saved to data files for review.
GROUND TEST AND ANALYSIS
Various experimental procedures were required to collect the ground and flight test data and validate the derived reaction equations. After solving for the theoretical equations, the first stage of the experimental investigation was to develop and conduct a ground test to verify the distribution of loads at the store CG to the six SRI points due to inertial loads only. The ground test data would therefore validate the original assumptions in the SRI reaction equations or support empirical modifications.
For the static ground test, a rigid test cell was assembled and a MK 84 bomb body with a conical tail f i n was mounted on a BRU-32. The rack was affixed to large steel beams, ensuring that limited flexibility was present in the testing apparatus. Additional lugs were welded at various locations on the MK 84; cables attached to hydraulic actuators applied loads to the store via the new lugs. The locations of the welded lugs were based on known distances fiom the store CG as determined by store mass properties testing. Loads were applied at these various lug locations to mimic inertial loads through the CG by creating known force and moment combinations. Loads and strain gage data from instrumented lugs and swaybrace assemblies were recorded and reduced as reaction forces.
Instrumentation and Equipment
Six additional lugs were welded to the bomb bodythree on each side of the store and vertically centered on the longitudinal axis. One lug on either side of the MK 84 was located at the store CG, and CG, intersection and the remaining two lugs per side were symmetrically placed 36inches forward and aft of this point. Each lug weld was rated for a minimum 10,000 lbs. of shear or tensile load and was attached perpendicular to the CG axes vice normal to the local store surface. Loads were applied to the welded lugs via braided steel cables. To simulate single axis loading or specific force-moment couples, one or more lugs were loaded simultaneously with one or more hydraulic actuators. Spreader bars were used to symmetrically load two lugs simultaneously with one actuator. Figure 3-1 shows a typical welded lug on the MK 84. The BRU-32 used in the ground test was fully instrumented with strain gages installed on the lugs and the swaybrace assemblies. Data were captured on an Astro-med computer system at 100 samples per second (sps) during all tests and saved digitally and on strip charts. Figure 3-2 shows the MK 84 and the instrumented SRI attached to the test cell. Multiple load cells were required to apply the loads fiom the hydraulic actuators. Loads were recorded and tracked with the voltmeters. A typical load cell is shown in figure 3-3, including the assembly attaching the load cell to a welded lug and the necessary wiring to record loads data. The MK 84 and BRU-32 were attached to a rigid, steel Ibeam test cell to provide a foundation for the test loads. The I-beams were assembled with a forklift and fastened Vol. 6-2733 9 together with 1-inch bolts and doubler plates. A total of eighteen 10-foot beams, seven 20-foot beams and two 8foot beams were used in the construction. After attaching the instrumented BRU-32 to the I-beams, the MK 84 was attached to the rack via standard USN loading procedures. Swaybrace rods were set in the extended position and not additionally tightened. Real-time loads data were captured at all test points at 100 samples per second (sps) as well as voltage and reaction data from the four instrumented swaybrace rods and two instrumented lugs. Calibrations were conducted before and after each load case. Figure 3- During each test, a load was applied to the store from 0 Ibs. to the maximum of 10,000 lbs. (approximately 5 Gs of translational acceleration) and back to 0 lbs. For the load cases requiring the use of two hydraulic actuators, 5,000 lbs.
of force were applied from each actuator and the outputs monitored using separate voltmeters to allow consistent and syinmetric buildup of load on the welded lugs.
Load case 1 simulated a longitudinal force at the store CG and was loaded along the x-axis. Two hydraulic actuators each applied 5,000 lbs. of force simultaneously to the two, middle welded lugs located symmetrically on both sides of the store. Load case 2 consisted of a pure lateral pull of 10,000 lbs. through the store CG via one of the middle, welded lugs. Load cases 3 and 4 consisted of a pure vertical pulls upward and downward, respectively, of 10,000 Ibs. through the store CG using the center welded lugs and a spreader bar with one hydraulic actuator. Load case 5 modeled a pure moment about the longitudinal axis (MJ by simultaneously pulling up with 5,000 lbs. on the middle, starboard lug and pulling down on the middle. port lug. As shown in figure 3-6, load case 6 was the only combined load to be modeled. A hydraulic actuator and spreader bar were used to pull the nose upward with 5,000 lbs. of force using the two forward, welded lugs. An additional 5,000 lb. lateral load was applied to the middle, welded lug on the port side. The four swaybrace rods were also instrumented with Vishay strain gage circuits (model number CEA-06-062UT-350) as shown in figure 3-9; these gages were STC for steel with 350 ohms resistance. The gages on each rod were connected in standard, four-circuit bridges to cancel the bending load and fiction measurements.
FIGURE 3-9: Swaybrace Rod with Strain Gages
Although the swaybrace rod is threaded into the swaybrace assembly, the rod is either engaged in a fully extended or fully retracted position. For stores with 30-inch lug spacing as in this test, the swaybrace rods are typically set in the fully extended position and not additionally tightened or torqued further after the store is loaded in the rack. Each rod includes a ball-and-socket pad that was in contact with the store: as with all ball-and-socket joints, moments could not be reacted in the swaybrace rods. For each load case listed in table 3-1, the buildup to the maximum external load was accomplished in two to three minutes while unloading took an average of less than 30 seconds.
The swaybrace gage data were combined in the bridge and recorded as one overall strain reading per swaybrace per time step. The lug strain data, when considered versus total external force, defined the linearity and sensitivity of each gage to the particular load type and was useful in determining which gages were to be included for the reaction force calculations. The raw strain data from the instrumented lugs and swaybrace rods were combined and converted to the reaction data for each interface point. Swaybrace rod preloads for each load case were recorded; lug preloads were calculated using the recorded swaybrace preload data and weapon weight. Using equations (25) through (33), the interface reactions were calculated based on the known preloads, external loads, mass properties and geometry of the store and instrumentation; the calculations were completed with the FORTRAN program created for this research. This program incorporates the equations outlined in the previous chapter for converting store accelerations into forces and moments and then distributing the forces and moments from the store CG to the SRI.
Model VeriJcation with Ground Test Data
The reaction data in the lugs along the longitudinal axis did not show an equal split between the lugs due to a longitudinal load as was assumed in equation (29). Using the measured data, it was calculated that an average of 38.5% of the longitudinal load was reacted by the forward lug while 60.6% of the data was reacted by the aft lug. The remaining 0.9% of longitudinal load was attributed to a bending reaction in the swaybrace rods that was not recorded due to the circuit design. The longitudinal lug reactions due to a longitudinal force P, ( During the buildup of the external load, the interface points either react to the loading force or the unloading force as predicted. It was noted, however, that during a lateral external force P, the two swaybrace rods opposite those reacting in compression did not unload but held the preload as a constant value. Also, the assumption of the swaybrace rods reacting the entire lateral portion of the lateral load and the lugs not reacting in a lateral response proved to be correct. One point not predicted with the classical analysis was the reaction of the swaybrace rods due to a yawing moment about the vertical z-axis. The data showed that a positive yawing moment (nose left) was laterally reacted in the forward lug in the starboard direction and in the port direction for the aft lug as predicted, although higher in magnitude. The forward starboard and aft port swaybrace rods reacted in compression, however, instead of the assumed reactions in the forward port and aft starboard swaybrace rods. Since the store is more rigid than the bomb rack itself, it was determined that the rack was twisting into the opposite swaybrace rods creating the unpredicted reactions. Due to the unusually high yawing moment imposed for this test, this behavior has not been seen, or at least not been recognized, before. The developed equations were corrected to allow the full yawing moment to be reacted by the lugs while the opposite swaybrace rods were given a very small and empirically determined amount of reaction due to rack twist based on the amount of observed The SRI reactions to the externally applied load during buildup, while all interface points still contained preload, was reduced in magnitude to allow half of the predicted reaction to go towards the reacting points (increasing in load) and half to go to the respective unloading points. Unlike actual flight tests, however, the dominance of the one specific load type in each ground test load case quickly overpowered the preloads and was fully reacted by the necessary points. The equilibrium adjustment shown by example in equations (25) through (33) was only necessary until all of the unloading reactions went to zero and the reactions to the applied load were fdly carried by the interface points already increasing in load.
Graphs of the measured and calculated reactions for all eight load cases were created. Three graphs showing the reactions in the forward and aft swaybrace assemblies and the vertical lug were created to describe each load case. Load case 1 shows the only occasion that the lugs reacted longitudinally; it is the only test case with an external longitudinal load. The distribution of approximately 38.5% and 60.6% of the external load to the forward and aft lug reactions, respectively, can be seen in figure 3-11 as well as the comparisons between the calculated and measured data. The total applied force for load case 1 was approximately 10,000 lb. in the aft direction. The lug reactions were shown to be primarily linear in response, especially in the critical high load areas. These results match previous flight testing which showed linearity in the response of the hook/lug reaction and proposed that a longitudinal load would be unevenly reacted between the lugs[ 181.
Ground Test Summaly
This chapter focused on the static ground test conducted to uncover the true distribution equations relating loads at the store CG to the six points in a SRI. A MK 84 instrumented with welded lug attachment points allowed external loads to be applied to the store to simulate pure forces and moments at the store CG. Collected test data included strain gage output from 42 gages affixed to the lugs and swaybrace rods in the bomb rack. The applied forces were also collected with respect to time. Test points included eight different load cases, including forces through the store CG along the x-, y-, and z-axes, moments along the three axes and one combined load case. The final, measured data was reduced to 21 point sets for each load case, including 10 points during build up and 10 points during the unload in addition to the point of maximum applied load for ease of comparison and calculation.
The swaybrace assemblies by design cannot react in tension and the lugs cannot react in compression, yet both can unload and reduce a current load state. The corrections to the reaction equations to allow concurrent loading and unloading (until the unloading reaction reached zero) matched well with observation. The direct solving of a full set of loading and unloading equations was impossible due to the parallel nature of the loads and reactions and an indeterminate system of equations.
The FORTRAN program created to calculate the interface reactions using the developed equations was corrected based on the above observations; predicted results generally matched the measured reactions with a few exceptions. It was noted that the only axial reaction exhibited by the lugs was in the presence of a longitudinal load but the reaction was not evenly split. The forward lug carried 38.5% of the total longitudinal load while the aft lug reacted over 60%. The lugs reacted laterally only during the application of roll and yaw moments as seen in load cases 5 and 8, respectively. The swaybrace rods reacted differently than predicted during the application of a yaw moment, although the reactions were quite low and were attributed to a twisting of the bomb rack. Most of these variations in reaction were shown to be possible in [18], which also stated deflections in the pylon would not noticeably effect lug reactions. Although not predicting a twisting in the rack as proposed in this research, it shows that structural problems might occur regardless of rigidity.
Unlike flight-testing where multiple load combinations on a store are always present, a single dominant load in one direction was unique to this ground test. With the exception of load case 2, each load case completely diminished the preloads in the unloading interface points very quickly and the reacting points carried the entire load. During the ground test, this was seen in the initial swaybrace rod and vertical lug reactions as predicted; calculated and measured data showed high correlation as reaction slopes increased early into the load buildup.
As this analysis focused on the high-reaction loads and results, SRI reactions near zero were negligible. The maximum reaction magnitudes acceptable during flighttesting are currently approximately 20,000 lb. and 50,000 lb. per swaybrace rod and vertical lug, respectively, although the ultimate loads for both interface points are much higher. To quantify error observations, it was decided that the critical reaction points for this research would be any reaction equal to at least 10% of the maximum reaction magnitude typical of flight-testing. To that end, reactions below 2,000 Ibs. in the swaybrace rods and 5,000 lbs. in the vertical lugs were considered to be in the noise of the data and consequently neglected in the final error analysis.
FLIGHT TEST AND ANALYSIS
Flight-testing was used to validate the overall inertial model of the SRI reactions. The analytical models of inertial forces at the store CG found in chapter 2 via Newton's Second Law and Euler's equations were first verified with the accelerations found from store instrumentation data. The SRI interaction relationships were derived empirically by modeling the SRI reaction equations found via the ground test and comparing the expected results with actual flight test data from carrier suitability flight-testing. Aerodynamic loads immediately after an arrestment or catapult were considered negligible when compared to the inertial forces present at this same time. Assuming typical arrestment airspeed of 150 knots (approximately 250 "/sec), the dynamic pressure at sea level is approximately ?h psi. Assuming a reference area of 165 in2 for the GBU-24 crosssection, the aerodynamic load is approximately 80 Ib. Compared to nominal landing accelerations immediately after arrestment of approximately 8Gs and -2Gs in the vertical and longitudinal directions, respectively, the percent of total load attributed to aerodynamic forces for a 2,000 lb.
class store is approximately %YO and 2%, respectively.
Arrestment Flight Test
As only the predictions of the worst case interface loads were necessary, data for larger, heavier stores were desired. Therefore, this research only considered a 2,000-lb. class store, the largest general class of store, with 30-inch lug spacing. Smaller stores have inherent problems due to radius and were not examined here. [3,19] Flight test data were available for carrier suitability testing of a GBU-24 on an F-14 aircraft. Rigid body assumptions were possible, as the tested GBU-24 consisted of a steel BLU-I09 bomb body; the BLU-109 has a store diameter and wall thickness of 14.5" and 1.125," respectively, and is approximately 14 feet long. [20] For the required equations of motion as derived in section 2 and corrected in section 3, data collection was dependent on the accelerometer data and their locations. The GBU-24 flight test program incorporated ten servo-accelerometers on the store. These accelerometers were selected to perform in the frequency range optimal for store analysis with 100 G amplitude.
The storehacklpylon assembly natural frequency is typically in the 20Hz -40range and generally less than 60%. Being much stiffer, the store has natural frequencies of vibration much higher than that of the assembly.
Accelerometers are generally placed on orthogonal blocks vice directly on the store, allowing all of the acceleration vectors to line up in one of the three orthogonal axes and not have to be individually corrected. Experience has shown that the accelerometers placed on the less rigid tail and nose sections of a GBU-24 are more difficult to align, calibrate and obtain accurate information than those placed directly on the rigid BLU-109 portion of the GBU-24. Additionally, the nose and tail sections also have an inconsistent geometry when compared to the main section of the BLU-109 bomb body, requiring geometric corrections when used with the remaining six accelerometers. Therefore, the six accelerometers on the main body of the GBU-24 were the only ones used for the calculations and predictions in this research. A triaxial accelerometer group located longitudinally at the store CG also allowed ease in the calculation of translational accelerations with minimal coupling effects f?om other motions or accelerations along or about other axes. The physical characteristics and lug spacing of the GBU-24 with a BLU-109 bomb body are shown in figure 4-1, as well as the accelerometer placement used for flight-testing.
initial arrestment as well as when the inertial loads have been significantly dampened, allowing a minimal time interval for data analysis.
I+Az. Within a few hundred milliseconds, accelerations near +9 Gs and -5 Gs created forces and moments of approximately 15,000 lbs. and 150,000 in-lbs, respectively. Once inertial load calculations were completed at the store CG, the FORTRAN program used the information to predict the reactions at the SRI. As with the ground test data, preload values must be incorporated to accurately predict interface reactions; for this project, preloads were measured and known. Unfortunately, in a real-time flight test environment with only accelerometers on the store, real-time preload information will not be available. Although the ambient conditions as the aircraft settled in for the arrestment approach were generally near steady state 1 G flight, the total load state was still approximately 0.2 Gs and 1.1 Gs in the x-and z-directions, respectively. This imbalance increased or decreased the reaction at some SRI points due to an initial non-zero load state. Figure 4-2 showed the ambient accelerations prior to the arrestment.
CG @-Ax
As errors in preloads can lead to inaccuracies in the interface reaction calculations, careful consideration must be given to their estimation. Fortunately, past preload data has shown that for similar flight tests, historical averages can be used without inducing large errors since preload values are at least an order of magnitude less than the total reactions observed at the peak accelerations.
The percentage errors at low reactions may be large if preloads are used incorrectly, but the results at those lower levels are not critical. Additionally, the swaybrace rods are generally in the extended position for all stores with a 30-inch lug spacing yielding approximately the same preload each flight given the same store weight. Flight maneuvers, however, effect preload due to redistribution until the reactions have a chance to settle and return to steady state conditions within the dynamic environment.
When preload estimation is necessary, specialized swaybrace rods may be used that allow use of a torque wrench. These swaybrace rods may be additionally torqued after store loading to an arbitrary, initial setting (typically 1,000 lb. each). The magnitude of the steady state preload may change after completion of each flight maneuver, but the swaybrace and lug preloads will typically redistribute to values near their original settings once normal, steady state flight is resumed. Using an average preload from empirical data, any errors will generally be less than an order of magnitude smaller than the peak reactions.
Model VeriJication with Flight Test Data
The FORTRAN program calculated the six interface reactions at each time step and compare the results to the reactions directly measured by the instrumented rack. Unlike the ground test points involving loads along a single axis, the interactions between the various directions of flight test loads required examination. Constant preload values were known from the measured data for this test and were entered directly into the analysis. The FORTRAN program was modified to divide each reaction into the contributing components before they were subsequently combined by superposition, allowing the interaction of loads along various axes to be seen and allow for better data prediction. The predictions of the swaybrace rod and lug reactions generally matched well with the measured critical reactions. Critical reactions were defined conservatively as those reactions above 10% of the maximum allowable reaction in the swaybrace rods and lugs, or greater than 2.000 lbs. compression and 5,000 Ibs. tension, respectively. The main areas of poor agreement between measured and calculated values were in the very low reaction range (less than 1,000 lbs.) of the swaybrace rods. This is attributed to the noise associated with near-zero values in the accelerometer data, as well as not being able to more accurately model the conditions involved in the final release of reaction loads within the SRI; this area was not of primary concern due to its very nature. The higher compressive reactions in the swaybrace rods were predicted more accurately in all cases, including good correlation in overall shape, general response and magnitude between the measured and predicted data. Similar agreement was found in the vertical lug reactions between measured and predicted data.
The average swaybrace rod reaction error was 239 lbs. or 9% of the measured reactions; all critical swaybrace errors were less than 900 Ibs. Only 6 of 154 critical swaybrace reaction errors were greater than 600 lbs., while only 10 errors were greater than 20% of measured values. The average lug reaction error was 443 Ibs. or 7% of the measured reactions; all critical lug errors were less than 1,500 lbs. Only 6 of 78 critical lug reaction errors were greater than 1,000 lbs., while only 8 errors were greater than 15% of measured values.
FIGURE 4-18: Reaction Errors at Critical Points ('YO)
In calculating the reactions, preloads were combined with the new reactions at each time step and the total load state examined. If lug or swaybrace rod fully unloaded, the reaction was defined as zero and could not be reduced 17 hrther. The total load state for a lug reaction must remain positive (tension) or zero; swaybrace rods must stay in a total state of zero or negative (compression) load.
Flight Test Summary
Flight tests were conducted incorporating a 2,000 lb. GBU-24 on an F-14 Tomcat aircraft. Immediately after aircraft arrestment, the dynamic inertial loads were shown to be dominant while the aerodynamic loads were negligible. The GBU-24 was instrumented with accelerometers; the BRU-32 included instrumented lugs and swaybrace rods. Accelerometer readings ranged from -1.8 Gs to +1.1 Gs along the x-axis, -0.8 to +0.8 Gs along the y-axis, and -4.5 Gs to +8.7 Gs along the z-axis. After modifying the FORTRAN program to accept flight test data. the forces and moments at the store CG were calculated fiom basic equations while linear and angular rates were integrated across consecutive time intervals. The FORTRAN program also accepted variations in the number and placement of accelerometers to provide flexibility for various stores and test requirements.
In the 300 milliseconds immediately following arrestment, forces and moments up to approximately 15,000 lbs. and 150,000 in-lbs., respectively, were found. The forces and moments at the store CG were distributed to each of the six the interface points. Comparing the calculated reactions to the measured data, good agreement in form was found for all interface points; critical lug and swaybrace rod reactions averaged less than 7% and 9% error, respectively.
CONCLUSIONS
A classical analysis was used to calculate the reactions at the SRI using only measured accelerometer data, known mass properties and geometry; strain gage data were not required. Translational and rotational accelerations were found at the store CG using accelerometer data recorded on the store. Using the calculated accelerations and rates, equations required for the inertial forces and moments at the store CG were derived with Newton's Laws, kinetics of rigid bodies and the simplified Euler's Equations of Motion.
The calculated loads at the store CG were distributed to the six interface points using the principle of superposition to sum the individual reactions at each interface. Elastic effects were considered negligible for this analysis. A FORTRAN program was written to calculate the inertial dynamic loads and their distribution to the SRI.
The principal outcome of this research was the development of the first real-time, fully dynamic analysis of store reactions in an aircraft bomb rack without reliance on strain gages or instrumented suspension equipment. Previous attempts at similar programs to assist in structural flight test planning resulted in very conservative predictions; ongoing efforts in military standards similar to [3] qualify the active interest in reaction prediction methodology. Calculating the SRI reactions within a confident error band during the planning stages or actual flight-testing will save critical time and money. This research documents the first usable, accurate and repeatable version of those interests.
The successful completion of this research also provides numerous benefits for many structural flight test programs. Time and cost savings are primarily realized through fewer required test flights due to a confident prediction of the results at the necessary test points and hence fewer required flights. Further savings are obtained by minimizing: preflight instrumentation (including store and rack); post-flight data reduction; and repeated and cancelled test flights due to instrumentation failures in strain gages and specialized racks. Relieving the dependency on strain gages for reaction measurements is another advantage of this research; gages are costly in application and require specialized suspension equipment. The gages are also very &agile and easily damaged during the typical installation procedures for larger stores, resulting in poor data collection, wasted effort in the delay or repetition of test flights and added costs.
Additional applications of this research include: incorporating the prediction routines as secondary reaction calculations in case of primary instrumentation failure; using the program to compute the critical reaction points prior to actual flight-testing, thus allowing the test team to target a specific flight envelope instead of wasting test resources on irrelevant test points or repeated flights; and implementation of this research into a store's development phase to provide the design engineers with unique and critical SRI loading conditions prior to store fabrication.
The present research has shown that, in a limited form, the SRI reactions can be predicted and calculated real-time. Future work should consider four key areas, including the addition of aerodynamic load calculations, incorporating the program into a real-time telemetry system for flight-testing, gathering a larger database of flight data to refine the reaction equations, and improving preload estimation. Enhancements in any of these areas would increase the model's usefulness and reliability as a reaction prediction tool for flight-testing as well as design and analysis. | 12,436.2 | 2003-03-08T00:00:00.000 | [
"Physics"
] |
catena-Poly[[[diaquacopper(II)]-μ-quinoline-2,3-dicarboxylato-κ3 N,O 2:O 3] monohydrate]
In the title compound, {[Cu(C11H5NO4)(H2O)2]·H2O}n, the CuII ion is five-coordinated by two O atoms and one N atom of two symmetry-related quinoline-2,3-dicarboxylate ligands, and two water molecules. The water molecules occupy basal and apical positions of the square-pyramidal coordination polyhedron. Each quinoline-2,3-dicarboxylate dianion bridges two adjacent CuII ions, forming a polymeric chain along [010]. The chains are further connected via O—H⋯O hydrogen-bonding interactions and quinoline ring π–π interactions [centroid–centroid distance = 3.725 (4) Å], generating a three-dimensional structure. Lattice water molecules participate in the crystal structure via O—H⋯O hydrogen bonds.
Experimental
All commercially obtained reagent grade chemicals were used without further purification. A mixture of copper chloride dihydrate (0.1708 g, 1 mmol) and 2,3-quinolinedicarboxylic acid (0.2171 g, 1 mmol) was added into 20 ml of water with few drops of ammonia solution, and then stirred for 1 h. After 2 days, blue crystals of the title complex were collected by filtration, washed with distilled water, and dried in air.
Refinement
All H atoms bonded to C atoms were positioned geometrically and refined using the riding model with C-H = 0.93 Å.
The H atoms of water molecules were located from a difference map and were restrained at distances O-H = 0.83 (1) Å.
The separation between H atoms in the same water molecule was restrained to H···H = 1.35 (1) Å. Cu and OW2 atoms were restrained to have similar displacement parameters (SIMU restraint; Sheldrick, 2008). Isotropic displacement parameters for H atoms were calculated as U iso (H) = 1.2U eq (carrier C) and U iso (H) = 1.5U eq (carrier O).
Figure 2
Crystal packing diagram for the title compound. All atoms are shown as isotropic spheres of arbitrary size. H atoms bonded to C atoms are omitted for clarity. The H-bonding interactions are shown as red dashed lines. -2,3-dicarboxylato-κ 3 N,O | 431.2 | 2012-10-20T00:00:00.000 | [
"Chemistry"
] |
Bi-hemispherical Canopy Reflectance Model with Surface Heterogeneity Effects for the Estimation of LAI and fAPAR from MODIS White-Sky Spectral Albedo Data
: Bi-hemispherical reflectance (BHR), in the land surface research community also known as “white-sky albedo”, is independent of the directions of incidence and viewing. For vegetation canopies, it is also nearly independent of the leaf angle distribution, and therefore it can be considered an optical quantity that is only dependent on material properties. For the combination leaf canopy and soil background, the most influential material properties are the canopy LAI (leaf area index), optical properties of the leaves, and soil brightness. When the leaf and soil optical properties are known or assumed, one may estimate the canopy LAI from its white-sky spectral albedo. This is also because a simple two-stream radiative transfer (RT) model is available for the BHR of the leaf canopy and soil combination. In this contribution, crown clumping and lateral linear mixing effects are incorporated in this model. A new procedure to estimate soil brightness is introduced here, even under a moderate layer of green vegetation. The procedure uses the red and NIR spectral bands. A MODIS white-sky albedo product at a spatial resolution of 0.05 ◦ is used as a sample input to derive global maps of LAI, soil brightness, and fAPAR at the local moments of minimum and maximum NDVI over a 20-year period. These maps show a high degree of spatial coherence and demonstrate the possible utility of products that can be generated with little effort by using a direct LUT technique.
Introduction
The correct quantitative interpretation of remotely sensed land surface reflectance data is hampered by the anisotropy of most surfaces on Earth, whether it be bare soils, vegetation canopies, or water bodies. This is why, after the correction of atmospheric effects, attempts are often made to remove or normalize these so-called BRF (bi-directional reflectance factor) effects. Various parametric BRF models are frequently used in this case [1] and are inverted on the basis of several directional observations collected by satellite missions during a period of less than two weeks. This is performed in order to prevent large temporal changes in an object's properties. The output product is mostly the bi-directional reflectance normalized to a particular standard situation, e.g., nadir viewing under a 45 • solar zenith angle; however, for energy balance studies, hemispherical reflectance products are of more interest, since these express the surface's total reflectance response to the incident radiation from the sun and the sky. The hemispherical reflectance for incident solar radiation, called the DHRF (directional hemispherical reflectance factor) [2], is most important in this respect, since optical satellite observations are primarily only possible under nearly cloud-free conditions, so that the direct solar radiation input dominates. This quantity is sometimes called "black-sky albedo" and is obviously a function of the solar zenith angle. This is very useful for realistic energy balance calculations, but for the retrieval of inherent surface properties it is still a complication. Once again making use of the parametric BRF model, with angular integration of the DHRF over the incident hemisphere, it is possible The paper is organized as follows. Section 2 presents several radiative transfer modelling concepts and bi-spectral image processing approaches. Section 2.1 presents the two-stream BHR model in which the soil background is included with the adding method. Section 2.2 discusses feature space plots of the red and NIR bands to illustrate the effects of soil brightness and LAI and presents a new concept to retrieve soil brightness. In Section 2.3, the effects of crown clumping and linear mixing on moderate-resolution pixels are considered. Here, three models of surface heterogeneity are also introduced and their manifestation in feature space plots is presented. The estimation of fAPAR based on the three heterogeneity models is described in Section 2.4, and Section 2.5 describes the application of a direct look-up table (DLUT) technique applied to global satellite data to generate LAI, soil brightness, and fAPAR maps. The results are presented in Section 3 and discussed in Section 4, while the conclusions are presented in Section 5.
Bi-hemispheric Reflectance Model
In the 1D turbid medium radiative transfer model SAIL [3], an adding method [4] is applied to calculate the bi-hemispherical reflectance of the combination soil and canopy, which is given by the simple expression below [5]: where r s is the reflectance of the soil background and the right term describes the multiple reflections between canopy layer and soil background. To clearly express the downward and upward flows of diffuse radiation through the canopy required to account for the soil's influence, the hemispheric canopy transmittance τ dd appears twice in Equation (1). For a small LAI, this transmittance approaches unity, so it includes the direct transmission of light without any contact with leaves. The adding method is a simplified implementation of what has been termed by other investigators a combination of the so-called "black soil problem" and "S-problem" [6,7] in their explanations of numerical 3D radiative transfer as a particle transport problem, similar to approaches applied in the numerical modelling of neutron transport in nuclear reactor physics. In the much simpler analytical two-stream radiative transfer theory applied in SAIL, the bi-hemispheric reflectance and transmittance of the isolated canopy layer (the basic solutions of the black soil problem), ρ dd and τ dd , are given by well-known expressions: (2) where L is the leaf area index (LAI). The quantities m and r ∞ are known as the diffusion exponent and the so-called infinite reflectance, which is the BHR of a hypothetical canopy with an infinite LAI. Both quantities are functions of single leaf reflectance, ρ, single leaf transmittance, τ, and the leaf inclination distribution function, LIDF, which is symbolized as f (θ l ), where θ l is the zenith angle of the leaf's normal. The leaf azimuth distribution is assumed to be uniform. Together, these quantities determine the diffuse backscattering coefficient σ and the net attenuation coefficient a as follows [3,7,8]: where θ l is in radians and γ = π/2 0 f (θ l ) cos 2 θ l dθ l . Table 1 shows some values of γ for a few common LIDF types. They range from 0 to 1, but since in Equation (3) they are multiplied by half the difference of leaf reflectance and leaf transmittance, the effect of γ on these quantities will never be large, since leaf Remote Sens. 2021, 13, 1976 4 of 22 reflectance and transmittance usually do not differ by more than 0.05. Note in column 2 of the table that δ is the Dirac delta function. The diffusion exponent m is the eigenvalue of the coupled system of two-stream differential equations [5,7] and is given as follows: where η = a + σ = 1 + (ρ − τ)γ and α = a − σ = 1 − ρ − τ denotes the single leaf absorptance. The infinite reflectance, r ∞ , finally is given by the following equation: Figure 1 shows an example of the spectra of r ∞ and m (unitless) as obtained from the leaf optical properties model Fluspect-B [9], which, regarding leaf reflectance and transmittance, is very similar to PROSPECT-5 [10], except that it still includes brown pigments, as in older versions of PROSPECT [11]. In this typical case of a green leaf, the biochemical inputs are the ones listed in Table 2. Table 2.
The substitution of Equation (2) in Equation (1) gives an analytical expression for the canopy BHR in which the soil background, with reflectance s r , is also incorporated. This yields the following quasi-linear combination of s r and r : Table 2. The substitution of Equation (2) in Equation (1) gives an analytical expression for the canopy BHR in which the soil background, with reflectance r s , is also incorporated. This yields the following quasi-linear combination of r s and r ∞ : Further elaboration of Equation (6) provides the BHR of the combination of the soil and canopy with a slightly simpler equation: The quantity r s = r s −r ∞ 1−r s r ∞ . expresses the spectral soil/vegetation contrast. From Equation (7), it can be concluded that the sensitivity of r dd tohe LAI will be most favourable for large values of r s ', i.e., when there is a large contrast between r s and r ∞ . Spectral regions in the red and the near-infrared (NIR) parts of the spectrum both allow a large soil/vegetation contrast, and therefore both are very sensitive to the LAI, since in the red, the r ∞ for green vegetation is very small (<0.05) and in the NIR it is high (>0.60, Figure 1) when compared to the moderate reflectance of the soil in these spectral regions; however, the NIR has the advantage of featuring a much smaller value of the diffusion exponent m, which means that saturation at high LAIs occurs much later than in the red band. It turns out that, in general, the reflectance in the red band is particularly sensitive to low LAIs, whereas the reflectance in the NIR band is relatively more sensitive to the LAI when it is high. Also, should the soil's reflectance in either the red or the NIR band be close to the corresponding r ∞ , then the soil/vegetation contrast in the other band will be extra high, so both spectral regions will always complement each other. These cases may occur for either very dark or very bright soil backgrounds. Figure 2 shows how various combinations of soil brightness and canopy LAI appear in a red-NIR feature space plot. In this paper, the soil's reflectance in the red is adopted as a measure of soil brightness. For reflectance values obtained from various crops observed in the field, the resulting quasi-triangular shape is also known under the name of "tasselled cap" [12]. In the publication of Kauth and Thomas [12], the tassels mark the various paths in the feature space that are followed by cereal crops during ripening. In Figure 2, the tassels are actually missing in this simulated case of exclusively green vegetation. The plot also confirms why the NDVI is a fairly good indicator of the LAI, since lines of constant NDVI go through the origin, and the brown lines of constant LAI in the diagram show roughly the same behaviour, although for the low LAIs located just above the bare soil line this is clearly no longer the case [13]. In Figure 2, all green lines of constant soil brightness appear to join together in a single vertical line. Figure 2, the tassels are actually missing in this simulated case of exclusively green vegetation. The plot also confirms why the NDVI is a fairly good indicator of the LAI, since lines of constant NDVI go through the origin, and the brown lines of constant LAI in the diagram show roughly the same behaviour, although for the low LAIs located just above the bare soil line this is clearly no longer the case [13]. In Figure 2, all green lines of constant soil brightness appear to join together in a single vertical line. This means that in this region of the red-NIR feature space, only a very specific value of the reflectance ( r ) in the red band would allow possible solutions for model inversion.
Red-NIR Feature Space Plots
By allowing a much larger range of soil brightness values, the model's "repertoire" can be extended, and it appears that the triangle then also fills the upper parts of the diagram, as shown in Figure 3; however, the fact remains that solutions of model inversion are only possible for red reflectance values that exceed a certain minimum value, namely the canopy r in the red. This means that in this region of the red-NIR feature space, only a very specific value of the reflectance (r ∞ ) in the red band would allow possible solutions for model inversion. By allowing a much larger range of soil brightness values, the model's "repertoire" can be extended, and it appears that the triangle then also fills the upper parts of the diagram, as shown in Figure 3; however, the fact remains that solutions of model inversion are only possible for red reflectance values that exceed a certain minimum value, namely the canopy r ∞ in the red.
From Equation (7), we may obtain r s e −2mL = r dd −r ∞ 1−r dd r ∞ = r , and therefore the LAI can in principle be solved by the following: so that the LAI can be estimated analytically from the simple equation L = ln(r s /r )/(2m).
Remote Sens. 2021, 13, 1976 7 of 22 in principle be solved by the following: so that the LAI can be estimated analytically from the simple equation = ( ′/ ′)/(2 ). This equation makes it clear that for the estimation of the LAI, knowledge about the soil's reflectance is just as important as the measured reflectance of the soil and canopy combination and the optical properties of the vegetation, as expressed by r and m; however, since the soil's reflectance, in particular its brightness, is usually unknown, one still has to find a way to estimate s r before one can estimate the LAI. Therefore, we will now try to exploit the relationship between the values of soil reflectance s r at two wavelengths.
This relationship is mostly very simple, since the ratio of the soil's reflectance at two wavelengths for a given soil type is usually constant [13][14][15]. One can derive Since ′ = ′ 2 , we find the following equation: This equation makes it clear that for the estimation of the LAI, knowledge about the soil's reflectance is just as important as the measured reflectance of the soil and canopy combination and the optical properties of the vegetation, as expressed by r ∞ and m; however, since the soil's reflectance, in particular its brightness, is usually unknown, one still has to find a way to estimate r s before one can estimate the LAI. Therefore, we will now try to exploit the relationship between the values of soil reflectance r s at two wavelengths. This relationship is mostly very simple, since the ratio of the soil's reflectance at two wavelengths for a given soil type is usually constant [13][14][15]. One can derive r s −r ∞ 1−r s r ∞ = r s ⇒ r s − r ∞ = r s − r s r ∞ r s ⇒ r s = r ∞ +r s 1+r ∞ r s . Since r s = r e 2mL , we find the following equation: This indicates that, theoretically, with a known or assumed leaf area index L and given values of m, r, and r ∞ , it is possible to derive the soil's reflectance, provided of course that the LAI is not too high.
We may assume that for a given soil type, its reflectance r s in the near-infrared band (N) is a constant factor S times its reflectance in the red band (R). We express this by N s = SR s . Extending this notation for red and near-infrared reflectance (R and N) to the other quantities gives the following: These expressions indicate that with an assumed leaf area index L and given values of m, r, and r ∞ at both wavelengths, it should be possible to estimate the soil's reflectance at these wavelengths, identified as R s and N s . In particular, the ratio N s/ R s can then be Remote Sens. 2021, 13, 1976 8 of 22 established, and if this ratio happens to be equal to S, we can conclude that obviously the correct LAI was guessed. When R s and N s are plotted in a diagram as a series of points as a function of the assumed LAI, we obtain a graph like that shown in Figure 4. Here, the red line is the locus of points of varying LAI that intersects the point of measured reflectance (the grey dot).
quantities gives the following: These expressions indicate that with an assumed leaf area index L and given values of m, r, and r at both wavelengths, it should be possible to estimate the soil's reflectance at these wavelengths, identified as Rs and Ns. In particular, the ratio Ns / Rs can then be established, and if this ratio happens to be equal to S, we can conclude that obviously the correct LAI was guessed. When Rs and Ns are plotted in a diagram as a series of points as a function of the assumed LAI, we obtain a graph like that shown in Figure 4. Here, the red line is the locus of points of varying LAI that intersects the point of measured reflectance (the grey dot). It is clear from this diagram that in this situation the matching LAI can be found from the intersection of one red line segment with the blue line. Next, the soil's brightness follows from its position along the blue line relative to the origin. The grey dot corresponds to the point where the assumed L equals 0. In that case, substitution in Equation (10) gives the red reflectance: This makes sense, since, for an assumed L of zero, the soil's reflectance must be equal to the measured reflectance. Note that L here does not indicate the actual LAI, but rather It is clear from this diagram that in this situation the matching LAI can be found from the intersection of one red line segment with the blue line. Next, the soil's brightness follows from its position along the blue line relative to the origin. The grey dot corresponds to the point where the assumed L equals 0. In that case, substitution in Equation (10) gives the red reflectance: This makes sense, since, for an assumed L of zero, the soil's reflectance must be equal to the measured reflectance. Note that L here does not indicate the actual LAI, but rather the one that would be needed to travel the path from the measured reflectance point at the position (R, N) to the soil line at (R s , N s ).
Including Crown Clumping and Surface Heterogeneity (Linear Mixing) Effects
The BHR model discussed so far only simulates radiative transfer for homogeneous turbid medium-type vegetation canopies with infinite extension in the horizontal plane. In reality, vegetation canopies like forests may exhibit crown clumping and/or other forms of spatial heterogeneity, and satellite image pixels may also contain mixtures of bare soil and dense vegetation, certainly if the spatial resolution is moderate or low.
In the two-stream BHR model, a crown clumping effect as in forests can be introduced by modulating the canopy layer's reflectance and transmittance as follows: where C v is the vertically projected crown cover fraction. These expressions were taken from the SLC (soil-leaf-canopy) model [16], which simulates crown clumping effects in one of the simplest possible ways. They are based on the concept of mixing the scattering properties of tree crowns with those of voids, which do not scatter light at all, but have a transmittance of unity. For the modified canopy reflectance including the soil background, Equation (1) is employed again and then reads: Note that for C v = 1 this is equivalent to the spatially uniform (turbid medium) model presented in Section 2.2. Also, it turns out that according to Equation (13) this kind of 3D mixing is non-linear.
Estimating r s from this equation however is still straightforward, since it can easily be derived from Equation (13) as follows: As opposed to the above model of the 3D mixing of soil and vegetation, the simplest model of surface heterogeneity resulting into mixed pixels is linear and assumes that a fraction f C of a pixel contains vegetation of a uniform composition and the complementary fraction 1 − f C contains only bare soil, having the same reflectance as the one underneath the canopy, r s . The final mixed-pixel reflectance then becomes the following: This equation represents a new hybrid BHR canopy reflectance model in which vertical and lateral mixing with the bare soil are accommodated. In principle, estimating r s from this model is still possible if it is the only unknown, but this then requires solving the following quadratic equation: where the coefficients of the quadratic formula are the following: It was found that the well-known formula for the pair of solutions r s = −b± is less convenient for use in practice, as in the limit of f C = 1, in which case a = 0, only the root of the smallest magnitude can remain finite, which then must be estimated by means of L'Hôpital's rule. Using the product of both roots, which equals c/a, it was found that the root of smallest magnitude can better be written in a less-known alternative form as follows: where a = 0 directly gives the correct solution −c/b. Note that Equation (18) can only yield a value of r s if the LAI and both cover fractions are known. If this algorithm is applied to reflectance values in the red and the near-infrared bands, as shown in Section 2.2 ( Figure 4), it turns out that the LAI and soil brightness can still be estimated from bi-spectral red-NIR reflectance values, provided again that both cover fractions C v and f C are given.
The simulated effects of crown clumping and lateral linear mixing on the red-NIR diagram of Figure 3 are shown in Figure 5. Here, only the upper left panel, where both cover fractions are equal to unity, still represents a homogeneous turbid medium canopy corresponding to the original in Figure 3. The other panels are modified versions of the upper left panel due to the effects of crown clumping (column direction) and incomplete linear fractional cover (row direction). From this figure, we may conclude that these nonlinear and linear mixing effects definitely cannot be neglected. By these effects, all points in the red-NIR diagram are drawn into the direction of the bare soil line, especially the points in the upper left corner of the diagram. This also implies that points in the upper left of the diagram can only indicate pixels with dense homogeneous green vegetation. These points have the highest ratios N/R, and therefore also the highest NDVI. A high NDVI therefore not only indicates the presence of a large proportion of green vegetation in the target pixel, but also that the pixel was homogeneous (high f C ) and that the vegetation resembled a turbid medium (high C v ), since otherwise the NDVI would have been smaller, as demonstrated in Figure 5 by the panels for the lower values of f C and C v . One can conclude that a high NDVI must indicate a combination of canopy parameters formed by a high leaf chlorophyll content, a high LAI, and high values of both vegetation coverage parameters f C and C v ; however, the downside is that a low NDVI may have a multitude of causes, namely low values of any of the parameters leaf chlorophyll, crown LAI, crown cover C v , or the pixel vegetation cover fraction f C . Due to the so-called saturation effect, in Figure 5, the lines of varying soil brightness (shown in red) for an LAI of 8 must be very close to the similar lines corresponding to an infinite crown LAI (not shown) since the corresponding reflectance values are very close to each other, certainly in the red band. In other words, based on reflectance values alone, NDVI values are also slightly influenced by the brightness of the soil background, and to reduce this influence, several alternatives for the NDVI have been proposed, like the soil-adjusted vegetation index SAVI [14], transformed SAVI (TSAVI) [15], and modified SAVI (MSAVI) [13]; however, from Figure 5, we can conclude that these attempts to reduce soil brightness effects are most suitable for homogeneous turbid-medium type vegetation canopies, since for more open canopies, with C v and/or f C less than unity, the slopes of the soil lines are again strongly influenced by the soil, in particular for high LAIs. The lower f C and C v , the more we find that soil lines orient themselves parallel to the bare soil line, and in that respect the utility of indices (VIs) like the perpendicular VI (PVI) [13] and weighted difference VI (WDVI) [17,18] increases in comparison to the NDVI, since these indices are constant along lines parallel to the bare soil line.
Due to the so-called saturation effect, in Figure 5, the lines of varying soil brightness (shown in red) for an LAI of 8 must be very close to the similar lines corresponding to an infinite crown LAI (not shown) since the corresponding reflectance values are very close to each other, certainly in the red band. In other words, based on reflectance values alone, one can hardly discriminate between an infinite crown LAI and an LAI of 8. This factor alone already substantially complicates the estimation of the LAI from reflectance data when the LAI is high. However, there are more factors, such as the obvious ill-posedness of the model inversion from red-NIR reflectance data, which is manifested here by the fact that multiple combinations of LAI, f C , and C v can produce exactly the same reflectance pairs of red-NIR reflectance. For instance, in one scenario, one might assume green vegetation with a high leaf chlorophyll content, a fixed crown LAI of 8, complete pixel coverage (f C = 1), and then both vertical crown cover C v and soil brightness might be estimated from red-NIR reflectance data. In another scenario, one might neglect crown clumping by assuming C v = 1 and vary pixel heterogeneity by considering f C as an unknown, as well as soil brightness. After model inversion in both of these scenarios, an effective LAI [19] could be assigned to the pixel that is equal to the product of crown LAI and both cover fractions, i.e., LAI eff = LAI × C v × f C , but in principle these alternative scenarios are just as applicable as the scenario of a homogeneous turbid medium canopy, and the outcomes in terms of effective LAI would probably be different.
Another issue is that, even with assumed values of LAI and leaf chlorophyll, red and NIR reflectance data are not adequate to estimate both cover fractions if soil brightness is also left free, since the number of estimated parameters would still exceed the number of reflectance values by one. Therefore, the application of three extreme models (or scenarios) that are invertible based on different assumptions is proposed. The assumptions common to all three of them are the following:
•
The LIDF is spherical; • Single leaf reflectance and transmittance in the NIR band are 0.52 and 0.44, respectively (0.04 leaf absorptance). In the red band, we use 0.07 and 0.01, respectively, which is common for green leaves; • The maximum crown LAI is 8;
•
The bare soil's spectral slope, defined by the ratio S = N s /R s , is given and equal to 1.2; • Soil brightness is always a free parameter that is to be estimated.
In the first model (I), both cover fractions are assumed to be equal to unity and the LAI is estimated. This model represents homogeneous surfaces covered with a uniform turbid-medium type of canopy.
In the second model (II), the crown LAI is assumed to be fixed at 8, the surface is homogeneous (f C = 1), and the crown cover fraction C v is estimated. This model represents homogeneous forests with variable crown density.
In the third model (III), the vegetation LAI is also fixed at 8, the crown cover fraction C v is unity, but the surface is heterogeneous. The fractional vegetation cover f C is estimated. This model represents green agricultural fields or grasslands interrupted by patches of bare soil within the pixel.
In all cases, soil brightness is estimated, as well as an effective pixel-level LAI, defined by LAI eff = LAI × C v × f C . Figure 6 shows red-NIR feature space plots for the three models, in which soil brightness varies together with the other free parameter of the corresponding model, so LAI, C v , and f C in models I, II, and III, respectively. where the estimated fractional cover fC is obviously approximated by a linear combination of R and N, whereas the estimated soil brightness, represented by Rs, is found as a ratio of two linear combinations of R and N. The linear combination of R and N that estimates fC suggests that, with a proper choice of the weights, the WDVI [17,18] is a good predictor of the fractional vegetation cover, and in model III this is also a predictor of the effective LAI.
Estimation of fAPAR
Since the infinite reflectance in the red band is known from its assumed value R , and the soil's reflectance follows one of the methods described in the previous section, one can derive all relevant quantities needed to estimate the fraction of absorbed photosynthetically active radiation, fAPAR, for diffuse incident radiation, while assuming that red light is representative for the whole visible or PAR region. In the homogeneous turbid In all three panels, the red lines show the effect of soil brightness variations, and the green lines the effects of the other free parameter of the respective model. It appears that all three models occupy similar triangular shapes, and in the overlapping area we can conclude that the three models are equally acceptable as candidates for model inversion.
For comparison, Figure 7 shows a 2D histogram of the red and NIR bands obtained from the MODIS white-sky albedo product at the local moment of maximum NDVI, which reproduces the predicted triangular shapes very well. where the estimated fractional cover fC is obviously approximated by a linear combination of R and N, whereas the estimated soil brightness, represented by Rs, is found as a ratio of two linear combinations of R and N. The linear combination of R and N that estimates fC suggests that, with a proper choice of the weights, the WDVI [17,18] is a good predictor of the fractional vegetation cover, and in model III this is also a predictor of the effective LAI.
Estimation of fAPAR
Since the infinite reflectance in the red band is known from its assumed value R , and the soil's reflectance follows one of the methods described in the previous section, one can derive all relevant quantities needed to estimate the fraction of absorbed photosynthetically active radiation, fAPAR, for diffuse incident radiation, while assuming that The models clearly differ by their degrees of non-linearity. Model I is extremely nonlinear in its response to LAI, the different responses in the red and the NIR, and in the soil lines, which show a large increase of slope with LAI. On the other hand, model III is almost perfectly linear, except in the upper left corner, where the soil line for unity f C has a much higher slope, since this is the slope corresponding to a homogeneous turbid medium canopy with an LAI of 8. Model II is non-linear, like model I, but the response to C v is much more linear than the response to LAI in model I, and the soil lines of model II do have increasing slopes, but less so than in model I. In model I, the red reflectance hardly changes between LAIs of 4 and 8, so the soil lines for these cases are vertical and do virtually overlap. Here, the NDVI increases with soil brightness, unlike the situation for low LAIs and more open canopies. Model I is so close to linear that it allows the direct estimation of fractional cover f C and soil brightness (R s ) from red-NIR BHR reflectance data if S is known: where the estimated fractional cover f C is obviously approximated by a linear combination of R and N, whereas the estimated soil brightness, represented by R s , is found as a ratio of two linear combinations of R and N. The linear combination of R and N that estimates f C suggests that, with a proper choice of the weights, the WDVI [17,18] is a good predictor of the fractional vegetation cover, and in model III this is also a predictor of the effective LAI.
Estimation of fAPAR
Since the infinite reflectance in the red band is known from its assumed value R ∞ , and the soil's reflectance follows one of the methods described in the previous section, one can derive all relevant quantities needed to estimate the fraction of absorbed photosynthetically active radiation, fAPAR, for diffuse incident radiation, while assuming that red light is representative for the whole visible or PAR region. In the homogeneous turbid medium model (I), the absorbed fraction of the incident hemispherical flux is given by the following equation: This equation shows that canopy absorptance increases with soil reflectance, which is caused by light reflected upward by the soil into the canopy that for a black soil would be lost by absorption in the soil and now contributes to absorption in the canopy layer.
The absorptances by the canopy and by the soil, and the reflectance from the top, together must obey the law of radiant energy conservation. To demonstrate this, Equation (20) may be rewritten as follows: where the last term is the effective absorptance of the soil, which is formed by the fraction of incident light at the canopy bottom that is not reflected by the soil and thus is given by the following equation: By combining Equations (21) and (22), one may find that the sum r dd + α d + α soil is equal to unity, so the law of radiant energy conservation is thus obeyed.
The above derivations apply to model I, but they can easily be generalized to models II and III. To include the effects of crown clumping and lateral mixing, the canopy absorptance of Equation (20) must be modified into the following form: In the literature, a linear relationship between fAPAR and the NDVI is often reported. These relationships are mostly based on experimental evidence, but model calculations [6,19,20] also support this idea. Taking the canopy diffuse absorptance in the red band as a proxy for fAPAR, the relationship with the NDVI may be calculated with models I-III for varying soil brightness levels and the LAI or cover fractions. The result is presented in Figure 8, which shows feature space plots of fAPAR vs. NDVI with varying soil brightness effects in red and the effects of the other free parameter in green. The near-linear relationships found earlier by other investigators are confirmed here for soils of moderate brightness (R s of approximately 0.25-0.40). For black soils, there is hardly any relationship, since in that case NDVIs are high regardless of the soil background, while fAPAR can still vary to a large extent. For dark soils, the relationships are strongly curved in all three models. Similar results have been reported by Myneni and Williams [19]. When soil brightness is increased under a constant LAI or C v , fAPAR increases while NDVI decreases, especially at high soil brightnesses. In model III, we see horizontal red lines in this case, indicating that in this model fAPAR depends only on f C . This is explained by the high fAPAR at a crown LAI of 8, but only for the fraction of the pixel that is covered with dense green vegetation, with minimum influence of the soil background on total fAPAR per pixel.
Retrieval of LAIeff, Soil Brightness and fAPAR from Red-NIR BHR Data
If no other information is available, the three scenarios of models I-III presented in the previous two sections are all equally likely to occur in reality for moderate-resolution satellite image pixels, although this may depend on the LAI. For instance, it is hard to imagine that a canopy can still resemble a turbid medium for LAIs < 1. It seems more likely that low LAIs are only possible in combination with clumping in plants or crop rows.
From Figure 6, one can conclude that the repertoires of these models do largely overlap and it can be shown that, provided a sufficiently large range of soil brightness is accommodated, the triangular shape shown in panel (c) for model III corresponds to the minimum repertoire common to all three models, and this provides a simple criterion to decide whether a given combination of red and NIR reflectance values still belongs to the common repertoire or not. In practice, Equation (19) of Section 2.3 can be applied to test whether the estimated values of fC and Rs are physically acceptable. If not, then one must decide to reject the input pixel for model inversion. Otherwise, inversion of models I-III should be possible, and the results would yield three solutions of soil brightness and separate solutions of LAI, Cv and fC, respectively. Since a crown LAI of 8 was assumed in models II and III, the effective LAI could be estimated by assuming equal weights: For soil brightness, simply the average of the three outcomes can be taken. Regarding fAPAR, Equation (23) is universally applicable to all three models, so also in this case a
Retrieval of LAI eff , Soil Brightness and fAPAR from Red-NIR BHR Data
If no other information is available, the three scenarios of models I-III presented in the previous two sections are all equally likely to occur in reality for moderate-resolution satellite image pixels, although this may depend on the LAI. For instance, it is hard to imagine that a canopy can still resemble a turbid medium for LAIs < 1. It seems more likely that low LAIs are only possible in combination with clumping in plants or crop rows.
From Figure 6, one can conclude that the repertoires of these models do largely overlap and it can be shown that, provided a sufficiently large range of soil brightness is accommodated, the triangular shape shown in panel (c) for model III corresponds to the minimum repertoire common to all three models, and this provides a simple criterion to decide whether a given combination of red and NIR reflectance values still belongs to the common repertoire or not. In practice, Equation (19) of Section 2.3 can be applied to test whether the estimated values of f C and R s are physically acceptable. If not, then one must decide to reject the input pixel for model inversion. Otherwise, inversion of models I-III should be possible, and the results would yield three solutions of soil brightness and separate solutions of LAI, C v and f C , respectively. Since a crown LAI of 8 was assumed in models II and III, the effective LAI could be estimated by assuming equal weights: For soil brightness, simply the average of the three outcomes can be taken. Regarding fAPAR, Equation (23) is universally applicable to all three models, so also in this case a simple average of the outcomes should suffice. For red-NIR pairs of measured reflectance values located below the bare soil line, none of the three models can provide a solution. In such cases we have to conclude that obviously the soil's spectral slope is less than that assumed, and the pixel is vegetation-free. Here, the red reflectance R s is accepted as the soil brightness output. The equal-weight solution suggested in Equation (24) can be considered as a starting point for more refined solutions, e.g., by adapting the weights according to a biome classification, as is done in the MODIS LAI processing chain [6]; however, to minimize possible spatial discontinuities, only the equal-weight solution is applied for the present paper, regardless of biomes.
For the application of the algorithm in practice, it has been proven to be beneficial to carry out the retrieval with a pre-computed direct look-up table (DLUT). Look-up table (LUT) solutions in radiative transfer model inversion problems are quite common and have been applied by numerous investigators [6,19]; however, these LUTs normally contain predicted reflectance values in a number of bands for a set of combinations of model input parameters, and the model inversion then consists of finding the vector of reflectance values that most closely resembles the measured reflectance data vector, possibly followed by some interpolation. Depending on the size of the LUT and its dimensionality, this may still take a considerable execution time. With a DLUT technique, however, the measured digital reflectance values are converted into an index, which is a direct pointer to the outputs stored in the DLUT. These outputs are the result of traditional model inversion techniques like numerical optimization. Once the DLUT has been generated, the application of this technique to images is nearly instantaneous regarding execution time, but it can only be applied in practice if the number of bands is limited to two or three. With the red and NIR MODIS bands as inputs, and a radiometric sampling interval of 0.001 reflectance units, a DLUT with about one million entries is sufficient to cover all possible combinations of red-NIR inputs. The generation of such a DLUT, which contains as outputs the effective LAI, average fAPAR, and average soil brightness (R s ) calculated from the retrieval results obtained for models I-III, takes about 25 seconds in MATLAB on an average PC (Intel i5 processor) without any speed optimizations. Application of the DLUT to a single MODIS CMG global scene (3600 × 7200 pixels) may take only 0.18 seconds, which is nearly instantaneous indeed. The application of the DLUT to long time series of MODIS CMG data or other large datasets should therefore not be a problem. It should be mentioned that this DLUT technique can only be applied for a fixed assumed value of soil spectral slope S and fixed green leaf optical properties. If S and/or the leaf optical properties vary pixel by pixel, the DLUT technique can no longer be used and one should count on a processing time of about 25 microseconds per pixel, which for a global CMG image would be 650 seconds or approximately 11 minutes; however, by skipping all water pixels, that time could easily be reduced by one third or to about 4 minutes.
Results
The Moderate Resolution Imaging Spectroradiometer (MODIS) MCD43C3 Version 6 Bidirectional Reflectance Distribution Function and Albedo (BRDF/Albedo) dataset [1] is produced daily using 16 days of Terra and Aqua MODIS data in a 0.05 degree (5600 m at the equator) Climate Modelling Grid (CMG). Data are temporally weighted to the ninth day of the retrieval period, which is reflected in the Julian date in the file name. This CMG product covers the entire globe for use in climate simulation models. MCD43C3 provides black-sky albedo (directional hemispherical reflectance, DHR) and white-sky albedo or BHR data at the local solar noon. Black-sky albedo and white-sky albedo values are available as a separate layer for MODIS spectral bands one through seven, as well as the visible, near-infrared (NIR), and shortwave infrared bands. Along with the 20 albedo layers, ancillary layers for quality, local solar noon, percent finer resolution inputs, snow cover, and uncertainty are also provided. The data can be downloaded for free from the Land Processes Distributed Active Archive Center (LP DAAC) website: (https: //lpdaac.usgs.gov/products/mcd43c3v006/ accessed on 1 March 2020). For this particular project, data were downloaded for all 16 day periods since 2000 until the beginning of 2020. Only two processed data sets were further employed, namely the MODIS white-sky albedo data in bands 1-7 at the pixel-wise situation of accumulated minimum and maximum NDVI, respectively, taking all observations from all years together as input. Figure 9 shows the two global NDVI maps taken from MODIS white-sky albedo bands one and two as red-NIR pairs. A total of 455 images at 16-day intervals was used as input. During the processing of all input files, two quality indicators ("BRDF _Quality" and "Percent_Inputs") supplied by the distributor, in addition to a snow index, were applied to only admit input pixels of sufficient quality as candidates. For the snow index, MODIS bands one and six (red and shortwave infrared) were employed in the formula SI = 255 × R / (R + M), where R and M are the digital numbers in bands 1 and 6, respectively. Only pixels with "BRDF _Quality" < 4, "Percent_Inputs" > 50 and SI < 110 were accepted. This resulted in nearly continuous maps of the minimum and maximum NDVI with a high degree of spatial coherence, as can be seen in Figure 9.
The map of minimum NDVI clearly shows where permanent green vegetation was present during the observation period, although African rainforests show up less promi- During the processing of all input files, two quality indicators ("BRDF_Quality" and "Percent_Inputs") supplied by the distributor, in addition to a snow index, were applied to only admit input pixels of sufficient quality as candidates. For the snow index, MODIS bands one and six (red and shortwave infrared) were employed in the formula SI = 255 × R/(R + M), where R and M are the digital numbers in bands 1 and 6, respectively. Only pixels with "BRDF_Quality" < 4, "Percent_Inputs" > 50 and SI < 110 were accepted. This resulted in nearly continuous maps of the minimum and maximum NDVI with a high degree of spatial coherence, as can be seen in Figure 9.
The map of minimum NDVI clearly shows where permanent green vegetation was present during the observation period, although African rainforests show up less prominently than Amazonian and Indonesian rainforests. The map of the maximum NDVI shows the places where dense green vegetation was present at least once during the whole period. This appears to be almost everywhere except in desert areas and regions with permanent snow cover, such as on Greenland and in the Himalayas.
Before discussing the results of applying the DLUT technique to the red and NIR bands of the MODIS data, first some remarks are in order regarding the settings of some important sets of input parameters, since they may have a tremendous effect on the results. As discussed briefly in Section 2.3 and Section 2.4, the repertoire of the models used depends on the assumed optical properties of green leaves or needles in the red and the near-infrared, and on the assumed soil spectral slope. This repertoire determines the retrieval rate, i.e., the percentage of pixels that passed a successful model inversion, since combinations of R and N falling outside the repertoire cannot be inverted, so the larger the repertoire, the larger the retrieval rate can be; however, this may come at a price, namely, less accurate results in terms of LAI, soil brightness, and fAPAR. This is because of the ill-posedness of the model inversion: in the model, several LAI, soil brightness, and leaf optical property combinations can yield the same combination of reflectances in the red and NIR bands, and if the wrong optical leaf properties are assumed, this will be compensated by incorrectly retrieved values for the LAI and soil brightness. The leaf or needle optical properties determine their reflectance and transmittance in the red and NIR spectral bands, and in turn, these determine the corresponding canopy infinite reflectance values,R ∞ and N ∞ , as explained in Section 2.1. These infinite reflectance values determine the position of the upper left corner of the triangular-shaped repertoires shown in Figure 6. The smaller R ∞ and the larger N ∞ , the larger the model's repertoire, since this is the direction away from the bare soil line. Additionally, the spectral slope of the bare soil line has some influence on the repertoire: the smaller this slope, the larger the repertoire will be.
With the initial assumptions for the inputs of the three models presented in Section 2.3, the results were not completely satisfactory, since a considerable number of pixels, located mainly in the rainforests, led to unsuccessful retrievals because their reflectance in the red band was less than the assumed R ∞ , so they fell outside the repertoire. In the 2D histogram of Figure 7, one can observe that several pixels had red reflectance values less than 0.01, so it was decided to adjust the single leaf reflectance to 0.02 and transmittance to zero in order to achieve a retrieval rate close to 100%. In this way, one obtains an R ∞ of 0.0067, and this turned out to be sufficient to reach a high retrieval rate.
Regarding the assumed bare soil's spectral slope, for the MODIS red and NIR bands, a fixed value of S = 1.2 was found to be adequate, since this almost perfectly coincided with the modal spectral slope found in the image of the minimum NDVI. A higher assumed slope would result into a lower retrieval rate, since more bare soil pixels would become non-invertible. Assuming a lower slope would enlarge the repertoire and improve the retrieval rate, but bare soils with a slightly higher slope would be mistaken for sparsely vegetated surfaces. In this respect, it should be noted that pixels below the assumed bare soil line actually were not considered to be non-invertible. Instead, for these pixels the LAI was simply set to zero, and the observed reflectance in the red band was adopted as the measure of soil brightness. This can be interpreted as a local posterior adjustment of the assumed bare soil's spectral slope. In the future, this may lead to a more refined use of maps of soil spectral slope in order to improve the retrieval results compared to the results obtained with a constant value of S.
Discussion
The results of applying the DLUT to the red and NIR bands of the images shown in Figure 9 are presented in Figures 10-12. The effective LAI maps for the situations of minimum and maximum NDVI are shown in Figure 10. From these maps one can conclude that they look more or less as expected, with maximum LAI values of about 4. They roughly follow the NDVI maps of Figure 9, with the exception of the rain forests, which show only moderate LAIs of about 2-2.5.
Remote Sens. 2021, 13, x FOR PEER REVIEW 19 of 23 Figure 10. Estimated effective LAI from the DLUT applied to the red-NIR bands of the images of Figure 9.
As for the white-sky albedo (BHR) the role of the 3D canopy structure was expected to be limited, this low LAI in rainforests was initially only attributed to woody material. This was not incorporated in the two-stream BHR model, but it potentially has a large impact on the NIR canopy reflectance via its effect on the infinite reflectance, N , since its absorption of NIR radiation reduces it enormously. The N for the assumed leaf with a single leaf absorptance of 0.04 was 0.67; however, according to Equation (5) in Section 2.1, the infinite canopy reflectance can be approximated by 12 − , so for a slightly raised effective leaf absorptance of 0.16 we would obtain 1 -2 × 0.4 = 0.2, which means a reduction by a factor of more than 3. This implies that adding some absorbing woody material to a leaf canopy can have a huge negative impact on the NIR canopy reflectance if the LAI is high, such as in forests. An inspection of the reflectance values in the rainforests revealed that in the NIR band they were substantially lower (about 50%) than in agricultural regions, but in the red band this reduction was even greater, and this cannot be explained by woody material alone, since woody material is brighter than green leaves or needles in the red band, so one may expect a raised reflectance in the red band. The very low red reflectance for forests can only be explained by the clumping effect, combined with a high LAI. This is also the main reason for the difference between the maps of NDVI and LAIeff. As for the white-sky albedo (BHR) the role of the 3D canopy structure was expected to be limited, this low LAI in rainforests was initially only attributed to woody material. This was not incorporated in the two-stream BHR model, but it potentially has a large impact on the NIR canopy reflectance via its effect on the infinite reflectance, N ∞ , since its absorption of NIR radiation reduces it enormously. The N ∞ for the assumed leaf with a single leaf absorptance of 0.04 was 0.67; however, according to Equation (5) in Section 2.1, the infinite canopy reflectance can be approximated by 1 − 2 √ α, so for a slightly raised effective leaf absorptance of 0.16 we would obtain 1 -2 × 0.4 = 0.2, which means a reduction by a factor of more than 3. This implies that adding some absorbing woody material to a leaf canopy can have a huge negative impact on the NIR canopy reflectance if the LAI is high, such as in forests. An inspection of the reflectance values in the rainforests revealed that in the NIR band they were substantially lower (about 50%) than in agricultural regions, but in the red band this reduction was even greater, and this cannot be explained by woody material alone, since woody material is brighter than green leaves or needles in the red band, so one may expect a raised reflectance in the red band. The very low red reflectance for forests can only be explained by the clumping effect, combined with a high LAI. This is also the main reason for the difference between the maps of NDVI and LAIeff. The maps of NDVI show high values in both forested and agricultural areas, whereas high values of the retrieved LAIeff at the moment of maximum NDVI are limited to areas of high agricultural activity, such as the Corn Belt in the USA, south of the Amazon rainforest, Europe, Australia, and North-East China.
The corresponding maps of retrieved soil brightness shown in Figure 11 are fairly similar for the situations at maximum and minimum NDVI, not only in desert regions, where this could be expected, but also in most vegetated areas. An exception was formed by the rainforests, which showed lower retrieved soil brightness values at the maximum NDVI. The reason for this is unknown, but it is not of crucial importance anyway, since the impact of soil brightness on densely vegetated forest scenes is very small. Very high soil brightness values greater than 0.5 occur in the Sahara Desert and Arabia. The elucidation of how stable these soil brightness maps are over time may be investigated by time series analysis applied to the MODIS white-sky albedo data; however, this falls outside the scope of the present paper. The corresponding maps of retrieved soil brightness shown in Figure 11 are fairly similar for the situations at maximum and minimum NDVI, not only in desert regions, where this could be expected, but also in most vegetated areas. An exception was formed by the rainforests, which showed lower retrieved soil brightness values at the maximum NDVI. The reason for this is unknown, but it is not of crucial importance anyway, since the impact of soil brightness on densely vegetated forest scenes is very small. Very high soil brightness values greater than 0.5 occur in the Sahara Desert and Arabia. The elucidation of how stable these soil brightness maps are over time may be investigated by time series analysis applied to the MODIS white-sky albedo data; however, this falls outside the scope of the present paper. Figure 11. Estimated soil brightness (in the red band) from the DLUT applied to the red-NIR bands of the images of Figure 9.
The last product discussed here is fAPAR, shown in Figure 12. This product was derived from the retrieved prime parameter of models I-III (LAI, Cv or fC) and soil brightness, Figure 11. Estimated soil brightness (in the red band) from the DLUT applied to the red-NIR bands of the images of Figure 9. The last product discussed here is fAPAR, shown in Figure 12. This product was derived from the retrieved prime parameter of models I-III (LAI, C v or f C ) and soil brightness, as explained in Section 2.4. The results look very similar to the maps of LAIeff shown in Figure 10, with moderate values in rainforests and higher values in areas with strong agricultural activity. Again, this can be attributed to disregarding woody material and underestimating clumping effects in forested areas, although lower values of fAPAR in forests are not against expectations, since visible light absorbed by trunks and branches does not contribute to the fAPAR going to leaves or needles. Nevertheless, the model would have to be reformulated in order to properly take account of the absorption by woody material. Figure 10, with moderate values in rainforests and higher values in areas with strong agricultural activity. Again, this can be attributed to disregarding woody material and underestimating clumping effects in forested areas, although lower values of fAPAR in forests are not against expectations, since visible light absorbed by trunks and branches does not contribute to the fAPAR going to leaves or needles. Nevertheless, the model would have to be reformulated in order to properly take account of the absorption by woody material. Figure 12. Estimated fAPAR (red band) from the DLUT applied to the red-NIR bands of images of Figure 9.
Comparing the results obtained so far with this first version of the algorithm to those of the MODIS LAI products [6], we may conclude that including biome-specific information about the distinction between forests, shrubs, and other types of vegetation would probably be beneficial for an improved result that better takes account of the spectral diversity amongst biome types. Biome maps might also be used for a more specific weighting of the results of models I-III, and perhaps by applying different values of maximum LAI in these models. In the MODIS LAI product [6], a classification into six biomes is included, and each biome is linked to extensive datasets of specific information which drive the numerical radiative transfer model outputs stored in look-up tables. Since the Comparing the results obtained so far with this first version of the algorithm to those of the MODIS LAI products [6], we may conclude that including biome-specific information about the distinction between forests, shrubs, and other types of vegetation would probably be beneficial for an improved result that better takes account of the spectral diversity amongst biome types. Biome maps might also be used for a more specific weighting of the results of models I-III, and perhaps by applying different values of maximum LAI in these models. In the MODIS LAI product [6], a classification into six biomes is included, and each biome is linked to extensive datasets of specific information which drive the numerical radiative transfer model outputs stored in look-up tables. Since the retrieval rate obtained was limited, a backup algorithm based on biome-specific assumed relationships of LAI and fAPAR with the NDVI was employed.
In a Bayesian approach, the JRC-TIP algorithm [19], which applies a two-stream turbid medium model to generate black-sky albedo simulations, uses statistical a priori information to combat the ill-posedness of the model inversion, but this information is applied globally and is non-specific with respect to biomes. An advantage of the Bayesian approach is that maps of the posterior uncertainties of the retrieval results can be provided as by-products. In the method of the present paper, this could be achieved in an approximate manner by comparing the results obtained with the different surface heterogeneity models I-III, since the agreement amongst these models about the outcomes would be a good measure of the uncertainty.
A validation of the presented product prototypes could take place by comparing them to time series of field measurements of LAI and fAPAR at a limited number of flux-tower locations representative for the various biomes on Earth. Also, comparison to several existing LAI and fAPAR products from MODIS (e.g., MOD15, MYD15 and MCD15) and other missions would be desirable, but since in this paper only a few first examples of the new products were shown for some rather arbitrary preliminary settings of the heterogeneity parameters in models I-III, this was considered to be premature, in addition to the considerable impact that it would have on the length of the paper. Also, the mentioned MODIS products have a different spatial sampling density and projection (sinusoidal), which would introduce considerable additional complications. Therefore, since the present paper was only meant to introduce a new simple approach, this task has been left for future work. Nevertheless, in this paper, this new approach has already been presented to enable young researchers in the land surface remote sensing community to experiment with it. To this end, the MATLAB codes of the model inversion, the DLUT generation, and its application to images are provided as Supplementary Materials.
Conclusions
A simple bi-hemispherical canopy/soil reflectance model has been introduced and extended with crown clumping and linear mixing effects. Assuming that the soil's spectral shape is given, soil brightness can be derived if all canopy parameters are known. If applied to red and NIR spectral bi-hemispherical reflectance values, this allows the simultaneous retrieval of soil brightness and LAI by using an efficient direct look-up table technique (DLUT). The fraction APAR in the red band can be provided as an important by-product. Since crown cover and linear mixing coverage are mostly unknown, it was proposed to use three invertible extreme model scenarios of surface heterogeneity and to average the results obtained from these models. The proposed methodology was tested with global MODIS white-sky albedo data in the red and NIR bands. It was shown that the method can be applied with a retrieval rate close to 100% and at a nearly instantaneous speed thanks to the DLUT technique. The global maps of effective LAI, soil brightness, and fAPAR produced are spatially continuous and show a high degree of spatial coherence. Refinements based on biome-specific information are recommended, especially to differentiate forested areas from other vegetation types. | 14,571.2 | 2021-05-19T00:00:00.000 | [
"Environmental Science",
"Mathematics"
] |
Proteomics for the Identification of Biomarkers in Testicular Cancer–Review
A large number of biomarkers have been proposed for the diagnosis of testicular cancer, representing putative molecular targets for anticancer treatments. However, no conclusive data have been provided. Proteomics represents a research field recently developed. It evaluates the large-scale analysis of the full protein components of a single cell, of a specific tissue, or of biological fluids. In the last decades, proteomics has been applied in clinical fields, thanks to modern technology and new bioinformatic tools, to identify novel molecular markers of diseases. The aim of this review is to argue the findings of recent studies in the discoveries of putative prognostic and diagnostic markers of testis cancer by proteomic techniques. We present here a panel of proteins identified by proteomics which might be used after validation for early detection and the prognostic evaluation of testicular tumors. In addition, the molecular mechanisms revealed by these proteomic studies might also guide the development of novel treatments in future.
INTRODUCTION
Testicular germ cell tumors (TGCT) is the most frequent cancer occurring in young men. TGCTs are classified into two subgroups: non-seminoma and seminomas germ cell tumors. Seminomas rate for 50% of testicular cancer and non-seminoma germ cell tumors rate for 40% of testicular cancer. The remaining 10% of testicular cancer is associated with tumors and they usually include both seminoma and non-seminoma components (1). The difference in the diagnosis between seminoma and non-seminoma is fundamental for the objective of treatment and of prognosis.
TGCTs originate from transformed gonocytes or undifferentiated spermatogonia, but the pathogenesis of TGCT remains unexplored (2). To date, accessible markers for the diagnosis and follow-up aftercare include a-fetoprotein (AFP), ßHCG and LDH. However, AFP and ßhCG have high specificity (90%) but often relatively low sensitivity. For this reason tumor markers alone are not able to detect many recurrences, indeed in about 40% of men with disease recurrence the levels of these markers are usually "normal." The LDH appears to have poor diagnostic performance (3). Therefore, the discovery of novel clinical biomarkers would clearly help the early detection and the monitor of the disease. The diagnosis of "cancer" can be challenging. In addition to histopathological interpretation and immunohistochemical stains for confirming the precise cell of origin (4), more recently global gene expression profiling has been applied in order to facilitate the treatment decisions and prognosis.
One key application for patients with the primary disease is precise prognosis, which helps to divide patients into different risk groups and select both treatment and monitoring strategies. Usually, the prognosis is based on clinical parameters such as age and tumor stage. Recently, considerable attempts have been made to incorporate molecular information in the staging process for accurate prognosis. Different tumors are classified in accordance to similar gene expression patterns. They represent a "molecular signature, " composed by several tens or hundreds of genes, in particular with regard to cell morphology or tissue characteristics. Global gene expression is most easily measured using cellular RNA; on the other hand, protein expression profiling provides a more dynamic view, offering additional informations on protein-protein interactions, post-trasductional modifications and finally on protein abundance (5).
Testicular cancer has been mainly studied at a genetic level. However, proteomics represents a promising technology that could allow novel insight into the disease at the molecular level to increase the understanding of their function. In the present molecular era proteomic is evaluated as a crucial point in personalized medicine to identificate specific target proteins for the pathophysiological state. Moreover, modern bioinformatic analysis offers information about the involvement of the proteins in the biological pathways of the tumor (6). Proteomics analysis of testicular tumor compared to normal testicular tissue may create a platform for enhanced understanding of differentially expressed proteins which might represent potential biomarkers for cancer. Protein expression profiling is a powerful tool in clinical practice, particularly in identifying cancer biomarkers to help the diagnosis and to choose a personalized treatment and monitoring of patients.
In the field of testicular tumors, particularly, the studies are few and outdated. Currently the main problem in the identification of protein markers is the small number of proteins which have been associated to TGCT.
PROTEOMIC TECHNOLOGIES APPLIED IN TGCT: FEATURES AND PERFORMANCES
Technological advances in proteomics have improved sensitivity and multiplexing ability of the method, as well as the possibility of identifying protein interactions. These advances can be of help to understand the molecular mechanisms involved in TGCT. The use of proteomics technologies offers an appealing approach to the identification and development of new tools to be used in clinical practice, identified by the simultaneous comparison of hundreds or thousands of proteins. The development and use of performant sample preparation techniques together with the increasing availability of proteomics technologies and recent technical advances in mass spectrometry (MS) enable identification and quantification of proteins involved in the diseases, although they are expressed at low abundance (7).
Up to now the most common proteomics technologies applied in the studies about TGCT include "gel-based" proteomics such as 2D-PAGE and 2D-electrophoresis associated with mass spectrometry (MS). SELDI-TOF and High-performance liquid chromatography (HPLC) associated with tandem mass spectrometry (MS/MS) moreover have been used in two studies. One additional proteomic study was performed in serum by SELDI-TOF.
2D Gel Electrophoresis
2D gel electrophoresis detaches proteins consistent with their isoelectric point. A second, independent separation step is then performed, dependent on mass (molecular weight), using sodium dodecyl sulfate polyacrylamide gel electrophoresis. This 2D-PAGE method can be performed to develop samples at high resolution and on a large scale. It might represent a primary screening method to form hypotheses and to guide further researches. 2D gel electrophoresis is the most frequently performed technique for proteomic analysis, although some limitations which are related to protein solubility (i.e., membrane proteins not easily solubilized), poor protein separation in the initial pH gradient (too basic or too acidic proteins) and low and high molecular weight (8,9).
To better identify the separated proteins, proteases may be used to digest bands or spots obtained 2D-electrophoresis. The smaller fragments are then ionized and analyzed by mass spectrometry (MS).
High-Performance Liquid Chromatography (HPLC)
HPLC is used with the aim to obtain a complete protein recovery, including small basic and hydrophobic types (9). In fact, protein separation is based on some specific protein properties (hydrophobicity, surface charges, specific amino acid sequences) (10). Connecting HPLC with a mass spectrometer it is possible to obtain the rapid separation and the comprehensive identification of components of a complex protein mixture, with the aim to deeply analyze a proteome.
Mass Spectrometry (MS)
MS-based techniques can be used to study complex protein mixtures previously fractionated by electrophoresis or HPLC. This technology provides precise mass values by the measurement of the mass-to-charge ratio (m/z) of the ions generated from the peptides and the proteins. In recent years a significant technical improvement of mass spectrometers has been observed. Modern mass spectrometers, such as time of flight (TOF), Fourier transform ion cyclotron resonance (FT-ICR) and Orbitrap detectors provide extremely accurate masses of analytes (11). Makarov invented the Orbitrap in 1999, as a mass analyzer that couples high resolution with high mass-accuracy, a significant m/z range and a high dynamic range (12,13). The high mass-accuracy of the Orbitrap significantly contributes to the amount of acquired data and the number of analytic approaches that can be connected to MS (14).
SELDI TOF-MS
Surface-enhanced laser desorption/ionization time-of-flight mass spectrometry (SELDI TOF-MS) has been used widely in biomarker discovery because its sensitivity. It requires moreover only a small amount of protein for analysis. This method decreases the sample complexity by studying proteins with a specific chemical property that allows them to be adsorbed in a known surface (e.g., a charged, hydrophobic, or functionalized affinity surface). The non-bound protein population is removed by extensive washing. The adsorbed sample is ionized and analyzed (15). (16). This strategy has been applied to identify differences in the proteomes obtained by non-neoplastic and neoplastic tissues. Protein extracts have been analyzed by 2DPAGE combined with mass spectrometry. The comparative study permitted the identification and detection of quantitative differences of expressed proteins between seminoma and non-seminoma tissue.
PROTEOMICS OF TGCT TISSUE
Glutathione S-transferase (GSTs) M3 protein has been downregulated in seminoma tissues. GSTs detoxification enzymes partecipate in the pathogenesis of cancers. GSTM3 is a critical GSTs variant and previous evidences showed that GSTM3 polymorphism is associated with an increased risk to develope a cancer (20,21). A lot of studies previously investigated the association of GSTM3 gene polymorphism with the risk to develop a lung cancer. Furthermore, a reduction in GST protein level was earlier reported in human TGCT (22), but the M3 isoform has not been previously identified in TGCT. Only proteomics permitted the identification of this specific isoform, as possibly associated with testis cancer. The M3 homodimer is a specific isoform specifically identified in the brain and in the testis (23). The GSTM3 gene is polymorphic and GSTM3 polymorphisms control the enzyme activity by the modulation of substrate binding (24). Distinct polymorphisms of GST-M3 enzyme are associated with a higher risk for TGCT formation (16). In particular, the study published in 2009, performed on a large population of patients who survived TGCT, reported that GSTP1 genotype influences the risk of developing a TGCT (25). M3 and P1 polymorphisms of GSTM3 represent promising markers for predicting the risk of TGTC formation.
In 2009 Leman et al. used a proteomics analysis to discover a pattern of proteins related with nuclear changes in seminoma cells (17). The changes in the form and in the size of the nucleus are trademarks of the cancer cell. Proteomics approach focused to mark a specific pattern of proteins related with the specific alterations in the nuclear structure of seminoma cells. Using high-resolution two dimensional gels, four nuclear matrix proteins have been identified in seminomas, but not in the normal tubules and three of these four proteins are part of gamma-tubulin complex component 6 (GCP6).
The γ-tubulin complex is a large multiprotein complex which plays an essential role in microtubule nucleation at the centrosome. C-tubulin ring complex (c-TuRC) is a key component of the centrosome which nucleates microtubules. GCP6 is localized in the pericentriolar material, a protein matrix composed by protein complexes involved in centrosomeassociated functions (i.e., microtubule nucleation). Moreover, GCP6 is needed for centriole duplication and Polo-like kinase4 (Plk4)-induced centriole over-duplication. GCP6 interacts in fact with Plk4 and it is phosphorylated by the same Plk4. Therefore, controlling centriole numbers GCP6 preserves cells to have the correct number of centrosomes and cilia. Excessive number of centrioles lead to tumorigenesis in flies (26) and has been associated with chromosomal instability in humans (27).
Another specific protein for seminomas is Cyclin-dependent kinase 10 (CDK10). CDK10 is a Cdc2-related kinase and is a key element in the advance from the G2 to M phase transition of the cell cycle. Two isoforms of CDK10 have been documented: HCDK10-1 and HCDK10-2 (17). CDK10 is highly represented in colorectal cancer where it takes part in the suppression of apoptosis and in the stimulation of tumor growth in vitro and in vivo. The modulation of CDK10 expression in colorectal cancer indicates that CDK10 is involved in cell growth and it is associated with a reduction in chemosensitivity. Finally it inhibits apoptosis through the upregulation of the expression of Bcl-2 (28). Both the CDK10 isoforms are expressed in the nuclear matrix of the seminomas, supporting the role of CDK10 in the cell cycle regulation that may induce testicular cancer.
In normal testis moreover seven specific nuclear matrix proteins have been detected, which are absent in seminoma tissue. Some of these proteins are testis specific: Y177 encodedlike protein 4, cytokeratins, glutamine synthetase, and StARrelated lipid transfer protein 7 (StarD7).
StarD7 has been reported for the first time as related with nuclear matrix and it is part of the family of StAR1related lipid transfer (START) proteins. These proteins are involved in lipid transport, metabolism and signaling (17). Overexpressed StARD7 gene has been specifically linked with colorectal cancer. StarD7 protein upregulation has been documented in differentiating cytotrophoblast suggesting that StarD7 might have a function in throphoblast differentiation through phospholipids uptake and transport (29). The loss of StarD7 protein in JEG-3 cells alters ABCG2 multidrug transporter level, cell migration, cell proliferation, and differentiation marker expression. StarD7, expressed in normal testis but not in seminoma tissue, might have a role in maintaining the differentiated form of the normal cells. The loss of StarD7 might induce cancer development.
The testicular microenvironment is a unique environment. Spermatogenesis and tumorigenesis at testicular level in humans are two biological processes which have got numerous similarities between them. It is very important to know the (19) common mechanisms between the two biological processes. In 2013 Liu et al. performed an extensive study for tumor marker identification by proteomic of testicular tissues. Using 2D-high performance liquid chromatography (HPLC)-MS/MS (LTQ Orbitrap Velos hybrid mass spectrometer) the Authors identified 7,346 proteins in testis tissue with normal spermatogenesis (18). The protein data were confirmed by immunohistochemistry and by comparison with previously published data from the Human Testis Proteome Database. These date have been confirmed by using of a GWAS study, using associated SNPs in case of differential expression of these proteins. Among these testicular proteins, six novel cancer/testis gene transmembrane protease have been characterized: serine 12 (TMPRSS12), tubulin polymerisation promoting protein family member 2 (TPPP2), protease serine 55 (PRSS55), double-sex and mab-3 related transcription factor 1 (DMRT1), piwi-like RNA-mediated gene silencing 1 (PIWIL1), and hemogen (HEMGN). The last four proteins have been proposed to exert a central role in spermatogenesis and cancer development, although they still haven't known specific functions in testis (18).
The Mab-3 doublesex-and Related Transcription Factor 1 (DMTR1) has been associated to prostate cancer. It is a controller of mitotic proliferation in germ cells, thanks to a specific zinc finger structural motif which is associated with cell cancer proliferation. A recent study confirmed these results in the human testis; nuclear expression of DMRT1 was reported in spermatogonia, but not in primary spermatocytes that have entered meiosis 1 or in more mature germ cells. DMRT1 expression was confirmed in the germ cells of testicular mixed germ cell-sex cord stromal tumor (MGC-SCST) and spermatocytic tumor but not in those of seminoma. For this reason the germ cells of MGC-SCST are related to spermatogonia, which espress DMRT1. The strong expression of DMRT1, together with the absence of TCLF5 in the germ cells of both MGC-SCST and spermatocytic tumor, suggests a premeiotic origin for both tumors (30).
The second protein is Piwi-Like RNA Mediated Gene silencing 1 (PIWIL1) which, together with the allelic variant rs10773777, was also detected in prostate cancer cases. PIWIL1 was also detected by proteomic approach as a specific protein of TGCT. This protein might play a pivotal function in RNA silencing during the modulation of tranlational activity. Previous studies reported that PIWIL1 plays an important role in cancer development, improving DNA methylation. Several cancer-germline genes have been defined to stimulate PIWIL1 as a part of oncogenic pathways involved in cell proliferation (31). More recently PIWIL1 expression was demonstrated in spermatocytes and spermatids. Up to 70% of TGCT samples express PIWIL1, which is not normally expressed in premeiotic germ cells. These evidences support that in many germ tumors is present an aberrant expression of PIWIL1. The enhanced expressions of piwil2 was found in seminomas and has not been reported in testicular nonseminomas tumors (31).
The third protein is Transmembrane (C-Terminal) Protease, Serine 12 (TMPRSS12). It was previously related with colorectal cancer for the variant rs11169552 (32). This protein has been demonstrated to be expressed in spermatids and spermatocytes (33). The Tubulin Polymerization-Promoting Protein Family Member 2 (TPPP2) and its variant rs1952524 was linked to liver cancer (34). Another protein is the Protease Serine 55 (PRSS55) along with the variant rs4404875, which has been mainly identified in Leydig and Sertoli cells and it is is associated with prostate and ovarian cancer (35). Finally Hemogein (HEMGN), which regulates proliferation and differentiation in hematopoietic cells, has been associated with thyroid cancer (36). Among the 300 proteins expressed in human testis, only 22.7% are TGCT-related proteins, of which only 65 proteins have been evaluated, indicating that other candidate proteins exist, although not still studied, The functional analysis of only 65 out 300 proteins might depend to still low sensitive methods in the field of proteomics. The six proteins previously cited might represent moreover important targets for personalized therapy in this kind of neoplasy (18).
In a very recent study, Castillo et al. identified 174 phosphorylated kinases in human testis by metal oxide affinity chromatography using TiO2 combined with LC-MS/MS. Protein phosphorylation is involved in the modulation of cell cycle, cell growth, cell differentiation and cell death. Two kinases have been studied in the testis phosphoproteome as candidates for further studies by immunodetection procedures. Immunodetection has been specifically used to study the potential function of cyclin dependent kinase 12 (CDK12) and p21-activated kinase 4 (PAK4) in testicular tissue. The in silico protein-protein interactions have been studied, and a functional analysis in a human embryonal carcinoma cell line has been performed. PAK4 is localized in human spermatogonia. Its function in preventing the activation of caspase is well-known. In embryonal carcinoma it has been observed that PAK4 protects cells from apoptosis. PAK4 inhibitors might represent an interesting pharmacological target for novel drugs modulating behavior of testicular cancer (19).
PROTEOMICS OF TCGC IN SERUM
Protein markers that distinguish cancer patients by healthy controls can be moreover searched in serum.
A single serum proteomic study regarding testis cancer was carried out. In 2010, Strenziok et al utilizing Surfaceenhanced laser desorption ionization time-of-Xight mass spectrometry (SELDI-TOF MS) identified the protein profiles of TGCT patients that are different in a highly significant degree from normal subjects (37). CM10 ProteinChip R array identified 138 peaks in a mass range of 3,800-10,000 Da that might repredent a "molecular fingerprint" to differentiate tumor serum sample from non-tumor serum samples. The spectra of proteins have been investigated by the proteomic platform "proteomic.net." Five peaks have been verified by CM10 ProteinChip R (6.48, 6.84, 8.15, 8.17, 8.92 kDa). There was no single peak that could discriminate the group of seminoma vs. control subjects, so a cluster classifications has been performed.
For statistical analysis, an artificial intelligence learning algorithm used three different bioinformatics methods to develop the training set for the decision trees, support vector machines, and neural networks and to differentiate between the two groups. Decision tree analyses developed the most powerful classifier with 89.4% specificity and 91.5% sensitivity (CM10 ProteinChip R , 95% confidence interval of 82.6-95.5%).
In this study the authors after the first step in investigation separating cancer from healthy controls with the definition of protein profiles of TGCT patients that differ in a highly significant degree from normal controls; the protein identification of peak masses was not necessary to differentiate cancer patients from healthy subjects.
Validation of these results may permit proteomic profiling to become a useful tool especially for aftercare follow up.
CONCLUSIONS
In the past 20 years molecular biomarker identification achieved importance in the field of personalized medicine, aimed to identify a cancer in the early stages and to develop novel therapeutical approaches. According to these premises, the discovery of novel specific markers would help the management of patients with testicular cancer.
Proteomics has proven to represent a promising platform for identifying biomarkers linked to testicular cancer.
The conventional tumor markers AFP, hCG, and LDH have demonstrated value in the clinical management of testicular malignant TGCT. However, their limitations in sensitivity and specificity prevent more universal application, especially in patients with seminoma. Tissue and serum biomarkers show exciting promises in the identification of markers for the diagnostics of TGCT. Furthermore, no proteomic studies have been permorfed aimed to detect markers of TGCT in semen, although it represents a promising source of putative biomarkers in different clinical situations (38)(39)(40).
Proteomic identification of TGCT-related proteins will allow to validate candidate markers for early detection and the prognostic evaluation of testicular tumors. In addition, the molecular mechanisms revealed by these proteomic studies might also guide the development of novel treatments in future.
AUTHOR CONTRIBUTIONS
DM and GG: conceptualization and writing-original draft preparation. DM and FP: literature analysis. FV and AP: writing-review and editing. AP: supervision. | 4,718 | 2019-07-12T00:00:00.000 | [
"Biology"
] |
Structure/Function Analysis of Nonwoven Cotton Topsheet Fabrics: Multi-Fiber Blending Effects on Fluid Handling and Fabric Handle Mechanics
Greige cotton (GC) has attracted interest in recent years as an eco-friendly, functional fiber for use in nonwoven topsheet materials. GC imparts favorable fluid management and sensorial properties associated with urinary liquid transport and indices related to comfort in wearable incontinence nonwovens. Nonwoven GC has material surface polarity, an ambient moisture content, and a lipid/polysaccharide matrix that imparts positive fluid mechanic properties applicable to incontinence management topsheet materials. However, a better understanding of the connection between functionality and compositional aspects of molecular, mechanical, and material property relations is still required to employ structure/function relations beyond a priori design. Thus, this study focuses on the relation of key indices of material fluid and sensorial functions to nonwoven topsheet composition. Greige cotton, polypropylene, bleached cotton, and polyester fiber blends were hydroentangled at 60, 80, and 100 bar. Greige cotton polypropylene and bleached cotton were blended at ratios to balance surface polarity, whereas low percentages of polyester were added to confer whiteness properties. Electrokinetic and contact angle measurements were obtained for the hydroentangled nonwovens to assess surface polarity in light of material composition. Notably, materials demonstrated a relation of hydrophobicity to swelling as determined electrokinetically by Δζ, ζplateau, and contact angles greater than 90°. Subsequently, three blended nonwoven fabrics were selected to assess effects on fluid management properties including topsheet performance indices of rewet, strikethrough, and fluid handling (rate and efficiency of transport to the absorbent core). These materials aligned well with commercial topsheet fluid mechanics. Using the Leeds University Fabric Handle Evaluation System (LUFHES), the nonwovens were tested for total fabric hand. The results of the LUFHES measurements are discussed in light of fiber contributions. Fiber ratios were found to correlate well with improvement in softness, flexibility, and formability. This study provides insights that improves the understanding of the multifunctional properties accessible with greige cotton toward decisions valuable to selecting greige cotton as an environmentally friendly fiber for nonwoven topsheets.
Introduction
For decades, most incontinence topsheet manufacturers have used polypropylene (PP), which is hydrophobic and less expensive than greige cotton. Petroleum-based PP has been the material of choice in the incontinence topsheet market, with approximately 50 million tons produced annually [1], compared to cotton's use, estimated to be approximately 2% (by volume/weight) of the total fiber consumption in nonwovens. Most of the cotton used, at present, in absorbent nonwovens is bleached [2]. However, given consumer demand for more environmentally friendly materials, there exists an opportunity for the increased usage of greige cotton fiber based on previously cited health and environmental issues [3]. Furthermore, the challenges for a PP fiber substitute are economical, process-related, and include properties that achieve optimal fluid handling while also improving on tactile properties. Additionally, it may be possible to reduce some of the environmental impact from PP by blending greige cotton fiber and achieve improved liquid handling and comfort.
The advantageous properties of the cotton may be described in terms of the fiber morphology and molecular composition. The primary cell wall and cuticle of the natural greige cotton fiber contain 1-2% pectins and waxes, as shown in Figure 1 [4]. During cotton fiber processing, pectins and waxes are removed by scouring at high temperature in the presence of caustic substances, exposing the secondary cell wall of the cotton fiber, which confers increased absorbency and hydrophilicity. On the other hand, when components of the fiber cuticle are not removed during nonwoven hydroentanglement, the greige cotton fibers retain the hydrophobicity associated with the cuticle waxes and are loosened, allowing an inlet of water to the hydrophilic cellulosic secondary cell wall. Thus, the resulting combination of hydrophobicity and hydrophilicity is a property useful for topsheet application.
Introduction
For decades, most incontinence topsheet manufacturers have used polypropylene (PP), which is hydrophobic and less expensive than greige cotton. Petroleum-based PP has been the material of choice in the incontinence topsheet market, with approximately 50 million tons produced annually [1], compared to cotton's use, estimated to be approximately 2% (by volume/weight) of the total fiber consumption in nonwovens. Most of the cotton used, at present, in absorbent nonwovens is bleached [2]. However, given consumer demand for more environmentally friendly materials, there exists an opportunity for the increased usage of greige cotton fiber based on previously cited health and environmental issues [3]. Furthermore, the challenges for a PP fiber substitute are economical, process-related, and include properties that achieve optimal fluid handling while also improving on tactile properties. Additionally, it may be possible to reduce some of the environmental impact from PP by blending greige cotton fiber and achieve improved liquid handling and comfort.
The advantageous properties of the cotton may be described in terms of the fiber morphology and molecular composition. The primary cell wall and cuticle of the natural greige cotton fiber contain 1-2% pectins and waxes, as shown in Figure 1 [4]. During cotton fiber processing, pectins and waxes are removed by scouring at high temperature in the presence of caustic substances, exposing the secondary cell wall of the cotton fiber, which confers increased absorbency and hydrophilicity. On the other hand, when components of the fiber cuticle are not removed during nonwoven hydroentanglement, the greige cotton fibers retain the hydrophobicity associated with the cuticle waxes and are loosened, allowing an inlet of water to the hydrophilic cellulosic secondary cell wall. Thus, the resulting combination of hydrophobicity and hydrophilicity is a property useful for topsheet application. The primary purpose of the absorbent topsheet is to facilitate the inlet of liquid to the acquisition and distribution layers, with rapid strikethrough and minimal rewet, providing efficient uptake of urine to the absorbent core as was first documented for disposable diapers [5]. The overall effect is to keep the skin dry and prevent dermatitis [6]. The topsheet is in direct contact with the user's skin and, thus, fluid transports away from the skin with the absence of leakage (efficient fluid handling) that is paramount to topsheet performance [7]. PP topsheets have low water retention on the surface, but have good bulk fluid transport by way of treatment with surfactants to improve wetting behavior [8]. Given PP's low cost and favorable fluid handling properties, it is understandable why it is the current material of choice for hygienic topsheets [9].
Another important goal of topsheet development is fabric comfort at the skin/topsheet interface. Various modelling methods involving both human subjects and instrumentation have examined the fabric-skin interaction in terms of friction (tribology), shear, and pressure for both woven and nonwovens [10]. Cottenden et al. [11] have described improved methodologies for modeling friction, pressure and shear at the skin-fabric interface with care sheets and incontinence pads. Studies have also been The primary purpose of the absorbent topsheet is to facilitate the inlet of liquid to the acquisition and distribution layers, with rapid strikethrough and minimal rewet, providing efficient uptake of urine to the absorbent core as was first documented for disposable diapers [5]. The overall effect is to keep the skin dry and prevent dermatitis [6]. The topsheet is in direct contact with the user's skin and, thus, fluid transports away from the skin with the absence of leakage (efficient fluid handling) that is paramount to topsheet performance [7]. PP topsheets have low water retention on the surface, but have good bulk fluid transport by way of treatment with surfactants to improve wetting behavior [8]. Given PP's low cost and favorable fluid handling properties, it is understandable why it is the current material of choice for hygienic topsheets [9].
Another important goal of topsheet development is fabric comfort at the skin/topsheet interface. Various modelling methods involving both human subjects and instrumentation have examined the fabric-skin interaction in terms of friction (tribology), shear, and pressure for both woven and nonwovens [10]. Cottenden et al. [11] have described improved methodologies for modeling friction, pressure and shear at the skin-fabric interface with care sheets and incontinence pads. Studies have also been reported on skin-fabric interactions as a function of epidermal hydration on friction of human skin against textiles [12,13].
An individual's evaluation of a piece of fabric often employs objective and subjective criteria that involve personal opinion and preference [14]. However, a fabric's color is not considered a factor in handle determination. Although there are longstanding traditional and cultural associations between whiteness, purity, and cleanliness, these are separate issues from fabric handle. There are a few reports correlating fabric whiteness with fabric hand. For example, Yenket et al. concluded "that a pronounced visual effect, color, generally had little influence on the tactile hand sensory properties of fabrics using trained or consumer panelists" [15].
However, consumer preference for whiteness is accepted by textile manufacturers who tailor processes using oxidizing chemicals, optical brighteners, or bluing agents [16,17]. As mentioned earlier, these reagents have an impact on the environment through the use of water, energy, and the production of hazardous waste streams. Thus, the present research has chosen to blend blue polyester fibers to address this issue alongside functional performance properties.
Over the past several decades, a few methods have been developed to quantify human touch perception. Tactile perception is a complex interplay of rubbing, squeezing, and stretching that involves a multidimensional analysis of the fabric. The conclusions from instrumental evaluation methods, such as the Kawabata Evaluation System for Fabrics (KES-F) and Fabric Assurance by Simple Testing (FAST), are not necessarily objective because of their dependency on statistical data [18], artificial neural network outputs [19], and relationships between measured fabric mechanical properties and the subjective hand preferences of human panels using reference fabrics [20].
Furthermore, the Kawabata method examines fabric hand in a limited single dimensional approach, and the standard fabrics used are solely based on 100% woven wool fabrics. In a handle assessment method based on extraction [13,21], a fabric is forced through a narrow opening, such as a ring or nozzle, creating deformations in multiple directions similar to that occurring during human hand interactions. However, the resultant quantitative assessment of fabric handle is based upon statistical principal component analysis and the conclusions of subjective assessment. Additionally, the random fabric deformations that occur in the extraction systems can lead to problems of reproducibility. In the Fabric Touch Tester [22], tactile comfort is evaluated based upon multiple fabric bending deformations, rather than the fabric responses during buckling deformation. Existing approaches outlined here do not quantify the recovery of the fabric following initial deformation, which is known to affect the tactile perception in subjective fabric hand evaluations [23].
Assessment of tactile comfort at the skin-fabric interface strives to improve the objective and subjective evaluations of fabric handle. The criteria for evaluating clothing comfort of woven and nonwoven fabrics has been reviewed [24], with the goal of bridging the divide between objective and subjective handle evaluations. This is at the heart of the fabric handle evaluation system that has been developed by the Leeds University Fabric Handle Evaluation System (LUFHES) [25]. This system mimics the multidimensional rubbing, squeezing, and stretching human interactions that play a role in handle determination. Figure 2b depicts the LUFHES method in operation, which demonstrates the compression buckling of a fabric cylindrical shell. Fabric handle indices are represented by using the different types of fabric deformations produced, which are quantified by the energy consumed during the cyclic deformation/recovery process.
Previously, research identified some of the fluid handling and fabric handle indices that are required for functional incontinence topsheets in hydroentangled greige cotton [26]. However, the structure/function relationship has not been elucidated sufficiently to predict cotton fiber substitution effects in topsheet design. Moreover, this needs to be done in light of recent results that also document the utilization of blue polyester fiber to improve nonwoven whiteness [27]. Thus, this paper addresses intimate blending and hydroentangling cotton fibers with polypropylene, bleached cotton, and polyester fibers to form nonwoven fabrics, which are further characterized at a molecular/material
Materials and Methods
Greige cotton fibers (True Cotton™) were obtained from T. J. Beall of Greenwood, Mississippi. These cotton fibers were precleared using a proprietary process which removed 99.99% of all impurities (field trash, stems, leaves, etc.) in a nonaqueous, mechanical process (http://www.tjbeall.com/natural-fibersnonwoven/true-cotton). Blue polyester fibers were supplied by Palmetto Synthetics LLC of Kingstree, South Carolina. Bleached cotton fibers were produced from a bale of True Cotton™ that were bleached by Tintoria Piana US in Cartersville, Georgia. The polypropylene fibers were purchased from Fibervisions ® of Duluth, Georgia. All fibers were received and used without further modifications.
All nonwoven fabrics were prepared at the USDA Southern Research Center in New Orleans, Louisiana. The manufacturing specifics are as follows: staple fibers were weighed and hand-blended, and then opened on an Opening and Blending Line, consisting of a fiber hopper, Hollingsworth WR cleaner, D106 distributor, and 310 Fine Opener and another D106 distributor. All fiber blends were processed through two times, to achieve optimal openness and blending.
The fiber blends were chute fed to a 40 inch-wide textile card fitted with 4 Cardmaster plates (Saco Lowell), followed by feeding the card web into a commercial crosslapper and needlepunch (NP) machine (Technoplants s.r.l., Pistoia, Italy). The NP processing parameters were varied to achieve the target weights (25-35 g/m 2 ). The NP fabrics were converted into hydroentangled (HE) nonwoven fabrics on a 1 m wide Fleissner pilot-scale hydroentanglement system (Trützschler Nonwovens GmbH, Dülmen, Germany) running at a constant production speed of 5 m min −1 . The hydroentanglement (HE) system utilized two pressure heads: one low pressure for fabric wet-out maintained at a constant pressure of 30 bar during fabric production; and the second is a high-pressure head maintained at 60, 80, and 100 bar during separate fabric production runs. The final bonding nozzle is suspended over a 24 mesh embossing drum to cast a 24 mesh pattern into the nonwoven material. Each strip on the bonding pressure heads consisted of 50 orifices per inch, with an orifice pore size of 120 µm. The water used for the HE fabric production was ambient temperature, which was approximately 25 °C. Following HE, the fabrics were fed directly through a gas-fired fabric-drying oven (Trützschler Nonwovens GmbH) at ~160-180 °C (depending on fiber melt points) and wound into rolls.
The hydroentangled fiber blend samples were analyzed for spectroscopic properties using a Konica Minolta CR-410 chromameter. The chromameter analyzed samples for color space properties in terms of CIELAB color coordinates (L*, a* and b*) using the 2° observer and the CIE standard illuminant
Materials and Methods
Greige cotton fibers (True Cotton™) were obtained from T. J. Beall of Greenwood, Mississippi. These cotton fibers were precleared using a proprietary process which removed 99.99% of all impurities (field trash, stems, leaves, etc.) in a nonaqueous, mechanical process (http://www.tjbeall.com/naturalfibers-nonwoven/true-cotton). Blue polyester fibers were supplied by Palmetto Synthetics LLC of Kingstree, South Carolina. Bleached cotton fibers were produced from a bale of True Cotton™ that were bleached by Tintoria Piana US in Cartersville, Georgia. The polypropylene fibers were purchased from Fibervisions ® of Duluth, Georgia. All fibers were received and used without further modifications.
All nonwoven fabrics were prepared at the USDA Southern Research Center in New Orleans, Louisiana. The manufacturing specifics are as follows: staple fibers were weighed and hand-blended, and then opened on an Opening and Blending Line, consisting of a fiber hopper, Hollingsworth WR cleaner, D106 distributor, and 310 Fine Opener and another D106 distributor. All fiber blends were processed through two times, to achieve optimal openness and blending.
The fiber blends were chute fed to a 40 inch-wide textile card fitted with 4 Cardmaster plates (Saco Lowell), followed by feeding the card web into a commercial crosslapper and needlepunch (NP) machine (Technoplants s.r.l., Pistoia, Italy). The NP processing parameters were varied to achieve the target weights (25-35 g/m 2 ). The NP fabrics were converted into hydroentangled (HE) nonwoven fabrics on a 1 m wide Fleissner pilot-scale hydroentanglement system (Trützschler Nonwovens GmbH, Dülmen, Germany) running at a constant production speed of 5 m min −1 . The hydroentanglement (HE) system utilized two pressure heads: one low pressure for fabric wet-out maintained at a constant pressure of 30 bar during fabric production; and the second is a high-pressure head maintained at 60, 80, and 100 bar during separate fabric production runs. The final bonding nozzle is suspended over a 24 mesh embossing drum to cast a 24 mesh pattern into the nonwoven material. Each strip on the bonding pressure heads consisted of 50 orifices per inch, with an orifice pore size of 120 µm. The water used for the HE fabric production was ambient temperature, which was approximately Materials 2018, 11, 2077 5 of 17 25 • C. Following HE, the fabrics were fed directly through a gas-fired fabric-drying oven (Trützschler Nonwovens GmbH) at~160-180 • C (depending on fiber melt points) and wound into rolls.
The hydroentangled fiber blend samples were analyzed for spectroscopic properties using a Konica Minolta CR-410 chromameter. The chromameter analyzed samples for color space properties in terms of CIELAB color coordinates (L*, a* and b*) using the 2 • observer and the CIE standard illuminant C. L* represents lightness and can be measured independently of color hue. A decrease in lightness is associated with a decrease in fabric reflectance. Normal to the L* axis (lightness) are the +b* to −b* axis and the perpendicular +a* to −a* axis, where b* represents the color yellow (90 • ), −b* blue (270 • ), +a* red (360 • ), and −a* green (180 • ). Each recorded L*a*b* measurement was the average of three reflectance readings, obtained by rotating the samples after each measurement.
The contact angle of a water drop on nonwoven fabrics was measured using a VCA Optima XE (AST Products, Inc., Billerica, MA, USA). Runs were performed with a 5 µL drop of distilled water syringed onto the fabric. The image of the drop was immediately captured and analyzed to yield a contact angle. Contact angles on twelve different areas were measured, and their average value was recorded.
The determination of the ζ-potential was carried out with the Electro Kinetic Analyzer (Anton Paar, Ashland Va.) using the Cylindrical Cell developed for the measurement of fibrous samples. When a fiber absorbs liquid and swells, the surface charges become farther separated, and the absolute value of its ζ-potential decreases. Two kinds of measurements were made on each sample: (1) swell tests to measure the rate and extent of fiber swelling (at a given pH) and (2) a pH titration in which the swelling is measured as a function of pH. All ζ-potential measurements were made in a 1 mM KCl electrolyte.
In the electrokinetic apparatus, the streaming potential is measured, and the zeta potential determined from the Smoluchowski equation [28,29]: where U is the streaming potential-the potential generated when an electrolyte is forced to flow over a stationary charged surface, p the pressure, ε γ and ε 0 the dielectric constant and the vacuum permittivity, η the viscosity, and κ is the conductivity of the measuring fluid. Surface conductivity of the fibrous samples was not taken into account. pH titrations were performed over a pH range of 11.0 to 2.5 to ensure recording both the isoelectric point (IEP) and the plateau potential. (The IEP is the pH at which ζ = zero, and provides insights into the surface association/dissociation processes.) Fabric handle was evaluated using the Leeds University Fabric Handle Evaluation System (LUFHES). Six nonwoven fabric specimens, three in the cross direction (CD) and three in the machine direction (MD), were tested. Cyclic compression buckling of 30%, cyclic shear deformation of 5 • twisting, and fabric-fabric self-friction properties of the six nonwoven fabrics were evaluated during the biaxial deformations of the fabrics. Six fabric handle descriptors and indices are defined based upon the energy consumed either to deform a fabric or to recover the deformation from the deformed state, as defined below.
Fabric sponginess (SP): the extent to which a fabric spontaneously recovers from deformation when the external deformation force is withdrawn (elasticity). Thus, the greater the TFHV value, the greater amount of energy is required to deform the fabric and to recover from the fabric deformation. An additional fabric handle parameter, referred to as fabric smoothness (SN), is defined as the dynamic fabric-to-fabric friction coefficient, or the coefficient of dynamic friction between the fabric and itself. In addition, fabric formability (30b), which is defined to represent a fabric's ability to be conformed to a 3D shape, could be calculated from the fabric compression buckling force measured in the tests.
Liquid acquisition, strikethrough, and rewet tests were performed by Marketing Technology Service, Inc. Kalamazoo, Michigan. According to EDANA/INDA WSP 70.3 and WSP 70.7, the liquid strikethrough time involves determining the time taken for a known volume of test liquid (simulated urine) applied to the surface of a test piece of nonwoven topsheet, which is in contact with underlying standard absorbent pads to pass through the fabric. Rewet testing was performed according to EDANA/INDA WSP 80.10 and WSP 70.8. In this procedure, the wet back of synthetic urine through a prepared sample onto a filter paper is determined. After applying a defined volume of liquid upon the prepared sample (strikethrough time test), a simulated baby weight is applied onto the specimen at a predefined speed and dwell time. Using a specific filter paper and an electronic balance, the amount of liquid that transfers back through the specimen's surface into the filter paper is determined. The liquid acquisition test evaluates in real time how effectively liquids penetrate into the absorbent core of the absorbent device at any flow rate. A 0.9% saline solution is applied to the specimen at a specific volume and flow rate. The dynamic change of the total volume of fluid acquisition and overflow liquids over the duration of the testing was obtained.
Design and Selection of Fabric Compositions
At the outset of the study, ten nonwoven fabrics (Table 1) were designed and prepared based on previous findings outlining the material surface polarity, liquid handling, fabric hand, and whiteness [26], with the goal of demonstrating the cooperative fluid handling, fabric hand, and whiteness profiles that greige cotton and polypropylene combinations exercise. Hydroentanglement of the blends with greige cotton confers combined hydrophobicity and hydrophilicity to the material through exposure of the cellulosic cell wall and loosening of the fiber cuticle waxes, as depicted in the SEM in Figure 3. Polypropylene and polyester also contribute hydrophobicity to the material's polar surface properties. Bleached cotton is included in the ratio blends contributing hydrophilicity.
Materials 2018, 11, x FOR PEER REVIEW 6 of 17 EDANA/INDA WSP 80.10 and WSP 70.8. In this procedure, the wet back of synthetic urine through a prepared sample onto a filter paper is determined. After applying a defined volume of liquid upon the prepared sample (strikethrough time test), a simulated baby weight is applied onto the specimen at a predefined speed and dwell time. Using a specific filter paper and an electronic balance, the amount of liquid that transfers back through the specimen's surface into the filter paper is determined. The liquid acquisition test evaluates in real time how effectively liquids penetrate into the absorbent core of the absorbent device at any flow rate. A 0.9% saline solution is applied to the specimen at a specific volume and flow rate. The dynamic change of the total volume of fluid acquisition and overflow liquids over the duration of the testing was obtained.
Design and Selection of Fabric Compositions
At the outset of the study, ten nonwoven fabrics (Table 1) were designed and prepared based on previous findings outlining the material surface polarity, liquid handling, fabric hand, and whiteness [26], with the goal of demonstrating the cooperative fluid handling, fabric hand, and whiteness profiles that greige cotton and polypropylene combinations exercise. Hydroentanglement of the blends with greige cotton confers combined hydrophobicity and hydrophilicity to the material through exposure of the cellulosic cell wall and loosening of the fiber cuticle waxes, as depicted in the SEM in Figure 3. Polypropylene and polyester also contribute hydrophobicity to the material's polar surface properties. Bleached cotton is included in the ratio blends contributing hydrophilicity.
Fabric Whiteness
Greater hydroentanglement pressures more effectively removed the yellow surface waxes on greige cotton fibers than lower hydroentanglement pressures. There is also a process-related reduction of L* (lightness), from a control high of 94.29 to a low in sample 10 (97% greige cotton) of 85.41 when hydroentanglement pressure was applied at 100 bar versus 60 bar, respectively as shown in Table 1. This trend, related to the effect of hydroentanglement pressure on yellowness, is especially evident in samples which contain increasing percentages of greige cotton. Similarly, as the percentage of greige cotton increased, the b* measurement also increases.
It has been previously reported by Easson et al. that small percentages of blue fiber can give the appearance of increased whiteness using the principles of additive color mixing [27] without the need for chemical agents. Simply stated, when blue and yellow fibers are intimately blended, the human eye cannot distinguish them. The two colored fibers are perceived as white, which is consistent with the colorimetry measurements which categorize the elements of reflective light responsible for the visual effect. This same principle is behind projection televisions which blend red, green, and blue lights to form picture images [30]. Thus, as seen with the samples containing blue polyester fibers, whiteness is increased.
Electrokinetic and Contact Angle Analysis in View of Topsheet Layer Function
Shown in Table 2 are the results of the zeta potential titrations for the ten nonwoven blends hydroentangled at 60 bar. The ζ plateau , isoelectric point, ∆ζ, swell ratios, and rate of swelling (k) have been previously employed to characterize compatible material interfaces of topsheets (TS) and acquisition distribution layer (ADL) materials found in marketed incontinence products [31]. As shown previously for cotton, the swelling of the material, as measured by the ∆ζ, is a function of the materials' expansion within the ion shear plane during pH titration [32][33][34]. Thus, based on polarity, porosity, percent moisture content, and the interaction of hydrophilic and hydrophobic fibers, a material will expand to some extent while undergoing zeta potential titration. The ζ plateau values for the ten blends range from −34 to −25 mv, and the influence of the hydrophobicity of polypropylene is evident, as the blends containing a larger percentage of polypropylene have the more negative zeta potential. This is consistent with the more hydrophobic contact angle values shown in Figure 4. Contact angle testing confirmed a finding first reported by Sawhney et al. and identified in GC topsheet [2,26], demonstrating that higher HE pressures remove the cotton cuticle waxes and increase the hydrophilicity of the material surface. However, seven of the ten nonwovens manufactured at 60 bar had measurable contact angles characteristic of a hydrophobic fabric as shown in Figure 4, and this correlates with a ratio of polypropylene in the blend that is greater than 12%. It is noteworthy that when polypropylene in fabrics with greige content of 60% and 76% is decreased by fifty percent of the fiber composition in the fabric, the contact angle is reduced from 118° to around 105°, and from 105° to 82°, respectively.
Analysis of the Fabric Tactile Properties
The results of the LUFHES tactile descriptor valuation are shown in Table 3. A comparison of total fabric handle values (TFHV) of the six nonwovens examined in this study is shown in Figure 5. Contact angle testing confirmed a finding first reported by Sawhney et al. and identified in GC topsheet [2,26], demonstrating that higher HE pressures remove the cotton cuticle waxes and increase the hydrophilicity of the material surface. However, seven of the ten nonwovens manufactured at 60 bar had measurable contact angles characteristic of a hydrophobic fabric as shown in Figure 4, and this correlates with a ratio of polypropylene in the blend that is greater than 12%. It is noteworthy that when polypropylene in fabrics with greige content of 60% and 76% is decreased by fifty percent of the fiber composition in the fabric, the contact angle is reduced from 118 • to around 105 • , and from 105 • to 82 • , respectively.
Analysis of the Fabric Tactile Properties
The results of the LUFHES tactile descriptor valuation are shown in Table 3. A comparison of total fabric handle values (TFHV) of the six nonwovens examined in this study is shown in Figure 5. From a structural standpoint, the blends analyzed for fabric handle indices may be grouped into three classifications: (1) samples 1 and 4 contain a higher percentage of polypropylene (PP) fibers (about 24-40%) relative to the greige cotton (60-76%); (2) Fabric 7 has a greater proportion of greige cotton fibers (90%) and relatively less PP fiber (10%); (3) In the third group (6,8,9), all the hydroentangled fabrics contain a relatively smaller proportion of polypropylene (PP) fibers (about 10-12%) and a relatively large proportion of greige cotton fibers (76-90%), with the presence of bleached cotton (3-10%) and very low levels (0-2%) of blue polyester fibers.
There are some trends in fabric smoothness. Fabrics 4 and 9 have the smoothest surface (SN = 0.675 and 0.730, respectively). This is consistent with fabric 9 being the softest fabric and fabric 4 containing a greater proportion of PP fibers (24%), as they might link to the formation of surface ribs/grooves formed during hydroentanglement process; both of them might have formed a smoother surface during the hydroentanglement process. Fabrics 1 and 6 have medium smoothness (SN = 0.902 and 0.841, respectively) consistent with a greater proportion of PP fibers (24% and 12% respectively) and the presence of a small amount of polyester. Fabric 7 and 8 have the least fabric smoothness (SN = 0.968 and 0.996 respectively), and this is consistent with these fabrics containing the greatest proportion of greige cotton fibers (90%), and polyester fibers that are absent.
There are some trends in fabric smoothness. Fabrics 4 and 9 have the smoothest surface (SN = 0.675 and 0.730, respectively). This is consistent with fabric 9 being the softest fabric and fabric 4 containing a greater proportion of PP fibers (24%), as they might link to the formation of surface ribs/grooves formed during hydroentanglement process; both of them might have formed a smoother surface during the hydroentanglement process. Fabrics 1 and 6 have medium smoothness (SN = 0.902 and 0.841, respectively) consistent with a greater proportion of PP fibers (24% and 12% respectively) and the presence of a small amount of polyester. Fabric 7 and 8 have the least fabric smoothness (SN = 0.968 and 0.996 respectively), and this is consistent with these fabrics containing the greatest proportion of greige cotton fibers (90%), and polyester fibers that are absent.
Formability
The fabrics 6 and 9 have the greatest formability (FMR = 0.35 and 0.37 mm 2 , respectively). Both fabrics contain polyester, and are principally composed of greige cotton fibers (86% for the fabric 6 and 93% for the fabric 9). Fabrics 4 and 7 have medium formability (FMR = 0.32 mm 2 ), they contain a lower proportion of cotton fibers (76% and 90% of cotton fibers, respectively) in comparison with fabrics 6 and 9. Fabrics 1 and 8 have the smallest formability (FMR = 0.27 and 0.29 mm 2 , respectively), and they are, coincidently, the fabrics containing smallest (60%) and greatest (90%) proportion of cotton fibers, respectively, without the presence of polyester.
Strikethrough, Rewet, and Fluid Acquisition of Penetrant-Treated Greige Cotton Fabrics
A group of three samples was selected to evaluate functional liquid handling properties central to incontinence management. In contrast with previously reported studies where the functional performance of 100 percent greige cotton was demonstrated, a ratio of greige cotton to polypropylene in the bi-component fiber blends was compared with a blend containing mostly greige cotton with lower percentages of polypropylene, bleached cotton, and polyester [26]. As shown in Figure 6, the rewet of the bi-component blend, with a ratio of 60:40 GC/PP, performed significantly better than the other two containing 90 percent GC.
The fabrics 6 and 9 have the greatest formability (FMR = 0.35 and 0.37 mm 2 , respectively). Both fabrics contain polyester, and are principally composed of greige cotton fibers (86% for the fabric 6 and 93% for the fabric 9). Fabrics 4 and 7 have medium formability (FMR = 0.32 mm 2 ), they contain a lower proportion of cotton fibers (76% and 90% of cotton fibers, respectively) in comparison with fabrics 6 and 9. Fabrics 1 and 8 have the smallest formability (FMR = 0.27 and 0.29 mm 2 , respectively), and they are, coincidently, the fabrics containing smallest (60%) and greatest (90%) proportion of cotton fibers, respectively, without the presence of polyester.
Strikethrough, Rewet, and Fluid Acquisition of Penetrant-Treated Greige Cotton Fabrics
A group of three samples was selected to evaluate functional liquid handling properties central to incontinence management. In contrast with previously reported studies where the functional performance of 100 percent greige cotton was demonstrated, a ratio of greige cotton to polypropylene in the bi-component fiber blends was compared with a blend containing mostly greige cotton with lower percentages of polypropylene, bleached cotton, and polyester [26]. As shown in Figure 6, the rewet of the bi-component blend, with a ratio of 60:40 GC/PP, performed significantly better than the other two containing 90 percent GC. As shown in Figure 7, it is apparent that a higher percentage of greige cotton facilitates improved strikethrough, since an approximate 40 percent increase in the rate of liquid uptake was observed between samples 1 (60:40, GC/PP) and 7 (90:10, GC/PP).
As shown in Figure 8, the rate of overflow and leakage were as much as 50 percent improved in the 90 percent greige cotton blend at the first insult of 20 mL. In the second insult, the leakage and overflow differences were decreased to 25 percent, which was still retained in the third insult in composition 7. As shown in Figure 7, it is apparent that a higher percentage of greige cotton facilitates improved strikethrough, since an approximate 40 percent increase in the rate of liquid uptake was observed between samples 1 (60:40, GC/PP) and 7 (90:10, GC/PP).
As shown in Figure 8, the rate of overflow and leakage were as much as 50 percent improved in the 90 percent greige cotton blend at the first insult of 20 mL. In the second insult, the leakage and overflow differences were decreased to 25 percent, which was still retained in the third insult in composition 7.
Discussion
At the outset of the study, ten nonwoven fabrics were designed and prepared, based on previous findings outlining the material surface polarity, liquid handling, fabric hand, and whiteness [26], with the goal of demonstrating the cooperative fluid handling, fabric hand, and whiteness profiles that greige cotton and polypropylene combinations exercise. Hydroentanglement of the blends with greige cotton confers combined hydrophobicity and hydrophilicity to the material through exposure of the cellulosic cell wall and loosening of the fiber cuticle waxes, as depicted in the SEM in Figure 3. Polypropylene and polyester also contribute hydrophobicity to the material's polar surface properties. Bleached cotton is included in the ratio blends contributing hydrophilicity.
Greater hydroentanglement pressures more effectively removed the yellow surface waxes on greige cotton fibers than lower hydroentanglement pressures. There is also a process-related reduction of L* (lightness), from a control high of 94.29, to a low in sample 10 (97% greige cotton) of 85.41, as hydroentanglement pressures decreased from 100 bar to 60 bar. This trend related to the effect of hydroentanglement pressure on yellowness is especially evident in samples which contain increasing percentages of greige cotton. Similarly, as the percentage of greige cotton increased, the b* measurement also increases.
In 2017 Easson et al. reported that small percentages of blue fiber can give the appearance of increased whiteness using the principles of additive color mixing [27] without the need for chemical agents. When intimately blended with light yellow greige cotton, certain shades of blue fibers cannot be distinguished from the yellow fibers by the human eye. The two colored fibers are perceived as white, which is consistent with the colorimetry measurements which categorize the elements of reflective light responsible for the visual effect. This same principle is behind projection televisions which blend red, green, and blue lights to form picture images [30]. Thus, as seen with the samples containing blue polyester fibers, whiteness is increased.
Considerations of material surface polarity are central to incontinence topsheet (TS) and acquisition distribution layer design (ADL) [31]. The use of electrokinetic measurements to demonstrate the relative contributions of TS and ADL to functional design has previously been demonstrated for heavy to light incontinence materials, and the ability to use greige cotton to modulate functional performance has been reported [26,31].
The material zeta potential (ζ-potential) is characterized through pH titration correlated to change in charge (millivolt increments), and ζ plateau is the point where charge equilibrium is attained on the material surface. Zeta plateau (ζ plateau ) can be either positive or negative, depending on its surface chemistry. Zeta potential analysis is highly applicable to absorbent materials, and provides a tool useful in the analysis of fluid dynamics in aqueous systems based on electrochemical double layer model [28].
The electrochemical double layer model has parallel properties to the fluid transport that occurs at the solid-liquid interface of incontinence material layers, i.e., there is transport of fluid through a material with a certain surface polarity that results in the swelling of a polar, porous material. Thus, zeta potential decrease (a change in the electrostatic potential at the shear plane of the electrochemical double layer) is caused by swelling of the fibers and outward movement of the aqueous shear plane where ions are in contact with the outer Helmholtz ion plane on the fiber surface [28,29]. An increase in swelling of the material occurs as the shear plane moves out to the more diffuse layer of ions surrounding the surface causing a decrease in ζ-potential.
The indication of a hydrophobic surface on greige cotton nonwovens, shown here, is consistent with incontinence topsheet functionality [26,33]. In addition, the zeta potential values are within an acceptable range, based on previously reported values for the TS of incontinence underwear and moderate or light incontinence products that contain polypropylene, cellulose, and polyester [31]. On the other hand, the ∆ζ values are within an acceptable range (0.019-0.15) for TS employed in incontinence underwear, moderate incontinence liners, and bed pads that contain polypropylene, cellulose, and modified acrylic acid [31]. As previously demonstrated, the degree of swelling of the TS and ADL in an incontinence product plays a role in designing functional performance parameters, such as strikethrough, rewet, and fluid transport to the absorbent core [33]. Rate of swelling (k) varied from 0.021 to 0.001 min −1 , and was more accelerated in materials containing higher levels of bleached cotton (greater than 3%). Rate of swelling is also mediated by the amphiphilic surface polarity of the material, i.e., a combination of bleached cotton, greige cotton, and polypropylene.
Fabric handle analysis using the Leeds University Fabric Handle Evaluation System (LUFHES) method was limited to six samples representative of a range of surface polarity. As stated earlier, the greater the TFHV value derived in the LUFHES, the more energy that is required to deform the fabric and to recover the deformed fabric. Thus, a lower TFHV value is more desirable for a softer fabric. The results of the LUFHES tactile descriptor valuation are shown in Table 3. Table 3 indicates that the standard deviation of the TFHV of these six fabrics were in a relatively large range (between 7% and 11% of their average TFHV). Such greater deviation was mainly due to the well-known anisotropy of hydroentangled nonwoven fabric properties in both machine direction (MD) and cross direction (CD) [34,35]. It is recognized that nonwoven fabrics have inherent anisotropic, and non-uniform properties in different directions. Thus, there is a significant difference between the average values of each index in different directions (i.e., frequently, the average value of each index in MD is about 1.2~1.6 times that in CD). The great difference of the objective measurement results in a different orientation in which a fabric may be positioned. This aspect of the assay may also be inherently one of the diverse subjective evaluation results among different evaluators in a human evaluation panel, and is consistent with the difference of the objective measurement results in different directions of a fabric, leading to the diverse subjective evaluation result in human evaluation panels [14,20].
Hydroentangled fabrics containing more greige cotton fibers and a small proportion of blue polyester fibers yield a softer fabric than those with higher PP. It is noticed that fabric 1, containing least greige cotton, is not the least soft fabric, especially in comparison with fabric 4. This is possibly due to the reason that polypropylene fibers have a relatively greater bending rigidity than cotton fibers, and might need greater energy in the hydroentanglement process to form fiber tanglement; thus, fabric 1, which has a higher proportion of PP fibers (40%), might have smaller hydroentanglement intensity (HI) (34b) than fabric 4, which has a smaller proportion of PP fibers (24%) in the level of the water jet pressure applied; smaller hydroentanglement intensity (HI) would lead to a relatively loose and bulky fabric, which should be relatively softer (34b). Previously, it was shown that a 50/50 greige cotton/polypropylene blend had a TFHV that is five times less than a commercial spunbond nonwoven, and this is predominantly due to higher softness and more sponginess, flexibility, and crispiness (18). A similar trend is also found in this study with the fabric stiffness (ST), sponginess (SP), and TFHV. For example, lower amounts of both polypropylene and bleached cotton tend to improve fabric THFV and flexibility when the fabric is predominantly composed of greige cotton.
Fabric formability (FMR) represents a fabric's ability to be conformed to a 3D shape. Thus, a greater FMR value represents the fabric's greater ability to be conformed to a 3D shape. The relative order of formability observed in this study is consistent with the presence of blue polyester fibers in those fabrics. This suggests the potential to utilize polyester to enable greater formability. Moreover, in the absence of polyester fibers, it is possible that either too small (60%) or too high (90%) amounts of cotton fibers could lead to smaller fabric formability, and fabrics containing a suitable proportion (76-90%) of greige cotton fibers enable a relatively greater formability. This suggests that cuticle waxes from greige cotton may increase formability. This finding also suggests that more malleable hydrophobic materials, such as wax and polyester, may improve fabric formability.
Optimal strikethrough, rewet, and fluid acquisition properties in an incontinence topsheet are essential for modulating levels of incontinent episodes in light, moderate, and heavy capacity categories [31]. As shown in Figure 6, the first rewet insult demonstrated sufficiently low rewet for all three samples, consistent with being undetectable by the skin [36]. On the other hand, the rewet values resulting from the second liquid insult to the material was within the margins of error of rewet previously reported for 100 percent GC, and twice the rewet capacity of a commercial spunbond sample [26]. The significant difference found in the rewet values between 60 percent greige cotton and 90 percent greige cotton is consistent with a higher ∆ζ and swell ratio found in the electrokinetic profile of nonwoven fabric 1 versus fabrics 7 and 9 ( Table 2). Swell ratio has previously been correlated with liquid gradient and polar surface functionality between and within incontinence product layers i.e., topsheet, acquisition, and distribution layers [31]. In the context of topsheet rewet functionality, improvement in rewet is a function of interfacial liquid uptake and retention between the topsheet prototype material (60:40, GC/PP) and the absorbent core. Moreover, a larger swell ratio in fabric 1 suggests a greater expansion of the hydrophobic channels and interstitial spaces distributed within the test sample, and between the sample and absorbent core as pressure (2.5 kPa) simulating body weight is exerted, and a capacity to retain more liquid. This is also interesting in light of previous determinations showing that polypropylene swells at a greater rate than greige cotton [33]. However, this may be due to its lower density and porous nature. In this study, materials with less polypropylene swell more. Thus, the polar gradient formed between hydrophilic and hydrophobic portions of the material work synergistically to improve absorbency and retention of liquid by increasing the web interstitial space and modulating pore size [37,38].
In contrast with rewet functionality found in the topsheet, designing optimal strikethrough is largely dependent on accelerating both the wicking and retention capacity of the material. Strikethrough properties are, in part, optimized in commercial polypropylene samples by the application of surfactants [8] to promote rapid inlet of urine. However, the strikethrough values were somewhat attenuated compared with previously reported ones for greige cotton, due to the absence of surfactant treatment. The improved strikethrough observed with the 90 percent greige cotton is consistent with the structure of greige cotton which, as discussed previously, has some amphiphilic properties characteristic of surfactants i.e., molecules with a polar and nonpolar face [33]. Moreover, the distribution of bleached cotton on the surface of the material would be expected to increase wicking.
One of the most interesting results of the fluid handling experiments was the improvement in liquid acquisition observed over commercial spunbond topsheet performance in the multi-dose liquid acquisition test. Thus, in contrast with previous studies where 100 percent greige cotton was comparable to spunbond polypropylene topsheets, it is evident, here, that GC/PP blends improve liquid acquisition properties over commercial spunbond polypropylene topsheets.
Conclusions
Greige cotton, polypropylene, bleached cotton, and polyester fibers, were blended in discrete ratios and examined for whiteness, tactile, and fluid handling properties for topsheet application. In Leeds University Fabric Handle Evaluation System (LUFHES) tactile analyses, nonwoven blends containing low percentages of polyester and polypropylene and high percentages of greige cotton and bleached cotton were shown to have the lowest total fabric handle value (TFHV). Fiber ratios demonstrated a relation of hydrophobicity to swelling, as determined electrokinetically by isoelectric point (∆ζ, ζ plateau , and contact angles greater than 90 • . The influence of the hydrophobicity of polypropylene is evident as the blends containing a larger percentage of polypropylene have the more negative zeta potential. Seven of the ten nonwovens manufactured at 60 bar, with a ratio of polypropylene in the blend greater than 12%, had measurable contact angles greater than 90 degrees and are hydrophobic, which is consistent with incontinence topsheet functionality. The rate of swelling (k) was greater in nonwovens containing a greater amount of bleached cotton and a blend of amphiphilic fibers. Blue polyester fiber did improve fabric whiteness and factored in favorably in handle determination, but whiteness itself was not a factor in handle. On the other hand, improved liquid acquisition may work synergistically with improved formability, as observed with fluid management and fabric handle, respectively. The mechanics of fabric hand and fluid management functioning synergistically could improve incontinence material leakage and promote a more efficient retention of fluid in the absorbent core. The analyses of three nonwoven fabrics for rewet, strikethrough, and fluid handling aligned well with commercial topsheet fluid mechanics. Finally, this research provides valuable insights into nonwoven topsheet material selection, and provides a basis for better understanding of the fluid and hand dynamics that greige cotton imparts for incontinence applications. | 11,068.4 | 2018-10-24T00:00:00.000 | [
"Materials Science",
"Engineering"
] |
Observation of giant spin-split Fermi-arc with maximal Chern number in the chiral topological semimetal PtGa
Non-symmorphic chiral topological crystals host exotic multifold fermions, and their associated Fermi arcs helically wrap around and expand throughout the Brillouin zone between the high-symmetry center and surface-corner momenta. However, Fermi-arc splitting and realization of the theoretically proposed maximal Chern number rely heavily on the spin-orbit coupling (SOC) strength. In the present work, we investigate the topological states of a new chiral crystal, PtGa, which has the strongest SOC among all chiral crystals reported to date. With a comprehensive investigation using high-resolution angle-resolved photoemission spectroscopy, quantum-oscillation measurements, and state-of-the-art ab initio calculations, we report a giant SOC-induced splitting of both Fermi arcs and bulk states. Consequently, this study experimentally confirms the realization of a maximal Chern number equal to ±4 in multifold fermionic systems, thereby providing a platform to observe large-quantized photogalvanic currents in optical experiments.
T he discovery of topological insulators reinvigorated the understanding of the electronic band structure and inspired generalization of the topological band theory concerning solid states [1][2][3][4][5] . This led to the discovery of quasiparticle excitations of the Dirac and Weyl fermions within solidstate materials characterized by a linear band crossing in metals along with the creation of a direct analogy between the said fermions and fundamental particles in high-energy physics [6][7][8][9][10][11][12][13][14][15] . On the other hand, quasiparticles within electronic band structures need not necessarily follow the Poincare symmetry pertaining to high-energy physics. Instead, they adhere to the crystal symmetry such that new types of fermionic excitations can be realized within solid states without having counterparts in highenergy physics [16][17][18][19][20] .
Multifold fermions protected by chiral crystal symmetry attracted extensive attentions recently 16,[21][22][23][24][25][26] . In comparison with Dirac fermion with zero topological charge and Weyl fermions with Chern number ±1, multifold fermions in chiral crystals host large Chern numbers and chiral Fermi arcs on their surface states (SSs). Since these symmetry-enforced multifold fermions locate at high-symmetry time-reversal invariant momenta, realization of long-surface Fermi arcs expanding throughout the Brillouin zone (BZ) becomes topologically guaranteed. These Fermi arcs are orders of magnitude larger and highly robust compared with those in any other Weyl semimetal. This affords a natural advantage over twofold degenerate Weyl fermions with regard to detection of Fermi-arc states. Identification of multifold fermions with large Chern numbers has previously been performed via observation of surface chiral Fermi arcs using angle-resolved photoemission spectroscopy (ARPES) [21][22][23][24][25] as well as the remarkable quantized circular photogalvanic effect 26 .
According to theoretical predictions, the largest topological charge from multifold fermions has a Chern number ±4 hosted by compounds with space groups P2 1 3 (No. 198), especially in compounds such as CoSi, RhSi, PtAl, CoGe, RhGe, PdGa, etc. 16,[21][22][23][24]26 . Protected by this topological Chern number of ±4, there exist four Fermi arcs crossing surface-projected multifold fermions at high-symmetry points. However, given the small spin splitting of Fermi arcs, all ARPES measurements performed to date have only confirmed the existence of chiral Fermi arcs connecting projected multifold fermions. That is, the exact number of Fermi arcs that exist has yet remained unclear. In other words, the theoretical Chern number of ±4 has so far not been experimentally verified by any surface-detection experiment. Realization of large spin splitting of the Fermi arc requires a very strong spin-orbit coupling (SOC) in the chiral crystals. Among all chiral multifold fermionic materials investigated thus far, PtGa demonstrates the strongest SOC. This paper reports results obtained using a combination of high-resolution ARPES, quantum oscillations, and state-of-the-art ab initio calculations to illustrate the giant spin splitting of topological states within a new chiral topological semimetal PtGa.
Results
Crystal growth and electrical transport measurement. In this study, high-quality PtGa single-crystals were grown using the selfflux technique, as discussed in Methods and Supplementary Note 1. The crystal symmetry was confirmed via rigorous single-crystal diffraction analysis. The estimated Flack parameter value of 0.03 (4) indicates the existence of a single structural chirality in the crystal. The samples demonstrated metallic behavior throughout the measured temperature range ( Supplementary Fig. 2). The observed large residual resistivity ratio (RRR = ρ(300 K)/ρ(2 K)) of 84 and giant magnetoresistance of roughly 1000% at 2 K (Supplementary Fig. 2) reflect the high quality of the crystals. Values of the carrier concentration and mobility were estimated to be 2.1 × 10 21 cm −3 and 4650 cm 2 V −1 s −1 , respectively, using fielddependent Hall-resistivity data ( Supplementary Fig. 3) and the longitudinal resistivity at 2 K.
Crystal structure and electronic structure. PtGa crystallizes in the non-symmorphic chiral space group P2 1 3 (no. 198) with lattice parameter a = 4.9114(3) Å (Fig. 1a). The corresponding simple cubic BZ is depicted in Fig. 1c with high-symmetric momenta-Γ, X, M, and R. Results of ab initio calculations and (001) Fermi-surface (FS) projection demonstrate the existence of hole and electron pockets at Γ and M, respectively, resulting in Fermi arcs to span over the entire BZ surface. Subsequently, giant spin splitting of the said Fermi arcs was observed post SOC incorporation, as depicted in Fig. 1d. Consequently splitting also occurs for the bulk bands. Owing to the crystal symmetry, a threefold degenerated point, the location of which coincides with that of Γ, follows the spin-1 Hamiltonian and hosts a topological charge of Chern number +2. Likewise, band degeneracy at the location of momentum R yields a double-Weyl fermion with Chern number −2, thereby causing the entire system to follow the "no-go theorem" (Fig. 1c, e, f). Considering SOC, the spin-1 fermionic excitation at Γ splits into a doubly-degenerated Weyl and fourfold degenerated Rarita-Schwinger-Weyl fermions with topological charges of Chern number +4. Meanwhile, the double-Weyl fermions at R transform into sixfold fermionic points with Chern numbers −4 as well as a trivial double-degenerated point (Fig. 1b, d-f). Therefore, it can be inferred that PtGa serves as an ideal platform to visualize the effect of strong SOC on quasiparticle excitations of multifold fermions in chiral-topology semimetals.
Fermi-surface topology from quantum oscillation. Quantum oscillations in single-crystalline PtGa were investigated in this study to detect spin-split FS pockets due to strong SOC, as illustrated in Fig. 1b Fig. 2b and Supplementary Fig. 6). We analyzed the corresponding fast Fourier transformation (FFT) to determine the dHvA frequencies and, for simplicity, only the 2-K data are depicted in Fig. 2d. Corresponding dHvA oscillations and FFT results obtained for all intermediate temperatures are shown in Supplementary Fig. 6. The different dHvA frequencies in the FFT results were ascribed to extremal areas of the FS by constructing the full FS with the help of ab initio calculations, details of that are discussed in the Methods section. Evidently, the energy bands of PtGa get spin-split owing to the strong SOC and non-centrosymmetric crystal structure and this results in the emergence of FS pairs having similar shapes but largely different sizes. The complete FS along with the BZ is illustrated in Fig. 2e, revealing that the FSs are mainly spread around the high-symmetry locations Γ, R, and M. In this study, all dHvA frequencies obtained via FFT could be easily matched with calculated k-space extremal FS cross-section areas with their normal vectors along the z-axis. For simplicity, however, FSs corresponding to only the Γ-point are depicted in the inset of Fig. 2d. The identified FSs for the other FFT frequencies are presented in Supplementary Fig. 8. The fourfold band crossing at the Γ-point generates two electron types nearspherical FSs (with extremal frequencies of 31.9 T for the α 1 and 93.96 T for the α 2 orbit) and two square-box-shaped FSs. For each of this latter FS, two extremal cross-sections were determined that match well with the experimental FFT results in the frequency range depicted in Fig. 2d (β 1 at 113.22 T and β 2 at 259.14 T; γ 1 at 492.6 T and γ 2 at 679.37 T). The observed large frequency difference of 187 T for the latter pair clearly reveals the giant spin splitting of FSs due to strong SOC. The effective mass m* of the various spin-split Fermi pockets is estimated from the temperature dependence of the corresponding dHvA frequencies (Fig. 2c) using the thermal damping factor of the Lifshitz-Kosevich formula, R T ¼ X=sinhðXÞ. Here, X ¼ 14:69 m* T=B and B is the magnetic field averaged over 1/B. The extremal cross-sectional areas A F , Fermi wave vector k F and Fermi velocity v F of the FS pockets shown in Fig. 2d are estimated using the Onsager relation Angle-resolved photoemission spectroscopy. Using low-energy high-resolution ARPES on the high-quality single crystal, we investigate the electronic band structure of PtGa. An FS intensity plot was obtained for the 1st BZ along with ARPES intensity plots along high-symmetry directions on the (001) surface, as depicted in Fig. 3a and c. The calculated FS is presented in Fig. 3b, whereas, the band structure combining surface and bulk states are shown in Fig. 3d. By comparing the ARPES spectrum against results obtained from ab initio calculations-including the electron band at Γ and electronic pocket at M-it is evident that the ARPES spectrum is well reproduced by calculated SSs. Owing to the relatively low photon energy, no bulk states were observed in the ARPES spectrum. As observed via our calculations, four spinsplit surface bands correspond to Fermi-arc-related states that originate from the Γ point. Two of these bands (green lines) extend to the M point at the right side, while the other two (blue lines) connect the left M point. The experimental data are in good agreement with the calculations. In Fig. 3c, four crossing points are observed between M-Γ-M, as indicated with orange arrows along the white line of Fig. 3a. Each crossing point contains two spin-split Fermi arcs. However, due to the finite ARPES resolution, the spin splitting of the Fermi arcs are difficult to distinguish along the M-Γ-M direction. Compared with the experiments, the calculated Fermi arcs have much more twisted paths, resulting more crossing points.
Spin-split Fermi arc. We closely look into the PtGa SSs to realize the giant SOC-induced spin splitting of the Fermi arc. On the (001) surface, the Fermi arc was observed to cross the entire BZ that helically wraps around the two M point with opposite chirality. Figure 4a Fig. 4b. To demonstrate the splitting of the Fermi arc, we acquired two ARPES spectra along two cuts, as indicated by the white and green lines in Fig. 4a. Because of the strong SOC in PtGa, we observed a distinct splitting of the Fermi arc from the band dispersion. The largest energy difference of the splitting in Fig. 4c and d is~0.2 eV. We also calculated the surface spin texture for ΔE ¼ E À E F ¼ À0:1 eV as shown in Supplementary Fig. 10. It is evident that each pair of neighboring surface states shows different spin textures, with almost opposite spin orientations. This indicates that each pair of neighboring surface states are split from one state. In Fig. 4e and f, we show the band structures along loop 1 and 2, as indicated in Fig. 4a. The loop 1 enclosing M, presents two right-moving surface band crossings E F , as indicated with black arrows. The band splittings is clearly observed in Fig. 4c, d and also shown in Fig. 4e. Therefore, each crossing contains two surface bands, suggesting a chiral charge of the M point as |C| = 4. The Γ point is enclosed by loop 2, which shows six-band crossings, including four left-movings and two right-movings. Here the right-and left-moving crossings denote opposite chiral charge. Therefore, one pair of right-and left-moving crossings cancels out and does not contribute to the chiral charge. Since each crossing contains two spin-split Fermi arcs, the net crossing count along loop 2 is four; and the resulting chiral charge of the Γ point is |C| = 4. This is the first experimental observation of SOCinduced spin-split Fermi arcs and the verification of the maximal Chern number of 4 in topological chiral crystals. Since multifold fermions are a generic feature of all the chiral topological crystals with no. 198 SG, there, indeed, exist four Fermi-arc SSs connecting the Γ and M points.
Discussion
In conclusion, this paper presents a comprehensive investigation performed using high-resolution ARPES, quantum-oscillation measurements, and state-of-the-art ab initio calculations to examine a giant SOC-induced splitting of Fermi arcs and bulk states. Owing to its large SOC, chiral PtGa demonstrates strong spin splitting, as observed via both dHvA quantum-oscillation analysis and surface ARPES measurements. The splitting of Fermi arcs connecting the time-reversal invariant points of Γ and M directly confirms a Chern number of ±4 for chiral multifold fermions that exist in this class of topological materials. Thus, the proposed study confirms realization of a maximal Chern number equal to ±4 for the first time in multifold fermionic systems. SOC can be considered as an effective parameter for tuning the sharpness of surface Fermi arcs, thereby paving the way for further observation and exploration of Fermi-arc-related phenomena in multifold chiral fermions.
Methods
PtGa crystal growth and structural refinement. PtGa single crystals were grown from the melt using the self-flux technique. A polycrystalline ingot was first prepared by arc melting stoichiometric amounts of constituent metals with 99.99% purity in an argon atmosphere. Subsequently, the single-phase ingot was crushed, placed in an alumina crucible, and sealed within a quartz tube. The assembly was then melted inside a commercial box-type furnace at 1150°C and maintained at that temperature for 10 h to ensure homogeneous mixing of the melt. The sample was then slowly cooled to 1050°C at a rate of 1°C/h followed by further cooling to 850°C at a rate of 50°C/h. Finally, the sample was annealed at 850°C for 120 h prior to being cooled to 500°C at a rate of 5°C/h. As observed, the annealing process has a major impact on the quality of the grown crystal. In electrical transport, the RRR value significantly increases (up to~80) with post annealing compared with the corresponding value (~24) for fast-cooled single crystals. The obtained single crystal measured~6 mm in diameter and 17 mm in length, as depicted in Supplementary Fig. 1. Single crystallinity of the grown crystal was first verified at room temperature using a white-beam backscattering Laue X-ray setup. Observation of of single single, sharp Laue spots clearly revealed the excellent quality of the grown crystal sans any twinning or domains. A representative Laue pattern, superposed with a theoretically simulated one is also depicted in Supplementary Fig. 1. Chemical composition of the PtGa crystal was verified via energydispersive X-ray (EDX) spectroscopy, results of which demonstrated good agreement with the target composition of PtGa. To analyze structural chirality of the grown crystal, rigorous single-crystal X-ray diffraction experiments were performed, results of which are discussed in Supplementary Information Note 1 structural characterization section. The refined crystal structure confirmed Form A in ref. 29 Magnetic and electrical transport measurements. The magnetization measurements were performed using a commercial MPMS3 from Quantum Design. In electrical transport, the longitudinal and Hall resistivity were measured using a low-frequency ACT option in a physical property measurement system (PPMS-9T, Quantum Design).
Angle-resolved photoemission spectroscopy. ARPES experiments were carried out at the Berliner Elektronenspeicherring für Synchrotronstrahlung (BESSY) (beamline UE112-PGM-1) with a Scienta Omicron R8000 analyzer, and at beamline 13U of the National Synchrotron Radiation Laboratory (NSRL) with a Scienta Omicron DA30 analyzer. The single-crystalline samples were oriented and finely polished on the (001) surfaces. The samples were Ar-ion sputtered for 30 min in ultra-high vacuum chamber, and then annealed at 680°C for 30 min.
First-principles calculations. The electronic structure calculations were performed based on density functional theory (DFT) by using the full-potential localorbital code 30 with a localized atomic basis and full-potential treatment. The exchange and correlation energies were considered in the generalized gradient approximation (GGA) level 31 . We have projected the Bloch wave functions into the atomic-orbital-like Wannier functions, and constructed the tight binding model Hamiltonian. With the tight binding model Hamiltonian, the SSs were calculated in a half-infinite boundary condition using the Green's function method 32,33 .
Data availability
The supporting data that is used to illustrate the findings of this study are available on request from the corresponding authors K.M. and C.F. upon request. | 3,789.8 | 2020-03-17T00:00:00.000 | [
"Physics"
] |
Ship Safety Officers' Perceptions and Attitudes Toward Near-Miss Management Systems
Unlike learning from accidents, learning from near misses is based on events that caused no injuries or damage. Therefore, reporting and investigating near-miss events in shipping could be considered a more convenient means of reducing accidents and safety improvements than accident investigations. However, to facilitate learning from near misses, an adequate and efficient Near-Miss Management System must be implemented on board ship. Since ship Masters and Safety Officers are responsible for the efficiency of the Near-Miss Management System (NMMS) on the shipboard side, their attitudes and opinions on implemented systems might be considered indicators of its quality. Therefore, the questionnaire was developed and distributed among Masters and Safety Officers to collect their perceptions of and attitudes toward Near-Miss Management Systems. Furthermore, the paper aims to examine the relationship between the respondents' ranks (Masters and Safety Officers), the type of ship they are serving
INTRODUCTION
A marine accident is an undeliberate and unexpected event or a sequence of events resulting in injury or death of people or damage to property and environment, which occurred directly with the operations of a ship (IMO, 2008a). It is necessary to report, investigate and analyze marine accidents, and disseminate lessons learned publicly to improve safety at sea. Although systems and equipment comprised of new technologies aiming to prevent accidents and improve safety at sea are introduced on ships, accidents still happen. The reason for it might be the human-technology interaction, whereas relatively new causes of marine accidents emerge, like inadequate knowledge of own ship systems, overreliance on technology, and complacency (Bielić et al., 2017a;Bielić et al., 2020). According to Baker and McCafferty (2005) and Ugurlu et al. (2015), human error is accounted for about 80-90 % of marine accidents. It must also be noted that the organizational climate, which can be simply explained as the way things are done on board ship, is one of the leading causes of human error, besides the already-mentioned technology-related human error (Hasanspahić et al., 2021a). Reporting adverse events that did not cause injuries or material damage is crucial for safety improvements on board ships. Therefore, near-miss events could be a valuable tool for preventing accidents in various shipboard operations. For example, mooring and unmooring operations could be KEY WORDS ~Maritime safety ~Personal Protective Equipment ~Accident prevention ~Near-miss ~Seafarers considered among the highest-risk ship operations. According to the UK P&I club report, mooring incidents are among the top seven types of insurance claims (DNV). Consequently, there is a need to implement updated mooring standards and practices in the shipping industry. As ships are getting larger, new materials and mechanical systems are introduced, and the number of crewmembers is being reduced, so mooring safety is seriously impaired (DNV). To develop new standards and procedures, it is necessary to receive feedback from seafarers and harbor workers who deal with mooring operations. Another input is accident reports, where something went wrong and there were serious consequences. For instance, one of the recently published accident reports was the fatality of a Chief Officer (C/O) on one general cargo ship during a mooring operation. The ship was moored alongside another ship and had a shipto-ship (STS) cargo operation. The ship needed to move forward to continue with the cargo operation and during the warping operation, C/O got struck in the head by the mooring line that broke under tension and deceased from the consequences of the injury. Among other factors affecting the accident, the investigation revealed an insufficient number of crewmembers assigned to carry out the warping operation and insufficient planning for the mooring and warping operation. Furthermore, crewmembers were inexperienced in STS bulk cargo operations and lacked time available for planning and preparation (MAIB, 2022). The unavailability of adequate risk assessment seriously impaired the safety of shipboard operation and "routine" warping operation turned fatal. However, the risk of such accidents could be mitigated if mooring operation near-miss events were reported, investigated, analyzed, and disseminated to all stakeholders. Then, lessons learned could be implemented in existing procedures and risks could be mitigated. For example, International Marine Contractors Association (IMCA), through Safety Flashes, disseminate important data about incidents, potential hazards, and the lessons learnt from them that can help prevent similar events in shipping (IMCA). There were several reports of near misses during mooring operations on ships with a high potential for serious injury (IMCA, 2014;IMCA, 2018;IMCA, 2022). They disseminate information about the potentially hazardous event, together with analysis (causation finding), recommendations for corrective measures, and lessons learned. In that way, stakeholders can learn from near-miss examples that did not cause harm and shipboard safety can be improved (e.g., amending risk assessments for mooring operations for data disseminated in Safety Flashes). Serious accidents could be avoided in this way and collective learning could be achieved without severe consequences.
However, quite often, there is a tendency to take a near miss as a positive signal and ignore its importance for possible safety improvements. Seafarers and all maritime stakeholders should understand that a near-miss event is not a success since no harm was done; it is a warning signal calling for a quick reaction (Dillon et al., 2016). It reveals weak spots in safety systems and allows the patching of holes in safety barriers before harm is done. Therefore, events like this should be recorded and reported since they constitute a near miss and share the same root causes as accidents. According to the International Maritime Organization (IMO), a near miss is "a sequence of events and/or conditions that could have resulted in a loss. This loss was prevented only by a fortuitous break in the chain of events and/or conditions. The potential loss could be a human injury, environmental damage, or negative business impact" (IMO, 2008b).
Correspondingly, serious marine accidents could be prevented by reporting, investigating, and analyzing near-miss events and, equally important, disseminating conclusions to all interested maritime stakeholders. Near-miss investigation and analysis is an excellent chance to improve safety because there is no need to wait for an accident causing harm and economic loss to happen. It might be said that the near-miss analysis is a "cheap" way of improving safety and reducing the number of accidents at sea. To encourage and improve near-miss reporting, the IMO issued Guidance on near-miss reporting (IMO, 2008b). Moreover, near-miss reporting needs to be implemented on board ships under the International Safety Management (ISM) Code Section 9 (reporting of hazardous occurrences) (IMO, 2010). However, despite the ISM Code and mandatory reporting, several studies found that seafarers do not report all observed nearmisses (Jones et al., 1999;Vepsäläinen et al., 2010;Hasanspahić et al., 2020). As Hasanspahić et al. (2020) found in their study, 95.5 % of seafarers consider that near misses should be reported, but only 38.5 % report each observed near miss. There could be several reasons for that, but the most significant ones identified in the literature are: blame culture (Phimister et al., 2003;Cooke and Rohleder, 2006;Wang, 2006;Erdogan, 2011;Lappalainen et al., 2011;Bhattacharya, 2012;Adamson, 2015), being ashamed (Vepsäläinen et al., 2010, Storgård et al., 2012b, Lappalainen et al., 2011, knowledge on near misses (Hasanspahić et al., 2020), inadequate leadership (Oltedal and McArthur, 2011;Bielić et al., 2017b;Hasanspahić et al., 2021b), near-miss reporting form complexity (Cooke and Rohleder, 2006;Wang, 2006;Erdogan, 2011;Lappalainen et al., 2011;Williamsen, 2013;Adamson, 2015;Hasanspahić et al., 2020), commitment from top management (Sanne, 2008;Oltedal and McArthur, 2011), seafarers' cultural differences (Sanne, 2008;Erdogan, 2011), turnaround on a particular ship (Oltedal and McArthur, 2011;Kongsvik et al., 2012), and "Nothing is wrong in my ship" approach (Safety4Sea, 2022). These barriers need to be overcome to efficiently use a Near-Miss Management System, improve safety at sea and develop a safety culture on board ship.
A Company needs to implement a Near-Miss Management System to efficiently and adequately deal with the near misses on board ship (Hasanspahić et al., 2020). As stated, near miss must be reported, investigated, and analyzed to find immediate and root causes, which will help draw conclusions and make suggestions for preventing recurrence and improving safety. Furthermore, the system should deal with the reporting barriers and guide seafarers to overcome them.
In addition, in research by Hasanspahić et al. (2020), near-miss reporting inequality was found between shipboard departments (Deck and Engine). Out of the 467 collected nearmiss reports with reporter's rank, 72 % were reported by Chief Officers, 14 % by Deck Officers, 9 % by chief engineers, 2 % by Masters and 3 % by other crewmembers. Therefore, in this study, we intend to detect if there is a difference in opinions on Near-Miss Management Systems among ranks and ship types. Accordingly, this paper aims to: • Investigate satisfaction of shipboard Safety Officers and masters with the implemented Near-Miss Management Systems in shipping because they are the users on shipboard side. Also, an efficient system might improve safety on board, and it is in the best interest of the seafarers to use it adequately; • Investigate relations between the seafarers' ranks (Master and Safety Officers) and attitudes toward Near-Miss Management Systems; • Investigate relations between ship type and attitudes toward Near-Miss Management Systems.
NEAR-MISS MANAGEMENT SYSTEMS IN SHIPPING -LITERATURE REVIEW
The efficiency of Near-Miss Management Systems implemented on board depends on the seafarers using it and Company management commitment from the shoreside. Previous studies on Near-Miss Management Systems identified five to eight phases (or steps).
Most studies consider identification the first phase in a Near-Miss Management System (Phimister et al., 2000;Oktem, 2002;Cooke and Rohleder, 2006;Meel et al., 2007;Gnoni et al., 2013;WSH, 2016;Hasanspahić et al., 2020). However, Rasmussen et al. (2013) recognized observation as the first phase, although it might be argued that to observe a near-miss event, one needs to identify it. Therefore, in their study, Craig et al. (2014) named the first phase awareness since their opinion is that seafarers need to be trained to identify hazards and near misses. Consequently, complete Near-Miss Management System efficiency and safety improvement depend on the seafarers' knowledge of near misses and hazards. If seafarers do not know what constitutes a near miss, the whole system is deficient, and it cannot be expected that safety will improve.
The system's second phase is reporting or disclosure (Phimister et al., 2000;Oktem, 2002;Cooke and Rohleder, 2006;Meel et al., 2007;Gnoni et al., 2013;Rasmussen et al., 2013;Craig et al., 2014;WSH, 2016). In the study by Lindberg et al. (2010), reporting is the first phase of a system, which could be considered adequate if all seafarers are proficient in recognizing near-miss events. Unfortunately, it was found that especially low-ranking seafarers have difficulties identifying near misses (Hasanspahić et al., 2020), and it can be considered that near misses cannot be reported without proper identification or processed further. Near-miss reporting could be done in two ways: 1) the seafarer observing the near-miss event reports it verbally to the Safety Officer or 2) the seafarer observing the near-miss event fills out the near-miss report form. In the first case Safety Officer fills out the report, while in the second case, he receives a filled-in report. In this phase, the Safety Officer's attitude will affect the reporting crewmember, and if it is negative, it could act as a reporting barrier and prevent near misses from being reported, thus downgrading shipboard safety. Along with identification, reporting could be considered a pillar of the whole Near-Miss Management System and, therefore, Hasanspahić et al. (2020) considered together with identification as the first phase.
The third phase of the system is prioritization or selection (Oktem, 2002;Meel et al., 2007;Gnoni et al., 2013;Rasmussen et al., 2013;Hasanspahić et al., 2020). Although some studies did not include prioritization in a Near-Miss Management System (Phimister et al., 2000;Cooke and Rohleder, 2006;Craig et al., 2014;WSH, 2016), it is a vital link within. For instance, companies with numerous ships employing hundreds or even thousands of seafarers could receive large numbers of near-miss reports in their offices. It would be impossible to investigate all of them; therefore, "minor" near misses (low-risk ones) should be resolved without performing an investigation and "wasting" resources. However, high-risk near misses must be investigated and analyzed to find root causes and learn from them. Therefore, rating near-misses (hazards that could have been caused) is critical to assess whether reported near-miss events should be investigated and analyzed.
Distribution could be considered the fourth phase of an effective Near-Miss Management System (Oktem, 2002;Meel et al., 2007;Gnoni et al., 2013;Hasanspahić et al., 2020), while Phimister et al. (2000) considered it the third phase. Some studies do not consider distribution (Cooke and Rohleder, 2006;Lindberg et al., 2010;Rasmussen et al., 2013;Craig et al., 2014;WSH, 2016), but without it, there could be no external investigation and analysis of the causes. If a near-miss report is prioritized and not distributed to the person in charge of safety within the Company, valuable knowledge could be lost and corrective actions not disseminated to a broader audience (if applicable).
The fifth phase of the system could be cause analysis or investigation (Oktem, 2002;Meel et al., 2007;Gnoni et al., 2013;Rasmussen et al., 2013;Hasanspahić et al., 2020). However, in the studies conducted by Cooke and Rohleder (2006), Lindberg et al. (2010), Craig et al. (2014), and WSH (2016), it is the third phase, and in Phimister et al. (2000) and Rasmussen et al. (2013), it is the fourth phase. All studies reviewed include cause analysis or investigation in the Near-Miss Management System. Its purpose is to identify immediate and root causes of a near miss that could trigger an accident in the future if no action is taken.
Solution identification is the sixth phase of the system (Oktem, 2002;Meel et al., 2007;Gnoni et al., 2013;Craig et al., 2014;Hasanspahić et al., 2020), while in Phimister et al. (2000), it is the fifth phase, and Cooke and Rohleder (2006) consider it the fourth phase. Studies by Lindberg et al. (2010), Rasmussen et al. (2013), and WSH (2016) do not incorporate this phase in their systems. This phase includes identifying adequate corrective actions that could improve onboard safety and prevent the recurrence of adverse events. It is important to stress that the solutions found should be practical and possible to implement on board ship. As seafarers will be the ones implementing corrective actions, it is suggested to discuss the implementation with crewmembers during a regular monthly safety meeting on board and ensure that the solution is adequate and efficient.
Corrective actions identified in the previous phase, together with the near-miss, should be disseminated in the seventh phase to the broader audience to increase safety awareness and prevent the occurrence of possible adverse events (Phimister et al., 2000;Oktem, 2002;Cooke and Rohleder, 2006;Meel et al., 2007;Gnoni et al., 2013;Craig et al., 2014;Hasanspahić et al., 2020). Lindberg et al. (2010) consider dissemination as the fourth phase of their system, while Rasmussen et al. (2013) and WSH (2016) do not include it in their systems. The value of dissemination in maritime safety improvement is immense. Without dissemination, the value of reporting, identifying and implementing corrective actions would be significantly reduced. Therefore, disseminating investigation results and findings, together with near-miss reports, is the basis of learning from incidents.
The final phase is resolution, during which the identified and implemented corrective actions are followed up, reviewed and evaluated to ensure their efficiency and applicability to the specific ship (Phimister et al., 2000;Oktem, 2002;Meel et al., 2007;Lindberg et al., 2010;Gnoni et al., 2013;WSH, 2016). Resolution should also include feedback to the initial near-miss reporter.
A Near-Miss Management System is important because it enables learning from someone else's experiences and, if adequately used, prevents accidents (Erdoğan, 2011). Identification and reporting are initial and the most critical phases of the system and any shortcomings during these phases will result in a flawed and inoperative system. Moreover, it must be noted that the other phases are also important, but dissemination is particularly important, especially for small shipping companies. It enables learning from near-miss events that occurred elsewhere and gives a chance for safety improvements based on solutions made by someone else. Near-Miss Management System literature overview provided an insight into the phases of the system.
METHODOLOGY
The main goal of this paper is to gain insight into the seafarers' satisfaction with the systems implemented on board their ships, together with the opinion on onboard locations where the most near-misses occur. Since Safety Officers and Masters are in charge of safety matters on board ships, this research aimed to collect and analyze their perceptions and opinions. In addition, the paper aimed to recognize the most common nearmiss categories and suggest measures to improve the existing Near-Miss Management Systems in shipping. Another important goal of the study was to examine the relationship between the respondents' ranks, shipboard departments, and the type of ship with attitude toward Near-Miss Management Systems.
An online questionnaire was developed to collect data on Safety Officers' and Masters' attitudes toward the Near-Miss Management Systems implemented on their ships (Hasanspahić et al., 2021c). A web link to the questionnaire was sent to several crew recruitment agencies. They were asked to forward the link to the Masters and Safety Officers, who, in return, could agree to participate in the survey or not. However, confidentiality and anonymity were agreed upon if they chose to participate. Also, the seafarers could forward the link to their colleagues and expand the number of potential respondents (virtual snowball sampling). The questionnaire contained 20 questions, which were as neutral as possible to avoid biased responses, and it was available online during 2019 and 2020. Moreover, a pilot survey was conducted to test the questionnaire and ensure its completion would not be time-consuming and complicated. The study was ethically conducted, and the protocol was approved by the Ethics Committee of the University of Dubrovnik on 19 November 2020 (EA 1459/20).
After receiving positive feedback from the pilot survey respondents, it was decided to continue with the questionnaire, consisting of two parts. The first part was composed of questions dealing with the respondents' demographics including nationality, age, rank, type of education, type of ship, time served in the current rank and total sea service time. The second part was composed of 11 close-ended and two open-ended questions dealing with Near-Miss Management Systems.
Descriptive statistics were used to analyze the responses, and the Chi-Square Test of Independence was used to examine the relationship between the responses. The programming language Python was used to examine the relationships between the variables. The first two variables were obtained by dividing the sample according to the rank (Master or Safety Officer) -V1, and the type of ship (tanker, cruise, dry cargo or other) -V2. The responses to the questions "Do you think that near-miss followup measures received from the Company are substantial and applicable to your vessel?" (Q3), "Do you agree that near-misses should be rated (given low or high priority) before sending them to the office (to the designated person)?" (Q10) and "Please, rate satisfaction with Near-Miss Management System in your Company" (Q5), were chosen as the following three variables. The following null hypotheses were tested: H0,1) There is no statistically significant relationship between seafarers' ranks and opinion on near-miss follow-up measures received from the Company (V1 vs Q3).
H0,2) There is no statistically significant relationship between seafarers' ranks and opinion on rating near-misses before distributing them to the Company (V1 vs Q10).
H0,3) There is no statistically significant relationship between seafarers' ranks and perceived satisfaction with the ship's Near-Miss Management System (V1 vs Q5).
H0,4) There is no statistically significant relationship between the type of ship and the opinion on near-miss follow-up measures received from the Company (V2 vs Q3).
H0,5) There is no statistically significant relationship between the type of ship and the opinion on rating near-misses before distributing them to the Company (V2 vs Q10).
H0,6) There is no statistically significant relationship between the type of ship and the perceived satisfaction with the ship's Near-Miss Management System (V2 vs Q5).
A total of 112 seafarers participated in the survey. Eight nationalities were represented in the survey sample. The majority of the seafarers were from Croatia (90 %), followed by the Philippines (2 %), Finland (2 %), Montenegro (2 %), Bulgaria (1 %), Greece (1 %), Serbia (1 %) and Ukraine (1 %). The ranks of the seafarers are presented in Figure 1. Most respondents reported being Chief Officers and Masters (51 %). Moreover, 75 % of the respondents belonged to the Deck department, while 25 % belonged to the Engine department. Furthermore, in terms of age, 21 % of the respondents reported being younger than 33, 49 % reported being between 34 and 42, 19 % between 43 and 51, and 11 % were older than 52. The majority of respondents were maritime college graduates (69 %), and the remaining 31 % finished maritime high school. The participants reported sailing on different ship types, among which oil tankers (33 %), LNG tankers (23 %) and cruise ships (21 %) were the most highly represented ship types, followed by bulk carriers (7 %), chemical tankers (5 %), container ships (4 %) and other ship types (7 %). Other ship types included Ro-Ro passenger ships, tug boats, AHTS (Anchor Handling Tug Supply ship), PSV (Platform Supply Vessel), FLNG (Floating Liquified Natural Gas) and FSRU (Floating Storage Regasification Unit) ships. In terms of years of service in the current rank, 70 % of respondents reported being in rank 6 years or less, while 30 % were in the rank for 7 years or more. Regarding their total sea service time, 39 % of the respondents spent 11 years or less at sea, 32 % between 12 and 17 years, and the remaining 29 % served for 18 years or more. According to the analysis of the seafarers' responses to the demographic questions, it can be concluded that the sample was composed of experienced and educated high-ranking officers in charge of safety on their ships.
RESULTS AND DISCUSSION
The second part of the survey was composed of the questions regarding Near-Miss Management Systems on the respondents' ships. The questions were grouped to facilitate the presentation of the results. The first two questions (Q1 and Q2) dealt with the respondents' attitudes towards the Marine Accident and Incident Investigation training. The questions and the descriptive statistics for the responses are presented in Table 1.
Question
Yes ( The IMO developed Model Course 3.11, Safety investigation into marine casualties and marine incidents, which deals with investigating accidents and incidents (including near misses). The investigation includes collecting and analyzing the incident data to draw conclusions that will determine immediate and root causes and the contributing factors and provide safety recommendations. In addition, some maritime training centers and shipping companies developed tailor-made courses based on the IMO Model Course 3.11. The subjects of the tailor-made courses include reporting, investigation, and learning from accidents and incidents. However, since they are not compulsory, some shipping companies are reluctant to invest money and train their employees even though this kind of training aims to improve safety awareness onboard ships, facilitate accidents/ incidents investigations, and enable the development of the safety culture at sea.
The majority of the participating seafarers responded affirmatively to Q1 (Table 1), meaning that most of them are formally trained to investigate incidents, find their causes and suggest corrective actions. However, according to the responses, not all the participants are trained (31 %) and there are still Masters and ship Safety Officers who have not received formal training in incident investigation. In addition, according to the responses to Q2, most of the respondents (69 %) believe that incident investigation training should be mandatory for shipboard Safety Officers and Masters. There are several benefits of introducing incident investigation training as mandatory, such as more efficient unwanted events' (incidents, near-misses) investigations on board ships and improved safety culture in shipping.
The analysis of the responses to Q2 showed that more respondents belonging to the Deck department have a positive attitude towards introducing mandatory incident investigation training (75 % of positive responses to Q2) than the respondents belonging to the Engine department (25 % of positive responses to Q2). Also, most of the ship Masters and Chief Officers favored mandatory training (Figure 2).
Questions and descriptive statistics regarding respondents' attitudes toward Near-Miss Management Systems implemented on their ships are presented in Table 2 From the analysis of the responses, it can be concluded that most respondents have a positive attitude toward the implemented Near-Miss Management Systems (Figure 3 and Figure 5). Previous studies pointed out that shore-based management commitment plays a vital role in a Near-Miss Management System. If management is not involved in resolving the reported near-misses on board ship and there is no feedback from the shoreside, it might negatively affect near-miss reporting and act as a reporting barrier (Oltedal and McArthur, 2011;Kongsvik et al., 2012). Therefore, shore-based Management must be involved in the reporting process and actively support ship Masters, Safety Officers, and other crewmembers to improve safety and protect the environment. In addition, each reported near-miss should be commented on and responded to, which could raise safety awareness onboard a ship and create a safe working environment.
However, the opinions on reporting are different (Figure 4). Most respondents believe that their crewmembers do not report all the observed near-misses, which might be a serious defect within the Safety Management System. Reporting barriers are already mentioned in the Introduction to this paper. However, one interesting fact found in previous studies was that only a minor number of reported near-misses came from the ship rating side (Storgård et al., 2012a(Storgård et al., , 2012bHasanspahić et al. 2020). According to Hasanspahić et al. (2020), about 3 % of collected near-misses were reported by ship ratings (sample size: 580 nearmiss reports). Therefore, near-miss education and training of the seafarers (active and future ones) could help improve safety awareness on board ships and help involve ratings in near-miss reporting. Also, the active involvement of Safety Officers and Masters is necessary to implement reporting culture on board ships and support ratings in their reporting efforts. In that way, ratings' active participation could prevent losing valuable nearmiss data due to underreporting.
The following question (Q6) was: "Do you encourage your crewmembers to report all near-misses?" Analysis of the responses to Q6 has shown that most Masters and Safety Officers encourage reporting (92 %), while only 6 % do not encourage it. In addition, 2 % of the respondents have stated that they do not know if they encourage reporting. Encouraging crewmembers to report is an important step in creating reporting culture on board ship, and it should include promoting the 'no-blame' culture. The role of the Masters and Safety Officers in reporting is vital since, without their support, a near-miss will hardly be reported. Unfortunately, although few, a number of respondents stated they do not encourage reporting. It is an alarming fact and the reason behind this should be investigated as well as how it could be changed.
Question 7 (Q7) deals with near-miss types, rating according to perceived safety importance by the respondents: "Please, rate below types of near-misses according to safety importance as per your own opinion and experience. (Each type can be given a value from 1 to 10; the same value cannot be given to more than one near-miss type.)". Ten near-miss types were offered, and respondents could rate each type by assigning values from 1 to 10, where 1 is the highest importance and 10 is the lowest. Figure 6 presents the near-miss types by perceived importance.
According to the response analysis, "Not using/inadequate PPE" is perceived as having the highest safety importance, while "Emergency exit/passage blocked" is perceived as the least important near-miss type. The analysis showed that the "Near-collision/grounding/contact" type is perceived as having relatively low safety importance (rated as the ninth out of ten), which is an interesting fact. If the chain of events were broken, this near-miss type could develop into severe accidents resulting in property damage, environmental pollution, and injuries or fatalities. However, the respondents perceived it as relatively unimportant.
The following question (Q8) aimed to discover the most frequent near-miss types, as perceived by the respondents: "Please, write below the most frequent near miss that you have experienced." As this was an open-type question, the respondents were supposed to type down their answers; out of 112 respondents, 91 provided their answers. Again, near-misses involving PPE were perceived as the most frequent near-miss types by 68.1 % of the respondents, followed by slippery areas (wet/oily surface) by 3.3 %, and faulty life-saving and fire-fighting equipment by another 3.3 %. The study conducted by the American Bureau of Shipping (ABS) collected 45,298 near-miss reports. According to their analysis, the most common types of near misses found in the collected database were struck by/ against / cut / crushed / strain/sprain (12.2 % of all the collected reports), followed by PPE (11.7 %) and equipment (11.3 %) (ABS). Therefore, perceived near-miss frequencies reported in this study could be considered accurate. Moreover, it could be concluded that adequate PPE usage, condition, or not using PPE at all present the major issue in shipping even though Safety Management Systems developed by shipping companies deal with PPE and identify each piece of PPE that is to be used during specific shipboard operation.
Question 9 (Q9) aimed to reveal the respondents' opinions on shipboard areas where most of the near-miss events happen: "Which onboard location, as per your own experience, is the location where most near-misses occur?" According to the analysis of the responses, the deck area is considered the area where most shipboard near misses occur (43 % of responses), closely followed by the engine room (35 %) (Figure 7).
The ABS study found that 37 % of all the near misses collected during their research in their database occurred in the deck area, followed by the engine room (20 %) and cargo area (14 %) [43] (ABS). However, this paper includes the cargo area in the deck area. Another study collected 309 near-miss reports from an Aframax oil tanker and analyzed them by onboard locations where a near miss occurred. It was found that 64.1 % of the reported near misses occurred in the deck area, followed by accommodation (16.8 %), engine room (13.9 %), navigating bridge (3.2 %), and ballast tanks (2.0 %) (Hasanspahić et al., 2022). The conclusion is that the deck area and engine room are areas where most of the near-miss events occur on board ships, and during work, seafarers should always apply extra care and follow the procedures. In addition, it is important to stress that any unsafe condition should be immediately reported to the supervising Officer, and work in that specific area should be suspended until safe conditions are restored.
As mentioned in Section 2, the third phase of the Near-Miss Management System is prioritization or selection. In that phase, near-miss reports are rated, giving them a high or low priority, which would decide whether they need to be investigated or not. However, from the authors' experience and informal communication with several ship Captains, near-miss events are not rated on board ships. Usually, they are prioritized in the Table 3.
Question and descriptive statistics on near-miss rating onboard ship.
office, which makes the system less effective since the near-miss rating is not an easy task, especially for persons not on board the ship where an event occurred. However, near-miss rating on board ships before distribution to the office makes near-miss management more difficult for shipboard personnel. Therefore, the authors asked the Masters and Safety Officers about their opinion on rating a reported near miss before sending it to the Company (Q10). According to the analysis of the responses, most respondents have a positive attitude toward near-miss rating (Table 3). However, shipboard Masters and Safety Officers should have adequate knowledge of risk assessment and incident investigation for an efficient and adequate rating. Therefore, prioritizing on board ship is dependent on Safety Officers' training.
Although a number of near-miss events in a specific period are hard to predict, some shipping companies require a certain number of near-misses to be reported annually. Therefore, Masters and Safety Officers were asked: "Do you think there should be a fixed number of reported near misses per vessel?" (Q11). Although only 72 respondents answered this question, it can be concluded that the experienced seafarers negatively perceived a fixed number of reported near misses (88.9 % of respondents answered "No" while only 6.9 % of answers were "Yes" and 4.2 % were "I don't know"). Question 12 (Q12) also deals with a fixed number of reported near misses, and the respondents were asked: "As per your own opinion, how many near misses should be reported annually per one vessel?" where four answers were offered. Again, there were 72 responses recorded, and most respondents (87.5 %) chose the answer "There should not be a fixed number of reported near misses per vessel per year." The answer "One reported near miss per week" was chosen by 8.3 % of respondents, while "One reported near miss per crewmember per year" gained 2.8 %, and "Ten reported near misses per month" gained 1.4 % of answers.
The respondent could comment on the questionnaire and the subject in Question 13 (Q13). Most of the comments referred to the fixed number of near-miss reports expected by the Company. For example, one LNG tanker Master stated: "If there is a fixed number of near misses to be reported, the whole point of promoting safety culture has failed." Another comment was: "Near miss is a random event. You cannot correctly predict a number of random events. There can be 100 near misses in one month, and none in the next month" (oil-tanker Third Officer). A cruise ship Staff Captain stated: "In case there is a dedicated annual number of reported near misses by ships, the quality of near-miss reports will be downgraded, and near-miss reports will be just another checklist to be filled. Quality should not be supplemented by quantity". From the responses and comments, it can be concluded that the fixed number of near-miss reports in a specific period is not well accepted in shipping, and experienced seafarers believe that it could downgrade safety at sea. However, different ship inspections might insist on checking near-miss reports, and if nothing is reported, it could be considered a non-conformity. Therefore, some shipping companies insist on a fixed number of near-miss reports (minimum number) to stay on the 'safe side' , but at the same time, this requirement might induce the fabrication of false reports (imagined ones). That could be dangerous, and analysis of such reports might be counterproductive for safety efforts.
One tugboat Master stated: "If this topic is overly regulated, it will just create paperwork and do nothing to help increase safety on board." As Bhattacharya (2020) found, a mandatory increase in reporting on one shipping Company did not decrease near misses and incidents, nor did it improve safety. He concluded that mandatory reporting might lead to reporting just to satisfy the requirements, not to improve safety and prevent future adverse events. Therefore, it could be concluded that the quantity might seriously endanger the quality of reports.
Furthermore, it needs to be emphasized that the number of reported near misses is usually relatively high while the system is being implemented, but if the suggested corrective actions are applied and if they are efficient, the number of reports should gradually decrease. That should be considered a sign of safety increment, not safety deficiency. Moreover, since safety is measured by its absence and not its presence, a reduction in the number of reports should be a sign of safety culture maturity and safety improvement. However, this should be corroborated by the absence of any recorded injuries and damages. Near-Miss Management Systems should be continuously monitored and evaluated to ensure purposefulness and efficiency.
number of reports should be a sign of safety culture maturity and safety improvement. However, this should be corroborated by the absence of any recorded injuries and damages. Near-Miss Management Systems should be continuously monitored and evaluated to ensure purposefulness and efficiency.
Some respondents suggested continuous use of nearmiss reports as a tool for safety improvement. For example, one respondent (container-vessel Chief Engineer) commented: "Near miss should be reported as per appearance and should be discussed on board to familiarize all the crewmembers as well as explain how to avoid (prevent) it".
However, attitudes are different. Filling in report forms makes near-miss reporting a 'bureaucratic' burden for some seafarers, and in their opinion, safety is not improved, and nearmiss management is inadequate to deal with safety improvement.
One LNG tanker Chief Engineer stated: "Near misses and most safety mechanisms have proved inadequate when considering that injuries and accidents are on the rise. Behavior-based safety, if implemented correctly (without additional paperwork or reducing current load) might be a step in the right direction".
The attitudes are different. While some Masters and Safety Officers consider near-miss reporting a step in the right direction, some consider it an unnecessary burden. However, most survey respondents believe that adequate and efficient near-miss management could improve onboard safety, and therefore systems should be continuously developed and upgraded with new knowledge. For example, one cruise ship Electrical Officer wrote: "I work for a Company where the nearmiss system has been implemented for some time now, and it is working well although there is room for improvement, e.g., disseminating high-risk near-miss reports and corrective actions taken to prevent future similar situations throughout a fleet. False accident reports are topics to discuss since I have experience with false reports whose intentions were to cover wrong procedures and people's reactions. Those things can lead new and lessexperienced people to wrong conclusions. I see why reports can be done that way, but accidents will continue to happen without honesty and open conversation. No matter what position on board we are, we all have a common goal: complete our contract fair and return safely. In my opinion, whatever is out there to help us be safe on the ship should be welcome and well-accepted among the crew. I believe this is one of the things (near-miss reports) that is good to help us at sea." The seafarers understand that data dissemination could solve numerous safety issues and therefore urge for communication and data sharing. Each shipping Company Management should act on this, improve existing incident/accident databases, disseminate data, and facilitate learning from incidents.
Finally, the Chi-square Test of Independence was done and the results are presented in Table 4.
CONCLUSIONS
From the analysis of the survey answers, it can be concluded that most of the surveyed seafarers consider Near-Miss Management Systems well-implemented and adequate. However, most of them believe that near-misses are still underreported although they encourage reporting and promote safety culture. Underreporting near-misses in shipping is a serious barrier to improving safety culture, and valuable knowledge is irretrievably lost. Therefore, each Company should promote reporting and thus improve safety on their ships. However, as seen from the survey responses, mandatory reporting might be considered another obstacle and negatively affect learning from incidents. Therefore, most of the surveyed seafarers believe that reporting should not be mandated, but all observed near-misses should be reported to improve safety.
Respondents consider near-misses dealing with Personal Protective Equipment as the most frequent ones and at the same time the most important. However, one might wonder how come PPE near-miss events are recurrent. Does it mean that the corrective measures implemented are not adequate or do seafarers report PPE near-misses just to comply with the mandated number of reports and satisfy the requirements? However, this type of near-miss event should be acted upon, and safety measures regarding adequate usage of PPE should be re-evaluated. Shipboard leadership, especially Masters and Safety Officers, should play an important role and encourage and promote PPE usage. In addition, shipboard training for ratings regarding near-miss identification and reporting could benefit onboard safety since previous studies found that ratings seldom participate in reporting due to various barriers.
Furthermore, the respondents believe that most of the near-miss events occur on deck and in the engine room, which seems quite reasonable. Therefore, lessons learned from previous near misses should be incorporated in permits to work and risk assessments. In addition, monthly safety meetings should be used as an opportunity to discuss lessons learned and promote safety awareness.
A statistically significant relationship between the respondents' ranks and types of ships with attitudes on nearmiss management was not found, suggesting the nonexistence of significant differences between seafarers. It is an important finding because it can be concluded that Masters and Safety Officers serving on different types of ships do not have significantly different attitudes toward Near-Miss Management Systems. Most of them consider attending training on incident investigation a positive detail that could improve their knowledge regarding near misses and help identify immediate and root causes.
The findings of this study could be used to help improve Near-Miss Management Systems in shipping and thus help improve maritime safety. Onboard safety is every seafarer's business, and all the crew should be actively involved in promoting safety. | 9,581.6 | 2023-04-20T00:00:00.000 | [
"Engineering",
"Environmental Science"
] |
The Impact of Ti/Al Contacts on AlGaN/GaN HEMT Vertical Leakage and Breakdown
Enhanced leakage paths below Ti-/Al-based contacts to GaN-on-Si HEMTs have been identified and studied. Through a novel use of the quasi-static capacitance–voltage technique, the depth of these preferential leakage paths was determined to be <inline-formula> <tex-math notation="LaTeX">$\sim 1.6~\mu \text{m}$ </tex-math></inline-formula>, extending down into the superlattice strain relief layer. Along these paths, the material resistivity was reduced by more than a factor of 100 compared to the uncontacted epitaxy. It is suggested that the cause of the additional leakage is decoration of dislocations. This result is important for understanding buffer transport, a critical parameter for breakdown, and buffer charge storage.
I. INTRODUCTION
A lGaN/GaN-on-Si high electron mobility transistors (HEMTs) are especially suited to power electronics due to their high breakdown field, high electron mobility, high carrier density, and compatibility with standard ≥150 mm Si process lines [1].These properties allow for high voltage blocking in the off-state and low on-resistance in the onstate, resulting in highly efficient operation, all at low cost.However, there are still some challenging issues which can impact operation such as current collapse (the accumulation of negative charge in the buffer in the off-state which subsequently increases the on-resistance [2]), as well as vertical leakage in the off-state which limits the maximum operating voltage, and hence device efficiency.
Reducing the resistivity in the top layers of the epitaxy has been linked to the suppression of current collapse due to the ease with which the trapped charge can be neutralized [3]- [6].This can be achieved, at least partially, by preferential leakage under the contacts which has been observed experimentally [7] and is routinely included in simulation [8], [9].Although these preferential leakage paths are required for low current collapse buffers, their origin and the depths to which they extend are not fully understood.Previously suggested origins include contact spiking, metal in-diffusion and dislocation decoration [3], [7], all of which are associated with increased vertical leakage.
Increased breakdown voltage has been achieved laterally, with field plates increasing the size of the gate-drain depletion region [10], [11], and vertically by increasing the thickness of the (Al)GaN layers [12].However, this thickness is limited by difficulties in managing the stress in GaN-on-Si growth due to the large lattice mismatch [13].Instead, further increases in breakdown voltage can be achieved by optimizing the breakdown field of the epitaxy [14], [15].It has been previously indicated that the choice of contact metallurgy impacts the vertical breakdown [16], with the implication that the Ti/Al based contacts (compatible with Si foundries) increased the vertical leakage.In this letter, we study the impact Ti/Al contacts have on the resistivity of the epitaxy and subsequent vertical breakdown.The quasi-static capacitance-voltage (QSCV) technique, applied to purpose designed structures, was used to measure the resistivity and depth of additional leakage paths introduced by the contacts.This is a result which is essential for full understanding of the current-transport in the semi-insulating buffer structure, responsible for breakdown and charge storage phenomena such as current collapse.
II. EXPERIMENTAL DETAILS
The vertical leakage structures used in this study were fabricated using a typical AlGaN/GaN-on-Si epitaxy architecture as shown in Fig. 1a.This consists of a >1 cm silicon substrate, AlGaN/GaN superlattice strain relief layer, carbon doped GaN buffer layer, unintentionally doped GaN channel and AlGaN barrier.The wafer was passivated with Al 2 O 3 and SiO 2 .The total thickness of the channel and GaN:C layer in this epitaxy was 1.3 μm with 1.9 μm of strain relief.The sheet resistance of the 2D electron gas (2DEG) was measured as 550 /sq.Full details of the Ohmic contacting process used on this wafer have been published previously [17].The AlGaN barrier was fully recessed followed by the deposition of Ti/Al/TiN contacts with a Ti/Al ratio of 0.05 which was annealed at only 550°C, delivering a contact resistance of ∼0.6 mm.All vertical leakage structures consisted of a fixed active area of 110×110 μm 2 isolated by a nitrogen implant with an energy of up to 375 keV to achieve an isolation depth of 550 nm.Inside this active area, Ohmic contacts of varying sizes were made This work is licensed under a Creative Commons Attribution 3.0 License.For more information, see http://creativecommons.org/licenses/by/3.0/through windows in the passivation layer.A breakdown field of 2.7 MV/cm was measured on a structure whose window area filled the active area (Fig. 1b) demonstrating this is already an excellent buffer.Lateral conduction paths have been suppressed in this optimized buffer.This has been confirmed by the invariance of back-gate 2DEG pinch-off voltage across contacted structures with various active areas when using the Si substrate as the back-gate [3] (not shown here).The presence of lateral conduction paths (such as a 2D hole gas which can extend up to 100 μm outside the active area [18]) would have the effect of reducing the pinch off voltage in smaller active areas, but that was not seen here.
For all measurements in this letter, a negative bias was applied to the substrate and the current was measured at the surface contact which was held at 0 V.This resulted in a field over the epitaxy of the same polarity as experienced in normal transistor operation.Two measurements were performed; current transient measurements with a substrate bias of −200 V were applied to >35 of each structure geometry.The current level at 30 s was used to assess the mean vertical leakage in each structure.Sampling the current 30 s into the transient ensured that the decaying displacement current spike from the step bias did not influence the leakage current measurement.The effect of surface charging on the transient was assessed using a guard ring structure [19].After a decay of ∼5 s, no surface effects were measurable and so this did not affect the results at 30 s.In addition, quasi-static capacitance-voltage (QSCV) measurements were performed.This technique permits the observation of dynamics which are too slow for a conventional capacitance voltage bridge.A continuous voltage ramp of −1 V/s was applied to the substrate down to −40 V and back.The current measured at the surface contact during this ramp was a sum of the leakage current and a displacement current generated by the ramp of I disp = C • dV /dt.As the sign of the displacement current depends on the direction of the ramp, by ramping in both directions, the displacement current component will change sign with the ramp direction.The measured current is always I meas = I leak ± I disp thus these two components can be distinguished and the displacement current then be used to evaluate the quasi-static capacitance of the structure.
III. RESULTS
The current transients from the structures with the smallest contact are shown in Fig. 2a along with the resulting distribution.A normal distribution was fitted to the data to extract the mean and standard deviation.This process was repeated for each geometry (shown in Fig. 2b), where the mean leakage current is seen to increase linearly with contact area.The ratio of standard deviation to mean was found to be largely independent of contact size, best seen on the log-log axes inset to Fig. 2b.
Following this, three structures of each geometry exhibiting mean vertical leakage were the subject of QSCV measurements.The raw measured current from the bidirectional continuous voltage ramp is shown in Fig. 3a along with the decomposition of the two constituent contributions in Fig. 3b, identified by the change in polarity with ramp direction.The mean displacement current was converted to capacitance by dividing by the ramp rate and, as shown in Fig. 3c, also monotonically increased with contact area.
IV. MODEL
Considering every structure has the same active area, and the contacts contact a 2DEG which extends to fill the entire active area regardless of contact area, the results of these measurements would at first sight be expected to be uniform across all geometries.However, as seen in Fig. 2b, the area of the contact has the effect of increasing the vertical leakage.This result is consistent with previous work [16] and with the ∼0.45 MV/cm shift in the vertical leakage characteristics seen at lower fields in Fig. 1b.This shift shows the presence of the contact increases the vertical leakage current, although the hard breakdown field is only decreased by ∼0.13 MV/cm, indicating the final failure mechanisms are similar.The lack of a plateau in these characteristics indicates that the leakage did not induce deep depletion in the Si substrate, and the Si can be treated as a ground plane [20].
Although the leakage currents fit a normal distribution, since variations in the dislocation density and in the leakage path conductances would both result in a normal distribution, it is not possible to tell from the shape which one, or a combination of both are the cause.
The capacitance between 2DEG and Si substrate would equally be expected to be invariant across the structures in the absence of extended leakage paths.We propose the model shown in Fig. 1c, where a region of reduced resistivity exists below the contact due to the metallization.The equivalent circuit includes an additional leakage path (ρ 2 ) in parallel with the existing resistors which results in the observed increase in vertical leakage with contact area.
This model results in a quasi-static capacitance increase.The structure is represented as two parallel equivalent circuit diagrams; around the contact, the epitaxy is simplified into a leaky dielectric layer, represented by a capacitor and a resistor.Underneath the contact an additional leakage path is introduced lowering the resistivity in the upper part of the epitaxy over a certain depth.Considering a potential divider, this lower resistivity means more of the voltage is dropped over the lower part of the stack giving rise to a higher displacement current which is dependent on C 3 (since now As the total capacitance of the structure is the sum of the capacitance under the contacts and the capacitance in the remaining active area, the capacitance increases with contact area.The exact capacitance change resulting from this model depends on the depth that the region of lowered resistivity (represented by ρ 2 ) extends.The equivalent circuit was used to model the expected capacitance at various depths with the results shown in Fig. 3c.The inference is that the depth of these leakage paths extends ∼1.6 μm down from the contact metal, stopping in the superlattice.Transmission electron microscopy (TEM) studies of other Ti/Al based contacts indicate only 5-30 nm of metal diffusion [21], [22] and < 100 nm of contact spiking [23], much less than 1.6 μm.TEM images of this particular contacting process showed no diffusion or spikes [17], which indicates a different cause.Considering the depth of these leakage paths, a more plausible possibility is the decoration of dislocations perhaps with the contact metal during annealing.These threading defects penetrate through the buffer and have been correlated with increased Ti/Al contact leakage in the past [24].
In order for this model to increase the capacitance under the contacts, the combined resistivity of ρ 1,upper and ρ 2 must be less than the effective resistivity of C 3 during the ramp.The effective resistivity of a capacitor, ρ C , in a voltage ramp is given by where is the permittivity, here about 10 0 .This places an upper bound of ∼10 11 cm on the resistivity of the upper part of the stack below the contact.The resistivity expected under the rest of the active area around the contact, ρ 1 , can be calculated from the x-intercept of Fig. 2b as 5×10 13 cm, which is consistent with previous estimations for the resistivity of C doped GaN [7], [25].This indicates more than a hundred times reduction in resistivity.Conventional, high frequency capacitance-voltage (CV) yields the same capacitance for all structures, equal to the expected capacitance of the entire stack.This occurs since the leakage time constants are too slow to respond to the lowest available measurement frequency of 100 kHz.Therefore, 1 RC = 1 ρ 2 < 100 kHz which imposes a lower bound on the resistivity and results in the limits 10 7 < ρ 2 < 10 11 cm for ∼1.6 μm under the contacts.
V. CONCLUSIONS
Specifically designed structures have been used to study the effect of Ti/Al based Ohmic contacts on vertical leakage.The result of vertical leakage increasing with contact area is consistent with previous work.The key result of this study indicates that these preferential leakage paths extend ∼1.6 μm into the epitaxy, ending somewhere in the superlattice.These paths lower the field required for substrate leakage and appear to define the hard breakdown field.Engineering the leakage due to contacts is critical for successful high voltage device operation and managing current collapse, buffer leakage and breakdown.This result indicates that these leakage paths are much deeper than may have been expected, and this will aid accurate simulation as well as efforts to improve device performance and breakdown.
Fig. 1 .
Fig. 1. a) The layout of the vertical leakage structure on a typical AlGaN/GaN epitaxy.Vertical breakdown measurements are shown in b) on structures where the contact filled the entire active area (Large contact) and also 5x5 μm 2 (Small contact).Measurement of the current density below 10 −5 A/cm 2 was limited by the instrument.c) A lumped equivalent circuit diagram of the epitaxy treating the stack as a leaky dielectric.Under the contact, shaded in yellow, an additional leakage path with resistivity ρ 2 is included.
Fig. 2 .
Fig. 2. a) Vertical current transients during a 200 V stress on structures with a 5×5 μm ¾ contact.The distribution at 30 s is inset.b) The mean vertical leakage current for each vertical leakage structure scales linearly with contact area.The same data is inset on log-log axes of the same scale and show the standard deviation divided by the mean is approximately constant.
Fig. 3 .
Fig. 3. a) The bidirectional continuous IV for the QSCV analysis on two different structures.This data is decomposed in b) into the displacement and leakage components based on the sign change with ramp direction.c) The capacitance of each vertical leakage structure geometry.A parallel system capacitance of 0.93 pF from an open calibration was subtracted from the results.The solid and dashed lines show the measurement data and model respectively.
Manuscript received August 15, 2018; accepted August 17, 2018.Date of publication August 29, 2018; date of current version September 25, 2018.This work was supported by the ECSEL-JU Powerbase Project under Grant 662133.The work of B. Rackauskas was supported by the U.K. EPSRC.The review of this letter was arranged by Editor T. Egawa.Belgium, and also with the Department of Electronics and Information Systems, Ghent University, 9000 Ghent, Belgium.
Color versions of one or more of the figures in this letter are available online at http://ieeexplore.ieee.org.Supporting data available at doi.org/10.5523/bris.3vjckq2z2mh5l2l0jk8h6ldjn3.Digital Object Identifier 10.1109/LED.2018.2866613 | 3,410.4 | 2018-08-29T00:00:00.000 | [
"Engineering",
"Physics"
] |
Designing a NAABSA Class Tanker Ship with Bottom Protection from Ground
Indonesia is the world's largest archipelago country with a high potential for economic development and top producer and exporter of palm oil. As an archipelago country, the most efficient cargo transportation routes are through rivers and seas. Designing and building tankers, taking into account the specifics of the operation, are relevant. The paper considers the issues of designing a tanker for the transportation of crude palm oil with a defined operation area and route. The general concept of the vessel is proposed, taking into account the restrictions on the navigation area and draft for operation in the river. Particular attention is given to the issues of strengthening the hull in terms of overall longitudinal strength, as well as the bottom and the propeller-steering complex in terms of interaction with the ground. An external structural protection (ESP) from the ground was developed, and comparative calculations of the stress-strain state of the compartment and ESP structures were performed. The effectiveness of the solution for protecting the hull from direct contact with the ground is shown, and outlined ways for the possible development of tanker projects for the conditions.
I. Introduction
The Republic of Indonesia is the largest archipelago country in the world. It is located between The Indian and Pacific oceans. Indonesia is ranked 4th in the world by population (after China, India, and the USA), with a population of 280 million people. Indonesia has a rich shelf zone with biological and other resources (Figure 1a), with a total area of 1.92 million km 2 and 13.6 thousand islands.
The share of the oil industry in Indonesia is $292 billion, and it is an important part of the country's economy (29%) -16th in the world and 1st in Southeast Asia. Oil and gas exports bring 20% of the country's income, ranking 3rd in the list of major exported goods after coal and palm oil [1]. Indonesia is the world leader in palm oil production, with 46 million tons worth $18 billion annually. Other major export commodities are rubber ($4 billion), coffee ($1.4 billion), seafood, etc. [1].
The planned route of the tanker is the port of Palembang (departure) and the port of Priok (Jakarta) (Figure 1b). The navigation range is 500 nautical miles, including the river section -54 miles, then narrow Bangka Strait -100 miles, and then in the open Java Sea -346 miles. The minimum depth of the Musi River is 4.5 meters [2], with tide up and downs that can reach up to 2 m. This determines the probability of contact with the ground and partial drainage. The draft of the tanker was adopted 4 m, taking into account the margin for structural protection from the ground. There are no bridges on the considered section of the river, and no height restrictions are required.
The narrow Bangka Strait separates the islands of Sumatra, and Bangka has the lowest depth of 7.5 m. The Java Sea is relatively shallow, and its average depth is 40 m [3]. The Java Sea is characterized by great thermal stability with an annual average temperature of 28 °C and deviations up to 3 °C. [3].
General Conception of the Tanker
Taking into account the conditions of the operating area, the dimensions of the tanker are length 90 m, width 15 m, side height 5 m, and draft 4 m, displacement 5037 t. The main cargo is palm oil. The possibility of modifying the tanker for crude oil transportation is also taken into account. Additionally, the issues of restrictions on the navigation area and the class of the vessel NAABSA1 were considered according to the Rules of the Russian Maritime Register of Shipping (RS) [4,5].
The vessel is a 2-screw, with a tank and a 3-tier superstructure in the stern. In the area of the cargo area, there is an elevated main deck (height of 1 m), double bottom (height of 1 m), and double sides (width of 0.9 m) equipped. The transverse spacing is 700 mm in the cargo area. The longitudinal spacing is 600 mm. Floors and frames through 3-4 spacings. The hull has 9 transverse bulkheads and 1 longitudinal one. The side view of the vessel and the cross-section diagram are shown in Figure 2.
In the first stage, loads and general requirements for the size of connections are determined according to the Rules [5]. The lining of the bottom and cheekbones is 10 mm thick. The flooring of the second bottom is 9 mm thick. Floors and zygomatic brackets with a thickness of 10 mm. Checking the overall strength of the hull showed that there is a shortage for the I limited navigation area (200 miles away), and it was decided to use steel with 235 yield strength for bottom and sideboard construction and high-strength steel with 315 MPa for the elevated main deck and its longitudinal beams [15]. Then the total longitudinal strength is provided with a margin of 12.1%.
Hull Reinforcements for NAABSA1 Class and Ground Landing
In the second stage, the task of strengthening the hull for safe contact with the ground and compliance with the NAABSA1 class was solved [5]. Additionally, the issues of the influence of the external structural protection of the hull were considered. Based on the results of calculations and in the absence of an ESP, it was obtained: -increase the thickness of the lining of the bottom and cheekbones to 11 mm, the flooring of 2 bottoms -up to 10 mm; -increase the longitudinal bottom beams No 2-8 from Diametrical Plane (DP) from the half-column 18a to 20b [9].
External Structural Protection Design
The operation of vessels in rivers increases the possibility of interaction with the ground. To avoid accidents and damage to the hulls, the bottom is most often damaged. Therefore, protecting the bottom from direct contact with the ground is important [6]. The works [11,12] show the advantages of a trapezoidal ESP with overhangs. The parameter of the ESP was achieved by following methods by Kulesh V.A. and Pham Trung Hiep [8][9][10], and for such a form, the decision was made in Figure 3a
FEM Analysis
The ship's hull compartment was modeled using the SolidWorks program [13]. The model was created as a fragment of ¼ compartment. The design of the compartment and its loading conditions were assumed to be symmetrical with respect to the diametral plane. The set of the compartment included different types of bracing sections: tees (keel, bottom stringers, carlings, beams, and bulkhead frame racks), bulb strips (frames, bottom longitudinal beams, above-deck longitudinal beams, idle racks, and horizontal bulkhead shelf).
The compartment boundary conditions shown in Figure 4a are fixed along the section in the diametral plane (no displacements and no rotations), fixed on the transverse bulkhead (no shifts or turns), reference geometry along the cross-section of the middle of the compartment (no rotations around the y axis). The mesh used is standard and solid, with a global size of 200 mm and a tolerance of 3.8 mm [14], shown in Figure 4b The analysis of the stress-strain state of the compartment under the action of soil was performed, and 2 cases are considered: 1. Tank without cargo (Figure 5a.); 2. Tank with cargo (Figure 5b.). The soil is different and has many properties, but for this study, we take the calculated nominal resistance of the soil; 300 kPa is assumed (we take an average of 100-600 kPa). The load on the compartment from the ground side was 8357 kN.
III. Results and Discussions
The calculation is presented in Table 1. The highest stresses in the floor are 192 MPa when the tank is filled. Stresses in double bottom longitudinal, stringer and stiffner connection also increased significantly caused by the pressure of the filled cargo tank. The maximum deflections along the horizontal keel are 0.19 mm (Figure 6a) and along the decking of the second bottom 0.23 mm (Figure 6b) for case 1 and case 2, respectively. Analysis of the results of FEM calculations showed: 1. Tanker with ESP under the most severe conditions (full draining at the highest weight) receives a relatively high level of stress in the bottom floors -up to 82% of the yield stress in the filled tank scenario; 2. The stresses in the ESP elements are less significant -up to 6% of the yield stresses in the vertical plate and in the internal bracket-3%; 3. For relatively large vessels, the conditions with a partial transfer of the reaction of the soil simultaneously to the hull and to the ESP are especially important and provide opportunities for obtaining acceptable design solutions in terms of dimensions, weight, and stresses. This design solution was also developed from previously researched by Prasetio et al. [9].
IV. Conclusion
As a result of the work, a tanker concept is proposed for the considered operating conditions. The necessity of increasing the overall longitudinal strength of the hull and the local strength of the bottom is shown. The directions and objectives of the development of this project are related to the technical and economic performance indicators, taking into account loading options and clarifying design concepts for protection from the soil. | 2,140.6 | 2023-06-09T00:00:00.000 | [
"Engineering"
] |
Expression of angiogenic factors predicts response to chemoradiotherapy and prognosis of oesophageal squamous cell carcinoma
The ability to predict patients' responses to chemoradiotherapy by analyzing pre-treatment biopsy specimens would be valuable for managing oesophageal squamous-cell cancer. To this end, the expression of p53, thymidine phosphorylase and vascular endothelial cell growth factor was analyzed by immunohistochemistry in 52 patients with oesophageal squamous-cell cancer prior to chemoradiotherapy. Treatment consisted of radiotherapy (40 Gy) and 5 day-infusion of 5-Fluorouracil (500 mg m−2 per day) combined with cisplatin (10 mg m−2 per day). Following treatment, imaging and endoscopic reassessment was performed to establish treatment response. Thirty-one patients underwent radical surgery and 21 patients were treated with an additional 20 Gy of radiotherapy. Of the tumours studied, 58% were p53-positive, 40% thymidine phosphorylase-positive and 44% vascular endothelial cell growth factor-positive. A clinical response was observed in 36 patients (69%) and was negatively associated with thymidine phosphorylase expression (P=0.02) and vascular endothelial cell growth factor expression (P<0.001). However, the 5-year survival rate was significantly lower only in patients with vascular endothelial cell growth factor-positive tumours (P=0.037). Multivariate analysis identified vascular endothelial cell growth factor as a significant independent prognostic factor (P=0.0147). These results suggest that expression of angiogenic factors has predictive value for the treatment response and outcome of patients with oesophageal cancer. British Journal of Cancer (2002) 86, 552–557. DOI: 10.1038/sj/bjc/6600129 www.bjcancer.com © 2002 Cancer Research UK
Despite improvements in surgical techniques, rapid fatal recurrence is common in patients with advanced oesophageal cancer (Isono et al, 1982(Isono et al, , 1990. Because surgical resection alone rarely results in long-term survival, efforts are now focused on combined multimodality treatments in an attempt to improve local control and eliminate micro-metastasis present at the time of surgery. Recently, although neo-adjuvant chemoradiotherapy (CRT) followed by oesophagectomy has become widespread after several favourable pilot studies were reported (Poplin et al, 1996;Stahl et al, 1996;Ancona et al, 1997;Ide, 1997), contradictory data have also been published (Bosset et al, 1997;Tamin et al, 1998). A major problem in this context seems to be that a lower rate of cancerrelated deaths after combined treatment is counterbalanced by a higher rate of treatment-associated mortality. Because only patients with potentially responsive tumours would benefit from such aggressive treatment, prediction of treatment response by means of tissue analysis is invaluable in the management of these patients with advanced oesophageal cancer. If non-responsive tumours could be identified, these patients could be spared the significant toxicity, time, and financial expense associated with intensive ther-apeutic regimens. Determination of appropriate pre-treatment factors necessary for prediction of patients' responses to CRT is vital and can be achieved by 'biological staging' using predictive biological factors. Therefore, a primary consideration in setting up CRT for patients with advanced oesophageal cancer is to identify markers serving as good predictors for treatment response.
Given the importance of alterations in the p53 gene and expression of angiogenic factors for progression of oesophageal cancer, it is reasonable to explore whether such markers may have predictive value for the patients' response to therapy. Indeed, genetic alteration of p53 or p53 protein over-expression has already been reported to be a good predictor for treatment response and survival in oesophageal cancer (Sarbia et al, 1994;Nabeya et al, 1995;Casson et al, 1998;Ribeiro et al, 1998;Yang et al, 1999;Kobayashi et al, 1999;Shimada et al, 2000).
Angiogenesis plays an essential role in the process of growth and metastasis of solid tumours (Weidner, 1995, Hanahan andFolkman, 1996). Among several angiogenic factors, vascular endothelial cell growth factor (VEGF) has been shown to be vital for pathological angiogenesis. VEGF induction and vascularization of solid tumours has been shown to play an important role in the response to chemotherapeutic agents and radiation therapy (Shintani et al, 2000;Veikkola et al, 2000;Volm and Rittgen, 2000). Immunohistochemical (IHC) analyses of oesophageal carcinoma have revealed that angiogenesis, as determined by micro-vessel density, is a prognostic factor (Inoue et al, 1997;Kitadai et al, 1998;Sato et al, 1999;Shih et al, 2000). Over-expression of VEGF protein is, therefore, at least partially responsible for the malignant potential in oesophageal cancer and represents a useful prognostic marker. Thymidine phosphorylase (TP,EC 2.4.2.4), which is identical to platelet-derived endothelial cell growth factor, is also a potent angiogenic factor (Griffiths and Stratford, 1997). In oesophageal squamous-cell carcinoma (SCC), IHC studies indicated that high TP expression was associated with angiogenesis, tumour progression and poor prognosis (Igarashi et al, 1998;Takebayashi et al, 1999). In head and neck SCC, a low percentage of cancer cells with nuclear TP expression in pre-treatment biopsies was associated with a high rate of complete regression after combined CRT . Although angiogenic factors were reported as prognostic indicators in oesophageal cancer after surgery, little information is available on their predictive value for the treatment response and their prognostic significance in patients receiving CRT.
In this report, we analyzed pre-treatment biopsy samples from 52 patients with primary oesophageal SCC by IHC to identify p53, TP and VEGF expression. We found that TP and VEGF expression were significantly associated with clinical responses to treatment. We also found that VEGF expression was an independent prognostic factor for patients with oesophageal SCC following CRT.
Patients and samples
For inclusion in this study, patients were required to have presented at the Department of Academic Surgery, Chiba University Hospital, between 1991 and 1999 with histologically-proven primary SCC of the oesophagus and to have been treated by external beam radiotherapy concurrent with chemotherapy. A review of the clinical records identified 61 patients satisfying entry criteria. Nine of these were subsequently excluded because of the small size of the biopsy (five patients), previous or synchronous malignancies (two patients) or the existence of distant metastasis at the onset of treatment (two patients). The other 52 patients with primary advanced oesophageal squamous cell carcinoma underwent a prospective, non-randomized trial of combination CRT. The patients consisted of 43 males (83%) and nine females (17%), with a mean age of 65+9.6 years. Pre-treatment evaluation included clinical staging according to the TNM classification (Sobin and Wittekind, 1997), determined by radiography, endoscopic ultrasonography and computed tomography examinations (Table 1). This study was reviewed and approved by the Chiba University School of Medicine Internal Review Board. Patient eligibility criteria included the following: (i) histologically confirmed SCC of the cervical and thoracic oesophagus; (ii) age limit of 80 years and Karnofsky performance status of greater than 70%; (iii) white blood cell count greater than 4000 cells mm 73 , haemoglobin greater than 10 g, platelet count greater than 100 000 mm 73 , creatinine less than 1.5 mg dl 71 and creatinine clearance greater than 50 ml min 71 , total bilirubin less than 1.5 mg dl 71 ; and (iv) informed consent according to the Declaration of Helsinki present. At least three biopsy samples taken from different areas of the tumour of these patients for IHC analysis were obtained before treatment and stored until assay.
Treatment plan
The chemotherapy schedule consisted of cisplatin 10 mg m 72 day 71 intravenous administration and 5-fluorouracil 500 mg m 72 day 71 in continuous intravenous infusion for 5 days. The radiotherapy dose of 2 Gy per day was initiated on day 1 of chemotherapy and continued daily for 5 days per week for 4 weeks, totalling 40 Gy. The target was the entire oesophagus as well as the supraclavicular lymph nodes for the upper and mid-third lesions. Coverage of the celiac lymph nodes was decided according to computed tomography examination. Of the 52 patients, 21 were treated with an additional 20 Gy of radiotherapy (to be a definitive CRT) and 31 were treated by transthoracic oesophagectomy. Postoperative treatment was not given. Resection of the oesophagus and the proximal stomach was performed by a combined right thoracic abdominal and cervical approach. Resection included excision of the para-oesophageal, paracardial, left gastric and celiac lymph nodes.
Definition of the response to treatment
Re-evaluation of the primary tumour was performed by computed tomography, endoscopy and gastrography 2 weeks after completion of CRT. The response to treatment was basically evaluated according to the General Rules for Esophageal Cancer proposed by the Japanese Society for Esophageal Disease (1998) and was categorized as either a complete or partial response, stable or The P-value was determined using Fisher's exact probability.
Molecular and Cellular Pathology
Angiogenic factors and prognosis of oesophageal cancer H Shimada et al progressive disease. This evaluation was based on a comparison of initial and pre-operative imaging studies. A complete response was defined as the disappearance of all signs and symptoms of the tumour. A partial response was defined as a reduction of 50% or more of the tumour volume and microscopic evidence of residual tumour in postoperative specimens. The sum of the perpendicular diameter of the lesion was used to calculate tumour volume. Stable disease was defined as less than a 50% decrease or less than a 25% increase in tumour volume. Progressive disease was defined as no significant change in tumour mass or more than a 25% increase in tumour volume. The patients who showed a response, complete or partial, were categorized as responders.
The remaining patients with either stable or progressive disease, were categorized as non-responders. All patients underwent clinical examination and imaging every 3 months for the first year after the end of treatment; thereafter, every 6 months. Thirty-four patients (65%) were followed until their deaths with a median follow-up period for survivors of 36 months.
Immunohistochemical staining for p53, TP and VEGF Paraffin-embedded tissue blocks of formalin-fixed three biopsy specimens from different areas of the tumour were processed for conventional histological assessment by hematoxylin and eosin (H&E) staining and IHC analysis by the avidin -biotin -peroxidase method (Hsu et al, 1981). p53, TP and VEGF protein over-expression in the biopsy specimens was detected by anti-p53 monoclonal antibody (DO-7, DAKO, Carpenteria, CA, USA), anti-human TP (Nippon Roche Research Center, Kamakura, Japan; Nishida et al, 1996) and anti-human VEGF (A-20, Santa Cruz Biotechnology, Inc., Santa Cruz, CA, USA) using conventional peroxidase methods (Ribeiro et al, 1998). In brief, 4 mm thick sections were deparaffinized in xylene, dehydrated through graded alcohol concentrations and incubated in citrate buffer (pH=6.0) for 5 min using a household microwave oven at 800 W. After microwave exposure, the slides were allowed to cool to room temperature. The slides were briefly washed with PBS and incubated for 15 min with 3% hydrogen peroxide in methanol to block endogenous peroxidase activity. The antibodies to p53, TP and VEGF were diluted 1 : 250, 1 : 500 and 1 : 100, respectively and incubated for 24 h at 48C. Biotinylated antimouse/rabbit antibody (DAKO) at a dilution of 1 : 500 was used as the second antibody. After washing, ABC (DAKO) was applied and diaminobenzydine was used for visualization.
The stained sections were evaluated at a high magnification (6400). Staining was considered positive for p53 when more than 10% of the cells' nuclei were strongly stained. Staining was considered positive for TP or VEGF when more than 10% of the tumour cells were strongly stained. Evaluation of this immunoreactivity of three biopsy specimens was performed without knowledge of the patients' clinicopathological factors by two investigators simultaneously (T Hoshino and A Takeda). When more than two of three biopsy specimens revealed positive immunoreactivity, staining was considered positive.
Statistical analyses
Fisher's exact probability test was applied to determine the significance of the difference between two groups. Actual 5 year survival rates were compared between the two groups. Survival probabilities were calculated by the product limit method of Kaplan and Meier. Differences between groups were tested using the log-rank test. The influence of each clinicopathologic variable on survival was assessed by Cox's proportional hazards model. All statistical analyses were carried out using the Stat View program (SAS Institute Inc., Cary, NC, USA), and all P values were considered to be statistically significant if 50.05.
Immunoreactivity and clinicopathological variables
p53 expression was detected on the cells' nuclei. TP expression was detected on the cell cytoplasms, on the nuclei and on the some tumour-infiltrating stromal cells. VEGF expression was mainly detected on the cell cytoplasms or the membranes of the carcinoma cells (Figure 1). The overall frequency of expression of p53, TP and VEGF, summarized in Table 1, was 58% (30 of 52), 40% (21 of 52) and 44% (23 of 52), respectively. By p53-IHC, significant differences between positive and negative groups were observed for the factors age (P=0.025) and tumour size (P50.001). By TP-IHC, no significant differences between the two groups were found. By VEGF-IHC, significant differences between the two groups were observed in terms of gender (P=0.03) and tumour size (P=0.019).
Response to treatment and prognosis
Overall responses, including complete and partial responses, were observed in 36 patients (69%), with no response in the remaining 16 (31%). No significant differences were observed between the response rates of p53-ICH-positive or negative tumours (67 vs 73%, P=0.45) (Figure 2). However, the response rate of TP-ICHpositive tumours was significantly lower than TP-ICH-negative tumours (43 vs 87%, P=0.02). A similar tendency was observed for VEGF-ICH-positive or negative tumours (43 vs 90%, P50.001). The overall 5-year survival rate was significantly higher in the responder group than the non-responder group (14.5 vs 6.3%, P=0.0178) (Figure 3a). A similar tendency was observed in each group of the patients having CRT only (Figure 3b) and the patients having CRT followed by surgery (Figure 3c). However, because the number of the patients in each group were not enough to reach statistically significant levels (P=0.069 and 0.089, respectively).
Prognostic relevance and multivariate analysis
Using univariate analysis, treatment modality, tumour depth, N factor and VEGF-IHC status yielded a significant estimate of prognosis (Table 2). In contrast, neither p53-IHC status nor TP-IHC status was informative for the prognosis after CRT in these oesophageal SCC patients. To determine independent prognostic values for patients' survival, a Cox's regression model was constructed using TNM factors and IHC status (
DISCUSSION
In this study, the clinical significance of ICH-positivity for p53, TP and VEGF in pre-treatment biopsy specimens was examined in 52 patients with oesophageal SCC prior to CRT. Our results indicated that both TP and VEGF but not p53 expression was associated with treatment response. VEGF expression was also identified as an independent prognostic factor. In contrast to previous reports (Ribeiro et al, 1998;Yang et al, 1999), we found that p53 expression was not a predictive indicator The P-value was determined using Log-rank test.
Molecular and Cellular Pathology
Angiogenic factors and prognosis of oesophageal cancer H Shimada et al of treatment response. We suggest three possible explanations for p53-IHC status not being associated with either treatment response or survival. First, p53-ICH-negative cells also include instances of loss of both p53 alleles or nonsense mutations. Second, there may be a discrepancy in sequence analysis between assessments from different locations, i.e. endoscopic biopsy samples and surgically resected specimens. Endoscopic biopsy samples do not accurately represent characteristics of all tumour cells. In our other series, we compared the p53-IHC and TP-IHC staining results from the biopsy and resected specimens. The sensitivity was more than 90% and the specificity was around 80% (unpublished data).
In this present study, because all patients received CRT, we could not compare the IHC staining results from the biopsy and resected specimens for the validity of the data. Third, there are significant differences in histology between the tumours examined in the previous reports and our study, because all of our cases were histologically proven to be SCC, whereas more than two-thirds of the previously-reported cases were adenocarcinomas. There might well be differences between the response rates of SCC and adenocarcinoma even with the same p53 mutations. The association of angiogenic factor expression with a high incidence of treatment failure may contribute to the resistance to therapy observed in both the TP-ICH-and VEGF-ICH-positive groups. The duration to treatment failure and the treatment response rates were significantly poorer in the VEGF-ICH-positive group compared to the VEGF-IHC-negative group, and thus would eventually lead to poorer survival, as reported previously (Kitadai et al, 1998;Sato et al, 1999;Shih et al, 2000). Blocking VEGF activity was reported to enhance the anti-tumour effects of ionizing radiation (Gorski et al, 1999). Those authors proposed a model in which induction of VEGF by ionizing radiation contributes to the protection of tumour vessels from radiation-mediated cytotoxicity. In our present study, both the TP-ICH-and VEGF-ICHpositive groups experienced significantly lower treatment response rates. Therefore we propose a new model in which VEGF and TP expression both contribute to the protection of tumour blood vessels from CRT-mediated cytotoxicity and thereby to treatment resistance.
One question raised by the present study is why TP expression was not a significant prognostic factor, despite the fact that it was significantly associated with response rate. We suggest two possible reasons for this result. First, TP expression in our series was not associated with TNM factors, in contrast to previous reports (Igarashi et al, 1998;Takebayashi et al, 1999). Second, the complete response rate of TP-ICH-negative patients was relatively low compared to VEGF-ICH-negative patients. In the multivariate analysis, because TP expression was significantly associated with VEGF expression (P=0.03, data not shown), consistent with previous reports (Fujimoto et al, 1998;O'Byrne et al, 2000), VEGF might be identified as an independent prognostic factor in place of TP. As the small number of patients enrolled in this study was a limitation, further larger scale studies are required to address this question.
The development of convenient and reliable biomarkers predicting which patients are most likely to develop recurrence of primary disease would allow intervention strategies to be specifically targeted to patients most likely to benefit from them. Such a capability would be cost-effective and would avoid treating patients with a low response potential, who do not react to the usual adjuvant therapy. The present study suggests that patients with locoregional advanced oesophageal SCC positive for angiogenic factors are less likely to benefit from neoadjuvant CRT with the usual regimen than patients who are negative for angiogenic factors. Among 52 patients, 13 had tumours both TP-and VEGF-IHC-positive and a further 21 both TP-and VEGF-IHCnegative. Only three patients in the former group of 13 responded to therapy whereas 20 patients of the latter group of 21 did respond (data not shown). Moreover, none of the 13 in the TPand VEGF-IHC-positive group survived 45 years. Therefore, monitoring angiogenic factors may be an important determinant for the differential application of therapy, not only for primary tumours but also for adjuvant therapy after definitive treatment of oesophageal cancer. It is essential to evaluate the prognosis separately in patients having neoadjuvant therapy with resection (n=31) and neoadjuvant therapy only (n=21) groups to confirm the prognostic value of VEGF-IHC. However, because a limited number of the patients in each groups, TNM factors and VEGF-IHC were assessed by multivariate analysis with treatment modality. Although neither p53-IHC nor TP-IHC were not independent prognostic factor, both VEGF-IHC and the treatment modality were selected as independent prognostic factors.
It is very difficult to develop an alternative treatment strategy for patients with tumours expressing angiogenic factors; however, radical surgery should at least be conducted without delay in these cases. VEGF-ICH-positive patients are deemed to have higher risks for recurrence and thus need more aggressive adjuvant therapy than the VEGF-ICH-negative group. Anti-VEGF therapy using anti-VEGF antibodies (Gorski et al, 1999;Lee et al, 2000) or anti-VEGF receptor therapy (Klement et al, 2000;Geng et al, 2001) may be useful in improving the effect of CRT and the prognosis of such VEGF-positive patients. Inhibitors of TP and prodrugs that are activated by TP Miwa et al, 1998;Takebayashi et al, 1999) may suppress the growth of TP-expressing tumours and may enhance the effect of CRT for patients with oesophageal SCC.
In conclusion, the present study indicates that monitoring the expression of angiogenic factors in biopsy specimens from patients with oesophageal SCC prior to treatment may have predictive value for their response to CRT and hence overall prognosis. | 4,621.4 | 2002-02-12T00:00:00.000 | [
"Medicine",
"Biology"
] |
Tunable unidirectional surface plasmon polariton launcher utilizing a graphene-based single asymmetric nanoantenna
We design and numerically investigate a graphene-based asymmetric nanoantenna microstructure that can be used to realize electrically controllable, unidirectionally propagating broadband surface plasmon polaritons. The device geometry facilitates the simultaneous excitation of two localized surface plasmons resonances in the whole structure, and consequently, the asymmetric nanoantenna can be considered as being composed of two oscillating magnetic dipoles, wherein the interference of the radiated electromagnetic waves leads to a unidirectional propagation effect. Our results indicate that our proposed active device is promising for realizing compactable, tunable, terahertz plasmonic light sources. © 2017 Optical Society of America OCIS codes: (240.6680) Surface plasmons; (240.0310) Thin films; (260.5740) Resonance; (350.5030) Phase. References and links 1. J. A. Schuller, E. S. Barnard, W. Cai, Y. C. Jun, J. S. White, and M. L. Brongersma, “Plasmonics for extreme light concentration and manipulation,” Nat. Mater. 9(3), 193–204 (2010). 2. P. Genevet, J. Lin, M. A. Kats, and F. Capasso, “Holographic detection of the orbital angular momentum of light with plasmonic photodiodes,” Nat. Commun. 3, 1278 (2012). 3. S. Wu, Z. Zhang, Y. Zhang, K. Zhang, L. Zhou, X. Zhang, and Y. Zhu, “Enhanced rotation of the polarization of a light beam transmitted through a silver film with an array of perforated S-shaped holes,” Phys. Rev. Lett. 110(20), 207401 (2013). 4. L. Li, T. Li, X. M. Tang, S. M. Wang, Q. J. Wang, and S. N. Zhu, “Plasmonic polarization generator in well-routed beaming,” Light Sci. Appl. 4(9), e330 (2015). 5. Q. Gan, B. Guo, G. Song, L. Chen, Z. Fu, Y. J. Ding, and F. J. Bartoli, “Plasmonic surface-wave splitter,” Appl. Phys. Lett. 90(16), 161130 (2007). 6. Q. Gan, Z. Fu, Y. J. Ding, and F. J. Bartoli, “Bidirectional subwavelength slit splitter for THz surface plasmons,” Opt. Express 15(26), 18050–18055 (2007). 7. J. J. Chen, Z. Li, S. Yue, and Q. H. Gong, “Efficient unidirectional generation of surface plasmon polaritons with asymmetric single-nanoslit,” Appl. Phys. Lett. 97, 04113 (2010). 8. D. Li, D. H. Zhang, C. Yan, T. Li, Y. Wang, Z. Xu, J. Wang, and F. Qin, “Unidirectional surface plasmon-polariton excitation by a compact slot partially filled with dielectric,” Opt. Express 21(5), 5949–5956 (2013). 9. J. Yang, S. Zhou, C. Hu, W. Zhang, X. Xiao, and J. Zhang, “Broadband spin-controlled surface plasmon polariton launching and radiation via L-shaped optical slot nanoantennas,” Laser Photonics Rev. 8(4), 590–595 (2014). 10. Y. Liu, S. Palomba, Y. Park, T. Zentgraf, X. Yin, and X. Zhang, “Compact magnetic antennas for directional excitation of surface plasmons,” Nano Lett. 12(9), 4853–4858 (2012). 11. L. Wang, T. Li, L. Li, W. Xia, X. G. Xu, and S. N. Zhu, “Electrically generated unidirectional surface plasmon source,” Opt. Express 20(8), 8710–8717 (2012). 12. T. Liu, Y. Shen, W. Shin, Q. Zhu, S. Fan, and C. Jin, “Dislocated double-layer metal gratings: an efficient unidirectional coupler,” Nano Lett. 14(7), 3848–3854 (2014). 13. S. Xiao, F. Zhong, H. Liu, S. Zhu, and J. Li, “Flexible coherent control of plasmonic spin-Hall effect,” Nat. Commun. 6, 8360 (2015). 14. J. Lin, J. P. B. Mueller, Q. Wang, G. Yuan, N. Antoniou, X. C. Yuan, and F. Capasso, “Polarization-controlled tunable directional coupling of surface plasmon polaritons,” Science 340(6130), 331–334 (2013). 15. I. P. Radko, S. I. Bozhevolnyi, G. Brucoli, L. Martín-Moreno, F. J. García-Vidal, and A. Boltasseva, “Efficient unidirectional ridge excitation of surface plasmons,” Opt. Express 17(9), 7228–7232 (2009). Vol. 7, No. 2 | 1 Feb 2017 | OPTICAL MATERIALS EXPRESS 569 #281660 http://dx.doi.org/10.1364/OME.7.000569 Journal © 2017 Received 28 Nov 2016; revised 3 Jan 2017; accepted 10 Jan 2017; published 25 Jan 2017 16. A. Pors, M. G. Nielsen, T. Bernardin, J. C. Weeber, and S. I. Bozhevolnyi, “Efficient unidirectional polarization controlled excitation of surface plasmon polaritons,” Light Sci. Appl. 3(8), e197 (2014). 17. M. Tymchenko, A. Y. Nikitin, and L. Martín-Moreno, “Faraday rotation due to excitation of magnetoplasmons in graphene microribbons,” ACS Nano 7(11), 9780–9787 (2013). 18. C. Chen, S. Rosenblatt, K. I. Bolotin, W. Kalb, P. Kim, I. Kymissis, H. L. Stormer, T. F. Heinz, and J. Hone, “Performance of monolayer graphene nanomechanical resonators with electrical readout,” Nat. Nanotechnol. 4(12), 861–867 (2009). 19. T. J. Echtermeyer, S. Milana, U. Sassi, A. Eiden, M. Wu, E. Lidorikis, and A. C. Ferrari, “Surface plasmon polariton graphene photodetectors,” Nano Lett. 16(1), 8–20 (2016). 20. T. Mueller, F. N. Xia, and P. Avouris, “Graphene photodetectors for high-speed optical communications,” Nat. Photonics 4(5), 297–301 (2010). 21. M. Liu, X. Yin, E. Ulin-Avila, B. Geng, T. Zentgraf, L. Ju, F. Wang, and X. Zhang, “A graphene-based broadband optical modulator,” Nature 474(7349), 64–67 (2011). 22. M. Liu, X. Yin, and X. Zhang, “Double-layer graphene optical modulator,” Nano Lett. 12(3), 1482–1485 (2012). 23. Y. J. Bao, S. Zu, Y. F. Zhang, and Z. Y. Fang, “Active Control of Graphene-Based Unidirectional Surface Plasmon Launcher,” ACS Photonics 2(8), 1135–1140 (2015). 24. M. D. He, K. J. Wang, L. Wang, J. B. Li, J. Q. Liu, Z. R. Huang, L. L. Wang, L. Wang, W. D. Hu, and X. S. Chen, “Graphene-based terahertz tunable plasmonic directional coupler,” Appl. Phys. Lett. 105(8), 081903 (2014). 25. A. Vakil and N. Engheta, “Transformation optics using graphene,” Science 332(6035), 1291–1294 (2011). 26. W. Gao, J. Shu, C. Qiu, and Q. Xu, “Excitation of plasmonic waves in graphene by guided-mode resonances,” ACS Nano 6(9), 7806–7813 (2012). 27. L. Zhu, Y. H. Fan, S. Wu, L. Z. Yu, K. Y. Zhang, and Y. Zhang, “Electrical control of terahertz polarization by graphene microstructure,” Opt. Commun. 346, 120–123 (2015). 28. M. Jablan, H. Buljan, and M. Soljačić, “Plasmonics in graphene at infrared frequencies,” Phys. Rev. B 80(24), 245435 (2009). 29. J. D. Jackson, Classical Electromagnetics, 3rd ed.; John Wiley & Sons: New York, 1999. 30. L. Zhou, C. P. Huang, S. Wu, X. G. Yin, Y. M. Wang, Q. J. Wang, and Y. Y. Zhu, “Enhanced optical transmission through metal-dielectric multilayer gratings,” Appl. Phys. Lett. 97(1), 011905 (2010). 31. J. Yang, X. Xiao, C. Hu, W. Zhang, S. Zhou, and J. Zhang, “Broadband surface plasmon polariton directional coupling via asymmetric optical slot nanoantenna pair,” Nano Lett. 14(2), 704–709 (2014). 32. Y. Yao, M. A. Kats, P. Genevet, N. Yu, Y. Song, J. Kong, and F. Capasso, “Broad electrical tuning of graphene-loaded plasmonic antennas,” Nano Lett. 13(3), 1257–1264 (2013). 33. L. Huang, Y. H. Fan, S. Wu, and L. Z. Yu, “Giant asymmetric transmission and optical rotation of a three-dimensional metamaterial,” Chin. Phys. Lett. 32(9), 094101 (2015). 34. C. Q. Li, L. Huang, W. Y. Wang, X. J. Ma, S. B. Zhou, and Y. H. Jiang, “Electromagnetically induced transparency in nano-structures made from metallic nanorod and split-ring-resonator,” Opt. Commun. 355, 337–341 (2015). 35. S. Wu, Q. J. Wang, X. G. Yin, J. Q. Li, D. Zhu, S. Q. Liu, and Y. Y. Zhu, “Hybridized effects of plasmonic quadrupolar and dipolar resonances on the perforated planar metallic film,” Appl. Phys. Lett. 93, 101113 (2008). 36. E. Prodan, C. Radloff, N. J. Halas, and P. Nordlander, “A hybridization model for the plasmon response of complex nanostructures,” Science 302(5644), 419–422 (2003).
Introduction
Surface plasmon polaritons (SPPs) are the electromagnetic waves traveling along metal-dielectric or metal-air interfaces that originate from the interaction between light and collective electron oscillations on metal surfaces.SPPs have attracted considerable attraction in terms of their application to subwavelength-optics microscopy, lithography beyond the diffraction limit, and miniaturized photonics devices for practical applications.Owing to their characteristics of localized field enhancement and subwavelength confinement, SPPs have been widely applied in plasmonic metamaterial and metasurface applications [1][2][3][4].In this context, the generation of high-efficiency unidirectional plasmon waves has formed a fundamental research issue in the area of the nano-optics, since such generation is vital to realize novel nanoscale optical devices.Researchers have previously realized unidirectional SPP devices by adding two optimized grating structures (waveguides) on the opposite sides of a slit, wherein the confined SPP propagation behavior can be analyzed in the terms of the SPP dispersion curves [5,6].Subsequently, asymmetric slits [7,8], holes [9], nanoantennas [10], tilted gratings [11], and dislocated double-layer gratings [12] based on the interference effect have been proposed for unidirectional SPP excitation.However, unidirectional devices generally need the fabrication of precise samples to satisfy the interference-effect conditions.Constraints within current nanofabrication technologies make it difficult to construct precise samples in experiments.Fortunately, it is possible to solve these problems through tailoring of the phase of the radiated electromagnetic (EM) waves by methods offering external control, such as those utilizing the spin-Hall effect [13,14], and those involving changing the polarization angle of incident light [15,16].
Graphene, a single layer of carbon atoms arranged in a honeycomb lattice, exhibits almost all the electrical properties and functions required for integrated photonic circuits [17,18].Because of high carrier mobility and wideband absorption, graphene can be potentially applied in ultrafast broadband photodetectors [19,20].More importantly, the Fermi level of graphene depends on the bias or chemical doping, and thus, interband transitions can be switched on or off by shifting the Fermi level above or below the threshold value (ħɷ/2).Via tuning of the interband and intraband transitions, graphene can be utilized in designing for electro-optic modulators [21,22].In this regard, tunable graphene-based unidirectional devices have recently been proposed by Fang [23] and He [24].However, these devices offer limited scope for integration due to their structural complexity.Thus, in this study, we design a graphene-based unidirectional SPP device composed of a single asymmetric antenna and Au/SiO 2 substrate.When compared with previously proposed devices, our designed structure can achieve the tuning unidirectional SPP propagation over a broadband wavelength range based on the interference of excited localized surface plasmons (LSPs).
Model and simulation
Figure 1 depicts the schematic of our proposed device.An asymmetric plasmonic antenna deposited on a graphene sheet is separated from a gold substrate by a 100-nm -thick SiO 2 spacer layer.The asymmetric plasmonic antenna is composed of three different metal strips with two non-identical cavities.The geometric parameters of the structure are shown in the inset of Fig. 1, where d1 = 1.45 µm, d2 = 4.5 µm, d3 = 2.5 µm, h1 = 150 nm, h2 = 50 nm, g1 = 20 nm, and g2 = 40 nm.A plane transverse magnetic (TM) polarization wave (whose magnetic field is perpendicular to the y-z plane) with a wavelength of 6.4 µm is normally incident on the sample plane.Plasmon waves propagated upon application of a voltage V between the graphene sheet and the back-gated Au are investigated by means of the finite element method (FEM).The anisotropicity of graphene can be expressed by means of a diagonal tensor.Graphene's out-of-plane permittivity is set to 2.25, and its in-plane permittivity [25] can be obtained as where σ g,i and σ g,r represent the imaginary and real components of the conductivity of graphene (σ g ), and ɛ 0 and t = 0.33 nm the vacuum permittivity and thickness of graphene, respectively.In the terahertz frequency range, the conductivity of graphene σ g can be described by means of the following Drude-like expression [26,27]: Here, f E and τ represent the Fermi energy level and carrier relaxation time, respectively.
Parameters E f and τ are given as ħv f (πn g ) 1/2 and μE f /ev f 2 [28], respectively, with the Fermi velocity v f = 10 6 m/s and carrier mobility μ = 3000 cm 2 /(V•s).In our calculation, the permittivity of the spacer layer SiO 2 is assumed as 2.25, and the optical constants of gold are modeled by the Drude model with a plasma frequency of 1.367 × 10 16 rad/s and collision frequency of 6.478 × 10 13 rad/s.
Results and discussions
Figure 2 shows the cross-sectional view of the y-component of the electric field distribution and power flow distribution for different Fermi energies.Utilizing the gate-voltage dependent optical conductivity of graphene, we can control the propagation characteristics of the SPPs.Figures 2(a)-(c) show the electric field intensity distribution with the increase in the Fermi energy of graphene.As the Fermi energy increases, the electric field intensity in the left side of the structure decreases.At the Fermi energy of 0.5 eV, the electric field intensity in the left region disappears, and thus, the SPP mainly propagates in the right side [Fig.2(c)].The corresponding power flow distribution also verifies this result [Fig.2(d)].Specifically, the excited SPPs field is highly delocalized, and it extends about four-fifths of incident wavelength to the air layer.Since the wavelength of SPPs is close to the incident light, the excited SPPs are Au-Air interface modes rather than Au-SiO 2 modes.Figure 2(e) characterizes the extinction ratio η as a function of the Fermi energy (here, the extinction ratio η is defined as the ratio of SPP power flow along the right and left sides of the asymmetric antenna).An extinction ratio of η = 61 is achieved at the Fermi energy of 0.5 eV.It is a definitely good performance for a unidirectional SPP launcher with an electrically tunable property.In the following, we would like to unveil the underlying that attributes to this interesting effect It is well known that a metal nanostrip and a metallic mirror separated by a dielectric spacer can be regarded as a magnetic dipole resonator.Under the magnetic dipole approximation, the radiated field can be defined as [29] ( ) where m represents the magnetic dipole, n the unit vector in the direction of r with respect to the position of the dipole, ɸ the radiated phase of the dipole, k spp = 2π/λ spp the wave vector of SPPs given by [ the radiated EM waves will interfere destructively along the positive direction of the x-axis while interfering constructively along the negative x-direction.This directional propagation of the EM waves leads to the unidirectional propagation of the SPPs.Here, 1 ( ) 4) and ( 5) yields the following expression Thus, unidirectional SPP propagation will occur when the relative phase between the two cavities equals / 2 π .Given that the plasmonic resonance of the metallic antenna depends on the geometric size and environment of the cavity (i.e. the gap surrounded by the metallic strip), the resonant phase and frequency of the antenna can be modulated based on the equivalent circuit model [32], when a tunable material such as graphene is introduced in the cavity.To verify that the phase of the cavities is a function of the Fermi energy, the electric field and phase for different Fermi energies in the two cavities are calculated as in Figs.3(a)-3(d).For the left cavity [Fig.3(a)], the resonant peak at the wavelength of 6.4 μm is blue-shifted with increase in the Fermi energy.The frequency shift Δɷ is given by the following formula [32]: Here, ɷ 0 denotes the resonance frequency in the absence of graphene, and L and L G the inductance of the metal pair and graphene, respectively.The inductance of graphene can be expressed as L G = -g/ɷ 2 ɛ 0 ɛ G t G = -g/ɤɷ 2 ɛ 0 E f t G (ɤ denotes a deduced coefficient for simplicity, and g denotes the cavity width), Further, C p represents the capacitance of the metal pair.The blue-shift of the resonance frequency results from decrease in the graphene inductance L G because of the injection of external carrier concentration.However, for the right cavity, the inductance of L G is large due to its larger size.From Eq. ( 7), we can infer that the resonance frequency in the right cavity [Fig.3(b)] does not significantly change with increase in the Fermi energy.Thus, the resonant-peak shift affects the resonant phase change in the left cavity.The resonance phases in the left and right cavities are depicted in Figs.3(c) and 3(d), respectively.As expected, the tuning behavior of the resonant phase agrees well with the simulation results of the electric field in the two cavities.We next plot the phases corresponding to the left and right cavities as a function of the wavelength in Fig. 4 for the Fermi energy of 0.5 eV.The resulting plot confirms that the designed structure meets the phase relationship corresponding to the unidirectional effect.At an incident wavelength of 6.4 µm, a phase difference of ~/ 2 π can be obtained, which is in good agreement with Eq. ( 6).In addition, by dynamically tuning the Fermi energy of graphene, we can achieve unidirectional SPP propagation over a broadband wavelength range (at incident wavelengths of 5.7-6.7 μm, see Fig. 6 in the Appendix).These results indicate that the unidirectional effect is caused by the interference of the radiated EM waves induced by the two cavities.Finally, we examine the electric field intensity distribution in the designed structure [Fig.5(a)].Figures 5(b) and 5(c) are the electric field intensity distributions with graphene in the left and right cavities, respectively, in both of which the electric field are greatly enhanced compared with those in outside.It is observed that the graphene sheets rightly locate at the strong field region in both cavities that ensures the tunability by changing the graphene's Fermi energy.However, by comparing the field intensity in two cavities, one can find that the left cavity has much stronger field enhancement than the right one, which would reasonably attribute to the different resonant effect according to different cavity parameters.So that, the stronger enhancement of field on graphene, the bigger resonance shift with respect to different Fermi energy (the left cavity), and vise versa (the right cavity), which agree well with the results in Fig. 3(a) and 3(b).Here, the normalized field intensity in the graphene region for left cavity is ~160 V 2 /m 2 , and the right cavity is ~8.17V 2 /m 2 .The tunability of the cavity is proportional to the field intensity in the graphene region.Thus, the left cavity can provide a significantly larger tunable range than the right cavity.Our analyses indicate that the interaction between graphene and the asymmetric antenna gives rise to the directional SPP propagation with tunable capability.
Fig. 1 .
Fig. 1.Schematic of the proposed structure composed of a graphene monolayer and an asymmetric metal nanoantenna positioned atop a 100-nm-thick SiO 2 spacer layer supported by a Au substrate.The inset shows the detailed geometric parameters of the asymmetric nanoantenna.
of the Fermi energy level of graphene, and the initial phases of to the asymmetricity of the device structure.Further, d denotes the distance between the two gaps (g1 and g2), while N denotes an arbitrary integer.Simple manipulation of Eqs. (
Fig. 2 .
Fig. 2. (a)-(c) The y-component of the normalized electric field distribution at 0.1 eV, 0.3 eV and 0.5 eV, respectively.(d) Plot of the x-component of the power flow distribution at 0.5 eV.(e) Electrically controllable characterization of the extinction ratio as a function of the Fermi energy level.
Fig. 3 .Fig. 4 .
Fig. 3. (a) and (b) Simulated electric field magnitudes at the center of the left and right cavities for varying Fermi energies, respectively.(c) and (d) Dependence of the phases of the left and right cavities on the Fermi energy, respectively.
Fig. 5 .
Fig. 5. (a) The cross-sectional views of the structure.(b) and (c) are the simulated electric field intensity distribution with graphene in the vicinity of left and right cavities, respectively.The incident wavelength is 6.4 μm, Fermi energy is 0.2 eV. | 4,446.4 | 2017-02-01T00:00:00.000 | [
"Physics"
] |
Computational Simulation of flight behavior of Flying Wing UAV
The aim of this work is to model and simulate a Flying Wing Aircraft and acquire flight data from two different simulations. The framework proposed constitutes of the development of a graphical model as well the use of a mathematical model of the aircraft. In order to simulate realistic flight conditions, we used a commercial flight simulation environment, X-plane 10. We performed the real time flight data collection from the simulator, as well as the software-in-the-loop feedback control through MATLAB/Simulink. Key-words: UAV, Software In the Loop, Hardware In the Loop, X-Plane, Simulink , Flying Wing
Abstract-The aim of this work is to model and simulate a Flying Wing Aircraft throughout a combination of two software and an automated system to compare the flight data acquired from two different simulations.The framework proposed constitutes of the development of both graphical and mathematical model of the aircraft aimed at flight data collection.In order to implement the Software-in-the-Loop (SIL), we used a commercial flight simulation environment, X-plane 10, which simulates the dynamics of the aircraft through the Blade Element Theory, and the software MATLAB/Simulink, which simulates and implement the control laws.Data transfer between X-Plane and Simulink was made possible through UDP (User Datagram Protocol), which enables process-to-process communication and works in conjunction with higher level protocols to help manage data transmission services.We collected the flight data from the simulator, as well as the responses of the mathematical model based on equations of motion and the aerodynamic derivatives, which were computed via the Digital DATCOM software.Both analytical simulation and the SIL simulation were performed with the same flight conditions, validating the longitudinal dynamic of the computational model.This framework allows further studies on aircraft simulation to be developed, reinforcing the collection, storage and processing of flight data.
I. INTRODUCTION
N the recent years, computational simulation has been an important tool for data acquisition, and in providing essential information about the behavior of mechanical systems [1].In the aerospace industry, simulations are used throughout the development of all aircraft, and they evaluate the control algorithm allowing for easy manipulation of the early model.This allows for data creation for the aircraft model, for faster development of the product, as well as minimizing the number of experimental flights.Thus, reducing the number of persons involved and consequently the final cost of the project [2].However, it is necessary to develop more reliable models of aircraft.
Several commercial simulators are available and used as a tool, known as Software-In-the-Loop (SIL), for the implementation of flight dynamics, navigation control, and for validating models in a precise fashion before field tests [2].SIL couples partially integrated software with an environment simulation and allows a direct information-technical communication between the two, for simple data creation, collection, and processing.
This also offers the possibility to execute tests before the hardware is available.Despite simulation-based testing being a very important part of the product development, Hardware-Inthe-Loop (HIL) simulation tests should be adopted whenever it is possible in order to validate together both the hardware and the software under realistic conditions [3].
In the specific case of Unmanned Aerial Vehicles (UAV) the mathematical model, as well as the implementation of SIL and HIL were describe by [4].It not only describes the physics of the system but also the behavior of the low-level autopilot, and the state estimation routines.UAVs are of high interest for prospective military and civilian applications due to their several advantages such as maneuverability, and lack of need for direct human interaction [5].The literature shows several works of multirotor aircraft modeling [6] [7].However, more detailed modeling of flying wings and their integration with SIL and HIL are still a growing field of high interest [8].Therefore, the literature does not provide many resources on this type of aircraft.
Since there is a lack of mathematical modeling that describes the behavior of flying wings because of their design complexity and aerodynamic instability, this paper presents the longitudinal dynamic modeling of a flying wing UAV through the derivatives of stability and control as well as the equations of motion, and through the Blade Element Theory.The mathematical model describes the system by its aerodynamic derivatives and equations of motion that establish the relationship among a set of variables, which provides an explicit expression of the systems behavior pattern and how it would work before building it.The behavior of each variable shows immediately in the equation, which in return allows the system to be analyzed and optimized.Whereas, through simulation modeling, it creates and analyses a digital prototype of a physical model to predict its performance in the real world with environment effects on it.
I
The purpose of the present work is to, first, mathematically and graphically model the flying wing.Second, to simulate both models aiming to compare and review the dynamic response from the introduction of longitudinal oscillations during cruise flight, using the commercial flight simulator X-Plane (Laminar Research©) and the mathematical software MATLAB/Simulink (MathWorks©).Finally, perform several flight simulation tests in order to apply refined adjustments to the aircraft parameters and to improve the control of flight stability.The results of both simulation methods were analyzed and compared.This model will allow the development of future projects from what has already been accomplished in this, thus helping the advance of new research with flying wing flight data creation, its software data storage and even a tool for embedded systems.
This paper is organized as follows: Section II introduces the software X-Plane and its Plane-Maker platform, and then the procedure used for the modeling through the Blade Element Theory.Section III describes the UDP communication between the X-Plane and MATLAB/Simulink, the necessary configuration for the interface, and the implementation of the control system through Simulink for the simulations of the UAV.Section IV demonstrates mathematically the flight dynamics of a fixed wing UAV through the derivatives of stability and control.Section V presents the results obtained and discussions.The last section concludes the importance of this work and presents the research that still needs to be done as future work.
II. PLANE-MAKER MODELING
The flying wing was modeled using the platform Plane-Maker, built into X-Plane 10, which enables the creation of any aircraft.With its graphical interface, the user can have a visual feedback of the in-flight behavior of the aircraft once all the physical characteristics are applied to the model.X-Plane 10 is a realistic flight simulator certified by the Federal Aviation Administration (FAA), and a viable tool due to its aircraft creation environment and its ability to simulate the aircraft dynamics.Through the combination of both these assets, we obtained the graphical model of the UAV.
The UAV chosen for this project was a commercial RC Flying Wing (Fig. 1), and its characteristics are shown below in Table I.This specific model did not exist in the X-Plane database; therefore, we modeled it from scratch (Fig. 2).The dimensioning of the reference model is very important because X-Plane separates the aircraft into small sections to calculate several aerodynamic factors on each, ensuring that the entire aircraft is being computed.This methodology is based on a theory called Blade Element, which is explained by [6].The model was created in the Plane-Maker according to the specifications of the actual model and following the guidelines recommended by the platform's manual [9].
The UAV is powered by a brushless electric motor and its parameters of the thrust and power are approximated according to the manufacturer's datasheet [EMAX 2822 Brushless Motors datasheet].Because X-Plane is an aircraft flight simulator on real scales, there is a difficulty in making the reaction time of the engines of a small UAV compatible with the real model.Several in-flight tests were performed to fine-tune the parameters, specifically in the engine and CG position.The suitable engine specifications are:
1) Maximum allowable Power: 166W; 2) Engine RPM: 12000; 3) Propeller: 10x5;
The static thrust of the propeller was computed via the program JavaProp.This feature describes the geometry of the airfoil to be tested, assigns the appropriate Reynolds number, generates the aerodynamic coefficients, and therefore thrust.A file is generated and then exported to Plane-Maker.
III. SOFTWARE-IN-THE-LOOP IMPLEMENTATION
MATLAB/Simulink is an environment for simulation of dynamic systems.It provides a customizable set of block libraries that enables the user to design, simulate, and test a variety of systems time-varying.This platform presents ready blocks for performing communications with external environments, allowing data exchange.Therefore, a block diagram was created in Simulink aiming to send flight control to the model built in the platform X-Plane.The controls details for the architecture used for this project can be found at [10].
UDP is a transport layer protocol, where each output operation of a process produces exactly one UDP datagram, causing one Internet Protocol (IP) datagram to be sent, an indepth description can be found in [6].The simulation constitutes of two computers that use an Ethernet network with IP addresses defined for each, which allows data exchange between MATLAB/Simulink and X-Plane, as illustrated in Fig. 3.Both software were configured for sending and receiving packets and data, following the guidance at [2].
The MATLAB/Simulink PC generates the error signal through a reference signal and feedback data provided by X-Plane.Then, it uses the error signal as input to the control law, thus generating a command signal for the motors that will be sent back to X-Plane.Finally, X-Plane PC receives the command data for the engine, executes the interactions and sends the new data position to the MATLAB/Simulink PC again [11].This communication procedure is explained in detail by [12] [13].
A. X-Plane Settings
After modeling the UAV in Plane-Maker, an extension *.acf file is generated, which is loaded into X-Plane to perform inflight simulations.An adjustment on the number of flightmodel per frame was necessary in order to establish the amount of time that the flight simulator calculates the forces in each frame.A value between 6 and 10 is recommended to ensure a better performance of the model during the simulations [2].
For the analysis of the simulations, obtaining the flight data is required.X-Plane has several ways to view this data, it can be displayed on the main screen of the simulator and sent via Ethernet using UDP.The configuration can be done in Setting / Data Input Output.For this work, the following output data was configured: 1) Angular position (, , ); 2) Scalar position (x, y, z); 3) Angular velocity (Q, P, R); 4) Scalar velocity (Vx, Vy, Vz).
Another important adjustment is the UDP rate/s and disk rate/s values to match the response frequency of the X-Plane with MATLAB/Simulink, which was set as 60.00 for both.
Each output generates a set of nine groups of 4 bytes each [14].The first four bytes represent the type of packet, the fifth byte is an internal policy, and the next group of four bytes represent the parameter label that is being sent as shown in Fig. 4. For each label, X-Plane sends a packet with eight data and all of them take over the format of single-precision floating point of 32 bits [3] [6].
To send commands to X-Plane, the packet must be built with the same architecture.
B. Control System
After modeling and adjusting all parameters of the Flying wing in Plane-Maker, following the same dimensions, weight and CG of the real aircraft, we implemented the control laws using the software MATLAB/Simulink, in order to test control algorithms and navigation of the aircraft.
The Flying Wing is a tailless fixed-wing aircraft that has no definite fuselage.Due to the lack of conventional stabilizing surfaces and the associated control surfaces, this type of aircraft suffers from the inherent disadvantages of being dynamically unstable and difficult to control.This means that the amplitude of any oscillatory motions induced by disturbances eventually increase with time infinitely relative to a steady-state flight condition as shown in Fig. 5.
Dynamic instabilities can be tolerated by a human pilot or by automatic controls for feedback controllers, which is known as "closed-loop stability".Therefore, to perform the flight tests for both analytical and graphical methods, a PID (Proportional-Fig.3. Simulation loop.Modified from [11].Fig. 4. Data packet from [6].Fig. 5. Graphical example of dynamically unstable aircraft motion from [15].Fig. 6.Graphical example of dynamically stable aircraft motion from [15].Integral-Derivative) Controller system for the control loops of pitch and roll was developed in Simulink as shown in Fig. 7.It allowed us to perform the simulations at a constant velocity, oscillating only the pitch attitude degrees of freedom.
C. Simulink Settings
In order to use the data received by X-Plane in the control and guidance laws implemented in Simulink, the packets need to be broken in variables.According to [6], this interface has five main blocks as illustrated bellow in Fig. 8.The blocks "Receiving data" and "Sending data" are responsible for closing the system control loop, and the blocks Graphical Interface and the Simulation Analyze act on every interface.
Simulink presents some blocks that send and receive data through the UDP protocol.The first block is the UDP receiver, it receives the packets sent from X-plane, it is necessary to configure it according to the guidance at [2].An unpacking block must be provided to extract the data, and the Byte Reversal to adapt the byte because X-Plane exports the data in big-endian format where the most significant one is the first byte, whereas MATLAB work in little-endian format, which the most significant byte comes last.Then, the control laws for a fixed wing aircraft are implemented through Simulink.The code interprets the UDP data.Fig. 9 illustrates the flowchart of the implemented program.And at the end the data is shown on the output displays.
The UDP Sending is similar to the receiving process.The data packet must be mounted with the same architecture that X-Plane sends commands as illustrated in Fig. 4.
IV. MATHEMATICAL MODELING
The dynamics for fixed-wing aircraft can be approximately decomposed into longitudinal motion.Aiming to obtain the mathematical model that describes the longitudinal dynamics, a series of considerations are made to obtain the equations, some are described below: 1) The Earth is the inertial reference system; 2) The aircraft is a rigid body and its weight is a constant; 3) Acceleration of gravity does not change with flight altitude; 4) The Reynolds number and Mach number effects are approximately constant; 5) The Sideslip Angle () is considered null; 6) The disturbances around the equilibrium are small, with small variations to the pitch angle (θ); 7) The deflection of the elevator does not alter any force, only the moment of pitch.
An aircraft usually has 6 degrees of freedom (6DoF), which can be described by equations of forces and moments acting on it and presents nonlinear dynamics [16].Using these nonlinear equations of the longitudinal dynamics, and the calculation of the forces and moments through the stability and control derivatives we can finally simulate the flight of the UAV.
A. Stability and control derivatives via USAF DATCOM
The stability and control derivatives determine the static and dynamic stability of the aircraft.The static stability refers to the direction of aerodynamic moments as the aircraft oscillates its nominal flight condition, where the tendency of restoring the initial position defines if the aircraft is statically stable.The dynamic stability refers to the dynamic behavior of the airframe in response to disturbances, where the aircraft is defined as dynamically stable if its response dampens the disturbance out over time as shown in Fig. 6.
The stability and control derivatives can be estimated or identified through wind tunnels tests or via flight tests [10].According to the literature, the most feasible method for obtaining these derivatives in cruise flight conditions is through calculations based on empirical data since wind tunnels is not always accessible.
The Data Compendium (DATCOM) has been widely used for the estimation of the aircraft derivatives.It was developed by the United States Air Force (USAF) and is based on empirical data (wind tunnel data history and in-flight testing of certain types of airfoils and fuselages) and simplified analytical methods, this method is known as semi-empirical.Given an aircraft configuration, declared in FORTRAN language, DATCOM easily provides the estimation of aerodynamic coefficients and their stability derivatives.
All the models parameters and specifications for the analytical modeling were the same as those used for modeling in Plane-Maker.The model was created in DATCOM following the guidelines recommended by the manual [17].Table II contains the numerical values of the longitudinal aerodynamic derivatives of the flying wing. is referred to as the longitudinal static stability derivative, is referred to as the
B. Aerodynamic coefficients for forces and moment
Aerodynamic forces and moments have a complex dependence on several variables, resulting in a complicated nonlinear correlation between each variable [10].These variables, known as aerodynamic coefficients, are dimensionless and depend on the derivatives of stability and control for a flight condition, and are described below: 1) Lift Coefficient: 2) Drag Coefficient: 3) Pitching Moment Coefficient: Where: is the elevator control deflection, and is the attack angle.
C. Aerodynamic Forces and Moment
As the aircraft passes through the air, a pressure distribution, or aerodynamic force, is generated around its body.The air velocity ( ), air density ( ) and the shape and attitude of the aircraft are all variables of the distribution of pressure acting on the aircraft.Hence, in order to capture the effect of the pressure with a combination of forces and a moment, the dynamic pressure ( ̅ ) is given by 2 and the variables that define the shape of the aircraft body are given by the planform area of the aircraft wing (), and the mean chord of the wing (̅ ).
The aerodynamic forces and moment are obtained from the dimensionless aerodynamic coefficients at a given flight condition as follows: A complete modeling of the engine with the propeller and its dynamic thrust are quite complicated and beyond our scope; however, a simplified model for the calculation of the propulsion force is the approximation by given as follow.
Assuming that engine provides enough torque to the propeller and the thrust delivered by the propeller is entirely dependent on the aerodynamics of the propeller, the thrust force can be calculated as described in [10]: Therefore, is the propeller radius, Ω is the rotational speed, and is an approximation at a given advance ratio value that is computed from the flight speed and the rotational speed of propeller at each instant.The propeller efficiency considered for the calculations was of 70%.
D. Linear State-space model of the UAV
The equations of motion for an aircraft are a set of 12 nonlinear, coupled, first-order, ordinary differential equations and are describe in [8].In order to produce reduced-order transfer function and state-space models more feasible for control system design, we linearized and decoupled the equations of motion.The transfer function model for the longitudinal dynamics are organized by the following variables: the pitch angle (θ), the angle of attack (), the pitch rate (), the altitude (ℎ), and the airspeed ( ).The control signals used to influence the longitudinal dynamics are the elevator ( ) and the throttle ( ℎ ).The nonlinear longitudinal equations of the system are described below: where: .̅ .( .+ . ) ] (12) ℎ = . () Considering that the objective of this work is to simulate modes of flight and control in cruise level flight, the state-space model linearized about the trim condition for the longitudinal dynamics is represented as follows.In other words, we solved the non-linear equations of the longitudinal dynamics for null derivatives with respect to the output and input state variables.The vector of states of the longitudinal dynamics ( 14) is formed by the airspeed, Euler angles, and altitude, which are the variables of interest for the chosen flight conditions.The control input vector (15) is composed of the deflection of the elevator control surface of the UAV plus the level position of the throttle.
The Jacobian of equations (8)(9)(10)(11)(12)(13) were derived and linearized at trim condition.Utilizing the state-space notation (16), we can rewrite the linearized system in the matrix form as described in section V.
= . + . (16) = . + . Where the vector of state and control of the linearized equation are: After performing the Jacobian, we obtained the linearized transfer function model and replaced all forces, moment, thrust force, the gravity acceleration (), and other variables that have been previously defined.
The detailed steps and calculations for the mathematical modeling of the longitudinal dynamics of a fixed wing UAV are described in [5].The analytical simulation was developed in a MATLAB language code.Through the datcomimport function, the stability and control derivatives of the model generated by DATCOM (Table II) are exported to MATLAB and introduced in the calculations of the coefficients and aerodynamic forces and moment.Then, all the values were replaced in the linearized equations of the longitudinal dynamics to obtain the state-space model.
V. RESULTS AND DISCUSSIONS
Based on the real RC Flying Wing specifications, the Flying Wing was successfully modeled in Plane-Maker environment, as illustrated in Fig. 2.Then, flight tests were performed in X-Plane as shown in Fig. 10, using the same flight conditions as on the analytical flight tests.
A model is based on approximate representation of some aspect of a real system, rather than exact values, therefore, its predictions are only as accurate as the data and assumptions used in its construction and operation.X-Plane separates the aircraft into small sections to calculate several aerodynamic factors on each, thus, the dimensioning of the reference model was fundamental to ensure that the entire aircraft was being computed.
X-Plane has proven to be a reasonably accurate tool in predicting the flight characteristics of an aircraft if its input parameters are correct.This is seen in the very close correlation of obtained flight test stall speeds to flight manual data, as well as the correlation of maximum level flight acceleration and speeds to flight manual data.
With the Flying Wing model in flight, it was necessary to adjust the PID Controller, in order to enter the cruise flight phase at a specific velocity and no roll attitude.Thus, once on a cruise flight, the aerodynamic coefficients, propulsive, and aerodynamic forces were obtained and could be compared directly with those calculated by the analytical method.This comparison is shown in Table III.It is important to emphasize that X-Plane considers the wing area as the wing area plus the area of the horizontal stabilizer, differently from the normal way adopted by the aerodynamic concept.Thus, this factor together with the fact that X-Plane also calculates the environment effects on the model, can be the reasons of the differences found.
Some advantages that X-Plane model has over the analytical model were found, such as a digital model may be more easily modified and recalibrated than the analytical model, which will allow increasingly accurate simulations as more data become available; the digital model should provide faster responses to complex maneuvers; and the digital model program has the capability to simulate changes in the aerodynamic performance resulting from environment effects.A series of tests were run on both the analytical and the digital models to compare solutions and to verify that the digital version was accurate and could produce the same results as the mathematical model.The results at trim conditions from both simulations were significantly close, which allow us to conclude that the digital prototype modeled in X-Plane is a reliable model of the Flying Wing aircraft enabling the implementation of HIL simulation as further studies.This method was also proved to be a solid tool for flight data creation, transmission, storage and processing, thus contributing significantly for the aeronautical industry.
The communication and processing flight data between MATLAB/Simulink and X-Plane required a lot of study in communication and information systems and revision of the content available in the literature.One of the biggest challenges was to understand the configuration needed to exchange data between the simulator and the external software (SIL).As well as platform response frequency compatibility for more realistic results.We obtained an acceptable integration between the X-Plane and Simulink platforms through UDP communication using a response frequency of 60.00 rate/s.
The aircraft has unstable modes in the longitudinal degree of freedom.These open loop instabilities make the aircraft challenging to fly manually.Fortunately, since the longitudinal dynamics is controllable, the unstable poles were shifted to the open left side of the s-plane during the design of the controller, ensuring closed-loop stability and enabling flight attitude control.In order to compare the analytical and X-Plane simulations, in-flight tests were performed using controllers to stabilize the unstable dynamic of the model enabling us to control necessary flight condition for the simulations, such as the altitude, velocity and roll attitude.
Therefore, the nonlinear equations of the longitudinal dynamic (8-13) were linearized at trim condition, as described below in S.I. units (19).From this, it was possible to rewrite the system in state-space notation (18).Thus, the control project was designed and optimized.The analytical simulation of the longitudinal dynamics of the aircraft when stimulated by unitary step showed a difference in the percentage of overshoot to the different amplitudes, according to Table IV.Whereas in X-Plane simulation the overshoot tends to decrease with the increase in amplitudes variation.The time until steady-state for both methods are relatively equal.Fig. 11 illustrates the comparison of the simulations described in Table IV, where the blue line represents the analytical response and the red line represents the X-Plane response.Note that the analytical method has smaller overshoots before stabilizing when compared with X-Plane method, which besides the overshoot it has also a small undershoot before stabilizing.The difference between the simulations can be justified by the different algorithms used by the platforms, being the MATLAB a purely numerical simulation, whereas X-Plane, being a flight simulator, that simulates the dynamics of the model in a more realistic way; as well as the difficulties in achieving perfect cruise flight at a constant speed in flight simulations.
CONCLUSION AND FUTURE WORK
Both platforms used for this project are great tools for simulation of early aircraft models to obtain flight data in different conditions, facilitating a prediction of its behavior and performance during the flight before manufacturing a prototype.The flight simulator X-Plane along with MATLAB/Simulink have shown high confidence level and reliability in modeling and simulating, providing strong expandability for flight data creation, collection and processing, and wide application in aircraft project development.
The mathematical model and its analytical simulations have validated the digital model, showing it is consistent and reliable.The model created in Plane-Maker opens up numerous possibilities for applications, enabling in depth studies of Flying Wings Aircraft design, also the testing of several techniques for control and navigation algorithms.Additionally, this simulation platform is a great tool for data generation for embedded systems, software data processing, and storage and simulation of flight recorders.
Developing both a simulation and an analytical model is a powerful method of getting answers to different sets of questions most effectively through one approach vs. the other.
The next development step, which will be tackled in future work, is the Hardware In The Loop (HIL) simulation and other control laws, which in turn, would receive the state variables generated from the simulator and calculate the control signals injected into the simulated plant at different flight conditions.The HIL simulator system simulates the aircraft characteristics, sensor and actuator modeling while communicating with the autopilot hardware.
The acquisition of data from the modeling of an aircraft, and its flight simulation as well as its data processing, could lead to a disruptive approach in aircraft design, potentially reducing time and costs necessary to manufacture a final prototype.
TABLE III COMPARISON
OF FORCES AND AERODYNAMIC COEFFICIENTS FROM BOTH
TABLE IV ANALYSIS
OF STEP RESPONSE OF DIFFERENT AMPLITUDES SIMULATING PITCH ANGLE VARIATION.RESULTS OF ANALYTICAL AND X-PLANE SIMULATIONS. | 6,414.8 | 2019-02-03T00:00:00.000 | [
"Computer Science"
] |
DYNAMICS OF SINGLE-SPECIES POPULATION-GROWTH - EXPERIMENTAL AND STATISTICAL-ANALYSIS
Abstract The logistic model, widely used for describing population growth, assumes that the per-capita rate of growth linearly decreases as the population size increases. Experimental data, however, suggest that often the per-capita rate of growth is not linearly related to population density. The theta model removes such linearity assumption by means of an additional parameter, θ; when θ = 1, the theta model reduces to the logistic model. We advance a method, the “jackknife” statistic, for estimating the rate of population growth (the largest eigenvalue and its variance) in the serial transfer system. Also, we propose a statistical method, PRESS, for quantifying the success of a given model in fitting experimental data. The criterion of success is the ability of a model to predict accurately new observations. One advantage of PRESS is that, contrary to what happens with other statistics such as R 2 , it tends to make a model less successful as the number of parameters increases (unless there is a disproportionate decrease in the bias of the new model). We have studied the rate of population growth in 25 genetically different populations of Drosophila melanogaster . The theta model provides a consistently better description of population growth in these populations than the logistic model. Moreover, the results indicate that the rate of growth is affected by the genetic constitution of a population.
INTRODUCTION
A widely used model describing population growth in a single species is the logistic model, proposed by Lotka (1924) and Volterra (1931). It predicts the rate of population growth as dN/dt = rN( 1 -N/K), where N is the population size, r is the approximate per-capita rate of increase achieved at low densities, and K is the carrying capacity or equilibrium population size. The model predicts that the per-capita rate of increase (N-'dN/dt) decreases from near r to 0 in a linear fashion as the population size increases; that is, the increase in intraspecific competition due to the addition of a new member is the same whether the total population size is small or large.
This linearity assumption can be removed in a variety of ways. For example, Schoener (1978) has produced some simple models of singlespecies population growth which include exploitative competition and interference competition. These models predict a faster-than-linear decline in percapita rate of increase. Moreover, there is empirical evidence suggesting that the linearity assumption of the logistic model may often be violated (Smith, 1963;Ayala et al., 1973).
More complicated versions of the logistic model are of course. possible, obtained by incorporating additional parameters into the logistic model (1). However, biologists are interested in simple models that capture the essential features of biological phenomena. Consequently, before the logistic model is ,abandoned, the benefits of more complicated models must be quantified in some meaningful way. Previous efforts have not always done this.
In this paper we ask whether relaxing the linearity assumption of the logistic model can yield a new model which can more accurately predict the dynamics of laboratory populations of Drosophila melanogaster. The robustness of the results is examined by appropriate statistical tests in a large number of independent populations. The experimental populations differ in their genetic constitution; hence, we also explore the possible contribution of genetic differences to population dynamics. A statistical test is described that determines whether the parameters estimated from the population growth equations are sensitive to the genetic constitution of populations. This is important because most theories of density-dependent selection assume that there is genetic variation in natural population for population parameters such as r and K of the logistic model (1).
A General Model of the Serial Transfer System
The Serial Transfer System (STS) of population growth is outlined in Fig. 1 (top). It is a discrete system of growth that allows for overlapping generations in the adult population. Two versions of the STS are possible, known as Type-1 and Type-2 experiments. In Type-l experiments, the population grows until it reaches its carrying capacity, which is thereafter approximately maintained for the duration of the experiment. Type-2 experiments are used to determine the rate of population growth at a given density (Ayala et al., 1973).
The procedure used in Type-l experiments is described in Ayala (1965). Here, we shall describe in brief outline how the procedure is used with one- week intervals between transfers. The population consists at any one time of four cultures of various ages. At the time of census, the youngest culture is one week old and contains egg-laying adults that have survived for one week since they were introduced in the culture. The remaining three cultures are 2, 3, and 4 weeks old and contain larvae, pupae, and newly emerged adults. At the time of census, the surviving adults in the youngest culture are counted and transferred to a fresh culture. At the same time, the newly emerged adults in the other three cultures are counted and are also transferred to the same fresh culture to which the adults were transferred; the oldest culture is, then, discarded. This process is summarized in Fig. 1 by arrows indicating transfer of adults from the four old cultures to the one fresh culture. The adults introduced in the fresh culture are allowed to lay eggs for one week, when a new census is taken. The procedure is repeated at one-week intervals.
At the time of census, the total number of adults, N,, is a function of the number of adults present in the leading culture 1, 2, 3, and 4 weeks ago. This relationship can be summarized as N,=f,Wt-,I +fi(N,-2) +fANt-3) +.fi(N+J, where fi(Nlei) is some unknown function that relates the number of adults that will emerge from an i-weeks-old culture with the number of adults that initially laid eggs in that culture. The exact form of these functions is of little general interest; as we shall show below, the analysis of population growth rates can proceed by an appropriate linearization of this model.
Experimental Determination of Rates of Population Growth at a Single Dens&
In order to investigate the usefulness of models such as (1) we need to obtain estimates of population growth rates at a variety of densities. This sort of information can be obtained by STS Type-2 experiments as shown in Fig. 1 (bottom) (Ayala et al., 1973). A specified number of adults, N*, are initially placed in a fresh culture. The suvivors are counted one week later and the emerging adults from this same culture over the following 3 weeks are recorded at one-week intervals. Each experiment of this kind yields the vector (Y,(N*), Yz(N*), Y,(N*), Y,(N*)), where Y,(N*) =fi(N*) + si, and ei is a random variable reflecting experimental error. Thus Y,(P) is the observed number of adults emerging (or surviving in the case of Y,(N*)) from an i-weeks old culture of egg laying by N adults during one week. Type-2 experiments can be repeated for multiple independent estimates of a given h(N*) and can be carried out at various densities.
In order to estimate the rate of population growth in the vicinity of N*. we look at a linear version of (2). N,=a,N,-, +azN,-2 +a,Nrp, +a,N,-,.
where each a, is a constant per-capita output of an i-weeks-old culture. Equation ( where the c's are constants which can only be determined if the four initial conditions are specified. Unfortunately there are no obvious initial conditions that can be used to obtain an explicit solution to (3). Moreover, these experiments are done at a variety of densities and it does not seem reasonable that the initial conditions used at one density will be compatible with the initial conditions at any other density. However, as t gets large in (4) a per-capita rate of population growth is obtained that is independent of the initial conditions. In such case the approximate per-capita rate of increase will be given by the one positive eigenvalue of (3), N,/N,_, IT 1. It is then possible to determine the weekly change in population size as a function of N*, as dN/At(N*) = (AN*) -N* = N*(A -1). Smith (1963) encountered a similar situation; i.e., where the initial conditions determine the outcome. In his experiments growing populations of Daphnia were maintained until a stable age distribution was attained. The relationship between density and rate of population growth was sensitive to the initial condition used to start the experiment. Repeatable results were obtained only when populations had attained a stable age distribution.
Statistical Estimation of Rate of Population Growth
If the Type-2 experiments (Section 2.2) are carried out m times at a density N*, then these m observations, [ Y'," can be used to estimate the a's in Eq. (3). Our procedure will be first to use all the observations to estimate the a's in (3) and then determine the largest eigenvalue of the resulting equation. It is also possible to estimate the largest eigenvalue for each observation and take the average of these m values. The results obtained by these two methods need not be the same. Given that the eigenvalues of (3) are functions of the population quantities ai, it seems more appropriate to estimate first the a,'s and then use these for determining the eigenvalues.
Therefore, we will estimate each ai by ii = l/m E=, q!'fN*, i = 1, 2, 3,4, where the argument N* has been deleted for simplicity. This yields one difference equation, N, = a^, N,-, + L&N~-~ + c?~N,-, + BjN,--,, from which an estimate of the largest eigenvalue 1 is obtained. There is, unfortunately, no simple way to estimate the variance of 1. One approximate solution is the jackknife statistic (for a review see Miller, 1974; for applications to population genetics see Mueller, 1979).
In order to calculate the jackknife statistic. the jth set of observations is deleted and the largest eigenvalue (as described in the previous paragraph) is calculated using the remaining data. This yields a new value x-j. One can then calculate m pseudovalues as sj=mX-(m-l)X-j, j = 1, 2 ,..., m.
The jackknifed estimate of the largest eigenvalue is simply the mean of these pseudovalues; 2 = (l/m) Cj sj. The variance of this eigenvalue is estimated by Var(X) = (l/m(m -1)) cj (sj -I)'. The pseudovalues may also be used to estimate m values of AN/At as N*(sj -1), j = 1, 2,..., m. These values of AN/At are necessary for the regression analysis described in the next section.
PRESS
One of our aims is to compare different models that predict the change in population size as a function of density. As mentioned in the Introduction, it is important to quantify the success of each model in some meaningful way. Two widely used statistics that measure how well a regression function fits a set of data are the proportion of variance explained by the model (R2) and the mean sum of squares (l/n RSS). R* will increase and (l/n) RSS will decrease as more complex elaborations of some basic model are examined. In the limit, if there are n + 1 observations it is always possible to obtain an nth order polynomial which passes exactly through all n + 1 points. This polynomial yields R* = 1 and (l/n)RSS = 0. Given such direct correlation between the number of parameters and goodness of lit, it is not clear what the best model might be. It is obviously necessary to establish criteria for deciding which model is best.
For the present purposes, we consider the best model that model that can predict new observations most accurately. The statistic PRESS (Prediction Sum of Squares; Allen, 1971a) provides a means for quantifying this property. We will describe this more formally. Consider linear models where p is the vector of parameters of the regression function (e.g., r and K of (1)). Similar arguments can be made for nonlinear models also. Suppose we employ the following partitioning of p = (p,, P2)T. Now let p be the leastsquares estimate of p and p be the least-squares under the assumption p2 = 0. If h is a vector of independent variables then we have two estimators of k'fi available: L'p and hrp. It can be shown (Walls and Weeks, 1969) that Var(Arp) > Var(S'p) whether or not p2 = 0. Obviously if p2 # 0 then hTB will be biased. Thus the following dilemma exists: the use of k'p as an estimate of L'p will have a small variance but may be biased, while use of Lrg will reduce this bias at a cost of increased variance. The problem is to strike a balance between variance and bias. In such a circumstance it may be best to use the estimator with the smallest mean-squarred error since this equals the variance plus the bias squared. PRESS provides an estimate of the mean-squared error. Let g(N, fi) re p resent the regression function, where N is the observed density and p is the vector of estimated parameters: then. PRESS is defined for a sample of n observations as and DC-i) has been estimated from the n -1 observations (U"', N(I)),..., (CY-", N"-"), (Ui"', N"+"),..., (U'"), N'"'). According to the present criterion, the best model is the one yielding the smallest value of PRESS.
Variation for Regression Parameters
Suppose a regression model has two parameters which must be estimated; that is, p = @,, p,). For model (l), /3, corresponds to r and p2 to K. In addition, there are I independent populations, for which p (i.e., p(I),..., p(l)) has been estimated. The question arises whether the 1 values of 0, and B, show significant heterogeneity. In our study, the I populations represent populations that are genetically different (each is homozygous for a different second chromosome). If we estimate r and K from Eq. (1) for each of these populations, the question is whether the genetic constitution of these populations has a significant effect on the estimated values of r and K. For linear models such as (I), an analysis of covariance can be used to answer this question. Analogous methods do not exist for nonlinear models. Since we shall be considering, in addition to the logistic model, a nonlinear model of population growth, we will adopt a method that is applicable to both linear and nonlinear models.
Suppose we examine the variation of parameter 8, over the 1 populations. The arguments that follow are unchanged when parameters other than 8, are considered; hence, we will drop the subscript 1 in the following discussion. Let the vector of these I parameters be p = (p('),...,/?(")r. For each pCi' we also have an estimated variance Sf . We assume that p has an multivariate normal distribution with parameters (p, C), where C = diag(ai,..., uf). Let 6 be some linear combination of the ps, C? = cf=, /z,B'~' = hrb. Using the methods of Gold (1963) and Goodman (1964), it can be shown that for all h', P[C; -L < u < C? + L ] > y, where L = fl-and P(xf < c) = y. Thus this method generates lOOy% simultaneous confidence intervals.
To determine whether there are significant differences between the ,8 values, the order statistics of the /.?, have been divided below into three groups. The choice of three is totally arbitrary. If a 95% simultaneous confidence interval on the difference between the means of any of these groups does not include 0, we conclude the ps are not homogeneous. This simultaneous inference scheme allows one to "hunt" for contrasts among the $s that will yield significant differences and still claim that this inference is being made at the 1007% confidence level.
THE MODELS
In addition to the logistic model we will estimate the parameters and calculate PRESS for the theta model (Ayala et al., 1973;Gilpin and Ayala, 1973): There are several reasons for examining this model. Previous results (Ayala et al., 1973) indicate that interspecific competition can be modeled best, when using the value of RZ as a criterion, with analogs of (5). Moreover, the theta model includes the logistic model as a special case, i.e., when 8= 1. Since our major interest is in relaxing the linearity assumption of (l), the theta model is a good option, because it allows for inflection points both less than and greater than K/2. The recent models of Schoener (1978) have inflection points always less than K/2. Only one linear regression need be performed in order to estimate PRESS (Allen,197 lb) for linear models such as (1). Nonlinear models, however, require repeated estimation of e -.
' " If each population has n observations, then calculation of PRESS for one population requires n iterations of the nonlinear estimation procedure. This is prohibitively time consuming. The number of computations can be cut in half by deleting two observations, (u"', N"'), (U(j), IV"'), at a time; estimating the new vector of parameters p'-"' on this set of n -2 observations; and then computing [tii) -g(N"', (jc-ii')]z, [U(j) -g(N"', /?I-"')]'. The bias, if any, introduced by this procedure would be to generate larger values of PRESS simply because less information is available for estimating p(-ii' than for estimating pCmi). This bias will only be present in estimates of PRESS for the theta model. Despite this bias, the theta model has a smaller value of PRESS most often and, hence, our conclusions will be unaffected by this bias.
In this study, the parameters of (5) are estimated using the algorithm of Marquardt (1963). Marquardt's algorithm uses a ridge-regression improvement at each iteration of the algorithm. This is particularly useful when there is a high degree of correlation between the parameters. Problems that occur when parameters are highly correlated are (a) round-off errors in the numerical procedure used may lead to inaccurate results; and (b) the estimated parameters have a very large variance. Ridge estimators are useful in combating these problems (Marquardt and Snee, 1975).
The results presented in Section 6 show that r and 0 are negatively correlated. This correlation is a consequence of the following. At each step of the nonlinear routine, the following linear regression problem is solved (Gallant, 1975 At the fth iteration, The next value of b is given by the standard solution to the above problem, fl,,, = (xTx)-'x'z.
This close correspondence between entries in the first and third columns of x causes r and 19 to be highly correlated.
EXPERIMENTAL TESTS
Drosophila melanogaster flies were collected at Strawberry Canyon, Berkeley, California. Crosses of individual wild males with males with balancer stocks produced a number of lines, each homozygous for the complete second chromosome (see Tracey and Ayala, 1974). A total of 24 nonlethal and nonsterile lines were selected for Type-2 experiments in order to measure the rate of population growth in each line. Type-2 experiments were also done with a random-heterozygous line (H) used as a standard or reference. The H line was produced by placing five virgin females and five males from each homozygous line in each of 10 cultures. These flies were allowed to mate at random and lay eggs for the next 5 days. F, progenies emerging from these cultures were placed in fresh cultures at the various densities (10, 50, 100. 250, 500, 750, and 1000) used in the experiments. The Fz progenies produced in these cultures were used to start Type-2 experiments. This procedure was repeated three times; i.e., each time that a new set of replicates was started with the homozygous lines.
As described in Section 2.2, N* adult flies were used to start each experiment. For each line, N* took on values of 10, 20, 50, 100, 250, 500, 750, and 1000. Six replicates were made for each homozygous line at each density, except at 1000 for which only three replicates were done. Exceptions were line 45, which had only two replicates at density 1000; and line 36, which had no observations at 1000 and only three at 750. The H line was replicated 12 times at each density, except 1000 at which only three replicates were made. Two replicates of each homozygous line (and four of the H line) were started simultaneously, so that a total of three sets of replicates were started at different times. The variance between replicates started at different times was always much greater than the variance between replicates started at the same time. Most of this increased variance is probably due to differences between batches of the culture medium, although variations in incubator temperature and other environmental variables may have also contributed to this variance.
All experiments were performed in 237 cc (half-pint) cultures with 40 CC of a standard cornmeal-molasses-agar medium. The cultures were kept at 23°C and ca. 70% relative humidity. All adults were between 7 and 14 days old at the start of each experiment, and had been raised at the same density condition as the density used in the experiment. An equal number of males and females were placed in each culture, in order to standardize the experimental conditions over all lines.
RESULTS
Tables I and II give the estimated values of the parameters for the logistic (1) and the theta (5) model for each homozygous line as well as for the random heterozygous (H) line. Standard errors and correlation coefficients are also included in these tables. As mentioned earlier, the strong negative correlation observed for the theta model between r and 8 is expected in view of the design matrix used in the nonlinear analysis. It should also be noted that there is, in general, good agreement between the estimates of K from models (1) and (5), but that the values of r are drastically different. Large differences between lines with respect to the parameter values are apparent in Tables I and II. Multiple contrasts between groups of homozygous lines obtained by the methods described in Section 3.2, are shown in Figs. 2, 3, and 4. Significant differences between homozygous lines exist for all parameters in both models. We believe that these differences are real phenomena, largely due to genetic differences among the populations. Table III shows the values of PRESS for each line according to each ,6,8,14,IS. 18,20. 25,30,37,SO;II = 9,13. 23. 33. 40,42,45;III = 2,3,7. 36,43. 52. In the case of K. the multiple contrasts are between I = I, 2. 6. 8. 9,14. 18. 20. 42;II = 3,IS. 23. 25,33. 37. 40,SO: III = 7. 13. 30,36,43. 45. 52. = 2,9. 14,25. 33,43,52;II = I. 6. 8,IS. 18,37. 50: 111 = 3,7. 13. 20. 23. 30. 36,40. 42. 45. In the case of K. the populations compared are the same as for K in Fig. 2. FIG. 4. Mean difference (with 95% confidence interval) between groups of populations for the parameter ~9 of the theta model. The populations in each group are the same as for r in Fig. 3.
DISCUSSION
The results presented in Section 6 show that relaxing the linearity assumption of the logistic model (l), as it is done by Eq. (5), yields a model that predicts new observations more accurately. The generality of this conclusion remains unknown, although nonlinear departures from the logistic model may be a common phenomenon (e.g., Smith, 1963;Ayala et al., 1973;Schoener, 1973;Thomas et al., 1980;Serradilla, 1979).
The inflection points dN/dt relative to N in all our experimental populations are at value of N smaller than K/2. Other detailed studies of population dynamics in Duphnia (Smith, 1963) and Drosophila (Thomas et al.. 1980;Seradilla, 1979) yield similar results. Schoener (1973) has made a crude estimate of the position of the inflection point for various species and has concluded that inflection points greater than K/2 are more common. His analysis includes two Drosophila populations with estimated inflection points greater than K/2. Given the high level of approximation used in Schoener's calculations, it would seem that a general conclusion cannot be reached until more detailed analyses, such as done in the present study, are available.
The desirable properties of mathematical models include generality and precision (Levins, 1966) as well as simplicity-i.e., having only the minimum number of necessary parameters (Ayala et al., 1973). Models should also have realism; that is, biological interpretation, in the case of population ecology. Hence, we are interested not only in a functional relationship, such as expressed in (5), that may describe "best" the nonlinearities observed in the experiments, but also in the biological phenomena responsible for the nonlinearities. Uncovering the underlying biological processes would make it possible to assess the applicability of models such as (5) to other organisms and to other populations. At present. only speculations are available. Schoener (1978) has shown that nonlinearity may simply result from the mechanism of feeding for a limited food supply. Gilpin et al. (1976) have suggested that nonlinearity may arise because quality resources are exhausted first in an environment with heterogeneous resources. A decision between these and other possible explanations must wait until experimental tests are performed addressed to ascertain the biological processes that account for nonlinearity in intraspecific competition.
There is one major reason why the values of r estimated from the theta and from the logistic model are quite different. In the logistic model, r determines the approximate rate of population growth at low densities. In the theta model, the rate of population growth is determined by the parameter t-0. When 0 is close to 0, the first two terms of the Taylor series expansion of (5) yield AN/At = rON ln(K/N). Thus if one compares r@ values estimated from the theta model and r values from the logistic, one sees similar predictions of the growth rate of the populations at low densities.
This study has demonstrated that the growth dynamics of single-species populations are sensitive to the genetic constitution of populations. We plan to explore in future papers the relationships between genetic variation, Darwinian fitness, and density-dependent selection. | 6,025.6 | 1981-08-01T00:00:00.000 | [
"Biology",
"Environmental Science",
"Mathematics"
] |
Sonic hedgehog signaling promotes angiogenesis of endothelial progenitor cells to improve pressure ulcers healing by PI3K/AKT/eNOS signaling
Background: Pressure ulcer is a severe disease in the paralyzed and aging populations. Endothelial progenitor cells (EPCs) are able to regulate ulcer healing by modulating angiogenesis, but the molecular mechanism is still obscure. Sonic hedgehog (SHH) signaling contributes to angiogenesis in various diseases and has been identified to modulate EPCs function. Here, we aimed to explore the significance of SHH signaling in EPCs function during pressure ulcers. Methods: The EPCs were isolated and characterized by the expression of DiI-acLDL and bind fluorescein iso-thiocyanate UEA-1. Cell proliferation was detected by cell counting kit 8 (CCK-8). The DiI-acLDL and bind fluorescein iso-thiocyanate UEA-1 were analyzed by immunofluorescent analysis. The angiogenesis of EPCs was analyzed by tube formation assay. The pressure ulcers rat model was constructed, the wound injury was analyzed by H&E staining and angiogenesis was analyzed by the accumulation of CD31 based on immunofluorescent analysis. Results: The expression of patched-1 and Gli-1 was enhanced by SHH activator SAG but reduced by SHH inhibitor cyclopamine in the EPCsThe PI3K, Akt, eNOS expression and the Akt phosphorylation were induced by SAG, while the treatment of cyclopamine presented a reversed result. The proliferation and migration of EPCs were enhanced by SAG but repressed by cyclopamine or PI3K/AKT/eNOS signaling inhibitor Y294002, in which the co-treatment of Y294002 could reverse the effect of SAG. Conclusions: Thus, we found that SHH signaling activated angiogenesis properties of EPCs to improve pressure ulcers healing by PI3K/AKT/eNOS signaling. SHH signaling may serve as the potential target for attenuating pressure ulcers.
angiogenesis [3,4].Meanwhile, endothelial progenitor cells (EPCs), which originated from the bone marrow upon acute injury, participate in the neovascularization [5].The fluid EPCs numbers are applied as a marker of endothelial disorder or remedy several diseases [6].Previous investigations have identified the crucial role of EPCs in diabetic ulcer healing by modulating angiogenesis [7,8].Importantly, it has been found that EPCs can improve pressure ulcer healing [9,10].However, the molecular mechanism underlying EPCsmediated ulcers remains elusive.
The Hedgehog pathway is crucial for precise morphogenesis and the formation of embryogenesis [11].The interplay of the hedgehog pathway and its receptor, Patched-1 (Ptch), results in activating transcription factor Gli, inducing its target gene expression [12,13].As an essential type of hedgehog pathway, sonic hedgehog (SHH) signaling plays a critical function in the formation of blood vessels.SHH signaling activation leads to neuroectoderm hypervascularization [14], and induces postnatal neovascularization [15].Exogenous SHH signaling can also actively enhance angiogenesis by principally regulating fibroblasts through activating angiopoietin-1 and VEGF [16].SHH treatment may significantly improve wound healing in diabetes by restoring nerve function and stimulating arteriogenesis [17].It also has been shown that SHH signaling is able to improve the role of diabetic EPCs [18,19].Moreover, PI3K/AKT signaling plays critical functions in EPCs senescence and proliferation, and previous studies have identified that the close correlation of SHH signaling with PI3K/AKT signaling [20][21][22].However, the association of SHH signaling with PI3K/AKT signaling in the regulation EPCs function and its role during pressure ulcer are still unclear.
In this study, we were interested in the role and the underlying mechanism of SHH signaling in regulating angiogenesis properties of EPCs during pressure ulcers.We identified a crucial function of SHH signaling in promoting angiogenesis properties of EPCs to improve pressure ulcers healing by PI3K/AKT/eNOS signaling.
EPCs isolation and treatment
EPCs were obtained from mononuclear cells (MNCs) as the previous reports [23].MNCs were discontinued into -complement medium (Gibco, USA) with endothelial growth factors (Sigma-Aldrich, USA), 10% FBS (Gibco, USA) and seeded on fibronectin (50 μg/mL, Sigma-Aldrich, USA)-pre-coated 6-well dishes (Sigma-Aldrich, USA).Cells were cultured at the condition of 5% CO2 and 37° C. Dead cells were removed after three days, and then the culture mediums were refreshed every three days.And the EPCs were identified and characterized by DiI-acLDL and bind fluorescein iso-thiocyanate UEA-1 using immunofluorescent analysis after two weeks.The SAG, cyclopamine, and Y294002 were purchased (Sigma-Aldrich, USA).
CCK-8 assays
The proliferation was assessed using CCK-8 assays.About 1×10 3 cells were plated in 96-well dishes and incubated for the transfection or treatment.The cells were added with a CCK-8 solution (KeyGEN Biotech, Nanjing, China) and culture for another 2 hours at 37° C. The proliferation was measured at an absorbance of 450nm by applying the ELISA browser (Bio-Tek EL 800, USA).
Transwell assays
Transwell assays analyzed the migration of EPCs by using a Transwell plate (Corning, USA) according to the manufacturer's instruction.Briefly, the upper chambers were plated with around 1 × 10 5 cells.Then solidified through 4% paraformaldehyde and dyed with crystal violet.The invaded and migrated cells were recorded and calculated.
Tube formation assays
The angiogenic capacity was analyzed by tube formation assays (BD, USA).The 24-well plates were coated with 100 µl Matrigel (BD Bioscience, USA) and incubated at 37° C for 2 hours.After the gel was solidified, HUVEC was suspended in culture medium as single cells and were plated onto the Matrigel in 24-well dishes.After incubation at 37° C for 24 hours, the formed tubes were captured by microscopy and quantified by ImageJ software.
Western blot analysis
RIPA buffer (CST, USA) was used to extract the total protein, followed by the quantification based on the BCA method (Abbkine, USA).The proteins at same concentration were subjected in SDS-PAGE and transferred (PVDF, Millipore, USA), followed by the incubation with 5% milk and with the primary antibodies at 4° C overnight.The corresponding second HRP-conjugated anti-mouse or anti-rabbit antibodies (Boster, Wuhan, China) were used for incubating the membranes 1 hour at room temperature, followed by the visualization by using chemiluminescence detection kit (Beyotime, Shanghai, China).The primary antibodies applied in this study comprise of patched-1 (Affinity, USA), Gli-1 (Affinity, USA), PI3K (Affinity, USA), Akt (Affinity, USA), eNOS (Affinity, USA), p-Akt (Affinity, USA), and β-actin (Abcam, USA).All antibodies were diluted in PBST solution at 1:2000 (v/v).
Pressure ulcers rat model
Sprague-Dawley (SD) rats (10-weeks old; 0.15-0.2Kg) were applied to construct the pressure ulcers rat model [24].The rats were performed anesthetization by intraperitoneally injecting pentobarbital (50 mg/kg).The dorsal hair was shaved and the area was cleaned using alcohol (75%).About 3-cm-full-thickness skin incisions were executed, and the autoclaved magnet disk was made on the areas.The incisions were filled with 4/0 size polysorbate sutures.After a regular two hours of clamping with magnet disk on the areas, the outer magnet was eliminated for 0.5 hours, and the removal/clamping recycles were repeated five times/day for 5 days.
Immunofluorescence analysis
The expression of CD31 was analyzed by immunofluorescence analysis.Slices were solidified at 4% paraformaldehyde for 30 min, treated with Triton X 100 (0.2%) for 10 min and treated with BSA (2%) for 30 minutes.The slides were hatched with the primary antibody overnight at 4° C, then hatched with secondary antibodies (Proteintech, Wuhan, China) for 1 hour at 37° C. The slides were stained with the Hoechst (Beyotime, Shanghai, China) for 10 min at 25° C. The Nikon microscope (Tokyo, Japan) was utilized to analyze the immunofluorescence.
Histological and immunohistochemical analyses
The slices of skin tissues (5 μm thick) were analyzed by Hematoxylin and eosin (H&E) staining.The photographs were captured by an Olympus BX60 microscope (Olympus Optical, Tokyo, Japan) at a magnification x200.The quantitative analysis was performed through a quantitative digital image analysis system (Image-Pro Plus 6.0).
Statistical analysis
Data were expressed as mean ± S.D of three independent experiments, and the statistical analysis was conducted by GraphPad Prism 7 (GraphPad Software, USA).The unpaired Student's t-test was used for comparing two groups when data confer to parametric distribution.For the analysis of datasets with non-parametric distribution, the Mann-Whitney U test was used for comparisons between two groups.p < 0.05 were considered as statistically significant.
SHH pathway regulates PI3K/AKT/eNOS signaling in EPCs
To understand the function of SHH and the underlying mechanism in EPCs, the EPCs were isolated and characterized by the expression of DiI-acLDL and bind fluorescein iso-thiocyanate UEA-1 using the using the immunofluorescent analysis (Figure 1A).Then, to evaluate the role of SHH, the EPCs were treated with the SHH activator SAG or the SHH inhibitor cyclopamine.Significantly, the expression of patched-1 and Gli-1 were enhanced by SAG but reduced by cyclopamine in the EPCs (Figure 1B).Meanwhile, the PI3K, Akt, eNOS expression and the Akt phosphorylation were induced by SAG, while the treatment of cyclopamine demonstrated a reversed effect (Figure 1B), indicating that SHH activates PI3K/AKT/eNOS signaling in EPCs.
SHH pathway enhances proliferation and migration of EPCs by PI3K/AKT/eNOS signaling
We then observed that the proliferation of EPCs was enhanced by SAG and repressed by cyclopamine or PI3K/AKT/eNOS signaling inhibitor Y294002, in which the co-treatment of Y294002 could reverse SAGmediated EPC proliferation in the system (Figure 2A).Similarly, the treatment of SAG induced but cyclopamine and Y294002 suppressed the migration of EPCs, while the co-treatment of Y294002 was able to inhibit SAG-regulated migration of EPCs (Figure 2B).
SHH signaling induces angiogenesis properties of EPCs by PI3K/AKT/eNOS signaling
Next, we further investigated the impact of SHH/PI3K/ AKT/eNOS signaling on the angiogenesis properties of EPCs.Remarkably, tube formation assays showed that the tube length was promoted by SAG but attenuated by cyclopamine and Y294002, in which the co-treatment of Y294002 significantly reversed the effect of SAG in the EPCs (Figure 3).
SHH pathway improves pressure ulcers healing in the rat model
Next, we assess the function of SHH signaling in the pressure ulcers rat model.Significantly, the treatment of EPCs attenuated the wound injury of the pressure ulcers rats, while the SAG treatment could enhance but cyclopamine treatment repressed this effect in the system (Figure 4A).We also observed that EPCs enhanced the expression of angiogenesis-related marker CD31 in the pressure ulcers rats, while the SAG AGING treatment could enhance but cyclopamine treatment repressed this effect in the system, suggesting that SHH pathway contributes to angiogenesis in the pressure ulcers rat model (Figure 4B).
DISCUSSION
Pressure ulcer is a severe disease and commonly affects the paralyzed and aging populations.EPCs contributes to the pressure ulcer healing, but the mechanism is still obscure.SHH signaling has been found to regulate EPCs function, but the role of SHH in pressure ulcer remains unclear.In this investigation, we found that SHH activates PI3K/AKT/eNOS signaling in EPCs.The treatment of SAG induced the proliferation, migration, and angiogenesis of EPCs, while the co-treatment of PI3K/AKT/eNOS signaling inhibitor Y294002 was able to inhibit these effects.Moreover, the results from in vivo pressure ulcers rat model demonstrated that SHH pathway contributes to angiogenesis in the pressure ulcers rat model.
Previous studies have found the critical role of EPCs in wound healing and angiogenesis.Simvastatin enhances EPCs neovascularization and mobilization to modulate diabetic rats wound healing [25].Strongly effective local treatment of EPCs remarkably activates full-thickness wound healing by stimulating angiogenesis [26].Meanwhile, SHH signaling contributes to the regulation of EPCs function.SHH signaling recoveries diabetic EPCs and contributes to cardiac repair in diabetic mouse model [18].SHH signaling promotes ischemiaassociated neovascularization through increasing EPCs function [27].SHH signaling stimulates VEGF production, migration, and proliferation of EPCs by PI3K/ AKT signaling [28,29].Furthermore, SHH signaling also presented the promising therapeutic potential in wound healing [30,31].In this study, we observed that SHH signaling enhanced proliferation and migration of EPCs and induced angiogenesis properties of EPCs.SHH signaling stimulated angiogenesis and improved pressure ulcers healing in the rat model.
Our data displays a critical role of SHH signaling in regulating angiogenesis of EPCs during pressure ulcers healing, uncovering the molecular mechanism of EPCsmediated ulcers healing.Moreover, PI3K/AKT/eNOS signaling presents important roles in EPCs modulation.Naringin promotes the tube formation and proliferation of EPCs by regulating the PI3K/AKT signaling [32].LncRNA WTAPP1 enhances angiogenesis and migration of EPCs by up-regulating MMP1 through MicroRNA-312/PI3K/Akt signaling [33].MicroRNA-9 stimulates angiogenesis of EPCs to promotes thrombi recanalization by targeting PI3K/ Akt/TRPM7 signaling [34].It has been identified the correlation of PI3K/AKT/eNOS signaling with SHH signaling.SHH signaling activation attenuates inflammation response to protect dopaminergic neurons by modulating PI3K/AKT signaling [35].SHH signaling promotes epithelial-mesenchymal transition by targeting PI3K/AKT signaling in ovarian cancer [36].In the present study, our data showed that PI3K/AKT/eNOS signaling was involved in SHH signaling-mediated proliferation, migration, and angiogenesis properties of EPCs.It presents a novel mechanism involving SHH signaling and PI3K/AKT/eNOS signaling in regulation of EPCs function.Nevertheless, it is possible that other regulatory signaling cascades exist in SHH-regulated angiogenesis during pressure ulcer.Further investigation is needed to identify detailed regulatory mechanisms including genetic and epigenetic regulators.
In summary, we concluded that SHH signaling activated angiogenesis properties of EPCs to improve pressure ulcers healing by PI3K/AKT/eNOS signaling.Our finding provides new insight into the mechanism by which SHH signaling contributes to the EPCs-mediated pressure ulcers healing.SHH signaling may be served as the potential targets for attenuating pressure ulcers.
Figure 4 .
Figure 4. SHH pathway improves pressure ulcers healing in the rat model.(A, B) The pressure ulcers rat model was constructed, and the rats were treated with EPCs, or SAG or cyclopamine-treated EPCs, respectively.(A) The wound injury was analyzed by H&E staining in the rats.The representative wound healing images were shown.N = 5. (B) The pressure ulcers rat model was constructed, and the rats were treated with EPCs, or SAG or cyclopamine-treated EPCs, respectively.The angiogenesis was analyzed by the accumulation of CD31 based on immunofluorescent analysis in the rats.N = 5. | 3,035.2 | 2023-10-09T00:00:00.000 | [
"Medicine",
"Biology"
] |
Dark Matter in the CP-violating NMSSM
In the Next-to-Minimal Supersymmetric Standard Model there is a strong correlation between the mass terms corresponding to the singlet Higgs and the singlino interaction states, both of which are proportional to the parameter $\kappa$. If this parameter is complex, explicit CP-violation occurs in the Higgs as well as the neutralino sectors of the model at the tree level, unlike in the minimal scenario. A small magnitude of $\kappa$ typically yields a $\cal{O}$(10) GeV lightest neutralino with a dominant singlino component. In such a scenario, the phase of $\kappa$, beside modifying the properties of the five Higgs bosons, can also have a crucial impact on the phenomenology of the neutralino dark matter. In this study we perform a first investigation of this impact on the relic abundance of the dark matter solutions with sub-100 GeV masses, obtained for parameter space configurations of the model that are consistent with a variety of current experimental data.
1 Introduction NMSSM, if the Higgs self-couplings, λ and/or κ, appearing in the superpotential are complex, the scalar and pseudoscalar interaction eigenstates mix together to give five neutral CP-indefinite Higgs states; see [66,67,68,69,70,71] for recent studies of the NMSSM Higgs sector with CP violation and [72] for a review. We henceforth refer to this model as the cNMSSM.
Several phenomenological scenarios emerging in the cNMSSM Higgs sector that are distinct from the NMSSM with real parameters (rNMSSM) have been studied in [73,74,75,76,77,78]. Importantly, the complex κ parameter associated with the singlet superfield also appears in the entry of the neutralino mass matrix that corresponds to the singlino weak eigenstate. The impact of a non-zero phase of κ on the phenomenology of the χ 0 1 DM has not been analysed in literature thus far. In this article, we take a first step in this direction, and investigate how the relic abundance of the χ 0 1 in the cNMSSM is affected by variations in this phase. We focus mainly on the (EW-scale) cNMSSM parameter space configurations that yield a sub-100 GeV DM, which can be predominantly singlino-like. We also test the consistency of these solutions with the most important latest experimental constraints, including the Higgs boson data from the LHC and the electron and neutron EDMs, and study some of their phenomenological implications.
The article is organised as follows. In the next section we briefly revisit the Higgs and neutralino sectors of the cNMSSM. Section 3 contains details of our numerical analysis of the model's parameter space with the focus on the DM observables. In section 4 we present the results of our analysis, and we summarise our findings in section 5.
The NMSSM with explicit CP-violation 2.1 The Higgs sector
The superpotential of the NMSSM is written as in terms of the singlet Higgs superfield, S, besides the two SU (2) L doublet superfields, of the MSSM. The above superpotential observes a discrete Z 3 symmetry, which is imposed in order to explicitly break the dangerous U (1) P Q symmetry, and renders it conformal-invariant by forbidding the µ H u H d term present in the MSSM superpotential. Here, the mixing between the H 0 d field and the H 0 u fields, necessary for each of them having a non-trivial vacuum expectation value (VeV) at the minimum of the potential, is instead generated by the λ S H u H d term. This results in a dynamic µ eff ≡ λs/ √ 2 term when the singlet field acquires a VEV, s, naturally near the SUSY-breaking scale.
The tree-level Higgs potential of the NMSSM is obtained as where g 1 and g 2 are the U (1) Y and SU (2) L gauge couplings. It is customary to define the trilinears proportional to the superpotential couplings as A λ and A κ above are the soft SUSY-breaking counterparts of the superpotential couplings, and all of these can very well be complex parameters, with the corresponding phases, e iφ A λ , e iφ Aκ , e iφ λ , and e iφκ . After spontaneous EW symmetry breaking, V 0 is evaluated at the vacuum, in terms of fields defined around their respective VEVs, v u , v d and s, as The potential then contains the phase combinations However, assuming vanishing spontaneous phases θ and ϕ, the last two phase combinations above can be determined up to a twofold ambiguity using the minimisation conditions of V 0 , leaving φ λ − φ κ as the only physical CP phase (see [75] for more details). The potential V 0 with complex phases leads to a 5 × 5 Higgs mass matrix, M 2 0 , in the H T = (H dR , H uR , S R , H I , S I ) basis, with the massless Nambu-Goldstone mode rotated away. After including the higher order corrections from various sectors of the model [73,75,79], the resulting Higgs mass matrix, M 2 H = M 2 0 + ∆M 2 , is diagonalised using an orthogonal matrix, O, as The masses of the five CP-mixed physical Higgs bosons thus obtained are ordered such that m 2
The Neutralino sector
As noted in the Introduction, the fermion component of S, called the singlino, mixes with the neutral gauginos, B 0 and W 0 3 , and higgsinos, H 0 d and H 0 u , to yield five neutralinos in the NMSSM. The symmetric neutralino mass matrix in the gauge eigenstate basis, , is written as with m W and θ W being the W -boson mass and the weak mixing angle, respectively. The neutralino masses and compositions at the tree level thus depend on the Higgs-sector parameters λ, κ, µ eff , v u , v d and the gaugino masses M 1 and M 2 . When any of these parameters is complex, the mass matrix in eq. (6) can be diagonalised by a unitary matrix N , to give D = diag(m χ 0 The neutralino mass eigenstates are then given by χ 0 i = N ij ψ 0 j , and are again ordered 1 , which is the lightest neutral SUSY particle and hence a DM candidate, is given by the linear combination Thus, the relative sizes of the soft gaugino masses M 1,2 , the µ eff -parameter and the κs term determine whether the χ 0 1 is gaugino-, higgsino-or singlino-like. For example, in the limit µ eff min[M 1 , M 2 ], the term [M χ 0 ] 55 = 2κs = 2 κµ eff λ in eq. (6) results in a singlino-dominated χ 0 1 for 2κ/λ < 1. Importantly, in SUSY models the charged higgsinos ( H + u and H − d ) and winos ( W + and W − ) also mix to form the chargino eigenstates, χ ± a (a = 1, 2). The mass matrix for the charginos is given by This implies that M 2 and µ eff have a lower bound of about 100 GeV, owing to the non-observation of a chargino at the LEP collider. Since no such constraint exists on the M 1 and 2κs terms, the χ 0 1 in the NMSSM can have a mass much lower than 100 GeV as long as it is predominantly binoand/or singlino-like. The presence of a certain amount of higgsino is, however, necessary to obtain a realistic relic abundance.
The existence of the singlino in the NMSSM, even in the absence of the CP-violating phases noted above, leads to some unique possibilities in the context of DM phenomenology, compared to the MSSM. In the limit of large tan β ≡ v u /v d and large m A ≡ λs sin 2β ( √ 2A λ + ks) (which effectively decouples the doublet-like H 3 and A 2 from the rest of the particle spectrum, so that m A 2 m A ), the masses of the two lightest CP-even scalars can be approximated by [80] where v ≡ v 2 u + v 2 d . Thus the mass of the lighter of these two (when the heavier one is required to be the H SM ) scales with κs, as does that of the singlino. At the same time, the mass-squared of the lighter pseudoscalar, which is almost purely a singlet, reduces to This correlation between the masses of the A 1 and the χ 0 1 implies that they can be naturally close to each other, thus opening the possibility of the former's self-annihilation via the latter. Evidently, while H 1 can also have a mass in the vicinity of m A 1 , it is more strongly constrained by Eq. (9) from taking values close to 2 χ 0 1 . Even if it does acquire the correct mass, the χ 0 1 annihilation via s-channel H 1 is p-wave suppressed, which would make its consistency with the thermal relic abundance difficult.
When the CP-violating phases of κ and λ are turned on, they enter the tree-level neutralino mass matrix independently of each other, unlike the combination φ λ − φ κ of the Higgs sector. In addition, M 1 and M 2 can also be complex parameters, which would be radiatively induced into the Higgs sector at higher orders. Given the composition of the χ 0 1 , the size(s) of the most relevant phase(s) would then affect not only its physical mass, but also its interaction strengths with other particles. Here, our focus on a sub-100 GeV DM, which is also preferably singlino-like (since lowmass bino-like solutions exist in the MSSM too and have been extensively studied), makes φ κ the most obvious choice to investigate the impact of.
The electric dipole moments of fermions
Beyond the Born approximation, various CP-violating phases are (co-)induced in the Higgs and neutralino sectors of the cNMSSM. Such phases are subject to constraints from the non-observation of the EDMs of the electron and the neutron. The most recent limits on these EDMs read |d e | < 1.1 × 10 −29 e cm [81]; |d n | < 1.8 × 10 −26 e cm [82].
The limit on the electron EDM above is based on the thorium monoxide experiment, and is more stringent than the one from the HfF + experiment [83]. In SUSY models, the one-loop EDMs of the charged leptons and the light quarks are induced by chargino and neutralino exchange diagrams. These should in principle constrain the CP-violating phases of M 1 , M 2 , λ and κ, appearing in the chargino/neutralino sectors at the tree level. In the SUSY spectrum generator code used for our analysis, details of which will be provided in the next section, calculation of the one-loop contributions to d e and d n , as well as to d τ , in the cNMSSM is currently implemented. Note that additional constraints also come from mercury [84] and thallium [85] EDMs, but these can generally be evaded if the masses of the first two generations of squarks are taken to be sufficiently heavy, as discussed in [66]. At the two-loop level, the Higgs-mediated Barr-Zee type diagrams can also contribute significantly to the electron and neutron EDMs. However, several studies have shown that even when these two-loop effects are taken into account, the phase φ κ is very weakly constrained by the fermionic EDMs [73,86,87,76], especially for smaller values of |κ|. This is in contrast with the other phases, especially φ Af (the phases of the Higgs-sfermion-sfermion trilinear couplings), which enter the Higgs sector at the one-loop level. Therefore, besides the reason noted above, we choose the φ κ as the sole representative CP-violating phase additionally to minimise the potential impact of these two-loop diagrams, which are not accounted for in our numerical code. Implementation of the complete set of contributions to the EDMs in our numerical code would go beyond the scope of this article, which aims to explore the DM properties when a non-zero phase appears in the neutralino sector.
Numerical Analysis
The radiative corrections to the tree-level Higgs and neutralino mass matrices make the parameters of the other model sectors highly relevant also. However, on the one hand, the NMSSM with grandunification-inspired boundary conditions is very tightly constrained by the current experimental results, and on the other hand, the most general NMSSM contains more than a hundred free parameters defined at the EW scale. Thus, in order to draw inferences for a particular sector of the model, it is imperative to make multiple assumptions about the free parameters that only impinge at higher orders.
For our numerical analysis, we therefore adopted the following (universality) conditions to impose on the parameter space of the Z 3 -symmetric cNMSSM at the EW scale: are the squared soft masses of the sfermions. The (less-often used) parameter Tf corresponds to sfermion trilinear couplings; usually these are taken to be proportional to the Yukawas, such that T ij t = Y ij u At, where i, j are generation indices. In our numerical code, we specified Tf directly at the low scale. We fixed all the elements of Tf to small values (1 GeV for the diagonal terms and zero otherwise), except for T which we left as a free parameter to be scanned over an extended range. The reason for this was to increase the probability of the consistency of m H SM , the dominant corrections to which increase proportionally to (At − µ eff cot β) 2 , with m H obs . We likewise scanned over wide ranges of M 1 and M 2 to allow maximal possible variations in the χ 0 1 composition. On the other hand, Mf and M 3 were fixed to sufficiently large values of 2 TeV and 3 TeV, respectively, so that the sfermions and the gluino could evade the direct search limits from the LHC.
As for the CP-violating phases, in light of the discussion in the previous section, we fixed However, the quantities that we choose for the solution of the (five independent) tadpole equations are The first three of these are standard choices familiar from the rNMSSM. However, once we break CP we have two additional non-trivial tadpole equations that must be satisfied, and it is logical to choose the complex part of the trilinear parameters, since these lead to the smallest impact on the spectrum and their magnitude will only be proportional to the violation of CP. With the only non-zero CP-violating phase being φ κ , this leads to As briefly noted in Sec. 2.1, the real parts of these trilinears are fixed as inputs: where now A λ , A κ are taken to be real. This means that at the tree level both the trilinear couplings pick up phases from the phase of κ (which are, however, small for A λ µ eff and A κ v d v u /s, and are modified at the higher orders).
To generate the particle spectrum for a given configuration of the final set of the free parameters, we incorporated the cNMSSM into the public fortran code SPheno-v4.0.4 [88,89] 3. The theoretical predictions of the following B-physics observables lied within 2σ deviation from their quoted experimental values.
5.
The |d e | and |d n | satisfied the experimental upper bounds given in eq. (11). 2 We ignored φκ > π, since we expected the real part of κ to be dominant by far, and hence the overall behaviors of the calculated observables to be approximately symmetric around φκ = π. This was nevertheless verified numerically for the sample value of φκ = 300 • .
3 While it is in principle possible to constrain these couplings using the latest combined measurements from the LHC (e.g., [103]) using a program like HiggsSignals-2 [104] instead we chose (for simplicity) to allow up to 20% deviation from the SM values. The figure of 20% corresponds roughly to the experimentally quoted uncertainties (we also checked that using a smaller value negligibly impacted our results); we were concerned that a combination was overly pessimistic regarding finding valid points, and not primarily concerned with tweaking the (heavier) HSM which otherwise plays little role in the dark matter properties. 6. The relic abundance of the χ 0 1 never exceeds +10% of the Planck measurement of Ωh 2 = 0.119 [108]. This allowance in Ω χ 0 1 h 2 is to crudely account for the rather large uncertainty in its theoretical estimation, due to the higher order corrections in SUSY models [109,110,111,112,113,114,115]. Note that MultiNest performs a multimodal sampling of a model's parameter space based on Bayesian evidence estimation. Our purpose for using this package was simply to scan the parameter space in a more efficient way than random sampling, rather than to draw Bayesian inferences about it. To this end, we defined a Gaussian likelihood function with a peak at Ω χ 0 1 h 2 = 0.119 and a width of ±10% of this value in MultiNest. Evidently, the scan collected a number of points far away from the peak also. From these, we removed all the points with Ω χ 0 1 h 2 > 0.131, but retained also the ones for which Ω χ 0 1 h 2 < 0.107 so as to accommodate alternative possibilities, such as non-thermal χ 0 1 production [116] or multi-component DM [117].
Low-mass DM in the cNMSSM
The scanned ranges of the nine free parameters (after fixing φ κ ) of the cNMSSM are given in Table 1. Evidently, these ranges cannot entail all possible configurations. They are, however, guided by some previous studies [20,33,35], and are extensive enough to fulfill the necessary conditions for this analysis, i.e., of yielding a H 2 or H 3 with a mass around 125 GeV, and a sub-100 GeV singlino/bino-dominated χ 0 1 . The upper cutoffs on the values of λ and κ are imposed to avoid the Landau pole. A 0 can in principle be both positive and negative, with a marginally different impact on the physical mass of the SM-like Higgs boson for an identical set of other input parameters in each case. Our purpose for using only its negative range was to enhance the efficiency of the numerical scanning code. Note that, at the EW scale κ and A κ are conventionally taken to be > 0 and < 0, respectively, in order to prevent negative mass-squared of the lightest pseudoscalar -see Eq. (10). But here, for 90 • < φ κ < 180 • , the real part of κ becomes negative, and hence A κ The left panel of Fig. 1 shows that a large number of parameter space points meeting all the conditions outlined in the previous section are ruled out by the latest limits on the cross section of the spin-independent DM-proton scattering, σ p SI , from the PandaX-4T Cmmissioning Run [118] (for comparison, the most recent exclusion contour from the XENON-1T experiment [119] is also shown). The right panel confirms the fact that φ κ is indeed very weakly constrained by the one-loop contributions to |d e |, since for almost all the successful scanned points its model prediction lies well below the experimental bound.
In Fig. 2 the χ 0 1 relic abundance is plotted as a function of its mass, separately for points from each scan with fixed φ κ . In this figure, the grey points in the background are the ones excluded by the PandaX-4T limits, and the coloured points further satisfy the following two conditions.
• Various searches targeting χ ± 1 and χ 0 i production, such as [122,123,124], are relevant here, since these states can decay to our (very) light χ 0 1 along with W/Z/H obs . These analyses quote limits of up to 750 GeV for winos with specific channels of decay. Searches for higgsinos are notoriously difficult due to their small production cross-section, and the limits from them are thus much weaker. The very latest ATLAS search [125] quotes limits of up to 210 GeV on a higgsino. A full recasting of these results using MadAnalysis [131,132,133,134] for all our good points will go beyond the scope of this study. We nevertheless imposed the simplistic requirement that the higgsino (wino) component of a given χ 0 i is less than 90% if it is lighter than 210 GeV (750 GeV). In the top-left panel of the Fig. 2, corresponding to the CP-conserving case, Ω χ 0 1 h 2 is generally quite small, except near m χ 0 1 ∼ m Z /2 and m χ 0 1 ∼ m H SM /2 for the H SM = H 3 scenario (blue points), where a few points show consistency with the Planck measurement within ±10%. A narrow peak of points also appears around m χ 0 1 ∼ 10 GeV, where, as we will see later, a very singlino-like χ 0 1 can undergo just the right amount of self-annihilation via the singlet A 1 . Recall that all the five Higgs bosons are ordered by mass and not distinguished by their CP-assignment, and thus the H 1 in the cNMSSM can be either one of the H 1 or A 1 of the rNMSSM. In the H SM = H 2 scenario (red points) the correct Ω χ 0 1 h 2 can be obtained for a wide range of m χ 0 1 , when it is near either m H 2 /2 or, more frequently, m H 3 /2. For φ κ = 30 • , in the top-right panel, a few points with m χ 0 1 between 10-20 GeV also appear within (or just outside) the Planck band (i.e., Ωh 2 = 0.119 ± 10%). This is not the case for the CP-conserving case above, although the overall picture looks very similar, and is a result of the slight modification in the χ 0 1 composition owing to the CP-violating phase. When φ κ is increased to 60 • (centre-left panel) some Planck-consistent points show up also around m χ 0 1 = 30 GeV. Most notably, however, it is possible to obtain the correct Ω χ 0 1 h 2 for the entire ∼ 5−40 GeV mass range when the sign of κ (and hence also of A κ ) is flipped, as demonstrated by the centre-right and bottom-left panels corresponding to φ κ = 135 • and φ κ = 180 • , respectively. In fact, for the latter phase, a sole point appears within the Planck band for m χ 0 1 in the ∼ 40−45 GeV range (which is, however, excluded by the LHC electroweakino searches). This is not observed for any of the other selected phases in this figure, but we will discuss below a point for which it is achieved around φ κ = 60 • also. The bottom-right panel presents a holistic picture, where one sees that nearly the entire sub-100 GeV range of m χ 0 1 with Ωh 2 = 0.119 ± 10% is covered by the good points from all of the scans.
For a closer analysis of the impact of the variation in φ κ on m χ 0 1 and its relic abundance, we selected four test points (TPs) from among those corresponding to φ κ = 30 • . The values of the corresponding scanned parameters, along with the spectra for φ κ = 30 • , are given in Table 2. For each of these TPs, m χ 0 1 is plotted as a function of φ κ in all the panels of Fig. 3. The heat map in the left column of the figure corresponds to m H 1 and in the right column to m H 2 . In this as well as the two figures that follow, the non-existence of a point for some values of φ κ in a given panel implies that SPheno did not produce an output on account of there being unphysical loop-corrected masses for some particles. The grey points imply inconsistency with one (or more) of the constraints 1-5 listed in section 3 and the two additional limits from the LHC noted above, while the coloured circles give Ω χ 0 1 h 2 > 0.131. The coloured boxes instead mean Ω χ 0 1 h 2 < 0.131 for that point, and a cross around a box reflects that Ω χ 0 1 h 2 lies within the Planck band. The TP1 in the top row of Fig. 3 is the single (red) point for the H SM = H 2 scenario with m χ 0 1 ∼ 40 GeV appearing within the Planck band for φ κ = 30 • . It does so, however, only for this specific value of φ κ . For almost the entire remaining range of the phase, this parameter space configuration is inconsistent with at least one of the enforced experimental constraints. The remaining three TPs belong to the H SM = H 3 scenario. For TP2 also, the Planck-consistent amount of self-annihilation of the χ 0 1 , via the H 2 , occurs only for φ κ a few degrees around 30 • . m χ 0 1 and m H 2 both reduce with increasing φ κ -the latter much slower than the former -until tachyonic masses appear in the particle spectrum for φ κ ≥ 80 • . In the case of TP3, as with the TP1, Ω χ 0 1 h 2 = 0.119 ± 10% is satisfied only for φ κ = 30 • , when the sharply falling m H 1 gets very close to 2m χ 0 1 12 GeV, as seen in the third row of the left panel. Beyond this value of φ κ , the Ω χ 0 1 h 2 drops for a few degrees, owing to excessive annihilation, before rising above the Planck bound again when m H 1 grows too small. Finally, TP4 is a representative point of the case when the Ω χ 0 1 h 2 falls within the Planck band for m χ 0 1 in the ∼ 40 − 45 GeV range, as hinted earlier. While this TP has also been taken from among the good points for φ κ = 30 • in Fig. 2, its Ω χ 0 1 h 2 lies below the Planck band for the original φ κ .
In the left column of Fig. 4 m χ 0 1 is again plotted as a function of φ κ for the TPs 1-4 (top row to bottom row), with the heat map now depicting the singlino fraction, N s , of the χ 0 1 . For TP1, the χ 0 1 has a negligible singlino component, but is instead entirely bino-like, with a small higgsino fraction just enough for the correct amount of its self-annihilation via the Z boson for φ κ = 30 • . On the other hand, the very large N s in the CP-conserving case for TP2 falls sharply with increasing φ κ , with the Planck-consistency occurring when it is just above 90% around φ κ ∼ 30 • . For TP3 the N s stays almost constant over the entire range of φ κ , while for TP4, as the singlino fraction as well as the mass of χ 0 1 drop slowly with increasing φ κ , its Z-mediated annihilation gradually reduces. It reaches a level sufficient to give the correct Ω χ 0 1 h 2 for φ κ ∼ 55 • − 65 • . The right column of this figure shows the BR(H SM → χ 0 1 χ 0 1 ). For TP1 it fluctuates between 1.5% and 2% for the allowed values of φ κ , and for TP2 it rises with φ κ but does not exceed 4%. For TP3 the BR(H SM → χ 0 1 χ 0 1 ) rises noticeably with φ κ (while m χ 0 1 , given by the heat map, drops), until it reaches the maximum of about 10% for 180 • , while for TP4 it is always insignificant.
In Fig.5 ) rises from ∼6% to about 17%, but this causes exclusion of φ κ > 144 • by the LHC data. Overall then, the phenomenology of the 4-body final states with invariant mass near that of H SM = H 3 for points analogous to the TP4 could be crucial for distinguishing the signatures of CP-conserving versus the CP-violating NMSSM. It will be the subject of a follow-up anaylsis.
Finally, since the χ ± 1 / χ 0 2 in all our TPs are higgsino-like and always heavier than 210 GeV, they are consistent with the current exclusion limits from the LHC. Besides, instead of decaying to W/Z/H obs , our χ 0 2 can decay dominantly to the lighter singlet-like Higgs boson(s) and thus keep evading detection in the near future. Likewise, the χ ± 2 / χ 0 5 are wino-like and heavier than 750 GeV for these four points. Nevertheless, in our follow-up analysis it would be interesting to test our scan points against latest results using a fast tool such as SModelS [126,127,128,129,130], and to process a handful using full recasting in MadAnalysis with the latest searches in [135,136] (see also [137] for a review of available recasting tools), as performed in, e.g., [138].
Summary and conclusions
The Higgs sector of the NMSSM can accommodate explicit CP-violating phases at the tree level, whereas in the MSSM such phases enter the Higgs potential only at the higher orders. While the measurements of the leptonic EDMs tightly bound the MSSM-like phase in Af , radiatively induced from the sfermion sector, they have been previously found to be much less constraining of the phase of κ. Importantly, this phase also appears in the tree-level mass term, 2κs, corresponding to the singlino interaction eigenstate. Therefore, if the χ 0 1 is singlino-dominated, its relic abundance can have a strong dependence on φ κ .
The cNMSSM contains 5 neutral CP-indefinite Higgs bosons in total, and any one (or more) of the three lightest of these can fulfil the role of the H SM , in specific regions of the model's parameter space. In this study, we have analysed in detail the quantitative impact of the variation in φ κ on Ω χ 0 1 h 2 , for scenarios wherein either H SM = H 2 or H SM = H 3 . The H 1 was required to always be lighter than 125 GeV to increase the prospects of self-annihilation of the singlino-like χ 0 1 solutions with mass 100 GeV, which was our main focus. For certain select values of φ κ , we performed numerical scans of the free parameters of the EW-scale cNMSSM, to find points consistent with a variety of recent experimental constraints, in particular the electron EDM.
In the overall picture that emerges from this analysis, for specific values of φ κ , nearly exact consistency with the Planck measurement of the DM relic abundance of the Universe is seen for certain m χ 0 1 that are precluded in the real NMSSM. Thus, while a large gap appears near Ω χ 0 1 h 2 0.119 for m χ 0 1 ∼ 10 − 30 GeV for φ κ = 0 • , this mass range starts filling up as the CP-violation increases, and gets almost entirely covered for φ κ ∼ 135 • . Evidently, this results from the subtle tweaks in the composition of χ 0 1 , so that its couplings allow just the right amount of its self-annihilation via one of the multiple potentially resonant sources available in this model, when the CP is violated. This inference was confirmed by a closer investigation of the four tests points selected out of the successful points from the scans. For each of these points we varied φ κ over the entire 0 • − 180 • , while fixing the other nine free parameters to their original values. This demonstrated how the different values of φ κ modify m H 1 and m H 2 , besides the mass as well as the N s of the χ 0 1 , to impact the consistency of these points not only with Ωh 2 but also with other experimental data. Furthermore, magnitudes of observables like the BR(H SM → χ 0 1 χ 0 1 ), the BR(H SM → H 1 H 1 ) and the BR(H SM → H 2 H 2 ) also show a dependence on φ κ significant enough that their dedicated inspection might help identify signatures of CP-violation in the NMSSM at the LHC. | 7,691.6 | 2022-01-25T00:00:00.000 | [
"Materials Science"
] |
Studies in Technology Enhanced Learning Technology and educational ‘pivoting’ in the wake of the Covid-19 pandemic: Introduction to the special issue
Department of Educational Research, Lancaster University, Lancaster, United Kingdom Technology and educational ‘pivoting’ in the wake of the Covid-19 pandemic: Introduction to the special issue Special issue Technology and educational ‘pivoting’ in the wake of the Covid-19 pandemic | More at https://doi.org/10.21428/8c225f6e.0a9292af Cover image JohnsonGoh via Pixabay. Publication history Online: 25 July 2022. Article type Editorial, not peerreviewed.
Introduction
The Covid-19 pandemic has, since early 2020, been associated with swift changes to how education is conducted across the globe. In attempting to maintain 'social distancing' and thereby prevent the further spread of the SARS-CoV-2 virus, the provision of teaching, learning and assessment has been re-mediated in ways that have placed digital technologies-and especially online platforms-at the forefront of public conversations and policy discourses about education systems to an unprecedented degree. The word 'pivot' has often been deployed, especially by policymakers and institutional managers, to highlight both the rapidity of the changes and the sheer extent to which it seems that long-entrenched educational practices are being supplanted by newly crafted alternatives. This development has been a double-edged sword for those scholars who devote their labours to research fields such as technology enhanced learning, distance education, online and open learning, and e-assessment. To be sure, the extent to which such fields have projected an aura of progress and relevance has long ebbed and flowed, with bursts of popularity punctuating periods in the wilderness-a picture complicated by a dramatic unevenness between different global contexts and the rapidity with which particular ideas are, and then suddenly are not, en vogue. Also https://doi.org/10.21428/8c225f6e.fd92e590 to be sure, any sense in which the scholars associated with these fields have wanted to evangelise their research objects has long been contested, with more recent generations of scholars seeming (relatively) more interested in theoretically driven critiques than their (relatively) solution oriented forebears (cf. Bligh, 2020). Yet, caveats notwithstanding, the situation has been an uncomfortable one. The very nature of the conjuncture in which public attention has been focussed on technological change in education has been one in which the surrounding context has presented difficult challenges to established scholarly wisdom. Finding themselves in the spotlight, researchers in these fields have needed to vacillate; perhaps even to worry whether they are being positioned to carry the can when everything goes wrong.
Such situations invite defensive mechanisms, of course, and one rhetorical manoeuvre was adopted remarkably quickly-the coining of exceptionalist terms. Probably the most widespread such term has been 'emergency remote teaching', meaning education newly undertaken online, at a distance, because of Covid-19. Emergency remote teaching is fundamentally distinct, we are told, from those more established forms of online education already growing in popularity over the preceding years: with the crucial differences concerned with condensed planning processes, the prosaic objectives driving the attendant change efforts, and the intention that online modalities are to be used only temporarily (Hodges et al., 2020).
A conceptual differentiation of present phenomenon from established research knowledge serves, of course, a number of useful functions. Where scholars lack answers, change initiatives stumble forwards chaotically, or stakeholder experiences leave a persistent bad taste, the response that what you are describing is not really online education can serve remarkably well. Yet it also brings deleterious consequences. Among other things, it serves to homogenise prior experiences (was online learning really always so well planned before?), discourages taking inspiration from knowledge that might have been partially valuable (haven't the facilitators and students of online learning always had novice experiences?), and invites a reactive focus on issues of training, support and technology-to the detriment of change, development and emerging vision. A narrow focus on the exceptional experiences, technologies and training provision associated with emergency remote teaching, for example, certainly seems circumscribed in its contribution back to the more established bodies of literature, and forward to debates about what might come next.
The purpose of the present special issue, therefore, was to invite a range of contributions that would be simultaneously similar and different. We wanted to present contributions which would focus on the nature and extent of educational change associated with the Covid-19 'pivot', but which would not be constrained, in so doing, to artificially circumscribe the scope or nature of their investigation. We wished to draw attention to the nature of the phenomenon we were studying-and so our call for papers explicitly used the vocabulary of "pivoting" and "Covid-19 pandemic"-yet we did not wish to strongly demarcate the work presented to an orientation within the exceptionalist literature. We certainly set up no expectations of using particular terms (such as "emergency remote teaching") or drawing on particular bodies of knowledge. The consequence, as elaborated below, is a mixed ecology, in terms of both vocabulary and scholarly inspiration-one that we hope better connects this 'specialised' field of enquiry to a range of debates both wider in scope and more longitudinal in implication.
Our initial call for papers elicited 13 proposals, of which 7 papers survived the inevitable attrition of peer review and the exceptional circumstances surrounding the writing and editing process (forced to withdraw part-way through, one correspondent invoked the irony of having been prevented from writing about the educational disruptions of the Covid-19 pandemic by… the educational disruptions of the Covid-19 pandemic). As has been our strategy since the beginning of the Studies in Technology Enhanced Learning project (Bligh & Lee, 2020), we attempted to generate a scholarly conversation between contributors to the issue. In this case, that was accomplished by trying to ensure that papers were peer reviewed, anonymously, by one other contributor-meaning that around half of the peer reviewing effort was undertaken by authors themselves-and by inviting all reviewers and members of the editorial board to contribute commentaries at the end.
In what follows this editorial, therefore, we present 8 entries in this special issue: seven full papers and a collected commentary. As we elaborate below, we cluster full papers into categories fundamentally concerned, in turn, with pedagogical values, change processes, and the position of technology. We conclude the issue with a collected commentary, comprising thirteen entries by 14 authors, which addresses a broader range of themes and considers the implications, moving forward, for both practice and scholarship.
The papers
Our first two papers each foreground, in different ways, the role of pedagogical values during the pandemic 'pivot'.
https://doi.org/10.21428/8c225f6e.fd92e590
The paper by Victoria I. Marín (2022), Student-centred learning in higher education in times of Covid-19: A critical analysis, juxtaposes the specialist literature on emergency remote teaching with that on student-centred learning-a phrase whose meaning encompasses a variety of constructivist approaches (like problem, project, case and inquiry-based learning) which have long been understood as relevant to digital pedagogy. Marín scrutinises the literature on emergency remote teaching, and finds that, where claims are made for supporting student-centred learning, this typically refers to specific, limited course design elements: such as student prompting and progress monitoring, the provision of tools, and some encouragement for students to create artefacts. What seems largely absent, conversely, is much ambition that students might take ownership of their own learning; for instance, by setting their own goals or sharing resources with each other. Remarkably, evaluative reports for these re-designed courses are overwhelmingly positivesometimes more positive than for the pre-pandemic courses that have been supplanted. Marín's analysis invites us to reflect on the ongoing relevance of firm pedagogical principles, how these have been sustained or attenuated during the period of the Covid-19 pivot, and the ongoing implications for re-designing educational provision in the future.
Our second paper, by Khadija Al-Ali (2022), is entitled To see or not to see; the withering boundaries of invisibility: A novice Kuwaiti tutor's experience of teaching online. Al-Ali draws mainly on the literature of feminist pedagogy, which has long critiqued issues of choice and power in classroom practices and which has, more recently, been applied to studies of online learning contexts. The paper presents an autoethnographic narrative from the vantage point of a "first-time online tutor in a Kuwaiti college during the Covid-19 pandemic". Al-Ali notices what she calls an "emerging context of invisibility", in which certain students choose to materialise their interests in unusual ways which, in turn, compel the tutor to adapt her teaching approach. The paper emphasises strongly the context-specificity of pedagogical values (like 'student empowerment'), and emphasises that we must be prepared for them to be manifest in very different ways as educational contexts change and develop.
Our next three papers encourage us to understand the Covid-19 'pivot' processually: through a lens of change.
The third paper, by Christos Petichakis (2022), is called Review of a pivoted fully online flipped learning modality to promote reflection for early career teaching staff development. The paper, which draws on the theory of situated learning, documents the re-design of a professional development course for teaching staff-contrasting the pre-Covid-19 design against the pivoted variant, and reflecting on the introduction of a flipped learning modality. The paper argues that the success of the initiative derives, in large part, from the principled application of a definite conceptual framework. Petichakis' work reminds us of the importance of being guided by definite objectives when engaging in change processes-even where, as in the case of the 'pivot', planning time is scarce.
The fourth paper, by Reya Saliba, Matthew A. Carey, and Rachid Bendriss (2022), is entitled Reimagining premedical foundation blended curriculum through design thinking: A qualitative study. The paper draws on the long history of Design Thinking, a tradition which originally emerged in fields such as architecture, design and art, and which has since spread across many disciplines, including healthcare and medical education, which is the context for Saliba et al.'s study. Saliba, Carey and Bendriss report on an initiative in which the Covid-19 pandemic was used as a stimulus to shift from a previous course design-based on practitioner shadowing and understood through the lens of experiential learning-to a new one based on student projects, which made use of an approach derived from Design Thinking. The paper argues that the approach positively stimulated interactions, both between tutors and students and within student groups. Moreover, Saliba et al. suggest that the approach has an ongoing potential, given the success of the new course design in helping students to think in more 'patient-centred' ways when pursuing their projects.
The fifth paper, by Dale Munday (2022), is called Hybrid pedagogy and learning design influences in a higher education context. The paper explores the issue of hybrid learning design, which has emerged as an important issue in the wake of the Covid-19 'pivot'. Munday's work explores the extent and nature of hybrid learning design as it has been achieved in practice: both by surveying academics from a range of UK universities and exploring institutional analytics from a single site. Munday's analysis suggests that the commonly understood features of hybrid learning are present in actual learning designs to only a fairly limited degree. Munday's work highlights the importance of a number of issues when pursuing hybrid learning: such as pedagogical focus, digital capabilities, institutional influence, and obtaining support from online professional networks external to the institution. Such issues will doubtless remain important for ongoing attempts at innovating teaching and learning modes.
Our final two full papers problematise the positioning of particular technologies during the Covid-19 pandemic 'pivot'.
https://doi.org/10.21428/8c225f6e.fd92e590
Our sixth paper, by Dave Gatrell (2022), is entitled Challenges and opportunities: Videoconferencing, innovation and development. While Gatrell draws inspiration from the literature on emergency remote teaching, he frames the issue in an unusually holistic way: noticing that there is much valuable insight that can be gleaned from the scholarship on pedagogy during previous public health crises, natural disasters, and protest-related disruptions. From this analysis, Gatrell draws the conclusion that it is important to understand situations such as the Covid-19 'pivot' as unfolding responses to ongoing dilemmas. Gatrell subsequently puts forward a longitudinal analysis, framed by activity theory, of his own support for university teachers adopting videoconferencing for synchronous teaching during the pandemic situation. The paper describes how a community of teachers was forged against this backdrop, for purposes of professional development in relation to uses of the tool, and the potential for subsequent institutional innovation that is already emerging as a consequence.
The seventh paper, by Liz Dovrat (2022), is called Perceptions of emergency remote teaching tools used during Covid-19 online teaching by an Israeli English for Academic Purpose (EAP) department. The paper brings together two strands of literature, that on emergency remote teaching and that on disciplinary pedagogy in the area of English as a Foreign Language, and notices that the latter conveys a long trajectory of debates of manifest resonance for the former-including the necessity of teachers managing shifts in identity when integrating technologies into their practice. Dovrat subsequently employs a social practice perspective to frame a study in which respondents from across a particular English for Academic Practice teaching setting are asked to reflect on the tools they used when 'pivoting' and how these continue to be integrated into their academic practice. The paper's findings show that teachers' views of particular technologies, and the extent to which they persevered with them, were deeply interwoven with how they were supported by local workgroups comprised of both other teachers and institutional administrators. Such findings challenge the prevalent argument that teachers simply used whatever technology was made available or prescribed within an institution.
We conclude the issue with a collected commentary, comprising 13 entries from fourteen authors, entitled Technology and educational 'pivoting' in the wake of the Covid-19 pandemic: A collected commentary (Bligh et al., 2022) Recurrent themes in the 13 concise entries include reclaiming a sense of history when discussing the Covid-19 'pivot', differentiating practitioner experiences from overblown rhetoric, reappraising educational sociality in light of sometimes harsh pandemic experiences, understanding the consequences of unprecedented change for long-established practices, examining moves to accommodate new modes of education (such as 'hybrid learning'), exploring the pandemic 'pivot' as interlocking and multifaceted processes of change, and maintaining a sense of trajectory-and rejecting a simple 'pivot back'. You are free to • Share -copy and redistribute the material in any medium or format • Adapt -remix, transform, and build upon the material for any purpose, even commercially.
Under the following terms: • Attribution -You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.
• No additional restrictions -You may not apply legal terms or technological measures that legally restrict others from doing anything the license permits.
Authors Brett Bligh
Brett Bligh is a Lecturer in the Department of Educational Research, Lancaster University, and co-Director of the Centre for Technology Enhanced Learning. His research interrogates the nexus of technology mediation, physical environment, and institutional change in higher education. Brett's work prioritises Activity Theory conceptions of human practice, and interventionist methodologies.
Kyungmee Lee
Kyungmee Lee is a Senior Lecturer in the Department of Educational Research, Lancaster University, co-Director of the Centre for Technology Enhanced Learning, and co-Editor of Studies in Technology Enhanced Learning. Her research targets the intersection of online education, adult education, and international education concerning issues of accessibility and inclusivity. Using a range of qualitative research methodologies and evocative academic writings, her current projects investigate the academic experiences of diverse non-traditional student groups in distance education settings. Kyungmee's scholarship emphasises concepts of discourse, knowledge and power, understood through a broadly Foucauldian lens. | 3,624.6 | 2022-07-25T00:00:00.000 | [
"Economics"
] |
DEVELOPMENT OF AN ANDROID-BASED REWARD SYSTEM TO ENHANCE THE ACTIVITY OF LEARNING
The students’ learning activity is an influential factor in the success of learning. New technology can be used to design applications to facilitate the learning process. The objectives of this study were (1) to produce a reward application system to improve the students’ activity, (2) to examine the feasibility of the developed application, and (3) to analyse the difference and the effectiveness of the application in learning compared to conventional learning. This study used the development research method adopting the 4D (define, design, develop, disseminate) model. The application was tested using ISO 25010 standards by media experts, teachers, and ten students of class X Axioo in Vocational High School also known as Sekolah Menengah Kejuruan Negeri 5 Surakarta. The effectiveness was tested using an experimental design. This study produced a proper and an efficient reward system application. The feasibility of the application for functionality, portability, usability and efficiency were 100%, 100%, 86.2% and 77% respectively. The results showed different effectiveness in learning activities between learning using the reward system application and using the conventional model.
INTRODUCTION
Education is the most critical part of progress indicators of a nation.Success in the education sector is directly proportional to the human resources, i.e., the more advanced the education of a country the higher the level of human resources.The Law of the Republic of Indonesia No. 20 of 2003 on National Education System article 18 explains that vocational education is secondary education that prepares learners primarily to work in industries.With the characteristics of vocational education that focuses on the achievement of skills and competencies, enabling learning in Vocational High Schools also known as Sekolah Menengah Kejuruan (SMK) is active student learning.Besides, the learning process is the integration of knowledge, skills, and attitudes (Baartman & Bruijn, 2011).Thus, in the learning process, students are actively involved in learning and can master the competence in the field.Teachers must be respected because the success of learning is highly dependent on professional educators.Teachers also have a role as a facilitator of learners in learning process activities.The choice of a learning model in a classroom determines the success of learning.The model can improve the students' learning outcomes.In the curriculum 2013, students are required to be independent and active in learning because the learning process does not continue to be centered on the teacher.
Learning is a process of changing behavior towards a better and relatively fixed behavior, and the effect of it may be changing knowledge, understanding, attitude, behavior, and skills.The liveliness of learners is a situation where students are active in learning.The activity of learning includes the involvement of students in the learning process as diverse as listening to the teacher's explanation, doing discussions, making reports, and completing the tasks.Learning activities of students include cognitive, affective and regulative activities (Vermunt & Verloop, 1999).The activity of learners in learning is an important and fundamental problem that must be understood, based and developed by every teacher in the learning process.Activity-Based Learning is a practical experience method for students to analyze and learn from the activity (learning by doing) (Lijanporna & Khlaisang, 2014).The power of the student's activity by nature will develop in a positive direction when the environment provides a good space for the development of that activity (Aunurrahman, 2009).In other words, the level of the students' activity is influential in learning achievement, as indicated by Ramlan, et al. (2014) that there is a significant influence on the learning achievement with the significance value of 0.001 which is less than 0.05.
In recent decades, the ownership of mobile Android devices is increasing.On September 3, 2013, Google announced that around 1 billion active mobile devices worldwide use the Android Operating System.The Android operating system is opensource and has multitasking capabilities to run multiple applications in Smartphone devices.It is the appeal of users, especially the price of Smartphones today is more affordable in the community.More and more people who own and use mobile devices open up opportunities for the use of mobile devices in education.
The use of mobile devices in the learning process is known as mobile learning.Using mobile devices allows students to not just stay in one place.Students can access something as long as there is internet from anywhere and anytime without the limitation of space and time.This is demonstrated by Edgar et al. (2016) who studied the use of Short Message Service (SMS) as a medium for evaluation of learning.His study is whether the use of SMS can increase the students' participation in learning.He used Blended Learning strategy, combining face-to-face classes with online activities.The results of the analysis showed that the students using the mobile messagebased learning evaluation significantly performed better learning than the students from the control group who did not get the mobile message-based learning evaluation.Nearly 20% of the experimental group was more effective in completing the assigned activities than the control group.Besides, the questionnaires answered by the experimental group showed that they were satisfied with the use of this technology.In addition, Paisal (2015) claimed that the students who use the internet in the learning activities will have positive attitude towards learning activities.Information and Technology based media can facilitate the students' learning better ( Suryanto & Kusumawati, 2017;Arief & Umniati, 2012) To encourage and motivate the students to play an active role during the learning process, the students need a reward in every active role undertaken by students.The reward can be an additional score given to students in an open and real-time manner.Thus there is a need for an application that stores and displays the additional scores, where the application can be monitored in real-time by the students.Android-based applications that are installed on the student's smartphones can do this.When one student gets a score because his activity in the class, it expected that it will motivate the other students to participate actively in learning as well to get the additional sore.The student activity at the time of learning process needs attention from a teacher so that the teaching and learning process get the maximum result.Thus finding effective and efficient ways to improve student activity is crucial.
METHOD
This study used the method of Research and Development with 4D models consisting of 4 development stages namely define, design, develop, and disseminate.The disseminating stage was used to develop an Android-based learning system application.The data collection methods were questionnaires and documentation.The defining stage produced an analysis of activities for development needs and product development requirements by user needs.The designing stage produced a design of reward system applications based on user requirements.The developing stage was feasibility testing of the application.The feasibility test was based on ISO 25010 that Basori, Development of an Android-Based Reward System to Enhance the Activity of Learning consisted of testing the application in the aspects of functionality usability, efficiency, and portability.The tests were conducted by media experts, teachers and ten students of class X Axioo in SMKN 5 Surakarta.The collected data were quantitative data.The aspects of testing refer to ISO 25010 which includes aspects of functionality, usability, efficiency and portability.
Meanwhile, the techniques of data analysis using data.percentages.
(1) From the percentage, then it was consulted with the Feasibility criteria of the application.adapted from Riduwan (2013).At the disseminate stage, experiments were conducted in the classroom using an experimental design with a posttest-only control group design.This experiment was conducted in SMK N 5 Surakarta.The data collection used a questionnaire of the learning activity.This experiment involved 39 students of class X MM2 and 39 students of class X MM3 as the experimental group and the control group respectively.The experiment assessed the comparison of the students' learning activity between conventional learning and learning using the reward system application.The experimental design of this study adapted from Sugiyono ( 2013) is presented in Table 3.The data analysis technique was descriptive statistical analysis using was t-test analysis and gain index analysis.The descriptive statistical analysis was intended to describe the level of interest of students of SMKN 5 Surakarta before and after the treatment of the reward system application with the frequency distribution table.The hypothesis testing used t-test analysis with significance level 0.05 with the criterion that Ho is rejected if tcount is more than ttable and Ho is accepted if tcount is more than ttable.While the gain index calculation used the gain calculation according to Hake (1999): (2) The result g was then referred to the gain index criteria adapted from Sundayana (2015).It is presented in Table 4 Table 4. Gain Index Criteria Value Criteria 0.7 < g < 1 0.3 ≤ g ≤ 0.7 0 < g < 0.3
RESULTS AND DISCUSSION
The results of this study are the realization an Android-based reward system application that can store and present the score data embodied in the form of medals so that by using this application, the teachers can simplify the provision of rewards for the active students.Students can also view and monitor the score in real-time and continuously without limitation of space, distance and time.
The concept of the hierarchical structure explains the program structure of the page in the application.The application of the reward system is divided into two branches, namely students and teachers.Teachers and students have different structures as shown in Figure 2. Students can only view lesson data, basic competence data, and scores data.Meanwhile, the teacher can see maple data, manage basic competence data and scores data.The structure of the program hierarchy is shown in Figure 1.To design the reward system application, PHP MySQL was used and to display utilizing bootstrap template, notepad ++ software was used.XMPP and Google chrome were used to build its Android applications utilizing Intel XDK software.The results of the implementation of the sequenced code are presented in Figure 2. Furthermore, the feasibility testing of the reward system application was addressed to media experts, teachers and students.The test was done by referring to ISO 25010 mobile application standards by testing the aspects of functionality, usability, efficiency, and portability.The results of the test are shown in Table 5.Table 5 shows the percentages of the aspects of functionality and portability performed by media experts by logging in the application as students and teachers.The aspect of functionality shows that the function of each page both login as a teacher and as a student.has the percentage of 100%.The portability aspect was conducted by installing the reward system application on various types of Android versions of ice cream sandwich 4.0 to nougat 7.0 and also testing on various screen sizes with portrait and landscape orientation then the percentage result shows 100%.
Figure 3 shows the efficiency aspect by logging in as students and teachers.The testers assess the response time required when running each page of the application.The results for the response time are the response time that is less than 3 second, between 3 to 9 seconds, between 10-12 second and over 12 seconds were 85%, 15%, 0% and 0% respectively.It means the users who feel very satisfied with the response time of fewer than 3 seconds are 85%.function of each page of the application both login as a teacher and as a student.Its percentage results show 100%.The usability test of the application includes several indicators of usability, satisfaction, ease of use and ease of learning of the overall total obtained 86.2%.The percentage results are categorized as very decent criteria.They are presented in detail in Table 6.While testing the efficiency aspect, teachers and students assess the response time of each application page when it is executed.The result is the percentages of page indicator requiring a response time of fewer than 3 seconds, 3 to 9 seconds, between 10 to 12 seconds and over 12 seconds were 69%, 29%, 1% and 1% respectively.It means that the users who were very satisfied, satisfied, less satisfied, and not satisfied were 69%, 29%, 1%, and 1% respectively.The results are shown in Figure 4.The disseminate stage is carried out using experiments in the class to obtain data about the effect of the application on the activity of students' learning.Table 7 shows the results of initial measurements before the treatment.Both the experimental and the control groups show relatively the same average scores of learning activities.However, the results of final measurements after the treatment by using the reward system application in the experimental group and the conventional model in the control group, there is a significant difference.The activity of learners had a significant increase between the two groups, but the higher level obtained by the experimental group.From the data, the posttest scores show that the conventional model reached the highest score of 160 and the lowest of 128 from 50 questionnaires with maximum, minimum, mean and standard deviation scores of 200, 50, 144.79 and 12.60 respectively.Table 8 presents the posttest data in the control group in the form of frequency distribution table.Table 8 shows that the most substantial frequency of the control group is 28.2% or 11 students located in the interval of 144 to 151 and the lowest frequency is 15.4% or 6 students in the interval of 160 to 167.Meanwhile, the activity data from the posttest of learning activity of students on learning using the reward system application shows the highest of 178 and the lowest score of 132 from 50 questionnaires with maximum, minimum, mean and standard deviation scores of 200, 50, 155.46 and 12.87 respectively.Tabel 9 presents the posttest score of the experimental group in the form of frequency distribution table.The data from Table 9 shows that the most significant frequency of the experimental group is 23.1% or 9 students in the interval of 168-175 and the lowest frequency is 2.6% or 1 student in the interval of of 160 to 167.Based on the distribution table, it can be determined the tendency of student learning variables, by groupping based on ideal value (Mi) and ideal Deviation Standard (SDi) with calculation steps adapted from Arikunto (2012) as follows: Mi = ½ (Ideal score + ideal lowest score) = ½ (178-132) = 155 SDi = 1/6 (ideal low score-ideal score) = 1/6 (178-132) = 7.6 From the calculation results then the determination of activity criteria is as follows: (1) High activity = X ≥ Mi + SDi that if the score obtained is more than 162.6, (2) Medium Activity = Mi-SDi ≤ X <Mi + SDi ie if the score obtained is between 147.4 up to 162.6; (3) Low Activity = X <Mi -SDi ie if the score is less than 162.6.Using these criteria, the activity calculation is presented in Table 10.In the experimental group, the students with the activity categorized as high, medium and low were 14 students with the percentage of 35.9%, 14 students with the percentage of 35.9%, and 11 students with the percentage of 28.2% respectively.In other hand, in the control group the number of the students with the activity categorized as medium and low were 19 students with the percentage of 48.7% and 20 students with the percentage of 51.3% respectively.There are not students with the activity categorized as high.Thus it can be concluded that the trend of the activity in the experimental group is in the medium and high category, but the control group is in the low and medium category.Meanwhile, to know the difference of effectiveness of the application on the students' learning activity, the experiment result data are analyzed using a t-test and a gain score.The results are presented in Table 11.Table 11 identifies the Tcount of the posttest scores from the experimental and the control groups are 3.790 with the significance of 0.000.It can be concluded that the significance si less than 0.05 and tcount is higher than ttable that is 3.790 higher than 2.642 (76), hence Ho is refused.It can be concluded that there is a difference in effectiveness between learning with and without the reward system application.The next step is to determine which is more effective using a gain score test.The results of the gain score test are presented in Table 12.Table 12 shows that the gain calculation in the experimental group and the control group is 0.302383 and 0.128941 respectively.The experimental group shows the medium criteria, but the control group indicates the low criteria.
From both calculation, it can be concluded that learning with the developed reward system application is more effective than the conventional learning.
CONCLUSION
The results of the study showed that teachers and students can use the developed reward system application through Smartphones with an Android operating system.The Android operating system to be used for the application is minimally Ice Cream Sandwich 4.0 and maximally Nougat 7.0.The developed application functions as an application to store and present data of reward for active students in the learning process.The application is realtime in-app data which is accessible anytime and anywhere.The results of reward system application testing conducted by media experts, teachers and students stated that the application has fulfilled the aspects of functionality, usability, efficiency and portability.The feasibility of the application for functionality, portability, usability and efficiency were 100%, 100%, 86.2% and 77% respectively.The results of the study also showed different effectiveness in learning activities between learning using the reward system application and using the conventional model.The data analysis shows learning with the reward system application is considered more effective than conventional learning measured by t-test and a gain score.
Figure 2 .
Figure 2. (a) App Icon Display (b) View of Login Page (c) Homepage Viewer
Figure 3 .
Figure 3. Diagram of the Efficiency Aspect to Media Experts: Response Time
Figure 4 .
Figure 4. Diagram of the Efficiency Aspect to Teachers and Students: Response Time
Table 1 .
Application Feasibility Level Criteria Criteria Percentage
Table 2 .
A Questionnaire Grid of the Learning Activity
Table 5 .
Test Results of Media Expert
Table 7
The experimental group is a group with a reward-based learning system, the other group is the control group with the traditional learning.Tabel 7. Descriptive Statistics Analysis of the Experimental Group and the Control Group
Table 8 .
Frequency Distribution of the Posttest
Table 11 .
The Difference in Effectiveness from the Students
Table 12 .
Comparison of Gain | 4,205.6 | 2018-04-17T00:00:00.000 | [
"Computer Science",
"Education"
] |
Ni2P nanocrystals embedded Ni-MOF nanosheets supported on nickel foam as bifunctional electrocatalyst for urea electrolysis
It’s highly desired but challenging to synthesize self-supporting nanohybrid made of conductive nanoparticles with metal organic framework (MOF) materials for the application in the electrochemical field. In this work, we report the preparation of Ni2P embedded Ni-MOF nanosheets supported on nickel foam through partial phosphidation (Ni2P@Ni-MOF/NF). The self-supporting Ni2P@Ni-MOF/NF was directly tested as electrode for urea electrolysis. When served as anode for urea oxidation reaction (UOR), it only demands 1.41 V (vs RHE) to deliver a current of 100 mA cm−2. And the overpotential of Ni2P@Ni-MOF/NF to reach 10 mA cm−2 for hydrogen evolution reaction HER was only 66 mV, remarkably lower than Ni2P/NF (133 mV). The exceptional electrochemical performance was attributed to the unique structure of Ni2P@Ni-MOF and the well exposed surface of Ni2P. Furthermore, the Ni2P@Ni-MOF/NF demonstrated outstanding longevity for both HER and UOR. The electrolyzer constructed with Ni2P@Ni-MOF/NF as bifunctional electrode can attain a current density of 100 mA cm−2 at a cell voltage as low as 1.65 V. Our work provides new insights for prepare MOF based nanohydrid for electrochemical application.
In recent years, the application of metal organic frameworks (MOFs) in the electrochemical field, especially electrocatalysis, has attracted considerable interest. The major appeals of the electrochemical application of MOFs are their large surface area, well-defined pores and tunable chemical composition 1,2 . However, the direct use of MOFs as electrode is often limited due to the poor intrinsic conductivity. So, MOFs are more often used as precursors to prepare advanced electrocatalysts through carbonization at high temperature 3 . The organic ligands in the MOFs are pyrolyzed to form graphitized/amorphous carbon matrix that could serve as freeway for electron flow. As a result, the MOFs-derived material through carbonization exhibited enhanced electrochemical activity for both hydrogen evolution reaction (HER) and oxygen evolution reaction (OER) due to the dramatically increased conductivity 4 . The electrocatalytic performance of the carbonized material can be further boosted by phosphidation 5 . However, the fascinating pore structure of MOFs is largely destroyed during the heat treatment, vanishing the advantages associated with the well-defined pore structure as well as high specific surface area 6 .
In recent years, MOF based nanohybrids prepared by integration of nanoparticles with MOF have attracted considerable attention 7,8 . Beside utilizing the super adsorption capability of MOF, such integration strategy could tuning the microenvironment of the nanoparticles to improve their catalytic activity 8 . Nevertheless, it is still very challenging to synthesize self-supporting MOF-nanoparticles nanohybrids, which is highly desired for the application in the electrochemical field.
On the other hand, OER is a thermodynamically sluggish multi-electron process with large over-potential, leading to energy consumption surge for electrochemical hydrogen production 9 . Moreover, it is a formidable challenge to completely avoid the formation of explosive H 2 /O 2 mixtures during water electrolysis 10 . Replacing OER with alternative electrochemical oxidation reaction offers an effective way to address these drawbacks 11 . So far, urea, methanol and hydrazine electrochemical oxidation have been explored as alternative anode reaction [12][13][14][15][16] .
Results and discussion
Materials characterization. To synthesize the Ni 2 P@Ni-MOF/NF electrode, the Ni(BDC)(DMF) MOF was first grown on NF by a solvothermal method at 140 °C and the sample was labelled as Ni-MOF/NF. Afterwards, the Ni-MOF was converted to Ni 2 P@Ni-MOF by a direct phosphidation process carried out at 300 °C. Powder X-ray diffraction (XRD) was first employed to get the composition and phase of the Ni-MOF/NF. Two sharp diffraction peaks at 44.9° and 52.2° were observed in the XRD pattern of Ni-MOF/NF (Fig. 1a), corresponding to the (111) and (200) planes of face-centered cubic nickel (JCPDS No. 01-1260). Apparently, the diffraction peaks of Ni-MOF were relatively weak. Enlarged XRD patterns on an expanded y-axis scale was provided in Fig. 1b Fig. S1 confirmed that 300 °C was not high enough to cause the decomposition of Ni-MOF.
The FTIR spectra of the nanosheets before and after phosphidation are shown in Fig. S2. Overall, the absorption bands of nanosheets became weaker after phosphidation. Nevertheless, the FTIR result verified the partial preserve of Ni-MOF. The absorption band around 3444 cm −1 was assigned to the stretching vibration of NH. The asymmetric and symmetric stretching modes of C = O appeared at 1582 and 1369 cm −1 , respectively 36,37 . The band at 1087 was attributed to the stretching vibration of C-O of carboxylic acid group. And narrow bands at 1025 and 748 cm −1 were attributed to δ(C-H) and γ(C-H) vibration of aromatic rings, respectively.
As revealed by the SEM characterization (Fig. S3), the Ni foam exhibited a highly open pore structure with pore size around several hundred microns. The high resolution SEM image revealed that the nickel skeleton was composed of relatively smooth large compact Ni grains. After Ni-MOF growth, the surfaces of Ni foam were uniformly covered with 2D nanosheets (Fig. 2). The Ni-MOF nanosheets exhibited a leaf-like morphology with lateral size 4-7 microns, which interlaced with each other to form a nest-like structure. The high resolution SEM image shown in Fig. 2c disclosed the fine vein-like structure on the surfaces of the as-grown Ni-MOF. After phosphidation process, the skeleton of Ni foam was still uniformly covered with nanosheets (Fig. 2d). But compared to Ni-MOF/NF sample, the population of the nanosheets was apparently decreased. As it can be clearly seen in the high resolution SEM images (Fig. 2e, f), the fine vein-like structures disappeared after phosphidation and the www.nature.com/scientificreports/ surfaces of the 2D nanosheets became smoother. Nevertheless, the sample still exhibited nest-like morphology composed of 2D nanosheets. The successfully incorporation of P elements through phosphidation was confirmed by the energy-dispersive X-ray (EDX) spectrum (Fig. S4).
For TEM characterization, the Ni 2 P@Ni-MOF nanosheets were scraped off the nickel foam and dispersed in ethanol with the help of vortex shaking instead of sonication to minimize the damage. The TEM image shown in Fig. 3a reveals the 2 dimensional nature of the Ni 2 P@Ni-MOF sample. The selective area electron diffraction (SAED) pattern displayed rings with bright spots, indicating the polycrystalline nature of the sample (Fig. 3b). High resolution TEM (HR-TEM) image revealed that the surface of the nanosheet was decorated with nearly monodispersed nanoperticles with size around 8 nm (Fig. 3c). The formation of Ni 2 P nanoparticles was limited by the available nickel atoms in the MOF nanosheets, similar to the formation of metal or metal oxide nanoparticles in layer double hydroxide 38 . The HR-TEM image of a single nanoparticle shows distinct lattice fringes with interplanar d-spacing determined to be 0.22 nm (Fig. 3d), which was attributed to the (111) plane of Ni 2 P. According to the elemental mapping image, the Ni, P, C and N elements were uniformly distributed in the Ni 2 P@ Ni-MOF nanosheet (Fig. 3e). The atomic ratio of Ni:P was ~ 5.6:1, proving the partial transformation of the Ni atoms in the Ni-MOF to Ni 2 P.
XPS was performed employed to investigate the surface chemical composition and electronic state of elements of the nanosheets before and after phosphidation. As shown in the XPS survey spectra (Fig. 4a), besides the Ni, C and N elements, new signal peak positioned at 133 eV corresponding to P element after phosphidation process. The high resolution Ni 2p XPS spectrum of Ni-MOF displayed four prominent peaks (Fig. 4b), the two peaks positioned at 855.6 and 873.5 eV are assigned to Ni 2p 3/2 and Ni 2p 1/2 , respectively 39,40 . The other two peaks positioned at 861.8 and 878.2 eV are the corresponding satellite peaks of Ni 2p 3/2 and Ni 2p 1/2 , respectively. After phosphidation, two new peaks corresponding to reduced Ni δ+ emerged at 852.8 and 870.0 eV 41 . Meantime, both the Ni 2p 3/2 and Ni 2p 1/2 peaks shifted to higher binding energy, suggesting the charge transfer from nickel to phosphorus atoms. The high resolution C1s XPS spectrum of Ni-MOF nanosheets displays two main peaks and the one positioned at 288.3 eV is assigned to C=O (Fig. 4c) 42 . The other peak positioned at 284.5 eV could be deconvoluted into two peaks corresponding to C-C (284.5 eV) and C-N (285.0 eV), respectively. After phosphidation, the peak related to C-N shifted 0.4 eV to higher binding energy, suggesting the charge transfer from carbon to neighboring atoms. The N 1s XPS spectrum of Ni-MOF nanosheets only displays one peak at 399.9 eV assigned to C-N-C from DMF (Fig. 4d). After phosphidation, the peak shifted slightly to lower binding energy to 399.6 eV, demonstrating the charge transfer from neighboring atoms to N. The N 1s peak could not be deconvoluted, suggesting that no graphitized nitrogen formed after phosphidation process. That was reasonable since the temperature of phosphidation process was too low to induce graphitization. The high resolution P 2p XPS spectrum exhibits two main characteristic peaks positioned at 129.3 and 134.7 eV (Fig. 4e), which are assigned to P-Ni and oxidized phosphate species 43,44 . Fig. 5a, the initial oxidation potentials of UOR for various electrodes were almost the same, 1.36 V (vs RHE). This result suggested that the active species for UOR of various electrode were the same as reported in previous study 45 . The Ni 2 P@Ni-MOF/NF electrode demonstrated the best UOR activity, followed by Ni-MOF/NF. The potential to drive the UOR at 100 mA/cm 2 is only 1.41 V, which is among the best when compared with the literature data (Table S1). The linear increase of the current density above 1.38 V (vs RHE) indicated the readily detachment of gases bubbles from the surfaces of the electrode 45 . The UOR performances of NiC/NF and Ni 2 P@NiC/NF were inferior to Ni-MOF/NF, which may be attributed to the encapsulation of nickel species by carbon. The current density reported in this work was normalized to the projected area of the electrode, which could not reflect the intrinsic catalytic activity of the material 46 . However, it was impossible to accurately determine the loading amount of the active material on the electrode. So, we tried to normalize the current density to the electrochemically active surface area (ECSA), which was regarded to be directly correlated with the active sites for electrochemical reaction 47 . The ECSA of the electrode could be calculated by using the following equation, where C dl is the double-layer capacitance of the electrode, C s is the specific capacitance of the material. The C dl s of various electrodes were determined from the CV curve obtained in a non-Faradic region (0.1-0.2 V) in 1 M NaOH solution ( Fig. S5 and Fig. S6a). Since C s s of various materials are unknown for, the ECSAs cannot be determined. The current density was therefore normalized to C dl by assuming that all materials have the same C s (Fig. S6b). Obviously, the Ni-MOF/NF electrode had the best performance, followed by Ni 2 P@Ni-MOF/NF.
Since OER was the competitive reaction with UOR at higher potential, the OER performance was also investigated (Fig. 5b). An obvious anodic peak corresponding of the oxidation of Ni 2+ to Ni 3+ was observed in the polarization curve of OER at the potential of 1.40 V (vs RHE). By comparing the LSV curve of UOR to OER, it is clear that the oxidation of urea occurs upon the formation of Ni 3+ species, demonstrating that UOR happens readily than OER. In sharp contrast to UOR, OER occurs at a relatively high potential (1.6 V vs RHE) after the completion of the oxidation of Ni 2+ to Ni 3+ , revealing that the real catalyst for OER is NiOOH 48 . It is noteworthy that the current density for UOR reached 850 mA/cm 2 before the occurring of OER, which is the highest ever reported to the best of our knowledge. Tafel analyses were conducted to compare the UOR kinetics of various catalysts. As shown in Fig. 5c, the Ni 2 P@Ni-MOF/NF electrode exhibited the lowest Tafel slope of 43.8 mV/dec, significantly smaller than other catalysts. The results indicated that the UOR with Ni 2 P@Ni-MOF/NF electrode occurred at faster kinetics than other electrodes. www.nature.com/scientificreports/ The charge transfer resistance is a critical influential parameter for electrochemical reaction, which is commonly negative correlated with conductivity of the electrode. In order to evaluate the charge transfer resistance during UOR, electrochemical impedance (EIS) tests were carried out at 1.42 V. The best fitting was obtained using the equivalent circuit shown in Fig. S7. As shown in Fig. 5d, all of the Nyquist plots exhibited a semicircle shape. The smaller diameter of the semicircle, the lower the charge transfer resistance. Obviously, the Ni 2 P@Ni-MOF/ NF electrode is not the smallest, which is slight larger than that of Ni 2 P@NiC/NF and NiC/NF. The fitting results indicated that the Ni 2 P@NiC/NF had the smallest charge transfer resistance (Table S2), clearly demonstrating that high temperature carbonization and phosphidation was the best way to enhance the conductivity. On the other hand, the results also indicate that conductivity is not the only determinant factor for UOR.
Chronoamperometry test was conducted at 1.4 V (vs RHE) to investigate the durability of Ni 2 P@Ni-MOF/ NF electrode. As shown in Fig. 5e, the current density gradually increased from 42 to 50 mA cm −2 in the early stage, which may be attributed to the activation of electrode. At the end of the stability test, the current density declined by only 1.9%, demonstrating the superior stability of the electrode. UOR involves gas evolution, the gas bubbles formed on the surface of the electrode need to be released promptly to avoid the blocking of the active sites 49 . The blocking effect will be exacerbated at constant applied potential, which leads drastic fluctuation of the current density 50 . It can be clearly seen from Fig. 5e that the current density fluctuation was negligible, suggesting the swiftly gases bubbles detachment from the electrode's surface. This fast detachment of gases bubbles was credited to the surface properties of the electrode and the porous structure of the nickel foam. The interconnected macroporous structure of nickel foam allows fast detachment of tiny gas bubbles, enabling it an ideal platform to prepare advanced electrodes for gas evolution reaction 51,52 . The LSV test was conducted again after durability test. As shown in Fig. 5f, the LSV curve overlapped with that collected before durability test. The results verified the excellent longevity of the electrode for UOR.
The Ni 2 P@Ni-MOF/NF electrode after stability test was first characterized by using XPS. As shown in Fig. 6a, the XPS peaks corresponding to the Ni δ+ species decreased dramatically after test. And the XPS peaks related to Ni 2p shifted to higher binding energy level. The result indicated the oxidation of Ni species to higher valance state in the material. No change was observed for N 1s peak (Fig. 6b) www.nature.com/scientificreports/ energy level, indicating the charge transfer form P atoms to neighboring atoms (Fig. 6c). The shifting to high energy level was also observed for the XPS peak corresponding to C=O (Fig. 6d). Overall, the XPS result indicated the occurrence of the oxidation of surface elements of the electrode. The morphology of the catalyst was further characterized by using SEM and TEM. As shown in Fig. S8a, no apparent morphology deterioration was found after test. And the lattice fringes of Ni 2 P could also be resolved as shown in the TEM image (Fig. S8b). The elemental mapping confirmed that the main components of the catalyst was still the same (Fig. S8c).
HER performance. Nickel phosphides such as Ni 2 P and Ni 5 P 4 have been extensively explored as electrocatalyst for HER 53,54 . So, the HER performance of Ni 2 P@Ni-MOF/NF electrode was evaluated by using LSV technique in 1 M NaOH containing 0.33 M urea. From Fig. 7a, Ni 2 P@Ni-MOF/NF electrode had apparent better electrocatalytic activity for HER compared with other electrodes. The overpotential at current density of 10 mA cm −2 (η 10 ) follows the order: Ni 2 P@NiC/NF (57 mV) < Ni 2 P@Ni-MOF/NF (6 mV) < NiC/NF (78 mV) < pNF (144 mV) < Ni-MOF/NF (165 mV) < NF (200 mV). The result clearly manifested that phosphidation and carbonization could dramatically boost the HER performance of the Ni-MOF. Although the η 10 of Ni 2 P@Ni-MOF/NF was larger than that of Ni 2 P@NiC/NF, the performance of electrode at large current density was inferior to the latter one. Moreover, the mechanical properties of Ni 2 P@NiC/NF were very poor, material fragments kept falling off the electrode during the electrochemical test. The HER activity of the Ni 2 P@Ni-MOF/ NF electrode, in terms of the η 10 value, is lower than many state-of-the-art NF based HER electrodes such as NiFe-MOF/ NF (134 mV) and Ni 2 P/Ni/NF (98 mV) (Table S3) 18,[55][56][57][58][59][60] . The Tafel slope of Ni 2 P@Ni-MOF/NF was 42.2 mV dec −1 (Fig. 7b), which was much smaller than other electrodes, NF (98.1 mV dec −1 ), pNF (79.3 mV dec −1 ), Ni-MOF/NF (76.6 mV dec −1 ) NiC/NF (39.3 mV dec −1 ) and Ni 2 P@NiC/NF (31.3 mV dec −1 ). The results indicated that the Ni 2 P@Ni-MOF/NF electrode had faster HER kinetics than other electrodes. Since the Tafel slope was smaller than 80 mV dec −1 , Heyrovsky step was suggested to be the rate-determining step of HER with Ni 2 P@Ni-MOF/NF electrode 61 . The Nyquist plots of various electrodes for HER are presented in Fig. 7c. The fitting parameters of impedance spectra of HER are summarized in Table S4. It can be clearly seen that the charge transfer resistance of the electrodes follows the same order as that for UOR. The results evidently confirmed that direct phosphidation process could significantly boost the conductivity of Ni-MOF. The electrochemical stability of Ni 2 P@Ni-MOF/NF electrode for HER was further tested by chronoamperometry at − 0.1 V (vs RHE) for 10 h. As shown in Fig. 7d, the current density gradually increased from − 25 to − 20 mA cm −2 in the first hour, which may be attributed to the slowly buildup of hydrogen gases bubbles on the surface of the electrode. At the end of the stability test, the current density still remained about 79.3% of the initial value. No obvious deterioration was observed for the LSV curves before and after stability test (Fig. 7e), substantiated the excellent durability of the electrode 62-65 . www.nature.com/scientificreports/ Urea electrolysis. Given the fact that Ni 2 P@Ni-MOF/NF electrode has excellent performance for both UOR and HER, we constructed an electrolyzer with Ni 2 P@Ni-MOF/NF as bifunctional electrode for urea assisted hydrogen production. As shown in Fig. 8a, the electrolyzer could deliver a current density of 100 mA cm −2 at 1.65 V for urea electrolysis. While a much higher potential of 1.91 V was required to drive water electrolysis at the same current density. The result clearly demonstrated the advantage of urea electrolysis over water electrolysis for hydrogen production. Compared with other reported electrodes, the Ni 2 P@Ni-MOF/NF was among the best ones for urea electrolysis (Table S3). The long-term stability for urea electrolysis with Ni 2 P@Ni-MOF/NF electrode at a current density of 10 mA cm −2 by chronopotentiometry (V-t). In Fig. 8b, the voltage increase in the first hour was attributed to the buildup of hydrogen gas bubbles on the surface of the cathode, consistent with that observed in Fig. 7d. At the end of the 20-h longevity test, the applied potential increased by only 1.1%, demonstrating the excellent longevity of the sample as bifunctional electrode for urea electrolysis. No appreciable change can be observed for the LSV curves obtained in 2-electrode configuration before and after stability test (Fig. 8c), advocating the excellent stability of the electrode. The results showed that Ni 2 P@Ni-MOF/NF was a promising bifunctional electrode for urea electrolysis.
Conclusions
In summary, Ni 2 P@Ni-MOF nanosheets were successfully grown on nickel foam through direct phosphidation of Ni-MOF nanosheets. The Ni-MOF structure was partially preserved as confirmed by XRD, TGA and FTIR characterization. The self-supporting Ni 2 P@Ni-MOF/NF exhibited excellent electrochemical performance for both UOR and HER. It only required 1.41 V and 66 mV (vs RHE) to deliver a current density of 100 mA cm −2 for UOR and 10 mA cm −2 for HER, respectively. The excellent UOR and HER performances of the Ni 2 P@Ni-MOF/ NF were attributed to both the enhanced conductivity and the fast release of the gases bubbles from the surfaces of the electrode. The electrolyzer constructed with Ni 2 P@Ni-MOF/NF as both anode and cathode could deliver a current density of 100 mA cm −2 in 1 M NaOH with the presence of 0.33 M urea at 1.65 V, which was 0.26 V lower than water electrolysis. Furthermore, the Ni 2 P@Ni-MOF/NF also demonstrated excellent longevity for urea electrolysis. Considering the low cost, easy preparation, long term stability and excel activity, Ni 2 P@Ni-MOF/NF electrode could be a promising bifunctional electrode for hydrogen production through urea electrolysis and to retrieve energy from urea-rich wastewater.
Synthesis of Ni-MOF/NF. Typically, 0.188 mmol NiCl 2 ·6H 2 O and 0.375 mmol BDC were dissolved in a mixture solvent containing 16 mL DMF, 1 mL ethanol and 1 mL water to form a clear solution. The NF was cut into small pieces (2 cm × 3 cm) and sonicated in 2 M HCl for 15 min to remove surface oxides. After washed with copious water and blown dry with pure nitrogen gas, it was then transferred into a Teflon-lined hydrothermal reactor with the above-mentioned solution. Subsequently, the hydrothermal reactor was heated in an oven at 140 °C for 48 h to grow Ni-MOF on NF. Finally, the sample was rinsed with ethanol and ultrapure water thoroughly and dried in air for later use.
Synthesis of Ni 2 P@Ni-MOF/NF. To synthesis Ni 2 P@Ni-MOF/NF, the Ni-MOF/NF and NaH 2 PO 2 with a mass ratio of 1:4 were placed on both sides of a porcelain boat. The porcelain boat was then put in a quartz www.nature.com/scientificreports/ tubing that housed in a tube furnace. The NaH 2 PO 2 was placed on the upstream of the gas flow. Afterwards, the phosphidation process was carried out for two hours at 300 ℃ under Ar flow. The sample was labelled as Ni 2 P@ Ni-MOF/NF. For comparison, the Ni-MOF was fist calcined at 600 ℃ for 2 h under the protection of Ar to get NiC/NF, which was subsequently phosphidized using the same procedure to obtain Ni 2 P@NiC/NF. Direct phosphidation of NF was also performed using the same procedure and the electrode was labelled as pNF.
Characterization. The morphology of the material was characterized by using scanning electron microscopy (SEM, TESCAN MIRA 3, Czech) equipped with an energy-dispersive X-ray spectrometer (EDX) and transmission electron microscope (HR-TEM, JEM-2010, Japan). X-ray diffraction patterns (XRD, Ulitma IV, Japan) were obtained on a PANalytical XPert instrument with Cu Kα radiation (λ = 0.1542 nm). X-ray photoelectron spectroscopy (XPS, Thermo ESCALAB 250XI, America) was used to study the composition and chemical state of the samples using an Al Kα X-ray source, and the binding energy was calibrated according to the reference C 1s peak at 284.6 eV.
Electrochemical measurements. Except the test with the electrolyzer, all electrochemical measurements were performed in standard three-electrode configuration on a CHI 760E potentiostat with Ag/AgCl and graphite rod used as reference and counter electrode, respectively. The potentials were reported against reversible hydrogen electrode (RHE) scale by converting the measured potential using the following equation, Electrochemical impedance spectra (EIS) were recorded in the frequency range of 100 kHz to 0.1 Hz at a voltage amplitude of 5 mV. The EIS spectra were fitted to obtain the charge transfer resistance, R ct . The electrochemical double-layer capacitance (C dl ) of electrodes were evaluated by using cyclic voltammetry at scan rates of 20, 40, 60, 80, 100 and 120 mV s −1 in a non-Faradic potential range. | 5,612.6 | 2021-11-01T00:00:00.000 | [
"Materials Science"
] |
Sequence Analysis of Bitter Taste Receptor Gene Repertoires in Different Ruminant Species
Bitter taste has been extensively studied in mammalian species and is associated with sensitivity to toxins and with food choices that avoid dangerous substances in the diet. At the molecular level, bitter compounds are sensed by bitter taste receptor proteins (T2R) present at the surface of taste receptor cells in the gustatory papillae. Our work aims at exploring the phylogenetic relationships of T2R gene sequences within different ruminant species. To accomplish this goal, we gathered a collection of ruminant species with different feeding behaviors and for which no genome data is available: American bison, chamois, elk, European bison, fallow deer, goat, moose, mouflon, muskox, red deer, reindeer and white tailed deer. The herbivores chosen for this study belong to different taxonomic families and habitats, and hence, exhibit distinct foraging behaviors and diet preferences. We describe the first partial repertoires of T2R gene sequences for these species obtained by direct sequencing. We then consider the homology and evolutionary history of these receptors within this ruminant group, and whether it relates to feeding type classification, using MEGA software. Our results suggest that phylogenetic proximity of T2R genes corresponds more to the traditional taxonomic groups of the species rather than reflecting a categorization by feeding strategy.
Introduction
The sense of taste is highly relevant for animal survival, as it probably evolved to provide animals with the ability to differentiate suitable from dangerous foods. There are five basic types of taste in mammals: sweet, salty, sour, bitter and umami. All act through a complex network of chemosensory receptors and signal transducers. Bitter taste iswell characterized for humans at both molecular and genetic levels, but little is known for ruminants, although they have the anatomical structures for taste perception and they make use of this important taste ability in their dietary choices [1][2][3][4]. Herbivores, ruminants included, have long been known to demonstrate preferences for different plant species and for parts of plants within species, so in a sense all herbivores are selective, maximizing nutrient intake and avoiding plant secondary metabolites [5]. According to the predominant type of feed ingested, herbivores can be classified into three feeding types: grazers (bulk and roughage feeders), browsers (selected diets containing at least 75% fruit, dicot foliage, and tree and shrub stems and foliage), and intermediate or mixed feeders (feeders that both browse and graze) [6].
Bitter taste has been extensively studied in various mammals and it is believed that it evolved to avoid the uptake of toxic substances, however, no strict correlation between bitterness and toxicity is observed [7]). At the molecular level, it is known that taste is sensed by taste receptor proteins present on the surface of taste receptor cells. Bitter taste receptors, in particular, are G-protein-coupled receptors (GPCRs) coded by a family of genes, TAS2R, that contain an average of 300 codons, and which are intronless. These characteristics make them easy to detect and analyze by DNA sequencing [3]. The repertoire of TAS2R (or T2R) is well described for several species, and shows rather large differences in gene numbers, from 15 genes in dogs to 54 in frogs, for example [2,3,8]. In the list of best (almost completely) described species are human, mouse [9,10], and chicken, consisting of 25, 34, and 3 functional genes, respectively [11,12]. Only scanty information on ruminant T2R genetics was available for cattle. Recently our team has reported eight T2R genes for sheep, applying a comparative genomics approach using cattle T2R data and the recently available sheep genome, followed by direct sequencing evidence using merino sheep DNA [13]. Using phylogenetic tools, we have also observed higher sequence conservation between the two ruminant species, sheep and cattle, than when comparing ruminants with other mammals.
In this present study, we have performed a larger sequence analysis to a collection of ruminant species belonging to different taxonomic families and habitats. Differrent foraging behaviors and diet preferences are represented within the group: five browsing species (fallow deer, moose, red deer, reindeer and white tailed deer), five grazing species (American and European bison, elk, mouflon, and sheep) and three intermediate feeding species (goats, musk ox and chamois). We aim to explore the phylogenetic relationships of T2R gene sequences within ruminant species. To achieve this goal, we described the first partial repertoire of T2R gene sequences for the chosen species and then studied the homology and phylogeny of these receptors within the ruminant group and in relation to previously described T2R sequences of nonruminants.
We analysed samples of the following ruminant species: American bison (Bison bison), chamois (Rupicapra rupicapra), elk (Cervus canadensis), European bison (Bison bonasus), fallow deer (Dama dama), goat (Capra hircus), moose (Alces alces), mouflon (Ovis ammon musimon), muskox (Ovibos moschatus), reindeer (Rangifer tarandus), red deer (Cervus elaphus), sheep (Ovis aries) and white tailed deer (Odocoileus virginianus). Sheep and goat samples were obtained from animals at the University of Évora and the Faculty of Veterinary Medicine of the University of Lisbon, Portugal, respectively. Reindeer samples were obtained from the Beitostølen region in Norway. Red and fallow deer samples were obtained at the Tapada de Mafra Natural Reserve (Portugal). European bison samples were obtained from animals from the Tatra National Park in Slovakia, while chamois and mouflon samples were obtained respectively in the Biokovo Mountain and Sibenik region in Dalmatia, Croatia. Moose samples were obtained from animals in the Uppsala region in Sweden, muskox samples from the Kangerlussuaq region in Western Greenland, and American bison, elk and white tailed deer from the USDA experimental herd (Ames, IA, USA). DNA from one individual of each species was used in this study. Genomic DNA was isolated from blood samples, using the Qiagen DNeasy Blood & Tissue Kit (QIAGEN, Venlo, the Netherlands), according to the instructions by the manufacturer.
Ethics statement
All the blood samples from which DNA was isolated were obtained during routine health monitoring, by specialized veterinary professionals. No animal experiment was performed; therefore, no specific ethical approval was necessary.
PCR and sequencing
Using the Primer3 software version 0.4.0 (http://frodo.wi.mit.edu/), PCR primers (Table 1) were designed for seven T2R genes (T2R3, T2R4, T2R10, T2R12, T2R13, T2R16, T2R67) that we have previously found in sheep [13]. Amplification was optimized to be able to use one primer set only. T2R genes are intronless, so primers were designed on the exon sequence having the functional sheep T2R gene sequences as template, in order to amplify most of the coding sequence of each T2R gene (800-900bp). Oligonucleotides were synthesized by Stabvida (Stabvida, Caparica, Portugal). PCR reactions using approximately 75ng of DNA were carried out in a Bio-Rad C1000 Thermal Cycler (Bio-Rad Laboratories, Munich, Germany), using standard conditions, as previously described [13]. Sheep DNA was used as positive control. PCR products were loaded on a 1.5% agarose gel to confirm the existence of a unique band with the expected size. The PCR products were purified using the QIAquick PCR Purification Kit (Qiagen, Venlo, the Netherlands), following the instructions from the manufacturer. The purified PCR products were then analyzed by direct sequencing (Sanger method) as a purchased service from Stabvida (Stabvida, Caparica, Portugal) and using the same primer sets as for PCR reactions.
Sequencing data analysis
Sequencing data were manually checked using Chromas Lite 2. [14,15]. Percent identity matrix was also produced by this software.
Phylogenetic analyses
The MUSCLE data obtained in clustal format were used for phylogenetic analyses using MEGA version 6 software [16], available at http://www.megasoftware.net. Only protein sequences of intact genes were included. Sequences of pseudogenes were excluded, as previously described by Dong 2009 [3]. Similarly to Dong (2009), we also selected for the neighbor-joining statistical method of analysis and the bootstrap consensus tree was inferred from 500 replicates [17,18]. Evolutionary distances were computed using the JTT matrix-based method [19]. For a second phylogenetic analysis, protein sequences from T2R orthologues already reported for human and selected non-ruminant animal species of different trophic groups were used. These animal species were: chimpanzee (Pan troglodytes), dog (Canis lupus), horse (Equus caballus), mouse (Mus musculus), pig (Sus scrofa) and rabbit (Oryctolagus cuniculus). Bos taurus was included as reference ruminant species. The sequences were obtained either from Emsembl database (release 73-September 2013, http://www.ensembl.org), or, when not available, by using the GenBank to obtain DNA sequences and then convert them to protein sequence. These sequences were analyzed together with our sequencing data (converted to protein sequences) as input to the MEGA 6 software and the same analysis criteria were used as for the first analysis.
T2R gene amplification and sequencing
PCR products of all seven genes analyzed were obtained for sheep, goat and mouflon, whereas, for the other species a lower number of genes were successfully amplified. PCR fragments obtained were of the expected length for all species using sheep fragments as control. Results of the PCR amplification of the seven different T2R analyzed in the 13 ruminant species are shown in Table 2. The sequencing results for each gene and species are shown in S1 Dataset. These sequences are also deposited in GeneBank (accession numbers KF898049-KF898092). These sequences were then converted to protein sequences and a percent identity matrix produced, excluding pseudogenes to this analysis (matrix presented in S1 Table). The identities are grouped by receptor gene. The primers used in the study were able to amplify sequences ranging from 81-100% in similarity to the ovine genes, from which the sequences of primers were designed. For some of the genes we observed 100% matching of the gene in Ovis ammon musimon to Ovis aries.
Phylogeny
A phylogenetic tree was built with the protein sequences for the obtained intact genes, using Neighbor-joining statistical method of analysis, with a bootstrap value of 500 (Fig 1). There is evident clustering by receptor genes. We also observe that some receptors are closer to each other than others. For instance, T2R4 and T2R16 originate from the same branch, which is separated from the other branch in the root of two other sub-branches, one for receptors T2R3, T2R10, T2R67, and another for T2R13 and T2R12. The phylogenetic relations between species are not constant for every gene, nevertheless, there is a clear trend for phylogenetic distance or divergence of T2R genes to correspond to the traditional taxonomic groups of the species, rather than to feeding types (grazers, browsers or intermediate feeders). Species of the Bovidae family/Caprinae sub-family (sheep, mouflon, muskox, goat and chamois) form a cluster separated from species of the Bovidae family/Bovinae sub-family (American bison, European bison) and the Cervidae family (deer, elk, white tailed deer, reindeer, fallow deer and moose). Analyzing each receptor separately we can see different phylogenetic patterns, with the Cervidade family closer to the root of the branch, and Bovidae further away. However, some interesting exceptions are observed. For T2R10 species of the subfamily Bovinae are closer to the Cervidae family than to the Caprinae sub-family of their own family (Bovidae), and for T2R67, there appears to exist a divergence for each species at a time, not in clusters, albeit keeping the same taxonomic proximities. Finally, O. aries, does not cluster with O. ammon musimon for every gene even though they are of the same genus. For example, in T2R10, O. aries is closer to C. hircus, or even to O. moschatus for receptor gene T2R16. Another interesting finding was that for T2R13 we were only able to find intact genes in the Cervidae samples. We could successfully amplify and sequence PCR fragments for other species but the resulting sequences have premature stop codons, indicatig pseudogenezation of this gene for those species.
A second phylogenetic analysis was performed, extending the comparison with the sequencing data published for the orthologous genes for other animals with even more different feeding Phylogenetic tree built using MEGA software for the sequenced T2R genes in the different ruminant species. The evolutionary history was inferred using the Neighbor-Joining method [22]. The bootstrap consensus tree inferred from 500 replicates [21] is taken to represent the evolutionary history of the taxa analyzed [21]. Branches corresponding to partitions reproduced in less than 50% bootstrap replicates are collapsed. The evolutionary distances were computed using the JTT matrix-based method [23] and are in the units of the number of base substitutions per site. The analysis involved 55 amino acid sequences. All positions containing gaps and missing data were eliminated. There were a total of 96 positions in the final dataset. Evolutionary analyses were conducted in MEGA6 [20]. strategies, including non-ruminants and within distant taxa (S1 Fig). S2 Table shows the alignment of the 103 sequences used and the 80 residues included in the final analyses, when excluding gaps and missing data. The correlation of the T2R genes grouping with the taxonomic relationships was maintained. For example, cattle clustered with American and European bison at all genes for which the sequencing data was available, confirming also this grouping of Bovinae at one branch.
Discussion
In the present study a sequence analysis of T2R genes is presented for the first time for a wide collection of ruminant species. Most of these species have very little genomic data available and we confirm that the approach herein described and developed using data available from other sequenced ruminants is a good comparative strategy to find missing genetic information; at least for features that are coded by considerably conserved genes, such as genes within the bitter taste receptor family. This is in accordance with previous genetic linkage mapping studies made with ruminants and humans [20,21]. In addition to obtaining the DNA sequence of several T2R we also studied the homology of T2R genes within the ruminant animal group and compared it to non-ruminant species, in order to study possible correlations with the different dietary habits of the different species. Our results show more evidence for phylogenetic distance or divergence of T2R genes corresponding to the traditional taxonomic groups of the species, rather than to the feeding types of these ruminants (grazers, browsers or intermediate feeders). Novel sequences for T2R genes are presented for the following ruminant species, building their first partial repertoire of T2R: A. alces, B. bison, B. bonasus, C. hircus, C. canadensis, C. elaphus, D. dama, O. virginianus, O. ammon musimon, O. moschatus, R. tarandus and R. rupicapra. The protein sequence identities among species for each sequenced receptor ranged from 81 to 100%, confirming evident clustering by receptor gene. The results can be also related to the fact that the same sheep gene-based primer sets were used for all species. Possibly genes with lower similarity levels, or with high similarities but containing gaps in the template sequence at the annealing point of the primers, could not be amplified by PCR and, therefore, were not selected for sequencing. We also cannot discard the possibility that some species might simply not have some receptors, as we know that different animal species have different number of T2R and/or different proportions of genes/pseudogenes in their T2R repertoires [2,4]. Interestingly for T2R13 we were only able to find intact genes for the Cervidae samples. We successfully amplified and sequenced PCR fragments for other species but the resulting sequences were pseudogenes. It is well known that some animal species have different number of T2R and/or different proportions of genes/pseudogenes in their T2R repertoires [2,4]. This result also suggests that this receptor in particular may not be necessary to function in some environments or for some species which are not sensitive to the bitter compounds perceived through this receptor.
In terms of phylogenetic relations between the sequences obtained, we observed that there is a correlation between taxonomic classification of the species studied and the obtained receptor sequences, following the general proximity levels of their genomes, i.e., from a phylogenetic analysis the divergence of species of the Bovidae family/Caprinae sub-family (sheep, mouflon, goat, muskox and chamois) from species of the Bovidae family/Bovinae sub-family (American and European bison) and from the Cervidae family (red deer, reindeer, fallow deer,moose, elk and white tailed deer) became obvious. However, we could find exceptions. For T2R10, species of the subfamily Bovinae are closer to the Cervidae family than to the Caprinae sub-family of the same (Bovidae) family. Also O. aries does not cluster with O. ammon musimon for every gene: for T2R10 O. aries is closer to C. hircus, for T2R16 to O. moschatus. We propose that these small alterations might be related to differences in diet or similarities in the environment and/or available food types, more specifically for the bitter substances that are recognized by those particular receptors. A database of bitter taste receptors and corresponding ligands has been built for humans [22]. Such information is not yet available for all these ruminants, thus further studies are needed to predict what food types are more responsible for these differences. However, it is noteworthy that recent studies have shown that in humans very similar receptors may have highly divergent agonist spectra [23,24], whereas dissimilar receptors can have considerable overlaps in their activating bitter compounds [25]. Therefore, sequence similarity level cannot accurately predict for TAS2R specificity. This might explain why we were not able to find a correlation between T2R gene similarity and feeding strategies of the different animals studied. Interestingly, the same correlation of the T2R genes grouping by taxonomic positions was maintained when previously reported sequences of T2R from mammals other than ruminants were introduced to the analyses. For instance, B. taurus genes, the main reference ruminant species for the known T2R gene repertoire, confirmed our phylogenetic results by consistently clustering with B. bison and B. bonasus.
As previously referred to in more extensively studied species, there is an ongoing evolutionary diversification of T2R receptors [26], where differences found in the T2R gene sequences and, therefore, protein sequences among species are likely related with the need to adapt to an environment with a different diet, and consequently food choices. Our study extends this hypothesis to different trophic groups of herbivores. However, we could not find a strong correlation of the sequence similarities with the feeding category of each ruminant (grazers, browsers or intermediate feeder), at least not stronger than the clustering by taxa. For instance, sheep are closer to goat in gene sequences than to European bison, but sheep and European bison are grazers while goat is an intermediate feeder. It is important to keep in mind that feeding types are indicative of feeding habits and that they can have no relationship with the actual content of bitter substances in the feeds. Ruminant evolution may be classified taking the physical and mechanical characteristics of the respective forages into account [27], where intake is related to relative forestomach capacity, and body weight. This classification is not without exceptions. For instance, muskox and reindeer cannot be placed into categories with other species. They have particular feeding habits, as they are adapted to survive and reproduce under the severe constraints of the Arctic. Muskox prefer a diet of graminoids, sedges and dicots [28] whereas reindeer prefer lichens [29]. Such different food sources certainly contain different bitter ligands for different T2R receptors.
We hypothesize that other genetic differences exist, for example at the level of number of genes or ratio of pseudogenes/funtional genes, and are not reflected at the level of ortholog functional genes. The number of (functional) T2R genes seems to be important for taste perception and it has been proposed that carnivores have fewer T2R genes, herbivores an intermediate number, and omnivores the largest T2R gene repertoire [30]. The number of taste genes has also been addressed for other taste receptors in carnivorous mammals. Jiang et al. (2012) have found pseudogenized TAS1R genes and suggested that T2R losses are consistent with altered feeding strategies when they could not detect intact T2R genes in the dolphin genome [31]. Sato & Wolsan (2012), on the other hand, recently hypothesized that factors underlying the pseudogenization of TAS1R1 in pinnipeds may be driven by the specific marine environment to which these animals are adapted, namely the feeding behavior of swallowing food whole without mastication (T1R1 + T1R3 receptor is distributed on the tongue and palate), and the saltiness of sea water (a high concentration of sodium chloride masks umami taste) [32]. Moreover, Li & Zhang (2014) have also proposed that the number of T2R genes in a species correlates with the fraction of plants in its diet, and supported the hypothesis that dietary toxins are the driving force behind the differences in T2R repertoires among species [33]. However, we can only speculate on the number of genes as, with the strategy used in this study, only a portion of the total repertoire of T2R genes of each species is unraveled and compared. For other taste receptors, it has been shown that function can be obtained using alternative strategies when a certain receptor is absent from the genome. For instance, in 2014 Baldwin and colleagues working in hummingbirds demonstrated that the widespread absence from birds of an essential subunit (T1R2) of the only known vertebrate sweet receptor can be substituted by the ancestral umami receptor T1R1-T1R3 heterodimer [34]. Behrens et al. (2014) have also addressed similar problems in birds, demonstrating that the small TAS2R gene repertoire of chicken and turkey compensates low gene number by large tuning breadths [12].
We demonstrate that using PCR primers from a related species within the same trophic group is a good strategy to find T2R genes in species for which the genomes are not yet available. Nevertheless, a strategy using degenerate primers could be used in future works to further complement the present repertoire. The knowledge of these sequences may also be helpful to understand the taste perception mechanisms in these animals, particularly if expression studies will follow. Moreover, data on bitter taste perception in ruminants may have a high impact in animal nutrition with important consequences for the optimization of feed utilization, hence contributing to more sustainable and efficient ruminant production systems.
Supporting Information S1 Dataset. DNA sequences obtained for each T2R gene and species, in FASTA format. Species: sheep (Ovis aries), goat (Capra hircus, reindeer (Rangifer tarandus, red deer (Cervus elaphus), fallow deer (Dama dama), European bison (Bison bonasus), moose (Alces alces), chamois (Rupicapra rupicapra), mouflon (Ovis ammon musimon), muskox (Ovibos moschatus), American bison (Bison bison), white tailed deer (Odocoileus virginianus) and elk (Cervus canadensis). (TXT) S1 Fig. Phylogenetic tree obtained with the sequencing data of T2R genes on our ruminants and the orthologous genes for other animals including non-ruminants, using MEGA software. Protein sequences were used for building the tree. The evolutionary history was inferred using the Neighbor-Joining method. The bootstrap consensus tree inferred from 500 replicates is taken to represent the evolutionary history of the taxa analyzed. Branches corresponding to partitions reproduced in less than 50% bootstrap replicates are collapsed. The evolutionary distances were computed using the JTT matrix-based method and are in the units of the number of amino acid substitutions per site. The analysis involved 103 amino acid sequences. All positions containing gaps and missing data were eliminated. There were a total of 80 positions in the final dataset. Evolutionary analyses were conducted in MEGA6. (PDF) S1 Biologia Experimental e Tecnológica) for their assistance in sample collection. The authors would finally like to thank COST actions FA-1002 Proteomics in Farm Animals (AM Almeida, M Bhide and K Hollung), FA1308 -Dairycare (AM Almeida and C Knight) and TD1101-RGB Net (AM Almeida and V Cubric-Curik), financed by the European Science Foundation (Brussels, Belgium) for rendering possible the fruitful scientific relations that led to this manuscript. | 5,453 | 2015-06-10T00:00:00.000 | [
"Biology",
"Environmental Science"
] |
Fault Management Cyber-Physical Systems in Virtual Storage Model
: On average, every two years, the amount of data existing globally doubles. Software development will be affected and improved by Cyber-Physical Systems (CPS). The number of problems remained even though developments helped Information Technology experts extract better value from their storage investments. Because of poor interoperability between different vendors and devices, countless numbers of Storage Area Networks were created. Network setup used for data storage includes a complex and rigid arrangement of routers, switch, hosts/servers, storage arrays. We have evaluated the performance of routing protocol Transmission Control Protocol (TCP) and Fibre Channel Protocol (FCP) under different network scenario by Network Simulator (NS)-3 Simulation. We simulated the Node Failure and Network Congestion issue with DoS attacks and a counter effect on the Packet Distribution Ratio and End-to-End Delay efficiency metrics with different nodes and speed of node mobility. The study is performed for the Simple Network Management Protocol (SNMP) on FCP routing. The results proved that the proposed method isolates the malicious and congested nodes and improves the Network’s performance.
Introduction
Virtualization of networking resources offers the service providers an innovative and automated means to provision networks for uninhibited growth and beyond physical boundaries. Simultaneously, the networking industry's discovery and evolution became feasible by the tremendous potential offered by decoupling data and control planes within the fold of traditional networks. The commencement of vendor locked networks' end indicates this shift, and it subsequently creates a platform for robust, manipulative and economic software-based tools that can perform related operations. As a result, vendor shipped boxes that carry out networking operations will no more be used in any networking invention and get outdated in every short time, enforcing the service provider for an expensive updation [1]. The world economy went through a quick transformation from production to services during the last 20 years. The service industry contributed 80% to the US economy and the remaining 15% by manufacturing, and 5% by agriculture. Cloud computing is the major contributor to the service industry and takes business computing to a new paradigm shift. By 2013, the global revenue in cloud computing is predicted to touch $150 billion from the $59 billion recorded in 2009. The applications can be accessed and configured at a minimal cost from any place based on the users' Quality of Service (QoS) parameters. Capital investments in hardware to execute the services of developers or incurring the operating charges of humans is redundant for developers with captivating brains for new internet services. The benefits offered by the cloud removes the low-level burdens of Information Technology (IT) companies in setting up hardware and software infrastructures. Thus, the users will be freed and drift their attention on invention and enhance the business value for their most preferred computing services [2].
Traditional storage management models were unsuccessful in influencing data management effectively, which was expected despite the increasing capabilities of storage systems. So, the overall result was over-provisioning and underutilization of storage resources [3]. The reality has always been different, but the promise made was fast and led to a shared storage model that would simplify storage requirements and data management. Based on the following findings, a modern data management strategy should be established to solve the real problems of data management and the complexities of conventional storage management as (a) The data administrator-the owner of the data, is the best Data management performer, (b) The physical storage administrator must manage the data administrator and (c) It offers software to summarize storage infrastructure information.
With the spread of the Network and the invention of technologies like Storage Area Network (SAN), Network Attached Storage (NAS), Directly Attached Storage (DAS), etc., data got scattered on the Network, and it became impossible to handle such an enormous sprawl of data. Daily emerging servers with Terabytes of storage provide high-capacity storage, but the task is to manage the data floods. Data of a node can be residing on various other nodes on the Network, but the question is how to recognize, access the data as a single unit, and make it feel that it still owns it [4]. Fig. 1 shows three locations where Storage Virtualization resides, namely Host/Server based, Network-based, and storage-based Virtualization. All the above-stated parameters work together for the storage virtualization environment and are interdependent as well. If any of the four stops working, no virtualization is present. If any parameter performance degrades, it will degrade the performance of the overall model [5].
There is a practical case: If storage is taking too long to respond because of any reason or any parameter discussed above, it is a violation of Virtualization. In today's rapidly growing world network, increasing complexity is unmanageable. It is becoming more and more prone to attacks like DoS and physical faults. So, in this paper, we propose a model for Virtualization in which the network layer will be equipped with a fault tolerance mechanism. Any fault or any problem on the Network will lead to the degradation of the virtualized model's performance. The cases at a particular situation may be responsible for the degraded performance by the Network: Storage virtualization describes the performance evaluation of the routing protocol in the model and checks the vulnerability of wired and wireless routing protocol. Fault tolerance is a crucial consideration for the Network's functionality. The SNMP [6] based fault tolerance system essentially switches the path to the localhost network to monitor the Network Interface Card (NIC) for transferring the packet. Both hosts are told on a network that NIC will receive the packet for a given host. Implement this fault tolerance scenario on protocols associated with Virtualization, i.e., TCP, FCP, etc. We can overcome the drawbacks and improve the performance of the virtualized environment. The method of equal load distribution regarding a threshold load is beneficial for getting optimal throughput. As a result, a method will be highly considered for this objective, i.e., to accept the Network's additional load. Storage Area Networks are used along with many algorithms and application, which are already available. As the data is spread out on many systems on the Network, optimization will be implemented on the network Layer and on the way, data is stored Network itself. A Fibre Channel (FC) transfer between the host and the storage virtually redirects all requests to the recipient. This approach does not depend on the operating system; indeed, it is unknown to the host operating system. This is true Virtualization through interoperability exists between our switch and storage array [7]. We evaluated the performance of routing protocol TCP and FCP under different network scenario by NS-3 simulation. There are various factors of data losses and performance degradation in the Fiber Channel network. Here, we have simulated the Node Failure, Network Congestion and Packet Loss issue and how many packets are missing inside the network. This paper simulated SNMP on FCP routing and proposed its Logical Unit Number (LUN) and node assignment method in case of failure or overloaded node and determining network performance under different network scenarios.
Overview of storage virtualization, fault tolerance deployment model and different services are presented in Section 1. Section 2 contains a background study of exited work done by a different researcher. Section 3 contains the preliminary study of current work-Section 4 compares the performance analysis of TCP and FCP routing protocol for various constraints. Section 5 discusses the implementation of SNMP with FCP only, LUN and node assignment method in case of failure or overloaded node and determining network performance under different network scenario, including simulation details and results. Section 6 outlines the conclusion remark and future direction.
Literature Review
As a decentralized distributed environment, Virtual Storage Network (VSN) provides a framework in which IT firms and government agencies worldwide use the storage Network. In a rapid deployment of software for many reasons, including managing 'sensitive software is essential to the critical operating data. A stable VSN must be built to support this mission-critical software and data [8]. A series of preliminary ideas presented are the basis for defining the event-oriented architecture proposed in this work. This work makes a particular focus on firewalls and routers [9]. Similarly, a fault tolerance solution for Voice over Internet Protocol (VoIP) is implemented using event-oriented middleware retrieving and saving states when a failure occurs [10].
The Author's [11] proposed network is linked by several factors: phase behaviour in local or transient explosions reveals differences in traffic inoculation behaviour, locality of reference data, missed application behaviour, and competency of the application phase. Scheduling problem-if several core programmes are operating concurrently on-chip until planned. Hence, there arose a multi-dimensional optimization issue, which exploits scheduled stability, decreases efficiency and resources. The Author suggested a multi-core network chip extract facility from the Multi-Core Chip Framework. The Chip network addresses the problem of linking several nuclei with a singlechip device.
The Author [12] proposed an on-chip network, a modern concept that is more advanced than in-circuit architecture. This paper includes features and procedures for Network-on-Chip (NoC) and its systems' tolerance for faults. The fault model and topology in this research were taken into account and clarified. During the transmission process, particular problems occurred with faults, and some of the faults stayed a lifetime, and such faults can be diagnosed using fault-tolerance models. The invented algorithm [13] of the C4.5 decision constructs a series of rules for classifying test cases into multiple separations so that missed test cases most likely fell short due to the same fault in the same partition. Depending on the inputs and outputs of the test case, distinct conditions for failures in test cases can be identified directly in the tree of judgment that reflects a rule modelling of distinct failure conditions. It is likely triggered by multiple failures, which lead to a specific probability of failure. To distinguish the reports using a heuristic similar to Tarantula, each partition is placed based on the declaration scope of both the failed and positive test cases. Such individual rankings are then compiled for a single ranking to be analyzed to evaluate the shortcomings.
Author [14] proposed a fault position approach based on the Back-Propagation (BP) network, one of the most commonly implemented variations in the neural network. A neural BP network has a basic structure, making computer programmes easy to execute. They are used to create a BP neural network to teach the network how to connect. Simultaneously, BP neural networks approximate complex non-linear functions. Analytical details are provided for each test case, along with the relevant success or failure. The coverage of many simulated events, each covering only one record within the programme, are then entered into the qualified BP network, and the output can be called the probability of each bug statement.
There are many probabilistic techniques applied to various debugging problems. In this work, they explained the Traceable Fault Location Model (TFLM) [15], which can be learned from the data, and probabilistically infer the location of the error. TFLM is trained against a corpus of error programs that will learn to recognize recurring error patterns. There will be a distribution of links with rich dependency structures, which are often computationally challenging to handle. They evaluated the fault location performance of TFLM, which used the Tarantula score as a feature in the probability model. Their research explains that learning TFLM isolates errors more effectively than previous statistical methods or directly uses Tarantula.
Author [16] proposed an infrastructure based on the interception of routine calls, and the fundamental idea is to introduce transparent middleware rent applications so that no modifications to the operating system or recompilation of the application are required. The proposed infrastructure is scalable and offers fault tolerance, proactive dependency, an adaptation of resources based on failures, and rapid fault detection and recovery. In this work, there was difficulty in providing a complete solution for fault-tolerant RT-CORBA real-time applications.
Preliminary Studies
This section presented the pre-elementary work, which is necessary to understand the presented work. Storage Virtualization was the technology of present need when data produced every day is growing exponentially. This technology gives a virtual yet effective way to store, manage, and extensive access data. Storage virtualization can be implemented in four areas, and for this research work, we have opted for the network layer, which comprises of Fiber Channel, a very high-speed network component fulfilling the high-speed data transfers. The Network is subject to scale with the number of nodes added, and risk also increases that some of the nodes in the path may be malicious or crashed, violating the promises made by Storage Virtualization [17].
The Fiber Channel Network
The network layer in Virtualization is the backbone process that combines both software and hardware network resources and the network functionality itself into a single entity called virtual networks, a software-based administrative model based on a fiber channel that works on FCP. The network layer for Virtualization includes network hardware such as switches and network adaptors, generally known as NIC, some network elements such as firewall or load balancers, Networks itself such as virtual Local Area Network (LANs) and containers like virtual machines, some storage devices on the Network and any network medium such as Ethernet or Fiber Channel (FC). There are other constraints for Virtualization to happen. In a virtualized environment, the host and the client are unaware of the Network they are using, the topology, the protocol, and other issues like storage devices. Here, the transport layer must transmit the data from host to client, and so, a protocol supporting the transport layer is also required. TCP is the most reliable protocol as it is a connection-oriented protocol, so it is usable in our paradigm [18].
Virtualization Architecture
There are three implementation approaches where Virtualization can be implemented. Choosing any one of the architectures out of available ones entirely depends on the need of the storage administrator and the storage resources themselves. However, the approach and implementation in this paper are based on network-based storage virtualization [19].
Host-based Virtualization
Host-based virtualization is a host-based programme; for example, Sun ZFS 7000 should be the host or process' fundamental or primary mission. Some operating systems have built-in volume control and are sold as a standalone feature in some others. LUN are supplied to the host machine with the normal physical device driver. Over the Disk Creation Driver sits an application layer that manages the I/O requests, mapping and provides metadata analysis.
Storage Device-Based Virtualization
A combination of multiple disks forms a virtualization layer called Redundant Array of Independent Disks (RAID). Indeed, they are just a collection of simple data storage devices and do not provide any virtualization on their own. Instead, they are presented as a single unit. The most straightforward disk arrays offer physical abstraction logic. RAID schemes bind several storage disks in one array and split the array into smaller volumes or fragments. In general, these tools do not offer advantages such as data transfer, refractive or heterogeneous storage replication since every vendor has its proprietary protocols.
Network-Based Virtualization
SANs are connected to a network-based system via Internet Small Computer Systems Interface (iSCSI) or FC networks to implement network-based storage virtualization. The most widely used network-based systems are available on the market.
Constraints of Virtualization
Four parameters are responsible for storage virtualization, and each has its place in the hierarchy and its work. Although the approach used here focuses only on the network first level, let's have a glimpse of all components together and their contribution to virtualization shown in Fig. 3.
Proposed Work
In a virtualized environment, the two issues are congestion and node/route failure. Congestion is the overloaded receiver and is a type of Denial of Service (DoS) attack. In an FC, two nodes communicate with each other via their FC port. There are certain conditions when the client FC port is busy or crashed, and hence there will be a delay in the network setup as the sender will wait for the client's FC port to become free. The proposal is to avoid network setup requests of FCP, i.e., F_CTL Frames, to broadcast to the Network to avoid crashed nodes. This can be done by using a network management strategy provided by the Simple Network Management Protocol, given that all nodes on the Network under consideration are SNMP enabled. The theory is that SNMP collects all of the information of the nodes within the network using its GET command and sets up a list of nodes on the network. If any crashed node is found in the path, GET_NEXT will find the next nearest node creating a path to the previous destination.
Method Used for Connection Establishment
A control protocol originated from Transmission Control Protocol/Internet Protocol networks is an SNMP. SNMP is described by a series of Comment Requests, where information exchanged between systems is specified and structured. Agent SNMPs live on management systems. The agent receives requests for information to be retrieved or changed by referring to Management Information Base (MIB) objects. The MIB objects are information units that provide the management system with information about the system and the Network. A subagent allows other MIB objects to be added dynamically without changing the agent. Simultaneously, these dynamic additions do not affect SNMP managers because they continue to work directly with the SNMP agent. A requested protocol is the type of protocol that SNMP uses to transfer data between the SNMP manager and user. The manager demands details of the MIB object, and the agent uses TRAPS to satisfy this order. Fig. 4 demonstrates an FC network that is SNMP enabled, which means all nodes on the Network have a mechanism to generate and receive SNMP commands and respond accordingly. All nodes communicate via FC port. FCP works on the mechanism of TCP, and so does all methods of TCP are used for connection establishment. The node discovery process in FCP is connection-oriented and communicates on a one-to-one mechanism. In the above diagram, SNMP is used for network discovery. Let us consider that Node_A want to communicate with Node_E, so the path traverses from node CXE. There is no other possible path, and Node_A generates a GET command to gather the route information. In reply, each participating node generates a TRAP command except those which are malicious. So here, Node_X doesn't generate a TRAP reply, and that is well recognized by Node_A and updated in its MIB base to avoid this corrupted node for the subsequent connection establishment. It will generate the GET_NEXT command at Node_X to get the address of the next nearest node, i.e., Node_E.
Detection of Malicious Nodes
An SNMP admin demands configuration information on a particular device. This request is compiled into a GET protocol data unit by the manager and transmitted to the agent through a communication service. The agent then pulls the requested MIB object information into a RESPONSE PDU after receiving the manager's request and returning it to the manager. The detection module first issues an SNMP command to check whether all nodes are SNMP enabled or not. If YES, then we will enter the SNMP module with a source and destination, leaving path constraints to the SNMP itself and if NO, then normal FCP will work for the rest of the session. SNMP module proposed here is only partial, and the capabilities of updating route tables alone are used. If all nodes are compatible, then, GET command is issued, which will create a routing table for that particular session. MIBs containing route details are stored in an array with a list of all nodes, their addresses, their status, 1 for available, and 0 for malicious nodes. Each time the GET and GET_NEXT command is issued; the routing table is updated with new MIB objects. If CURRENT_NODE is alive, i.e., equal to 1, then an FCP session will be initiated and led to normal FCP transfer; otherwise, the next nearest node is called GET_NEXT be new CURRENT_NODE. GET_NEXT will retrieve the following address from MIB Base, which is dynamically updated. There is verification; if TRAP doesn't respond means the route has a malicious node, then the route will be tested and updated; and will be forwarded to issue a new GET command. This verification is done by TRAP; if TRAP responds in PDU, then again, standard FCP transfer will start.
Proposed Algorithm
Step 1 Call procedure SNMP (GET, TRAP, GET NEXT) Step 2 Input and Output Variables: GET, TRAP, GET NEXT.
Step 3 For irritation i = 0 to nodes on FCP Header . . . Do Step 4 If (node == to agent) then Step 5 Write to MIB Base (); Step 6 MIB = get nodes (); Step 7 End If Step 8 If (Current_Node == 1) Step 9 Issue TRAP // for data access Step Step 20 Else Step 21 Return; Step 22 FCP Session Initiated Step 23 Normal FCP Transfer Step 24 End If Step 25 End
Performance Analysis of TCP/FCP in a Virtualized Environment
The Basic principle for data access in a storage virtualization environment is based on TCP. Storage virtualization technology uses FC as its backbone and FCP as its working protocol over a fiber channel, and data are transmitted in the form of packets from one source to another destination. Still, these packets are split into a stream of bytes when they are to be transmitted over the Network. FCP works on the mechanism of TCP, or we can say FCP is TCP, supporting much higher data rates.
Simulation of Peer-to-Peer-Setup
Here, we set up a simple simulation of how the data packets between two nodes are transmitted and determine the Network's performance. A network is created with the following simulation parameter in the NS3 platform, as shown in Tab. 1. The above implementation is run on the simulator for 15 s. Analysis of the ratio and performance of TCP in an ideal environment is done. As shown in Fig. 4, ideally, TCP gives 100% packet delivery, which will be not feasible in coming implementations. EED is ignored in Fig. 5.
Node Failure at a Particular Time
In a network, nodes are prone to crash anytime and hence can interrupt the data transmission. If any intermediate node crashes, then TCP will find an alternate path but, in this setup where only two nodes are available, data transmission will be interrupted. It can be noticed that data rates are maintained ideal in all three cases. Here, the receiver crashes at a given time interval, and the sender doesn't get Acknowledgement (ACK) and retransmit in proper time; the final session gets closed. The result made the receiver crash at 3, 5, and 8 s, keeping the rest of the parameters the same as that of simulation 1 and found that the Packet Delivery Ratio (PDR) is still 1. And because the number of packets sent by the sender is received, and then no packets are transmitted as the session expires, shown in Fig. 6.
Network Congestion by Artificial Load Generator
In this scenario, a TCP grid is set up, and a synthetic load generator is used. Packet size is reduced by 50% on each session established. A synthetic load generator is used, which will produce an unknown number of fake packets and will increase the load on the Network, and a condition will come when original packets are lost. A random load is generated along with 15 nodes present on the Network, as shown in Fig. 7, and a gradual packet loss is seen. As the simulation keep running, the amount of load generated increases. PDR is decreasing because of two reasons: first, the amount of data sent decreases by 50 bps, and second, artificial load keeps increasing, and both end up being 0.
DoS by Receiver
DoS attacks are when a receiver cannot get the data packet sent by the sender or unable to generate ACK for the received packet. In this work, 15 nodes are included in the network, and the simulation is executed for 15 s while maintaining the previous simulation's metrics. The receiver node is made to crash at certain time intervals, causing it to receive data and produce data loss at that particular time. It should be noticed here that the TCP receiver crash will not generate ACK, but the sender is provided with a false ACK. TCP Receiver crashes at time intervals, as shown in Fig. 8. A TCP connection is permanently closed at these intervals, but an untrusted node acknowledges these data packet to be received and continues to maintain the TCP connection with the help of ACK. These packets were lost at the receiver's end, and new packets are received, which are acknowledged. Data packets sent at each time slot are Constant Bit Rate (CBR). Only one less packet is sent at the last time interval and the total packets sent are 999. At time intervals 2, 3, 7, and 8 each, one packet is lost. At time 4 s, 3 and 2 packets are lost, making 9 packets lost in the whole simulation. It can be noticed that PDR is relatively low as compared to that of the previous simulation because of the ignorable load at the receiver.
e. Summary of TCP Simulations
Tab. 2 summarizes the PDR of each simulation scenario we have performed until now for TCP.
Performance Evaluation FCP
FC networks have very high speed where two nodes communicate via their FC port. As data transmission rates are very high, the amount of data sent every Sec. is also high. Considering the case with TCP in the previous section, if FCP stops for just a second, the amount of data loss will be much higher than TCP. So, in this section, we have simulated various scenarios for FCP and created a network with the following Simulation parameter (Tab. 3) for performance evaluation of FCP based on PDR and EED. This is because high data rates provide high-speed data transmission, but the risk of network/node failure also increases the data loss.
Fixed Memory Sized Receivers
We created a scenario in which 15 nodes involved in the Network, with each node having a maximum of 10 packets. FCP assigns a new node dynamically if the current node is unavailable because of any reason. We transmitted 100 packets with a constant bit rate, and the results are shown in the graph below. Simulation parameters are stated in Tab. 3. As the parameters describe, each node participating in the Network has a fixed size memory, thus can receive only a fixed number of data packets (Fig. 9). When a node is full of memory, the FCP session will be closed with that receiver node, and the sender will look up for next node with available memory and maintain a new session with a new node. A delay will be seen hereafter 10 packets, new nodes are searched for, and a session is to be initiated. It can be noticed that the data rate is continuously maintained the same for all nodes.
Congested Receiver
High data rates of FCP are appropriate for large data transmissions. We set up a scenario where the FC Port of receiving node is already overloaded, and the sender transmits packets at a constant bit rate. There will be a bunch of packets grouped at the receiver's end, making it congested. This may also be because of a low speed receiving buffer. Congestion will keep increasing for each session maintained. Loss is also taken as a parameter here, depicting several data packets lost in each session. As shown in Fig. 10, when the number of data packets sent increases, the low speed receiving end will collect many packets. The amount of sent data packets is kept constant as shown by graph line-Sent. With an increasing number of sessions, a line indicating the received packet falls because of increasing congestion.
Artificial Load on the Network
In this simulation Fig. 11 scenario, a function LOADGEN is created, generating random fake packets on the Network, increasing with each session. Added with the packets sent by the sender, these fake packets will increase the total number of packets at the receiver. With each session, the artificial load keeps on increasing, and so the sent packets. Received packets decrease due to congestion offered by this hiked data rates.
Summary of Results for FCP Based Simulations
Based on the above simulations, all FCP results can be summarized in Tab. 4 based on PDR. The investigations of TCP and FCP by different stimulation parameters and measuring the performance metrics under different network setup. From this comparison, we observed that both protocols have their advantage and disadvantage. TCP is a connection-oriented byte stream providing reliable data transmissions with high PDR but with a limited transmission speed preferred for small data transmissions. On the other hand, FCP is best suited for extensive data as it offers a relatively high speed of up to 100 Gbps, but data loss in case of any fault on the Network is very high than that of TCP. In a virtualized storage environment that works on FC and uses FCP, data losses are unacceptable. In Virtualization, data integrity is significantly similar to data rates.
Performance Analysis with Deploying Proposed Model a. Implementing Fault Tolerance on FCP
The analysis from various models of different scenarios for FCP is summarized and explored in depth. We have simulated the DoS attack and determined the impact on PDR and EED by varying the number of nodes, malicious nodes, and nodes' mobility speed. The attacks were performance metrics. We also simulated the attacks. Tab. 5 shows the simulation parameters used to create the scenarios.
In Fig. 12, it can be observed that when data transmission rates increase, there is a relative fall in the PDR of the Network. This is because the number of packets sent every second rises with the increasing speed, and if a node crashes just for a second, a high amount of data will be lost. When mobility speed ranges from 10-20 Mbps, there is a low loss factor. After this range, loss rises, and hence PDR graph falls rapidly.
The number of malicious nodes in the Network is created randomly between one to five. The distribution ratio of packets decreases as we add malicious nodes in the Network, as shown in Fig. 13. This is a straightforward action plan since the number of nodes in the network raises the risk of an active path being part of the malicious node. Keeping the transmission rate the same, the number of nodes has increased regularly, and the PDR graph is falling. The reason for this fall down is the addition of new nodes, which contains malicious nodes also. As new nodes are stopped adding, the PDR graph observes an increase and will maintain average PDR with time. With these losses in FC in such a given situation, it will be desirable to avoid such circumstances for better PDR. We implemented and created a function named SNMP and tested it with this entire scenario. SNMP function is run in NS-3. The following are some simulations with the FCP_SNMP module. PDR for normal FCP decreases gradually with the increase in malicious nodes, as shown in Fig. 14. With SNMP, PDR decreases, but with a low amount and for a specific time, it keeps drowning; after that, PDR starts pushing up as the MIB base contains a list of malicious nodes and avoids these nodes.
In Fig. 15, it is seen that PDR increases with SNMP, but another issue is the time needed or a delay in the formation and updating of the MIB base. Fig. 16 shows the average end to end delay. For normal FCP, delay first decreases with the increasing number of nodes and then rises. FCP crash means malicious nodes that show minor delay as the number of nodes is very few, not working. FCP_SNMP shows high delay because it first creates a routing table and then gets the new node's address from this MIB base or route table. As discussed earlier, mobility speed affects the PDR, and the Average End-to-End Delay (EED) for FCP is dropped as the mobility speed increases, and then a prolonged increase is observed. Delay for FCP Crash is low because of fewer nodes and negligible data to transfer. FCP_SNMP shows a very high delay because of fast data delivery and quick search and updating of the routing table.
From Fig. 17 and the table above, we can understand that the malicious node and high transmission rate degrade the Network's performance. The FCP protocol uses an FC port to communicate very quickly with another network node. SNMP is used to update the GET route table and to decide network status and topology information. Malicious node blocks the route, triggers ineffective data transmission, and reduces PDR since FCP is waiting to re-establish a connection. We considered performance measures such as an EED, packet delivery ratio and overhead routing to implement detection techniques. Therefore, we have introduced the SNMP module on the source and target node. This method does not change the FCP further functioning but introduces future delays due to the pre-processing. This method is helpful to isolate malicious nodes and corrupt the network route from another transmission protocol. Figure 17: Route discovery against malicious nodes
Conclusion and Future Work
Storage virtualization is the technology of present need when data produced are growing exponentially, says a study. The study simulated the behaviour of the FCP network under various scenarios. Increased data rates also degrade the performance as the number of packets sent per second increases and so the loss. A malicious node can easily interrupt the route generated by the source node and degrade the Network's performance. The proposed method isolates the malicious node and the congested node as well. It improves the performance by creating and maintaining a MIB Base, which can hold the details of the active node and based upon this, the route is set. It is observed that while an increasing number of nodes on the Network increased data transmission speed, an increase in sessions resulted in increased packet loss. Due to these factors, the overall Performance of FCP decreased, which is not appropriate for Virtualization.
In the future, we will work on the network's security constraints and optimize the rest of the three constraints of storage virtualization. It will be a milestone if fault tolerance can be implemented on the Host Bus Adapter (HBA) that connects the devices to the processor.
Funding Statement:
The authors received no specific funding for this study.
Conflicts of Interest:
The authors declare that they have no conflicts of interest to report regarding the present study. | 8,085 | 2022-01-01T00:00:00.000 | [
"Engineering"
] |
Optimal Formulas for Subnational Tax Revenue Sharing
We develop an analysis of optimal formulas for subnational tax revenue sharing for two cases of interest: when local public goods (lpg’s) show spillovers and are inter-regional perfect and imperfect substitutes. Our analysis could be relevant to understand the determinants of feasible competing alternatives for the design of tax revenue sharing systems. Our study shows that: 1) Inter-regional spillovers should be taken into account in the design of formulas for subnational revenue sharing when lpg’s are perfect inter-regional substitutes but should not be taken into account to supply lpg’s that are imperfect substitutes; 2) The distribution of preferences for lpg’s and population should be considered in formulas for revenue sharing when subnational governments provide both lpg’s that are perfect and imperfect substitutes; 3) For local public goods that are imperfect substitutes, the share of tax revenue is first increasing and then decreasing with increases in the population of the district; 4) For perfect substitute-lpg’s, the share of tax revenue in the district is increasing and concave with the district’s population; 5) The distribution of income should not be considered in the design of formulas for lpg’s that are perfect and imperfect substitutes. Other empirically relevant comparative analyses are considered in the paper.
Introduction
Tax revenue sharing seeks to distribute revenue for central and subnational governments from a given tax base.Some of the advantages of tax revenue sharing recognized in the literature include: 1) Tax harmonization be-
Optimal Tax Revenue Sharing for Local Public Goods that Are Imperfect Inter-Regional Substitutes
Consider an economy with individuals living in districts i and -i .In our economy, the central government is responsible for collecting tax revenue and allocating (through the use of formulas) the shares of tax revenue for all districts.For simplicity of the analysis, we assume that local governments use the resources allocated to them by the central government to provide local public goods.In district , , i i ∀ there is a representative individual with an endowment i e and his indirect preferences are given by ( ) − were equal to one then the local public good would be a pure national public good.The budget constraint of the individual is given by ( ) where t is a proportion- al income tax implemented by the central government.In this economy, there is inter-regional heterogeneity of preferences, that is , , and without loss of generality we assume i i e e − > .We consider a benevolent social planner ruling the central government that seeks to maximize the nation's social welfare ( ) where , i i N N − are the populations of districts i and -i .The problem of policy design for the central government is to maximize Ψ by choosing the size of the pro- portional income tax t the budget B, and the formula for sharing revenue in districts i and -i , ( ) ξ ξ − to finance local public goods , i i g g − .The budget constraint of the central government is where B is the central government's budget.The distribution of tax revenue in the economy is determined by formulas for revenue sharing such that the budget constraints of subnational governments are given by i g B ξ = and ( ) ξ ∈ is the share of the budget allocated to finance the local public good of district i .For simplicity of the analysis, districts i and i − only supply, respectively, i g and i g − .In this section, we analyze the case of local public goods that are imperfect substitutes, which are defined by a finite elasticity of substitution, while for the case of local public goods that are perfect substitutes (analyzed in Section 3) the elasticity of substitution is infinite 6 .Formally, the problem of tax revenue sharing for the central government when local public goods are imperfect substitutes is: ξ ξ − for local public goods that are imperfect inter-regional substitutes of districts i and i − are given by: ( ) ( ) Proof 5 Our choice of the utility function is for simplicity of the analysis.Moreover, this type of utility function is common in the literature see [8] and [9], among many others. 6To distinguish local public goods that are perfect vs imperfect inter-regional substitutes, we use the elasticity of substitution ( ) ( ) To see that our preference relation for i g and i g − characterizes local public goods that are imperfect inter-regional substitutes, note that the elasticity of substitution between i g and i g − is positive but finite since ( ) ( ) In the case of perfect inter-regional substitutes, the elasticity of substitution is infinite (see Section 3).
The first order conditions for the government's problem are: Rearrange terms to obtain conditions (3), ( 4) and ( 5).
3) The share of tax revenue allocated in district * , i ξ , also satisfies the following: ξ takes into account the nationwide distribution of social marginal costs and benefits of local public goods.For this reason we can't find an alternative feasible allocation , i i g g − in which we can benefit at least one individual without hurting someone else in this economy.Some other interesting results of our analysis are the following: for a local government in district i supplying an imperfect substitute-local public good, the optimal formula for tax sharing, * ξ , is constant (see proposition 2.2) and it depends only on the distribution of parameters related with the intensity of preferences of individuals for local public goods in districts i and -i , that is , , , , and the distribution of the population in the economy (see condition 4).
An issue of interest for the design of a policy of tax revenue sharing is the relationship of * ξ with the dis- trict's population.Our analysis suggests that . This is relevant for policy design since empirical evidence suggests the use of linear formulas between * ξ and i N (see [10]).However, our analysis suggests that the effect of an increase in the population of district i over the share of tax revenue to be allocated to district i is contingent to the level of * ξ .For sufficiently low values of * ξ (for ( ) an increase in i N should lead to an increase in * ξ .However, for sufficiently high levels of * ξ (for ( ) ) an increase in i N should lead to a fall in * ξ .The explanation of this result is straightforward: from the optimality condition is the net social marginal benefit for residents of district i of increasing * ξ where is the gross social marginal benefit for residents of district i of increasing * ξ (an increase in * ξ leads to a higher level of the local public good supplied by district i which increases the wellbeing of residents of that district).Moreover, a higher * ξ means that ( ) falls which leads to a lower supply of * i g − and the ben-efits for residents of district i from consuming * i g − fall by ( ) leads to a net increase in the social marginal benefits of residents of district i which leads to a higher share of tax revenue allocated to this district (the opposite occurs if ( ) Other interesting outcomes from the characterization of * ξ are: the formula for * ξ does not depend on the distribution of spillovers i k and i k − (see proposition 2.3b).The explanation of this outcome is that the marginal social benefits and costs of a change in * ξ do not depend on i k and i k − (see condition 7).A similar explanation is given to the outcome that the distribution of income should not be considered in the design of formulas for lpg's that are imperfect substitutes since ).Finally, proposition 3 says that an increase in the intensity of preferences of residents in district i and of residents of district i − for the local public good provided by district i (that is increases, respectively, in i γ and i θ ) should increase the share of tax revenue allocated to district i .This is the case because higher values of i γ and i θ increase the nation wide's social marginal benefits of the public good provided by district i .
Optimal Tax Revenue Sharing for Local Public Goods that Are Perfect Inter-Regional Substitutes
In this section, we analyze the case for formulas for subnational tax revenue sharing when local public goods are perfect inter-regional substitutes which means that the elasticity of substitution between i g and i g − is infinite.To distinguish our notation from our previous section, we denote B as the equilibrium level of the budget and ξ as the allocation formula of tax revenue for district i .For this case, is the indirect utility for a resident in district i where , ( ) For this economy, the problem of policy design is 8 : ˆˆ, , ln 1 1 Proposition 3. The optimal budget * B and formulas for distribution of tax revenue ( ) for local public goods that are perfect inter-regional substitutes in districts i and i − are given by: s s N e N e N e s s 7 To see that lpg's are perfect inter-regional substitutes, note that the direct preferences of the representative individual of district i are where , )( )( )
Proof
The first order conditions for the government's problem are: Rearrange terms to obtain conditions ( 10), ( 11) and ( 12).
3) The share of tax revenue allocated in district Proposition 4 shows some comparative static results on the determinants of formulas for revenue sharing for local public goods that are perfect inter-regional substitutes.In particular, * ξ is non decreasing with the population of the district, however, this relationship is not linear since In addition, the distribution of tax revenue for district i is increasing on the spillovers of the public good provided by district , i i k , and decreasing in the spillovers of the local public good supplied by district -, i i k − , (see propositions 4.3.band 4.3c).These last results are very intuitive: a higher level of i k increases the nationwide social marginal benefits of *i g (by the effect of the spillovers of *i g over district i − ) which leads to an increase in the tax revenue allocated to district i .A similar effect occurs with an increase in In terms of the practical significance of this paper, propositions two and four highlights the relevance of taking into account the inter-regional heterogeneity of preferences and the distribution of the country's population in the allocation of tax revenue among sub-national governments, since the heterogeneity of preferences and population are the main determinants of formulas for local governments supplying local public goods that are perfect and imperfect inter-regional substitutes.In particular, one common determinant of formulas for revenue sharing is the district's population.Our analysis shows that the relationship between the funds allocated in the district and the district's population is not necessarily linear (a common practice in some countries, see [1], [10]) and in fact it could also be negative (an increase in the district's population could lead to a lower share of the resources devoted to that district, see proposition 2.3a).Finally, the least intuitive finding of the paper is that the distribution of income in the federation should not determine the distribution of tax revenue sharing among sub-national governments.As mentioned before, in our paper this finding is explained by the fact that both the marginal social benefits and costs of allocating funds to some district in the federation do not depend on the household's income.Although this result should be subject of further analysis to test whether this finding is general or closely related to the parametric functions used in this paper.
Concluding Remarks
In this paper, we study the optimal design of formulas for subnational revenue sharing when local governments provide public goods with spillovers in two cases of interest: when local public goods are perfect and imperfect inter-regional substitutes.Even though, tax revenue sharing policy is implemented in many developed and developing countries there is little formal research on the design of formulas for subnational revenue sharing.In this paper we seek to contribute to fill this gap.Our main contribution is to distinguish the determinants for optimal formulas for subnational revenue sharing when local governments provide local public goods with spillovers that are perfect and imperfect inter-regional substitutes.This analysis has the potential to provide relevant information for policy makers on the determinants and some properties of formulas for subnational tax revenue sharing.
Our analysis suggests the following: 1) Spillovers should be taken into account in the design of formulas for revenue sharing when lpg's are perfect inter-regional substitutes but should not be taken into account for imperfect inter-regional substitutes; 2) The distribution of preferences for lpg's and population should be considered in formulas for revenue sharing when subnational governments provide both lpg's that are perfect and imperfect substitutes; 3) For local public goods that are imperfect substitutes, an increase in the population of the district leads to an ambiguous outcome in the share of tax revenue in the district: for sufficiently high (low) initial values of the share of tax revenue, an increase in the district's population leads to a fall (increase) in the share of tax revenue allocated to the district; 4) For perfect substitute-lpg's, the share of tax revenue in the district is increasing and concave with the district's population; 5) The distribution of income should not be considered in the design of formulas for lpg's that are perfect and imperfect substitutes.Other empirically relevant comparative analysis are considered in the paper.
Although the paper provides insights for the design of revenue sharing systems, it does not consider political institutions (such as electoral competition, the interaction between the executive and legislative bodies), the role of special interest groups and other issues raised by political economy models that might be central in shaping incentives of policy makers.Analysis on revenue sharing systems should also incorporate the role of mobility of households and firms, and adopt a systemic view of the optimal composition of the tax structure and spending of subnational governments and the central government.Future research on this topic should address these issues.
Appendix
3) The share of tax revenue allocated in district * , i ξ , also satisfies the following: . This follows trivially by the fact , , , , , ξ satisfies the following: , it is simple to see that From (4), e) * ξ is non decreasing on i θ and 2 * 2 From (4), Proposition 4. The optimal budget * B and formulas for distribution of subnational tax revenue ˆ, 3) The share of tax revenue allocated in district 3) The share of tax revenue for district i , * ξ , also satisfies the following: Moreover, it is simple to show
e ν ξ is the indirect 1 .
utility function of the representative individual of district i on feasible local public goods (a similar expression is given for The optimal budget * B and formulas for distribution of tax revenue
8
Use the household's budget constraint and the government's budget con- straint to characterize the indirect preferences denoted by The indirect preferences of the representative individual of district -i are given by
Proposition 4 . 4 . 1 )
The optimal budget * B and formulas for distribution of subnational tax revenue ( ) The implied distribution of local public goods,( )
ik−
which makes more attractive to allocate more resources in district i − and reduces * ξ .The distribution of income should not be considered in the design of formulas for lpg's that are perfect substitutes since see condition 4.3c).As in the previous section, the marginal social benefits and costs of a change in * ξ do not depend on i e and i e − (see condition 11).Proposition 4 also shows that * ξ has an ambiguous relationship with the distribution of the intensity of preferences for local public goods.To be specific,
Proposition 2 . 2 . 1 )
The optimal budget * B and formulas for distribution of tax revenue The implied distribution of local public goods,( )
4 . 1 )
The implied distribution of local public goods, Submit or recommend next manuscript to SCIRP and we will provide best service for you:Accepting pre-submission inquiries through Email, Facebook, LinkedIn, Twitter, etc.A wide selection of journals (inclusive of 9 subjects, more than 200 journals) Providing 24-hour high-quality service User-friendly online submission system Fair and swift peer-review system Efficient typesetting and proofreading procedure Display of the result of downloads and visits, as well as the number of cited articles Maximum dissemination of your research work Submit your manuscript at: http://papersubmission.scirp.org/ | 3,963.4 | 2016-07-19T00:00:00.000 | [
"Economics"
] |
Anti-Inflammatory Activity of 4-(4-(Heptyloxy)phenyl)-2,4-dihydro-3H-1,2,4-triazol-3-one via Repression of MAPK/NF-κB Signaling Pathways in β-Amyloid-Induced Alzheimer’s Disease Models
Alzheimer’s disease (AD) is a major neurodegenerative disease, but so far, it can only be treated symptomatically rather than changing the process of the disease. Recently, triazoles and their derivatives have been shown to have potential for the treatment of AD. In this study, the neuroprotective effects of 4-(4-(heptyloxy)phenyl)-2,4-dihydro-3H-1,2,4-triazol-3-one (W112) against β-amyloid (Aβ)-induced AD pathology and its possible mechanism were explored both in vitro and in vivo. The results showed that W112 exhibits a neuroprotective role against Aβ-induced cytotoxicity in PC12 cells and improves the learning and memory abilities of Aβ-induced AD-like rats. In addition, the assays of the protein expression revealed that W112 reversed tau hyperphosphorylation and reduced the production of proinflammatory cytokines, tumor necrosis factor-α and interleukin-6, both in vitro and in vivo studies. Further study indicated that the regulation of mitogen-activated protein kinase/nuclear factor-κB pathways played a key role in mediating the neuroprotective effects of W112 against AD-like pathology. W112 may become a potential drug for AD intervention.
Introduction
Alzheimer's disease (AD) is a complex neurodegenerative disorder with clinical characteristics including memory loss, dementia, and cognitive impairment [1], and represents a very important public healthcare problem with a serious economic burden for the society [2]. There are many contributing factors for AD. The main pathological features of AD include extracellular senile plaques (SPs) containing amyloid beta (Aβ), intracellular neurofibrillary tangles (NFTs) composed of hyperphosphorylated tau protein, the loss of synaptic and neuronal function, and neuronal death [3]. The causes for the vast majority of AD cases are unknown and satisfactory therapeutic and preventive measures for AD are unavailable.
The abnormal and excessive production, accumulation, and aggregation of Aβ is regarded as important causal factors in the pathogenesis of AD. Aβ is a peptide of 36-43 amino acid residues that results from βand γ-secretase-mediated cleavage of transmembrane amyloid precursor proteins (APP) [4]. The deposition of Aβ is considered to
Effects of W112 on Spatial Learning and Memory Abilities in Rats
To evaluate the spatial learning and memory ability of Aβ 25-35 -induced AD rats and the protective effects of W112, the Morris water maze (MWM) test was executed. Figure 2A,B showed that in the place navigation test, the escape latency time of the model group was significantly longer than that of the control group (p < 0.05). Compared with the model group, the escape latency time of W112 groups was significantly shorter (p < 0.05), indicating that W112 could improve spatial learning ability. In the probe test, the numbers crossing the platform in the model group were significantly less than those in the control group (p < 0.05), and the numbers crossing the platform in W112 groups were significantly higher than those in the model group (p < 0.05), indicating that W112 could improve cognition ( Figure 2C,D). These results indicate that W112 exhibits a neuroprotective role against Aβ-induced cytotoxicity. The MTT assay was used to detect the effect of W112 on PC12 cells activity. n = 6. ## p < 0.01 vs. the control group; ** p < 0.01 vs. the model group.
Effects of W112 on Spatial Learning and Memory Abilities in Rats
To evaluate the spatial learning and memory ability of Aβ25-35-induced AD rats and the protective effects of W112, the Morris water maze (MWM) test was executed. Figure 2A,B showed that in the place navigation test, the escape latency time of the model group was significantly longer than that of the control group (p < 0.05). Compared with the model group, the escape latency time of W112 groups was significantly shorter (p < 0.05), indicating that W112 could improve spatial learning ability. In the probe test, the numbers crossing the platform in the model group were significantly less than those in the control group (p < 0.05), and the numbers crossing the platform in W112 groups were significantly higher than those in the model group (p < 0.05), indicating that W112 could improve cognition ( Figure 2C,D).
Effects of W112 on Aβ25-35-Induced Tau Hyperphosphorylation
The tau protein is a principle neuropathological hallmark of AD. The hyperphosphorylation of tau protein seriously damages the microtubule structure and affects the synthesis, release, and transport of neurotransmitters, and eventually leads to the occurrence of AD. In the present study, we evaluated the effects of W112 on tau hyperphos-
Effects of W112 on Aβ 25-35 -Induced Tau Hyperphosphorylation
The tau protein is a principle neuropathological hallmark of AD. The hyperphosphorylation of tau protein seriously damages the microtubule structure and affects the synthesis, release, and transport of neurotransmitters, and eventually leads to the occurrence of AD. In the present study, we evaluated the effects of W112 on tau hyperphosphorylation both in vivo and in vitro via Western blot. Figure 3A-D showed that the levels of phosphorylated tau at thr181, thr205, and Ser396 sites were significantly higher in the Aβ 25-35 -induced cell model (p < 0.01), while W112 treatment reduced the levels of tau hyperphosphorylation (p < 0.01). Furthermore, in vivo, as shown by the results in Figure 4A-D, W112 treatment also reduced the levels of phosphorylated tau at multiple sites induced by Aβ [25][26][27][28][29][30][31][32][33][34][35] in the hippocampus of a rat model (p < 0.01). Based on the rat model, we further used an immunohistochemistry (IHC) assay to detect the level of tau phosphorylation at the thr181 site in the hippocampus. Positive staining of phosphorylated tau was significantly decreased after W112 treatment in the hippocampus ( Figure 4E). The results displayed that W112 could significantly prevent tau pathology in Aβ 25-35 -induced cell and rat models.
Effects of W112 on the Aβ 25-35 -Induced Neuroinflammation
To investigate the anti-neuroinflammatory activity of W112, we detected classic inflammation-related factors, such as TNF-α and IL-6, via Western blot. Compared with the model group, treatment with Aβ 25-35 significantly increased the expression of TNF-α and IL-6, while W112 markedly suppressed the production of TNF-α and IL-6 both in vitro (p < 0.01; Figure 5A-C) and in vivo (p < 0.01; Figure 5D-F).
Effects of W112 on the Aβ25-35-Induced Neuroinflammation
To investigate the anti-neuroinflammatory activity of W112, we detected classic inflammation-related factors, such as TNF-α and IL-6, via Western blot. Compared with the
Effects of W112 on Inhibition of NF-κB Signaling Pathway
To assess the effects of W112 on inhibition of NF-κB signaling, we evaluated the level of p-NF-κB/NF-κB by Western blot assay. The level of p-NF-κB/NF-κB in the model group was markedly higher compared with the control group, while W112 treatment significantly decreased p-NF-κB/NF-κB level both in vitro (p < 0.01; Figure 6A,B) and in vivo (p < 0.01; Figure 6C,D), suggesting that W112 treatment suppressed the phosphorylation of NF-κB signaling.
Effects of W112 on Inhibition of NF-κB Signaling Pathway
To assess the effects of W112 on inhibition of NF-κB signaling, we evaluated the level of p-NF-κB/NF-κB by Western blot assay. The level of p-NF-κB/NF-κB in the model group was markedly higher compared with the control group, while W112 treatment significantly decreased p-NF-κB/NF-κB level both in vitro (p < 0.01; Figure 6A,B) and in vivo (p < 0.01; Figure 6C,D), suggesting that W112 treatment suppressed the phosphorylation of NF-κB signaling.
Effects of W112 on MAPK Signaling Pathway
To further investigate the molecular mechanisms of W112-mediated intervention of Aβ25-35-induced AD-like pathology, we evaluated the effects of W112 on MAPK signaling by Western blot assay. The phosphorylation levels of p38, extracellular signal-regulated kinases 1/2 (ERK1/2), and c-Jun N-terminal kinase (JNK) were significantly up-regulated by Aβ25-35 compared with the control group, while W112 effectively reduced the abnormal ratios of p-p38/p38, p-ERK1/2/ERK1/2, and p-JNK/JNK both in vitro (p < 0.05 or p < 0.01; Figure 7A-D) and in vivo (p < 0.05 or p < 0.01; Figure 7E-H).
Discussion
With global aging, the prevalence of clinical AD is 2-3 times higher every 10 years. Rajan et al. reported that, starting in 2022, the number of people aged 75-84 suffering from AD will exceed those aged 85 and over [17]. With the change of the population burden of clinical AD, this will bring greater social, personal, and economic pressure to families and societies. The pathogenesis of AD is still inconclusive, and there are many hypotheses, which are very complex. Only four drugs are commonly used to treat AD, three of which are cholinesterase inhibitors (including donepezil, galantamine, and rivastigmine) and one is memantine, a non-competitive N-methyl-D-aspartic acid (NMDA) receptor antagonist. These drugs have been shown to be effective only for mild to moderate AD, which makes treatment options for AD very constrained. In 2021, aducanumab was approved by the FDA as the first anti-amyloid monoclonal antibody. Aducanumab was reported to cross the blood brain barrier (BBB) and then selectively bind with Aβ aggregates [18]. Although aducanumab has received both praise and criticism since its approval, its approval at least gives hope for AD drug development.
Triazoles have attracted more and more attention because of their wide range of biological activities. Triazole-based compounds are now under study for the treatment of a variety of central nervous system (CNS) diseases. The study of Gitto et al. revealed 1,2,4-triazole-based compounds could inhibit α-syn aggregation to prevent Parkinson's disease [19]. Wu et al. worked on 1,2,4-triazole derivatives and found a series of anticonvulsant compounds by using epilepsy models during both in vivo and in vitro studies [20]. Recently, triazole-based compounds have been explored for the possibility of treating AD. Wang et al. synthesized a series of novel triazole derivatives as a multi-functional agent for AD therapy and determined that the compounds demonstrate multiple effects including anti-neuroinflammation, selective inhibition of cholinesterase, and neuroprotection [21]. In the current experiment, we mainly focused on the effects of W112 on Aβ 25-35 -induced AD-like pathological changes and the molecular mechanism. Kollmer et al. reported that brain-derived Aβ amyloid fibrils fold differs sharply from Aβ 1-40 fibrils that were formed in vitro, and these findings underscored the importance of using patient-derived amyloid fibrils when investigating the structural basis of the disease. However, they also claimed that it would be a premature conclusion to state that in vitro formed fibrils were necessarily different from patient fibrils [22]. Millucci et al. also supported that the conditions of Aβ aggregation in the brain were different from those in the in vitro experiment and the actual aggregation kinetics would differ. However, it is very likely that this amyloid fragment also rapidly aggregated in the brain, and that changes in the actual aggregation process could result in the formation of aggregated structures that may have powerful effects on synaptic activity [23]. Therefore, in the current study, we still used humanized Aβ [25][26][27][28][29][30][31][32][33][34][35] to induce aggregation and mimic the neurotoxic role, as it represents the biologically active region of Aβ 1-40 or Aβ 1-42 and causes enhanced neurotoxicity. We first evaluated the neuroprotective effect of W112 in the cell model and its effect on improving learning and memory abilities in the rat model induced by Aβ [25][26][27][28][29][30][31][32][33][34][35] . The results showed that W112 could exhibit a neuroprotective role against Aβ-induced cytotoxicity and improve the learning and memory abilities of Aβ-induced AD-like rats. These results prompted us to further study the molecular mechanisms involved.
At present, despite the amyloid hypothesis still under investigation, it is the most mature mechanism to explain the pathogenesis of AD. The cerebral amyloid pathology appears 20-30 years earlier than the emergence of clinical AD symptoms [24]. Aβ peptides aggregate from monomers to oligomers and deposit as SPs in the extracellular, which eventually leads to the destruction of synaptic function, atrophy of neurons, and neurodegenerative changes. APP transgenic mice evidently reveal deficits in learning and memory, behavioral abnormalities, synaptic alterations, and SPs [25]. Intracellular Aβ oligomers can affect normal transmission, increase neuronal excitability of hippocampal neurons, and cause synaptic damage [26]. Mitochondria-associated membranes are an intracellular site of APP processing and Aβ is produced at mitochondria-endoplasmic reticulum contact sites, which may contribute to AD pathology [27]. Hyperphosphorylated tau represent another hallmark lesion of AD. Normal tau proteins play a vital role in neurons because of their binding and stabilizing of microtubules and regulating axonal transport. In AD pathological conditions, tau is hyperphosphorylated and aggregates into NFTs as a possible cause of memory loss and synaptic dysfunction. The microtubule-binding region of tau in cerebrospinal fluid is specifically increased and highly associated with the cognitive and clinical symptoms of AD [28]. The P301S mutant human tau transgenic mice show synaptic pathology and microglia proliferation in the hippocampus at 3 months old and synaptic dysfunction at 6 months old, which finally causes neurodegeneration [29]. The early accumulation of tau in the parietal hippocampal network is an important reason for the disorder of spatial orientation in AD [30]. The amyloid hypothesis believes that tau is a downstream target and Aβ drives tau pathology. Tau transgenic mice crossed with APP transgenic mice show that NFTs are substantially enhanced in the limbic system and olfactory cortex [31]. Aβ oligomers can cause intracellular Ca 2+ elevation and activate the Ca 2+ -dependent calmodulin kinase IIα, which is associated with increased hyperphosphorylation and mis-sorting of tau [32]. Aβ pathology can further promote the development of tau pathology in AD by increasing the spread of pathological tau [33]. Unfortunately, despite anti-Aβ drugs reducing SPs or Aβ accumulation, most have not been shown to modify cognition in humans. Due to too many failures of anti-Aβ drug development, more and more studies are beginning to re-focus on the tau and Aβ relationship. Fá et al. reported that high concentrations of Aβ or tau alone reduced synaptic plasticity and memory, and the same result also occurred when sub-toxic doses of oligomer Aβ were used in combination with oligomer tau at sub-toxic doses [34]. Gulisano et al. supported the idea that Aβ and tau might act at the same level or on different targets, but eventually converge on a common molecular mechanism [35]. In our current study, we found that Aβ [25][26][27][28][29][30][31][32][33][34][35] increased the levels of phosphorylated tau at multiple sites and W112 treatment significantly reduced the levels of tau hyperphosphorylation, both in vitro and in vivo studies. Our results (Figures 3 and 4) proved that Aβ can cause tau pathology, next we tried to further explore the underlying molecular mechanisms.
Increasing evidence shows that neuroinflammation is an active contributor to AD progression. Proinflammatory cytokines, including TNF-α and IL-6, are up-regulated in the brains of AD patients and in AD transgenic mice [36,37]. The excessive Aβ production and the hyperphosphorylated tau are both accompanied by the presence of inflammation; moreover, neuroinflammation increases the severity of the disease by exacerbating Aβ and tau pathology. Aβ-induced activation of the NLRP3 inflammasome significantly increases interleukin-1β (IL-1β) levels to enhance the progression of AD [38]. TNF-α regulates BACE-1 transcription, which results in an increased production of Aβ and further promotes TNF-α release [39]. Lipopolysaccharide injection affects inflammatory cytokine (TNF-α, IL-1β, and IL-6) production, accompanied by Aβ deposition in mouse brains [40]. The relationship between neuroinflammation and Aβ pathology is of significant concern, but few studies have paid attention to the interconnections existing between tau pathology and neuroinflammation. Neuroinflammation also plays a key role in NFTs formation. In addition, aggregated tau can further enhance inflammation and amplify neurotoxic injury. Astrocyte proliferation, microglia activation, and pathological neuroinflammation are observed in tau transgenic models [41]. Tau pathology has a direct positive correlation with neuroinflammation in the parahippocampus of AD patients examined by positron emission tomography [42]. Hyperphosphorylated tau trigger neuroinflammation in an NLRP3-dependent manner to activate IL-1β levels and impair spatial memory [43]. In this study, we found that Aβ 25-35 could promote the release of proinflammatory cytokines, and W112 prevented the over-production of TNF-α and IL-6 both in vitro and in vivo studies. The results revealed that the mechanisms of W112 preventing the pathological process of AD may be related to the "Aβ-tau-neuroinflammation" axis.
In the nervous system, NF-κB plays an important role as a transcriptional regulator and has post-translational regulatory activity. The activation of NF-κB in the brain induces neuroinflammation, and impairs neuronal survival, differentiation, neurite growth, and synaptic plasticity, which affects the development of AD. NF-κB is activated in Aβ plaquesurrounding areas in neurons from patients with AD [44]. In Aβ-induced microglia, NF-κB was up-regulated and the production of TNF-α and IL-6 was increased [45]. Inhibition of NF-κB signaling significantly repressed neuroinflammation and ameliorated Aβ plaque load and cognitive impairment [46]. MAPK expressed in the CNS mediates neuronal proliferation, differentiation, and cell survival. The most famous MAPK enzymes are ERK1/2, JNK, and p38 families. The pathological role of MAPK cascades in AD has been reported. Aβ in the hippocampus blocked the long-term potentiation via activation of the kinases JNK and p38 [47]. P38 MAPK specifically deleted from neurons in the brain of AD transgenic mice could decrease Aβ and tau phosphorylation load and improve the cognitive function [48]. Blockade of p38, JNK, and ERK1/2 inhibited the release of TNF-α and IL-6 induced by Aβ in BV2 cells [49]. Several studies have shown that NF-κB can be activated by the MAPK pathway. MAPK regulated the transcriptional activity of NF-κB in primary human astrocytes via acetylation of p65 [50]. MAPK inhibitors can inhibit NF-κB phosphorylation and reduce TNF-stimulated IL-6 gene expression [51]. Triazole derivatives have been shown to play an anti-inflammatory role via inhibiting NF-κB activation and MAPK phosphorylation [52,53], but its effects in AD need to be further explored. In our present study, we found that W112 may potentially inhibit MAPKs/NF-κB signal pathways to reverse Aβ-induced AD-like lesions both in vitro and in vivo.
Preparation of Aβ 25-35
Aβ 25-35 (Aladdin Biochemical Technology, Shanghai, China) was dissolved in distilled water and incubated in a 37 • C incubator for 96 h to induce aggregation as previously described [54].
Cell Culture and Viability Assays
The PC12 rat pheochromocytoma cells were cultured in RPMI 1640 medium (Gibco, CA, USA) supplemented with 10% fetal bovine serum (FBS, Gibco) and 1% penicillinstreptomycin (Beyotime, Shanghai, China) at 37 • C in a humidified 5% CO 2 atmosphere. Cultured cells were treated with Aβ 25-35 (20 µmol/L) in the absence and presence of W112 (5, 10, 20 µg/mL) or donepezil (Yuanye Biological Technology Co., Ltd., Shanghai, China) as a positive control for 48 h. Cell viability was evaluated by MTT assay as previously described [47]. The absorbance of each group was measured at wavelengths of 570 nm and 630 nm.
Animal and Treatments
Healthy male Sprague-Dawley (SD) rats (200-220 g) were provided by Yisi Company (Changchun, China) and kept in standard laboratory conditions (temperature 23 ± 2 • C, 12-h light/dark cycles) with food and water. The rats were randomly divided into six groups (n = 12 in each group): control, Aβ, Aβ + W112 (0.5 g/kg bodyweight), Aβ + W112 (1 g/kg bodyweight), Aβ + W112 (2 g/kg bodyweight), and positive group (donepezil, 1 mg/kg bodyweight). The skulls of the rats were drilled with small burr holes on two sides (1.0 mm caudal to the bregma, 1.5 mm lateral to the midline). Aβ 25-35 (15 nmol per rat) was intracerebroventricular (ICV)-injected at a depth of 3.0 mm in the Aβ, donepezil and W112 groups, and sterile normal saline was similarly injected in the control group. W112 groups received intragastric administration of 0.5, 1, and 2 g/kg W112, respectively, once daily for 28 days after the surgery. The control group and the Aβ group were treated with saline in the same way daily. The ethics approval of this study was granted by the ethical committee of the medical faculty of Inner Mongolia Minzu University (M2020015).
MWM Test
The MWM test was carried out under protocols detailed in previous reports [55]. The test was conducted to assess learning and memory performance. In brief, rats were trained to swim to reach the platform in a pool for 4 consecutive days and data of the escape latency were recorded. On the fifth day, the probe test was performed and the times of crossing through the original platform position were monitored by the WMT-100s Morris Water Maze video analysis system (TECHMAN, Chengdu, China).
Immunohistochemistry
The levels of tau phosphorylation at the thr181 site in the hippocampus from each group were detected by IHC. Briefly, the brain sections were cut at 5 µm thickness and incubated with primary antibodies against thr181-phosphorylated tau antibody (CST, Beverly, MA, USA) overnight at 4 • C. The next day, the slices were incubated with the second antibody and detected with diaminobenzidine tetrahydrochloride (Zymed, South San Francisco, CA, USA).
Statistical Analysis
Data were analyzed using SPSS 20.0 software (SPSS Inc., Chicago, IL, USA) and expressed as mean ± standard error of the mean and statistical analysis by one-way ANOVA, and the statistical significance standard was p < 0.05.
Conclusions
In summary, our current results showed that triazole derivative, W112, ameliorated Aβ-induced hyperphosphorylation of tau and reduced the production of proinflammatory cytokines, including TNF-α and IL-6, through significantly inhibiting MAPK/NF-κB signaling pathways both in vitro and in vivo studies. Thus, it is suggested that W112 may be a promising therapeutic strategy to prevent AD. | 4,890.2 | 2022-08-01T00:00:00.000 | [
"Chemistry",
"Medicine"
] |
Good things come to those who mate: analysis of the mating behaviour in the menstruating rodent, Acomys cahirinus
Background The Egyptian spiny mouse (Acomys cahirinus) is the only known rodent to exhibit true, human-like menstruation and postpartum ovulation, and is an important new model for reproductive studies. Spiny mice do not produce a visible copulatory plug, and calculation of gestational age is therefore restricted by the need to use mated postpartum dams. The current inefficient method of monitoring until parturition to provide a subsequent estimate of gestational age increases study duration and costs. This study addressed this issue by comparing the mating behaviour of spiny mice across the menstrual cycle and proposes a more accurate method for staging and pairing animals that provides reliable estimates of gestational age. In experiment 1, mating behaviour was recorded overnight to collect data on mounting, intromission, and ejaculation (n = 5 pairs per stage) in spiny mice paired at menses and at early and late follicular and luteal phases of the menstrual cycle. In experiment 2, female spiny mice were paired at the follicular or luteal phases of the menstrual cycle to determine any effect on the pairing-birth interval (n = 10 pairs). Results We report a broad mating window of ~ 3 days during the follicular phase and early luteal phase of spiny mice. Males displayed a discrete ‘foot twitch’ behaviour during intromission and a brief copulatory lock during ejaculation. Litters were delivered after 40–43 days if pairing occurred during the mating window, compared with 46–48 days for spiny mice paired in the late luteal phase. When pairing occurred during the late luteal phase or menses no mating activity was observed during the recording period. Conclusion This study clearly demonstrates an effect of the menstrual cycle on mating behaviour and pregnancy in the spiny mouse and provides a reliable and more effective protocol for estimating gestational age without the need for postpartum dams. Supplementary Information The online version contains supplementary material available at 10.1186/s40850-022-00112-1.
Background
Copulatory behaviour in rodents and most other mammals is tightly linked to their reproductive physiology, behavioural estrus, and the timing of ovulation [1][2][3]. This relationship is particularly useful in gestational studies in laboratory rodents where copulation and pregnancy are highly predictable and can be confirmed noninvasively. The Egyptian spiny mouse, Acomys cahirinus, is a rodent native to North Africa that has recently been used as a laboratory rodent in biomedical research. While limited information is available regarding A. cahirinus reproductive biology, recent research has highlighted several reproductive characteristics that are rare in rodents. Acomys cahirinus produce small litters with precocial young [4] and a relatively long gestation (~ 39 days; [5]), cannot become pseudopregnant [6], and, more recently, were the first rodents to exhibit a natural, human-like menstrual cycle [7] and a menopause-like transition [8]. Interestingly, unlike most rodents, A. cahirinus also do not produce a visible copulatory/seminal plug [9], which provides significant challenges when using mated females for gestational studies as ejaculation, and therefore early pregnancy, cannot be confirmed non-invasively.
However, based on observations in our breeding colony, female spiny mice also experience a postpartum ovulation, where ovulation and mating occur within 48 h of parturition [10]. Using this knowledge, estimates of gestational age are currently calculated from the date of delivery of the previous term litter [11,12]. While this method provides a useful and relatively accurate estimate of gestational age, it is time-consuming, costly, and inefficient. It requires the monitoring of breeder pairs until parturition (a minimum of 40 days from pairing virgin females and no guarantee that females will have mated immediately upon pairing) plus the subsequent timing of pregnancy to the required stage of gestation.
Observations from our laboratory have also identified variable pairing-to-birth intervals that ranged from 40-46 days from when females were paired at unknown stages of the 9-day menstrual cycle. This suggested that stage of the menstrual cycle influences mating behaviour that results in these variations in gestational outcomes in A. cahirinus. Although copulatory behaviour in this species has been described [13], in a study that also preceded the discovery of menstruation in A. cahirinus, no attempt was made to correlate the phase of the reproductive cycle with the timing of copulation and birth. In this study, the data recording methods used were not clearly detailed, and no video or photographic evidence of mating behaviour was provided. With the current access to improved recording technologies, a more accurate and effective method for timing mating, insemination and pregnancy can now be achieved. In this study, we have built upon the previous study by investigating the effect of menstrual cycle stage on receptivity, copulation, and birth outcomes in the Egyptian spiny mouse using digital recording equipment and used these data to propose a more effective method for estimating gestational age in the spiny mouse.
Experiment 1
Male spiny mice approached females from behind prior to mounting (Fig. 1, Supplementary videos 1-2) and females either moved away, sometimes climbing the wire cage top to prevent unwanted advances or were receptive and allowed the male to mate with her.
Mounting was usually initiated rapidly after introduction and was defined as the male clasping the female's back with his forepaws. Mount latencies were similar between the early follicular, late follicular, and early luteal phase (p > 0.05), and generally under 10 min ( Fig. 2A).
Fig. 1
Mounting behaviour during the lights-on (A, B) and lights-off (C, D) periods in spiny mice. Male spiny mice approach females from behind and place front paws on the middle of the female's back. Females will either allow mating to occur or move to prevent unwanted mating. Frames 2B and 2D are zoomed in images (2X) of 2A and 2C respectively However, during the late luteal phase mounting was only observed in one of the five pairs and occurred 85 min after introduction. No additional copulatory behaviours were seen in this pair, and no copulatory behaviours were seen when pairing occurred during menses in all pairs.
Intromission was a discrete event in which the female stood still while the male climbed further along her back with no apparent biting of the female's back or nape. Males also display a clear, and previously unrecorded, foot twitch during the brief period of intromission (~ 1 s; Supplementary videos 1-2) followed by a ballistic dismount.
Intromission was seen in all (5/5) late follicular phase pairings, but only in 3 each of the 5 early follicular and 5 luteal phase pairings (Fig. 2B). Intromission latencies were similar between the early follicular, late follicular, and early luteal phases (Fig. 2B, 3A; p > 0.05), and the first intromission typically occurred approximately one hour after introduction ( Fig. 2B; p > 0.05). Similarly, number of intromissions were also similar across these groups and generally occurred in a series of between 6 and 18 intromissions (11.3 ± 3.4 SD; Fig. 3A; p > 0.05). All pairings in which intromission was seen also resulted in ejaculation, which was confirmed by the presence of spermatozoa in the vaginal lavage. There were no behavioural signs that intromission or, by its association, ejaculation had occurred in any of the late luteal or menses phase pairs. Intromissions were confirmed as non-ejaculatory as females smeared after each intromission event contained no spermatozoa.
Ejaculation was distinguished from intromission by the difficulty of the mating pair to separate (Supplementary videos 3-4) following a discrete, but extremely brief, copulatory lock (~ 1 s). Copulatory locks were also confirmed to be ejaculatory events by the presence of spermatozoa in vaginal smears of females immediately following locking events ( Supplementary Fig. 1). In all 3 phases of the menstrual cycle where mating was successful, ejaculation occurred after a similar interval of between 7 and 24 min (16.3 ± 5.3 SD) after the first intromission ( Fig. 3; p > 0.05) and always followed a series of non-ejaculatory intromissions.
Most pairings resulted in a single ejaculation (8/11) during the observational period, however, based on observed locking events, three males ejaculated twice (2 in late follicular, 1 in early luteal) with a mean interval between ejaculations of 36 ± 11 min. No visible pelvic thrusting was evident during either intromission or ejaculation and genital grooming followed all intromission and ejaculation sequences. In all pairings for which ejaculation was observed, vaginal lavages from the following morning revealed no spermatozoa or visible seminal plugs. Also, litters in this experiment were born with a mean pairing to birth interval
Experiment 2
In experiment 2, only 4/5 females paired in the follicular phase and 1/5 paired in the luteal phase gave birth within 45 days of pairing. The female paired in the luteal phase which gave birth before 45 days had vaginal cytology containing a mixture of cornified epithelial cells (CECs), leukocytes and nucleated epithelial cells (NECs) (Fig. 5A) that was more typical of an early luteal phase vaginal smear in A. cahirinus. The remaining 4 females from the luteal phase pairings all showed vaginal cytology typical of the late luteal phase at pairing (Fig. 5B).
Litters were born between 39 and 43 days after pairing between the follicular and early luteal phase (mean 41.7 ± 1.2 SD; p > 0.05) by all females paired in the follicular and early luteal phases. In contrast, the pairbirth interval in 4/5 females paired in the late luteal phase was significantly longer (46-48 days after pairing, mean 46.6 ± 0.9 SD; p < 0.05), and the remaining female did not give birth.
Discussion
This study has described an improved analysis of spiny mouse copulatory behaviour from that published by Dewsbury and Hodges [13] and provided new data on the relationship between mating success and the spiny mouse menstrual cycle in our captive colony. In their original study, Dewsbury and Hodges argued that male copulatory patterns cannot be predicted from knowledge of the female estrous cycle. Our study challenges this assertion. We have demonstrated the clear presence of a broad 'mating window' in which mating can occur during the early follicular, late follicular, and early luteal phase phases of the menstrual cycle, but not during the late luteal or menses phases. In mammals, gonadal steroids are known to affect sexual behaviour and mating receptivity [14]. This relationship is true in female rodents [15], and also women, where copulation is significantly more frequent during their ovulatory window [16,17] when testosterone and dihydrotestosterone (DHT) concentrations are the highest [18]. Although androgen concentrations during the menstrual cycle in A. cahirinus have not been reported, our observations that females may also be receptive to mating several days either side of ovulation suggest Unexpectedly, neither the pair-birth interval nor the number of pups born were significantly different across cycle phases where pups were born. This is in contrast to hamsters and mice where a significant difference in litter size is attributed to timing of mating [19,20]. While time between the early follicular and early luteal phases of the menstrual cycle may differ by up to 2 days in the spiny mouse [7], it appears this time may be too little to observe a noticeable difference in litter size and pair-birth intervals. Further, a strong inverse relationship between litter size and gestation length has been reported in mice, rats, and gerbils [20][21][22]. However, differences are only apparent when comparing higher-and lower-order pregnancies in these species. Given the naturally small litter sizes in spiny mice (1-5; [4]) and higher-order pregnancies generally being born to multiparous dams (unpublished data from our colony), this may explain the similar pair-birth intervals we have observed.
Moreover, individual pairs established during the early follicular to early luteal phases of the cycle produce litters after a similar period to those reported in postpartum female spiny mice (41 days ± 4.4 days (SD); unpublished data from our colony). While litters were born > 45 days post-pairing in luteal phase females, these are most likely resultant of mating and pregnancies from the subsequent cycle, rather than the paired cycle. Thus, our method of timing pairing to a particular phase of the menstrual cycle, as outlined here, resulted in similar gestational outcomes to mated postpartum dams, while reducing study duration and costs.
Another distinction between our study and that of Dewsbury and Hodges [13], is the use of vaginal lavage to confirm behavioural cues for, and the timing and location of, ejaculation within the female reproductive tract. No spermatozoa were present in vaginal lavages taken immediately after individual, non-locking, intromission events. Despite interruptions to mating activities to obtain vaginal smears, and the potential stress involved in this procedure, spiny mice pairs rapidly resumed sexual activity after each consecutive vaginal lavage. This is an interesting and important observation for future mating studies because, although spiny mice are susceptible to stress (unpublished data from our colony), it appears that the mating drive is strong enough to overcome any stress caused by the disturbance of removing females briefly for vaginal lavage. Spermatozoa were seen in vaginal smears of females immediately after a copulatory lock, suggesting intravaginal ejaculation. However, as post-coital reproductive tracts of females were not examined in this study, we cannot rule out the possibility of intrauterine or intracervical insemination, as occurs in some other species like pigs and the camelids [23]. Our mating behavioural analysis reveals several similarities to Dewsbury and Hodges [13]. We observed no pelvic thrusting during either intromission or immediately prior to ejaculation, and a series of intromissions always preceding ejaculation. We also observed no obvious lordosis in female spiny mice during intromission or ejaculation; a feature that is typical of murid copulation [24]. Instead, we observed a distinct male foot twitch behaviour, which was not reported by Dewsbury and Hodges [13]. Interestingly, a similar behaviour, 'thumping' , was reported in Mongolian gerbils in which either individual taps its hind feet against the cage floor immediately following coitus [25]. However, the cause of this behaviour is ambiguous as it presents in both sexes in other non-coital settings and has been considered a sign of stress [26]. In contrast, the male spiny mouse foot twitch that occurred only during coitus, we interpret as a behavioural response of males to pre-ejaculatory penile insertion during mounting on a female not presenting any clear lordosis. While multiple intromissions have been suggested as a necessary requirement to trigger ovulation in several rodent species [1,27], A. cahirinus present with spontaneous, rather than induced ovulation [6,7]. Therefore, the multiple intromissions observed here are more likely a prerequisite to stimulate ejaculation, rather than to stimulate ovulation in A. cahirinus.
We also confirm the presence of a copulatory lock and no obvious copulatory plug in A. cahirinus. In an extensive review of 118 mammalian species, Dewsbury [1] categorised mating behaviour into 16 categories; the most common patterns being # 9 (no lock, intravaginal thrusting, multiple intromissions and ejaculations) and #13 (no lock, no intravaginal thrusting, multiple intromissions and ejaculations). Interestingly, all species in these two categories were either primates or rodents, with most rodents falling into category #13. From this, it appears that mating behaviour in A. cahirinus is broadly similar to other rodent species but with the addition of a copulatory lock. Seminal plugs are common in rats and guineapigs, but only observed in a few mouse species [15], and these non-plugging species generally have copulatory locks and reduced or underdeveloped accessory glands (reviewed by [28]). Further, Voss [15] argued that if there is a causal relationship between the presence of copulatory locks and the absence of plug formation, they 'must serve much the same function(s) as the plugs they presumably replaced' .
Male spiny mice have a normal complement of accessory glands typical of many murid rodents [29], with a large well-developed seminal vesicle but a comparatively small prostate and coagulating glands. Hartung and Dewsbury [28] have suggested that well-developed accessory glands are required for rodent seminal plug formation, but no copulatory plug has been observed in spiny mice. However, coagulation studies using spiny mouse accessory gland secretions [29], especially mixing of extracts from the seminal vesicles and the coagulating glands, shows coagulum formation. Together, this suggests the possibility of a covert post-ejaculatory seminal plug in spiny mice, perhaps deep within the vagina against the cervix or within the cervical canal.
The functional role of copulatory plugs in rodents has been debated for centuries [30] and several hypothesis have been suggested. These include prevention of insemination by rival males, assisting sperm transport, induction of pseudopregnancy and prevention of sperm leakage from the vagina [15]. None of these hypothesis are likely to apply in A. cahirinus considering the extremely brief lock compared to true locking species [31], and the demonstrated inability to induce pseudopregnancy in this species [6]. An alternative explanation is that, despite the brief ejaculatory lock, spiny mice deposit most or all of the ejaculate directly into the cervix or uterus like camelids and pigs [23]. Although spermatozoa were seen in vaginal lavages immediately following the brief locking events, these spermatozoa may be flowback through the cervix or leakage from the penis during withdrawal from the vagina. Considering this, if ejaculation does occur within the cervix or the uterus, formation of a small, very temporary, seminal plug may assist in maintaining spermatozoa at the site of ejaculation. Future studies of female spiny mouse reproductive tracts following coitus may provide answers to these questions on the site of insemination and presence of post-coital seminal plug, and provide new information on sperm concentration, survival, and transit through the female tract.
Conclusion
This study has extended and improved the analysis of spiny mouse mating behaviour reported by Dewsbury and Hodges [13] and, importantly, it provides a comprehensive description of mating behaviour across the recently discovered menstrual cycle of this species. We have defined a more reliable and efficient method for staging gestational age in female spiny mice by describing a discrete mating window where mating behaviour and pairing-birth intervals are highly predictable. Ejaculation was seen in 11/15 pairs (73%) during the mating window, and an average of 2 pups were delivered from 10/11 (91%) pairs between 38 and 43 days later. Moreover, females paired during the late luteal or menses phases were not receptive to mating, and males displayed no mating behaviour during these phases. This study has confirmed its original aim of providing a method for more reliably estimating gestational and fetal age in pregnant spiny mice and will improve gestational studies by reducing both the research time and financial costs of using postpartum animals.
Ethics approval and animal use
All experimental procedures carried out in this study were performed according to the ARRIVE guidelines, and adhered to the Australian Code of Practice for the Care and Use of Animals for Scientific Purposes. All animals (n = 70) were sourced from our breeding colony and approved for use by the Monash Medical Centre Animal Ethics Committee (MMCB 2019/13BC). Sexually mature spiny mice (3 -9 months of age) were housed in sex-segregated groups up to nine per cage or as breeding pairs under a 12:12 h light:dark cycle at 25-28 °C and humidity of 30-40% [4]. Cages were lined with wood shavings and plastic tunnels, cardboard boxes or tissue paper provided as environmental enrichment. Food (rat and mouse cubes; Specialty Feeds, Glen Forest WA) and water was provided ad libitum, with weekly supplements of fresh carrots and celery.
Experimental design
Two experiments were conducted in this study. The first experiment was designed to record and characterise the mating behaviour of spiny mice at different stages of the menstrual cycle. The second experiment was used to determine the effect of menstrual cycle stage on spiny mouse pairing-birth intervals.
Experiment 1 (n = 25 pairs; 50 animals)
The aim of this experiment was to compare behavioural cues for copulation with those outlined by Dewsbury and Hodges [13], and in particular, mounting, intromission and ejaculation. Spiny mice (n = 5 per phase) were paired at the early follicular (day 3-4), late follicular (day 4-5), early luteal (day 5-7), late luteal (day 7-9) and menses (days 1-3) phases of the cycle (Fig. 6). Being a naturally crepuscular species [32], the mating activity of each pair was recorded during the night for at least 6 h, commencing at 5 pm and ending the following morning at 8am.
Males and females were separated and returned to their respective home cages the following morning and females were lavaged vaginally on the morning of separation to check for signs of semen or a copulatory plug. Each female was then monitored for 30 days after pairing, when females were visibly pregnant and pups were palpable (~ 30 days; unpublished data from our colony). Females that were visibly pregnant were removed from their home cages and housed in a single cage to litter down; those that were not pregnant were left undisturbed in their home cage.
Recording
Mating behaviour was recorded in a clean cage, with fresh bedding and enrichment (as required by ethics), using a 1080p wireless security camera system (cctv-wfcla-4c-4b; UL-Tech, Australia) (Fig. 7). At least 6 h of footage was examined after pairing.
Mating behaviour
Mating behaviour was measured as described by Dewsbury and Hodges [13] with time to live birth (pairingbirth interval) and number of pups born as additional measures. The following behaviours were included in the behavioural assessment: mount latency (ML; time from pairing to first mount), intromission latency (IL; time from pairing to first intromission), intromission frequency (IF; number of intromissions prior to ejaculation), ejaculation latency (EJL; time from first intromission to ejaculation), ejaculation interval (EJI; time from Cahirinus lasts approximately 3 days and is followed by the follicular phase where several follicles are matured under the influence of estradiol (E2). Following a surge of E2, the dominant follicle ruptures releasing the mature oocyte, and subsequently forms a progesterone (P4) secreting corpus luteum. Following ovulation, the luteal phase begins, which encompasses the implantation window whereby a healthy blastocysts implants into a receptive uterus to establish pregnancy ejaculation to the next intromission) and ejaculation frequency (EJF; number of ejaculations observed). Females (n = 1 pair from each cycle stage) were also subjected to vaginal lavage immediately after copulatory behaviours were observed. These behaviours were considered to be ejaculatory events and vaginal lavages were performed to determine if spermatozoa was present in the vagina to confirm ejaculation.
Experiment 2 (n = 10 pairs; 20 animals)
The aim of this experiment was to determine the pairbirth interval of animals paired in the follicular and luteal phase when animals are left together until pregnancy was visible, rather than removed the following morning of pairing. Stage of the menstrual cycle was determined [7] in sexually mature, cycling females (n = 5 pairs per phase) prior to pairing during the follicular or luteal phases (Fig. 6) of the menstrual cycle. Animals in this experiment were paired at 1700 h and left together until there were visible signs of pregnancy as described in experiment 1.
Statistical analysis
All data were analysed using Prism 8 software (Graph-Pad). Data from experiment 1 were tested for normality using the Shapiro-Wilk test before further analysis. Oneway analysis of variance (ANOVA) was used to compare data between cycle phases, Tukey's multiple comparison test was used as post-hoc analysis, and Grubb's test was used to identify any outliers. Data from experiment | 5,408.2 | 2022-02-25T00:00:00.000 | [
"Biology"
] |
Threshold analysis regarding the optimal tax rate and tax evasion. Empirical evidence from Taiwan
For a long time, governments of all countries have attached great importance to the development of underground economic activities. The reason is that the characteristics of the underground economy are hidden and the information disclosure is not sufficient, which not only distorts the economic data indicators, but more importantly, the existence of the underground economy has led to the loss of a large amount of tax base, affecting the long-term economic development of the country. Whether raising the tax burden rate boosts the tax revenue or expand the scale of the underground economy. In this paper, we use Kuznet Tax Curve (KTC) method to analyze the relationship between GDP and TTR/DTR/ITR. We find that the tax base erosion rate of indirect tax is lower than that of direct tax. In addition, we explore the relationship among economic growth, tax rate and tax revenue and adopt SUR-OLS method and Threshold approach to estimate the response of economic growth on total tax revenue(TTR), direct tax revenue(DTR) and indirect tax revenue (ITR) in Taiwan from 1991-2020. Our empirical research shows that when DTR tax rates are between 12.59% and 13%, an increase in income leads to a decrease, not an increase, in DTR, leading to severe tax base erosion. That is, the relationship between GDP and DTR presents a N-shaped relationship. However, ITR does not exist any tax rate threshold effect. Obviously, with the increase of GDP, ITR also increases. This reflects that the difference of tax structure between direct tax and indirect tax plays a key role in the relationship between tax rate and tax base erosion.
Introduction
In the 1950s, Kaldor [1] and Cagan [2] mark the beginnings of preliminary research of hidden economic activity. Since then, more and more literatures have focused on discussing the relationship between undeclared income and tax erosion. Smith [3] defines the UE as the production of market-based goods and services, whether legal or illegal, escapes detection in the official GDP estimates. The formal economic theory of tax evasion can be dated to Allingham and Sandmo [4] who pointed out that when the tax burden rate increases, taxpayers must make a measurement between the expected benefits of tax evasion and risk bearing. Welch and Goyal [5] and Kostakis et al. [6] demonstrate that predictability may be time-varying and that the impact of predictors may be evolving over time. In general, there are two common modeling tools to deal with parameter instability in PR models: structural change and threshold model. In this study, we took Taiwan as a case study to explore the relation between TTR/ DTR/ITR and GDP over the period from 1991 to 2020. Concurrently, we adopt Hansen's approach to measure the size of tax base erosion over the same period using Tax Kuznet Curve (TKC) approach and select the parameter "tax burden rate" as a threshold variable to capture the response of income variation upon tax revenue. Our paper is organized as follows. In section 2, reviews relevant literature on underground economy and tax evasion, section 3 deals with the methodology and model, section 4 differs from existing empirical methods, we take Taiwan as a study case, using Kuznet Tax Curve (KTC) approach and the SUR-OLS method to calculate the gap between the actual tax revenue in 2020 and estimated tax revenue, then we acquire the amount of tax base erosion in 2020. Further, we estimate the relationship between TTR/DTR/ITR and GDP. Section 5, referring to Wu et al. [7], we construct threshold analysis model and take Taiwan as an study case to propose an empirical analysis and discuss the tax burden rate threshold effect for taxpayers and its reaction on TTR//DTR//ITR. Section 6 recapitulates concluding remarks and outlines policy implications.
The literature review
According to Schneider and Enste's [8] survey, during the last decades the underground sector was nearly three-quarters of the officially recorded GDP in Nigeria and Thailand, but it amounted to a noteworthy 15% in the OECD countries as well. Pissarides and Weber [9] estimated the underreporting of income by using the data of the household food expenditure survey in the United Kingdom. The study found that the real income of self-employed households was 1.55 times the declared income, and then estimated that the size of the British underground economy accounted for about 5.5% of GDP. In addition, Johansson [10] studied the income leakage of Finnish households and found that the income under reporting rate of self-employed households was 25% to 30%. Milorad and Williams [11] indicate that 22.6% of all employees in Montenegro are unregistered employees. In addition, 17.5% of all formal employees received under reported salaries from their employers in order to avoid paying taxes. Wang et al. [12] use a cash deposit ratio approach and a currency demand approach to estimate UE size. In addition, Giles et al. [13] depict that an increase in the effective tax rate has a greater effect on the UE than a decrease in New Zealand. Zhiqin and Qunli [14] outline that tax burden rate is positively related to the underground economy. In addition, Bhattacharyya [15] finds clear evidence for the U.K. that the UE has a significant effect on the consumer expenditure. Another, the night light images are taken by the operational line scan system (OLS) carried by the US military meteorological satellite program (DMSP) from 1992 to 2013. When the data were released, the abnormal lights, background noise and other non urban lights had been eliminated and could be directly used for relevant research. Similarly, Elvidge et al. [16] propose in 1996 that there is a strong correlation between night light and population, GDP and power consumption data. However, NOAA has no night lighting data of Taiwan from 1992 to 2013. The above is the situation of tax base erosion in some countries.
Similarly, Schneider and Enste [8] point out, at least two thirds of the income earned in the shadow economy is immediately spent in the official sector, revealing UE and the official sector might thus be complements. Specially, Hindriks et al. [17] consider that some tax payers may collude with the inspector so the inspectors underreport the tax liability of the tax payers in exchange for a bribe. However, Jorge and Mark [18] denote that the effect of increased enforcement effort in a given mode has an ambiguous effect on compliance in the targeted mode as well as the untargeted mode. Li Yiting and Tang Ruyin [19] pointed out that high inflation will erode the purchasing power of cash, so people will tend to hold less cash, reducing the incentives for people to engage in underground economic activities. Another, Giovanni et al. [20] propose the effect on evasion and government revenue of two policy instruments: a tax on cash withdrawals (TCW) and a tax rebate conditional on having the receipt. Their research shows the tax rebate reduces evasion but it is costly if tax evasion is low. Din [21] considered the heterogeneity of tax sources and found a positive relationship between the personal tax (direct tax) rate and the underground economy from the data of Malaysia, but a negative impact on the sales tax (indirect tax) rate.
In addition, Cranor et al. [22] analyze a large field experiment conducted with the Colorado Department of Revenue to study the presentation of financial incentives and social norms in tax delinquency notices, their research suggests that attention to seemingly minor decisions about the wording of notices sent by tax authorities can increase tax payments and reduce administrative costs associated with taxpayer delinquency. Advani [23] described that tax compliance varies with personal characteristics, and male and young people are more disobedient. Friedman et al. [24] found in a transnational analysis that higher tax rates are related to less underground economies. However, the empirical results show that this relationship is not very stable. It is worth mentioning that most of the domestic and foreign current literatures are discussed in the linear model, and the possible nonlinear relationship between them has seldom been discussed.
According to the traditional research, it is assumed that the impact of GDP on tax revenue exists nonlinear characteristics due to different degrees of tax rate. In order to verify this nonlinearity and alleviate the potential endogeneity of the traditional regression model. Granovetter [25] and Granovetter & Soong [26] propose the threshold model. In the spirit of Granovetter's threshold model, the "threshold" is the number or proportion of others who must make one decision before a given actor does so. In addition, Bick [27] applies non-dynamic (static) panel threshold regression that propounded by Hansen [28] on a balanced panel data from 40 developing countries. Kremer's [29] findings reveal a threshold inflation of 2.53% for industrial countries and 17.22% for nonindustrial countries. showing the relationship is significantly positive below the threshold and significantly negative above the threshold for the industrial countries. Gonzalez et al. [30] consider a nonlinear panel model which is called the panel smooth transition regression (PSTR) model. Their research generalized the PTR model by allowing the regression coefficients to change smoothly when moving from one "extreme" regime or state to another. However, the PTR model separates the observations into several sets or groups based on the value of the threshold variable with sharp "borders" or thresholds.
In recent research, Baumann et al. [31] design a variant of an optimal stopping task that allowed people to quantitatively characterize the deviations of human behavior from optimality and found that humans apply a simplifying strategy, where thresholds are linearly increased over time. Tariq et al. [32] examine the nonlinear relationship between financial development and economic growth in Pakistan using the threshold regression model for the period 1980-2017. They research indicates that economic growth responds positively to financial development when the level of financial development surpasses the threshold value of 0.151. However, when financial development lies below the threshold value (that is, 0.151), its impact on economic growth is negative. Yang et al. [33] extend Hansen's [34] constant threshold regression model by allowing for a time-varying threshold which is approximated by a Fourier function. Least-square estimation of regression slopes and the time-varying threshold is proposed, and test the existence of threshold effect and find there is little efficiency loss by the allowance for Fourier approximation in the estimation procedure even when there is no time-varying feature in the threshold.
In addition, Belarbi et al. [35] adopt Buffered threshold panel data model(BTPD) to examine the combined effects of oil dependence and the quality of institutions on economic growth. To do so, they introduce a new buffered threshold panel data model and apply it to 19 oil rent-dependent countries over the period 1996-2017, their research show that the relationship between growth and oil-dependence is not linear. Another, Zhang and Kim [36] establish a model of FDI location and explore to examine the threshold role of institutional quality in determining the relationship between labor costs and FDI location, using data from 14 South and Southeast Asian countries during 2000-2017, evidence shows that effects of labor costs on FDI are nonlinearly decreasing because their institutional quality is improved above threshold values. Zhiqi's [37] research contributes to the literature by distinguishing small-scale taxpayers from general taxpayers in terms of the optimal sales threshold for VAT. their research analyzes how the optimal sales threshold varies with changes in administrative and compliance costs and in tax rates. Furthermore, Yan et al. [38] use the threshold model to analyze the nonlinear characteristics between PSA and CO2 emissions under different degrees of government intervention. Similarly, Yu and Fan [39] requires that the variables affecting threshold are given or predetermined, but, in most applications, it is difficult to explore the factors which affect the threshold value in advance. Wang et al. [40] construct a threshold effect model, sets the institutional environment as the threshold variable, and empirically analyzes the impact of Internet development on the supply efficiency of government public services.
There are other literature on discussing Taiwan's underground economy. Wang et al. [12] examine the asymmetric response of the underground economy in Taiwan to the fluctuation of tax rate and measure the UE size from 1962 to 2003 using cash ratio approach and currency demand approach and find an increase in indirect or direct tax has a greater effect than the corresponding decrease. Ho and Tsai [41] examine the difference in the impact of different tax sources on the scale of Taiwan's underground economy, and found that business tax had a significant positive relationship with the underground economy, while income tax had a slight positive relationship. Lin et al. [42] use the OLS regression method based on Bai and Perron [43] to analyze and obtain the endogenous threshold of tax burden rate and discuss how the tax burden rate affects Taiwan's underground economy.
Our paper differs from the traditional literature, we adopt the threshold regression model to obtain the endogenous tax burden rate to explore the relationship between GDP and TTR/ DTR/ITR. Table 1 lists recent relevant documents on Taiwan's underground economy article.
Methodology, hypothesis
The Laffer curve is a threshold effect that describes the "inverted U-shaped" relationship between tax rates (tax burden rates) and total government revenue, inspired by the above literature and theory, this paper follows the threshold model setting of Bai and Perron [43] and takes endogenous variable" tax rate" as the turning point of interval change to estimate the response of Taiwan's economic growth to total tax, direct tax and indirect tax from 1991 to 2020. That is to say, different intervals in the model are divided by threshold variables greater than a certain threshold. The hypothesis of this paper is to use the threshold method to analyze, endogenously explore whether there exists a threshold value of the tax burden rate that alters the relationship between the total tax/direct tax/indirect tax. However, testing this hypothesis requires the estimation of a non-linear model. One traditional method that solves this type of non-linearity and heterogeneity is estimating a panel threshold regression (PTR), developed by Hansen [28]. The PTR assumes that analogous individuals should belong to one group. Thus, one can divide the individuals in the sample into several groups based on observables. But this is not the focus of this article. Due to the relationship between tax revenue and tax burden rate (tax rate). However, compared with the previous literature, most of them discussed the relationship between tax rate and underground economic size from a linear model. In this paper, we refer to the threshold model framework of Hansen [34] and Odedokun [44], selecting the tax rate as the threshold variable to explore whether there exists a threshold effect of tax rate on tax revenue, and whether the effect of dependent variables on tax revenue is different under high and low tax rates. Since the SUR-OLS method estimates the parameters of all equations simultaneously, so that the parameters of each single equation also take the information provided by the other equations into account. In general, the SUR-OLS estimates are consistently better than the OLS (equation-by-equation). Furthermore, the SUR-OLS estimator takes the correlation between the error terms into account, hence, SUR-OLS is a robust methodology for predicting (Cadavez & Henningsen [45]). As is well known, Taiwan's inland have convenient transportation links, taxpayers live in the same environment of tax laws and regulations. Hence, it has the heterogeneity of variance, and the residual has the characteristics of contemporaneous correlation. In view of this, in order to reduce the standard error, this paper uses "seemingly unrelated regression" (SUR-OLS) to test and analyze.
Model
The analysis of the EKC seeks to confirm whether wealth accumulation stimulates environmental degradation or contributes to improving its quality (Kaika & Zervas [46]). According to this approach, if GDP per capita is less than the level of the turning point, wealth accumulation contributes to environmental degradation; conversely, if GDP per capita is higher, environmental quality improves. In this setting, our research sets a theoretical model of the inflection point of Tax Kuznets curve(TKC) as follows. Eq (1) describes the indirect utility between tax burden and economic growth. We assume that utility function is separable in these two arguments, R and T, with the additive-separable function and additive preferences. Such that: In Eq (1), s 1 , s 2 , γ, δ > 0, where s 1 is coefficient, s 2 reflects the impact of real income on utility, γ reflects the impact of tax burden on utility, F represents the government's subsidy to taxpayers below a certain income threshold or tax exemption threshold, τ m represents marginal tax rate system, β is income declaration rate of taxpayers, R denotes the level of income, T is the tax burden. Hence, we set the tax burden paid by the taxpayer can be expressed as Hence, the higher the income declaration rate of taxpayers, β, the greater the T. Consider the character of progressive income tax rate system, we adopt the sustained-growth version of Guo and Lansingís's [47] nonlinear tax structure and postulate τ t as In Eq (3), R * t denotes a benchmark level of income that is taken as given by the representative household. In our model with endogenous growth, R * t is set equal to the level of per capita output on the economy's balanced growth path (BGP), where R * t R t ¼ θ > 0, for all t. Hence, the marginal tax rate τ mt , defined as the change in taxes paid by the household divided by the change in its taxable income which is given by where 0 < τ t , τ mt < 1, R t þF R t θ ð Þ ≧τ mt , as mentioned, R represents real income, T denotes tax burden for people, reflecting the adverse impact of tax burden on the people's indirect utility. Moreover, we assume that the marginal disutility of tax burden remains unchanged. In order to eliminate the impact of structural effects, we suppose that only one commodity model is used for analysis. In this situation, firms produce aggregate output, Y, we set a constant returns to scale technology of the Cobb-Douglas type. Therefore, a country's incomes Y is expressed as Eq (5): In Eq (5), λ is the conversion coefficient, P represents the commodity price, with λ2(0,1). F (K,AL) denotes aggregate production function, where K denotes aggregate physical capital and L represents aggregate labor employed in production, A represents the technical level, with A > 0, α 2(0,1). Eq (6) reflects the value of marginal tax burden upon taxpayers equal to the demand of reverse tax burden, which is given by:Γ Also, the value of marginal tax revenue levied by government can be expressed as follows.
Through the supply-demand production function, the expression of the Kuznets curve can be obtained through Eqs (6) and (7) Furthermore, the following formula can be obtained by calculating the derivative of optimal tax revenue/burden T.
Clearly, the inflection point of tax burden is R = δ. This shows that when economic growth reaches a certain level, there will be tax base erosion. This means that people begin to evade taxes in an attempt to reduce their tax burden. Eq (9) is a convergence function, its value is greater than zero. That is, if n positive convergence functions are added together, the function obtained should also be convergent. Based on the theoretical models derived from Eqs (1) to (9). we seek to use empirical analyses to discuss the existence of TTR/DTR/ITR-to-GDP ratio/ Kuznets Curve and further discuss whether the Kuznets Tax Curve/ TTR/DTR/ITR-to-GDP ratio exists in Taiwan covering the 1991-2020. Obviously, if these Kuznets curve does not exist, revealing that with economic growth, tax revenue will also increase.
Empirical analyses between TTR/DTR/ITR and GDP
In this paper, we take Taiwan as a case study and use Simultaneous equations model and SUR-OLS approach to exploit the cointegration relationship among the GDP, variables TTR, DTR, ITR for Taiwan over a time period ranging from 1991 to 2020. To capture the synchronous correlation between heterogeneity and residuals in the model, our research employs SUR-OLS approach to measure the correlation among those variables, determining whether the stochastic component contains a unit root or not. The results of unit root tests are presented in Table 2, which demonstrates that all the variables appeared stationary at the firstdifferenced form under 5% significant level, depicting the logged variables are I(1). We next utilize the SUR-OLS regression method evaluating the residual term and estimate whether the residual term conforms to no sequence autocorrelation.
We then adopt Johansen Cointegration to test whether there exist a long-term equilibrium relationship between TTR/DTR/ITR-to-GDP. In Table 3, Trace test result shows that there exists a set of cointegrating vectors at the 5% level, and Max-eigenvalue test also indicates the same result.
Owing to the Q-statistic proposed by Box and Pierce [48] is rather weak in large samples, Ljung-Box [49] proposes another modified Q-statistic suitable for small samples. However, Box & Jenkins [50] consider that it is necessary to diagnose whether the parameters have overfitting and also confirm whether the residuals have serial correlation. Below, the results of Ljung-Box Q test are shown in Fig 1, which reveals the probability values of Q-statistics from the first period to the sixteenth period are all significantly greater than the 5% significance level. On the other words, the residuals estimates of model 1 to model 3 in Table 4 have no sequence autocorrelation. The Ljung Box verification results can be seen from Fig 1(a)-1(c), the Ljung Box test statistics from Phase 1 to Phase 12 were all 5% higher than the significant level, indicating that there is no autocorrelation among the three variables, TTR, DTR, ITR in the residual items from Phase 1 to Phase 12.
We next exploit the Histogram-Normality test and Heteroscedasticity test. In Table 4, we use Breusch-Pagan-Godfrey to diagnose residual heterogeneity, which show the p-values of Fstatistic, OBS * R-squared and Scaled explained SS of all models are all significantly greater than 5%, denoting that the residuals from model 1 to model 3, in Table 4, do not exist residual heterogeneity.
Note that in Table 4, the p-values of F-statistic, OBS * R-squared and Scaled explained SS of model 1 to model 3 are significantly greater than 5%. In Table 4, model 1 to model 3 correspond to the three models in Table 5 in an orderly way. Standard errors in parentheses: *** Fig 1(a)-1(c), the Ljung Box test statistics from Phase 1 to Phase 12 were all 5% higher than the significant level, indicating that there was no self correlation between the residual items of Phase 1 and Phase 12.
https://doi.org/10.1371/journal.pone.0281101.g001 means the first-order difference passes the stability test at 1% significance level, ** means the first-order difference passes the stability test at 5% significance level. Further, in model 1 of Table 5, we discuss solely the nonlinear relation between TTR-to-GDP, where variable GDP 2 represents GDP squared, variable GDP 3 denotes GDP tripled. Including variables GDP, GDP 2 , GDP 3 , debt and consumer price index (CPI), all data are denominated in million TWD. Further, we establish the correlation among TTR, GDP, square GDP, triple GDP, debt and CPI as follows: where ε t = φ 1 ε t−1 + φ 2 ε t−2 + σ t Case 1: Eq (10) declares that TTR increases with the increase of GDP, reaching a significance of 10%, see model 1 of Table 5. That is, as the debt variable is included, TTR also increases with the increase of GDP, reaching a significance of 10%. However, as the consumer price index(CPI) variable is added. TTR also increases with the increment of GDP, but it does not reach the significance of 10%. In Table 6, we denote that TTR-to-GDP represents Nshaped curve relationship.
where ε t = φ 1 ε t−1 + φ 2 ε t−2 + σ t Case 2: Eq (11) demonstrates that DTR increases with the increment of GDP, see model 2 of Table 5, denoting the corresponding regression coefficient is 0.061, depiciting the increment of GDP, to a certain extent, resulting in the increase of DTR. However, the coefficient does not pass the 10% significance test. Further, if the variables debt and CPI are added to the model. It shows that DTR increases with the growth of GDP, whereas these two coefficients fail within the significance test of 10%. In Table 6, we show that DTR-to-GDP presents N-shaped curve relationship.
where ε t = φ 1 ε t−1 + φ 2 ε t−2 + σ t Case 3: Eq (12) illustrates that ITR increases with the increase of GDP, see model 3 of Table 5, reaching a significance of 1%. Even though variable debt is included, ITR also increases with the increment of GDP. Moreover, as variable CPI is added. ITR also increases with the increment of GDP. From the results of Table 6, our empirical research depicts that on the basis of the existing ITR-to-GDP, adding variable debt or CPI, the relationship between ITR-to-GDP presents a N-shaped relationship.
Obviously, in Table 5, we show that under low inflation, the promotion effect of consumer price index(CPI) upon total tax is not significant, our empirical result is in line with Khan et al. [51] argument.
Kuznet tax curve analysis (TTR/DTR/ITR and GDP)
Further, according to the statistics of DGBAS, Taiwan. Taiwan's GDP in 2020 is 19,766,240 million TWD, and the actual total tax revenue is 2,398,667 million TWD. However, according to Tax Kuznets curve of TTR-to-GDP ratio, when Taiwan's GDP in 2020 is 19,766,240 million TWD, the TTR should be 2,589,189 million TWD, revealing the total tax base evasion amount is 190,522 million TWD, accounting for 0.009638 of GDP in 2020. Our empirical results declare that Taiwan's total tax evasion rate in 2020 is 0.9638% (see Fig 2).
Similarly, Taiwan's GDP in 2020 is 19,766,240 million TWD, and the actual direct tax revenue is 1,324,208 million TWD. However, according to Tax Kuznets curve approach of DTRto-GDP ratio, when Taiwan's GDP in 2020 is 1,324,208 million TWD, the DTR should be 1,458,258 million TWD, revealing that direct tax evasion is 134,050 million TWD, accounting for 0.6781 percent of GDP in 2020. Our empirical results depict that Taiwan's direct tax base evasion rate in 2020 is 0.6781% (see Fig 3).
Furthermore, Taiwan's GDP in 2020 is 19,766,240 million TWD, and the actual indirect tax revenue is 1,074,459 million TWD. However, according to Tax Kuznets curve of ITR to GDP ratio, when Taiwan's GDP in 2020 is 19,766,240 million TWD, the ITR should be 1,117,707 million TWD, revealing the indirect tax evasion is 43,248 million TWD, accounting for 0.002187 of GDP in 2020. Clearly, our empirical results demonstrate that Taiwan's indirect tax base evasion rate in 2020 is 0.2187% (see Fig 4). It can be seen from the above analysis that the tax base erosion rate of indirect tax is lower than that of direct tax. The main reason may be that indirect tax is levied by withholding at source, which is difficult to evade taxpayment for taxpayers.
Discussion and recommendation
In this paper, we take Taiwan as a case study and use Simultaneous equations model and SUR-OLS approach to exploit the relationship among the GDP, TTR, DTR, ITR for Taiwan over a time period ranging from 1991 to 2020. The difference with Wang et al. [12] is that this paper uses the data from the Taiwan General Accounting Office database to bring into the theoretical model established in this paper and uses empirical analysis methods to estimate the amount of tax base erosion of total tax/direct tax/indirect tax in 2020, and compares the rate of tax base erosion to find that the rate of tax base erosion of indirect tax is smaller than the rate of tax base erosion of direct tax, this section adopts the Tax Kuznet Curve analysis, and the results obtained are consistent with those obtained in the next section according to the threshold theory analysis (indirect taxes have no threshold effect, but direct taxes have threshold effect). Our research opens up a new path for the research of TTR/DTR/ITR to GDP, and also fills in the theoretical gaps on these issues The findings and implications will offer indicative guideline for the study of the relationship among TTR/DTR/ITR and GDP.
Empirical analyses
To explore whether Taiwan's tax burden rate has a threshold value that changes the relationship between the tax burden rate and the economic growth of the underground sector. Different from Lin et al. [42], this paper, after controlling the influence of other variables (excluding tax burden rate), discusses how GDP and (total tax/direct tax/ indirect tax) fluctuate between 1991 and 2020, and measures whether there exists a threshold effect among TTR/DTR/ITR and GDP. We follow the threshold regression model of Bai and Perron [43], and take tax burden rate as the threshold variable, Check whether the tax burden rate has different effects on Taiwan's total tax revenue/direct tax revenue/indirect tax revenue in different high and low intervals under the control of other control variables, including government debt and consumer price index. Owing to different individuals have different thresholds, it is necessary to emphasize the determinants of threshold. Eq (13) is the list and definition of variables that being used to test the taxation threshold effect on economic growth. The model is written as follows: where I (E) is an indicator function. When event E occurs, I (E) = 1, otherwise I (E) = 0, the residual term e t = [e 1,t , e 2,t ], y 0 1 , y 0 2 , α, β are parameters to be estimated, t is threshold tax variable, (TTR/DTR/ITR) i,t is the explained variable, Debt i,t and CPI i,t are explanatory variables, and T � [t 0 , t 1 ] is the spatial parameter of t 1 t 2 , t 2 T. The threshold variable q i,t is either smaller or larger than the threshold t 1 that illustrate by slopes y 0 1 , y 0 2 and y 0 3 . I(�) is the indicator function, which takes the value 1 if the argument in parenthesis is valid, and 0 otherwise. The ε i,t is assumed to be identically and independently distributed (iid) with mean equal to zero and variance is finite, that is e it � [0~σ 2 ].
In this study, we set the dependent variable as TTR/DTR/ITR and use the income square term and income cubic term as explanatory variables to capture the nonlinear impact relationship between these variables and total tax/direct tax/indirect tax Hansen [34]. Owing to over parameterization will also reduce the degree of statistical freedom, resulting in inefficient regression estimation results. Hence, our research selects the tax rate as the threshold parameter. The empirical results can be obtained as follows: Case 1: TTR-to-GDP According to Eqs (10) and (13). When the tax burden rate is below 12.5%, the increase in GDP at this stage will produce a positive effect on total tax revenue. When the tax burden rate is between 12.5% and 13%, at this stage, the increase in GDP causes the total tax revenue to fall instead of increasing, indicating that the total tax base is being eroded. However, when the tax burden rate is greater than 13%, the increase in GDP at this stage will have a positive effect on total tax revenue. That is, the relationship between GDP and TTR presents a N-shaped relationship (see Fig 5).
Case 2: DTR-to-GDP Similarly, according to Eqs (11) and (13). When the tax burden rate is below 12.6%, the increase in GDP at this stage will produce a positive effect on direct tax revenue. However, when the tax burden rate is between 12.6% and 13.4%, at this stage, the increase in GDP causes the direct tax revenue to fall instead of increasing, indicating that the direct tax base is eroding. Moreover, when the tax burden rate is greater than 13.4%, the increase in GDP at this stage will have a positive effect on direct tax revenue. That is, the relationship between GDP and direct tax revenue presents a N-shaped relationship (see Fig 6).
Case 3: ITR-to-GDP Next, according to Eqs (12) and (13), we take the tax burden rate as the threshold variable, our empirical result reveals that indirect tax has no threshold effect, that is, with the increase of GDP, indirect tax revenue also increases. Clearly, the relationship between GDP and indirect tax revenue demonstrates an ╭-shaped relationship (see Fig 7).
Discussion and recommendation
Our empirical research can be summarized as follows: (i) When the total tax rate is below 12.5%, taxpayers are willing to pay even if the tax rate increases because the "expected benefit" of tax evasion is less than the penalty cost of being caught. However, when the total tax rate is between 12.5% and 13%, taxpayers measure the "expected benefits" of tax evasion to outweigh the penalty costs of being caught. So total tax revenue in this range will decrease as GDP grows, not increase. In fact, according to statistics from Taiwan's Ministry of Finance, the actual average total tax rate from 1991 to 2020 was 12.7%, which is between the threshold total tax rate of 12.5% and 13%, indicating that within this tax rate range, there is tax base erosion in total tax revenue. As mentioned earlier in this article, according to statistics from Taiwan's DGBAS. In 2020, Taiwan's GDP is NT$19,766.24 billion, and its actual tax revenue is NT$239.8667 billion. However, according to the tax Kuznets ratio curve (TTR-to-GDP), when Taiwan's GDP in 2020 is 197.6624 billion TWD, TTR should be 258.9189 billion TWD, showing that the total tax evasion in 2020 is 190.522 billion TWD, accounting for 0.009638 of GDP, which is consistent with the empirical results of the threshold theory mentioned above. (ii) When the direct tax rate is lower than 12.6%, taxpayers are still willing to pay taxes even if the tax rate is raised because the "expected benefit" of taxpayer evasion is less than the penalty cost of being caught. However, when the average direct tax rate is between 12.6% and 13.4%, taxpayers measure the "expected benefits" of tax evasion outweighing the penalty costs of being caught. Consequently, direct tax revenues within this threshold will decrease rather than increase as the economy grows. Our empirical results show that the tax base erosion rate of Taiwan's total tax revenue in 2020 is 0.9638%, which is consistent with our estimated results based on the threshold model. (iii) Finally, our empirical results show that from 1991 to 2020, Taiwan's indirect tax does not have a threshold effect, that is, indirect tax revenue increases with the growth of GDP.
Conclusion
Our research differs from the traditional methodology, we adopt SUR-OLS method, Tax Kuznet Curve (TKC) approach and Threshold model to estimate the response of GDP on total tax revenue(TTR), direct tax revenue(DTR) and indirect tax revenue (ITR) in Taiwan from 1991-2020. In empirical research, we select the parameter "tax burden rate" as a threshold variable to capture the response of income variation upon tax revenue. Our research contributed to the literature on Threshold analysis regarding the optimal tax rate and tax erosion as follows: First, according to the Kuznet Tax curve (KTC) approach, we find that Taiwan's total tax base erosion rate in 2020 is 0.9638%, indicating that the tax base erosion rate in Taiwan is not so severe. Our research show the empirical results of Threshold model and Kuznet Tax Curve approach are consistent. Second, our empirical result shows that total tax and direct tax have threshold effect, but indirect tax has no threshold effect, which reflects that the difference of tax structure will affect the robustness of empirical results when conducting empirical research on tax burden rate and tax base erosion. Third, according to the Kuznet Tax Curve model, we estimate the indirect tax base erosion rate in 2020 is 0.2187%, which is lower than the direct tax base erosion rate of 0.6187%. Obviously, our empirical research shows that indirect tax revenue without "tax rate threshold effect" is more effective in reducing the tax base erosion rate than direct tax with tax rate threshold effect. This reflects that the difference of tax structure between direct tax and indirect tax plays a key role in the empirical study of tax burden rate and tax base erosion. Finally, threshold models are widely used in economics. However, there are limitations to assuming that the threshold is stable or time-invariant. this paper adopts a piece wise in variable analysis method, not a piece wise in time analysis method, if "time" is taken as the threshold variable, its significance is to analyze the time point before and after the structural change of the tax burden rate. In the future, we can consider further examining the impact of fiscal policy change on taxation by taking "time" as a threshold variable. | 8,611.4 | 2023-03-31T00:00:00.000 | [
"Economics"
] |
Growth and distribution : a revised classical model
this paper discusses distribution and the historical phases of capitalism. It assumes that technical progress and growth are taking place, and, given that, its question is on the functional distribution of income between labor and capital, having as reference classical theory of distribution and marx’s falling tendency of the rate of profit. based on the historical experience, it, first, inverts the model, making the rate of profit as the constant variable in the long run and the wage rate, as the residuum; second, it distinguishes three types of technical progress (capital-saving, neutral and capital-using) and applies it to the 3 http://dx.doi.org/10.1590/0101-31572018v38n01a01 Revista de Economia Política 38 (1), 2018 • Brazilian Journal of Political Economy, vol. 38, no 1 (150), pp. 3-27, January-March/2018 * emeritus Professor at Getulio vargas Foundation, brazil. e-mail<EMAIL_ADDRESS>submitted: 16/may/2017; approved: 11/september/2017. f Paper presented to the conference of the research Network macroeconomics and macroeconomic Policies, “Inequality and the Future of Capitalism”, berlin, November 1, 2014. revised in september 2017. the model of this paper was originally developed in the book (bresser-Pereira, 1986). I express gratitude for the comments by adalmir marquetti, adam Przeworski, Daniel saros, eckhard hein, Fabio anderaos, Fletcher baragar, Gilberto tadeu lima, luiz antônio oliveira lima, José luiz oreiro, michael Keaney, takashi Yagi and Yoshiaki Nakano.
is possible to connect historically the types of technological progress with the wage rate and the profit rate, and, so, with functional distribution of income.On the other hand, since economic development is essentially a historical process, it is possible to distinguish stages in this process, and define the stylized facts that characterize it.
In this paper I will present a classical revised model of growth and distribution.It is a classical model for two positive reasons -because it deals with the classical concepts of labor and capital, and uses the classical approach to distribution of income, assuming one form of income as given or determinant and the other as residual -, and for two negative reasons: because, unlike the Keynesian model, it is not concerned with criticizing Say's law, and because, unlike neoclassical economics, it does not adopt a hypothetical-deductive method.It rather adopts a historical-deductive approach that generalizes out of historical experienceparticularly form the new historical facts which are behind the economic and institutional change processes that make economic development to occur in stages of phases.It does need not to criticize Say's law and introduce demand in it because it is not interested in explaining cyclical crises, although it offers a basic explanation for the upturn of the long cycles or waves.It is a revised model because it starts from the inversion of the classical theory distribution.While in the classical model wages are the constant, and the profit rate, the residuum, in this model the profit rate is the constant (for economic and institutional reasons), and the wage rate, the residuum.After arguing for the reasonableness of this inversion in the first section of the paper, in the second section I will present three types of technical progresses.In the third section I will shortly discuss the abstract relationships among the model's main variable.That will enable me to present in the fourth section four historical stages of capitalist development according to these variables, among which the determining one is the type of technical progress.2
Profit rate as a constant
To be simple, the model here presented assumes a closed economy, generalized competition, no state, one commodity, total and marginal output-capital relation equal, and just two agents: capitalists, receiving profits, R, and workers, wages, W. Income, Y, is the sum of wages and profits.Capitalists could be differentiated into entrepreneurs making a profit and rentiers receiving interests, but although I make the distinction in developing the argument, it is not necessary to the model.In the same vein, although I use the state and thus institutions in the argument, it may be ignored in the simpler version of the model.Expenditure is equal to income, and the sum of consumption, C, and investment, I.The functional distribution of income, is R/W = m,3 the profit rate is R/K = r, where R are total profits, and K is the stock of capital.
Economic growth is defined by the increase of productivity and of income per capita.Since I am not looking for the causes of economic growth, but for stylized facts which characterize growth, I can assume that the economy is growing, i.e., that income per capita and the productivity rate are increasing.Given the assumption that the labor force, L, in increasing at the same rate of population, N, the productivity rate, Y/L = y, and the increase of income per The model does not describe any specific capitalist economy, but has as reference the first developed national states, particularly Britain.In the model, as in Kaldor (1956) and in Sraffa (1960), the long term profit rate is assumed to be constant, except in one specific historical phase -Competitive Stage (1815/25 -1875/95) -, in which it is falling from a high level, which prevailed during the 'industrial revolution', to a 'reasonable' level since then.The classical model of Smith, Ricardo and Marx assumed the wage rate constant, corresponding to the cost of reproducing of the labor force.This cost could change historically, but this assumption is inconsistent with the extent that the real wage rate increased in the more developed countries since mid nineteenth century.On the other hand, the classical economists, using different arguments but all involving a fall in productivity, predicted that the profit rate would decline in the long run.This prediction as well proved to be false.Since mid nineteenth century the profit rate remained basically constant, at a 'reasonable', 'satisfactory', or 'satisficing'4 level, i.e., the level that capitalists require to borrow and invest, or that they invest from their retained profits.It varied strongly according to economic cycles, and responds to exogenous shocks, but in the long run it remained constant.
Why does it make sense to assume a constant rate of profit?Essentially it does because, on one hand, a satisfactory profit rate is a condition of existence or survival for the capitalist economic system; on the other, because there is not an economic alternative to capitalism.For some time it was thought that a command economy could be the alternative, but even whilst this belief was alive it was a distant belief.Given this lack of alternative, capitalist societies will have to preserve the profit rate.The capitalist system can only survive if a reasonable profit rate is assured to active capitalists or entrepreneurs -a rate reasonably above the interest rate received by rentiers.On the other hand, although capitalist economies and societies are characterized by instability and conflict, they are, in the realm of each national state, a cooperative undertaking.The existence of nation-states presupposes a broad political agreement.Capitalists fight for profits, but they know that a reasonable wage rate is essential for political stability and a sustained aggregate demand.Correspondingly, workers are asking permanently for higher wages, but they know that their wages cannot reduce the profit rate below a given level because this will endanger the capital accumulation and growth process.
Since classical economists believed that the productivity of labor would decline in the long run, 5 their bottom line was the wage rate.Yet, in so far as this prediction failed to be true, the alternative bottom line is the profit rate to be constant.While the constant wage rate proved a false prediction, and a third alternative -an increasing profit rate in the long run -makes no sense in a competitive economy, wages increasing in real terms in the long run does make sense.When an economy, in its cyclical process of growth, experiments high and sustained rates of growth, the wage rate will tend to increase.Theoretically the wage rate may increase up to the moment in which the economy achieves abundance, i.e., up to the point that people have the full freedom to chose between income and leisure, and overwhelmingly decide for the later.In practical or historical terms, the average wage rate will increase till the bottom line represented by a satisfactory profit rate.From this point on, a profit squeeze process will materialize, and the economy will be experiencing crisis or a threat of crisis, which will only be overcome if the profit rate is reestablished.Since economic agents need that the economy works, they either take the required policy and institutional measures to reduce the wage rate or to increase aggregate demand, or they wait that the market system processes the crisis and reestablishes the profit rate. 6n Marx's falling tendency of the rate of profit theory, the possibility that the countertendencies would effectively neutralize such tendency was considered as a possibility.In the theory that I am presenting, the long term constancy of the rate of profit is based on the institutional defense of the rate of profit by the capitalist class and government.Even if there was an economic alternative to capitalism, this defense would be fierce on the part of capitalists.Since there is not such alternative, capitalist eventually obtain the cooperation of the other social classes in the institutional process of protecting the rate of profit and the process of capital accumulation.Besides the market mechanisms that, till a certain extent, assure the way out of the cyclical crisis, governments are supposed to provide the institutional reforms and policies that will assure that this outcome is achieved, and, in doing so, it will keep the long run rate of profit at a reasonable level, consistent with investment and growth.
Types of technical progress
Technical progress is defined by the increase of the productivity of labor.There is technical progress when productivity is increasing, or, in other words, when workers are being able to increase their average value added.Thus, technical progress involves not only the introduction of new methods of production and new products, but also the transference of labor from activities with lower to activities with higher valued added.Yet, the process of labor productivity increase will be accompanied by changes in the productivity of capital, or the output-capital relation, Y/K (which Marx called technical composition of capital). 7Given the fact that, concomitantly with labor productivity increase, the capital productivity may decrease, remain constant, or increase, we have three types of technical progress, which are defined by the behavior of the productivity of capital.If the productivity of capital is decreasing (Y 2 /K 2 < Y 1 /K 1, where i indicates time), technical progress will be capital-using, or we will have 'mechanization'.If the product-capital relation is constant, technical progress will be neutral.And if the productivity of capital is increasing, technical progress will be capital-saving.
In the case of capital-using technical progress, income will be increasing at a smaller rate than capital: In the case of neutral technical progress, where Y 2 /K 2 = Y 1 /K 1 , income will be increasing at the same rate as capital: In the case of capital-saving technical progress, where Y 2 /K 2 > Y 1 /K 1 , income will be increasing at a higher rate than capital: How can we have, out of rational investment decisions, a situation in which technical progress involves a decreasing output-capital ratio?Or, in other words, which is the microfoundation for the choice of a capital using technique?Capital-using technical progress is typical of the 7 In this paper I used the concept of output-capital relation, Y/K, which Marx called 'technical composition of capital', avoiding the use of the 'organic composition of capital' concept which rather complicates than simplifies the argument.In the growth literature, capital-output relation is more often used, but I prefer its inverse, the output-capital relation, because when one says that this ratio is increasing, this means that capital productivity in increasing.
early stages of industrialization and capital formation, when mechanization or the successive substitution of different machines for different forms of labor is taking place.Whenever the costs involved in buying and operating a machine (and the respective production process) are smaller than the use of man-power, it will be rational for the capitalist to invest in this machine.As the business enterprise substitute capital for labor, the production costs will be reduced and the total productivity of labor will increase.Yet, the machines available have different productivities, and replace different types of labor.If we suppose that the business enterprises face a decreasing investment opportunities curve having, in the vertical, the cost reduction achieved, and in the horizontal axes, the respective machines and correspondent production processes available, the business enterprises will, first, invest in the more efficient machine, which replaces one kind of labor; second, they will buy the second best machine, which replaces a different type of labor; and so on, up to the point where breakeven is achieved. 8Despite the fact that, in this decision process, the productions costs decreased for the business enterprises as different kinds of labor were successively replaced by different kinds of machines with decreasing productivity, each new machine replacing different kinds of labor will reduce, the overall output-capital relation or the productivity of capital will decrease.Take, for instance, the choice of techniques in an economy that has only agricultural production, and that replaced all labor that was possible for a highly efficient machine, the tractor.Now, the second machine available which is economical or just became economical (i.e., reduces costs to the entrepreneur) is a harvesting machine.All farmers will have to buy or rent it, but, as they hold a lower output-capital ratio, the total output-capital ratio of the economy will fall, despite the fact that their costs were reduced.It is true that, in a given moment, a new machine replacing a type of labor which had not yet been mechanized, which is not less but relatively more efficient than the ones which had previously replaced other forms of labor, may be invented and made available to business enterprises.In this case, in which our cost-machines curve does not hold, mechanization will not cause the fall of the productivity of capital.Yet, this situation will be rather the exception than the rule.The tendency is that the inventions and specially innovations (the actual adoption of the invention) take place in sequence in such a way that the first innovations involve high output-capital ratios and the following, increasingly smaller ones.In this case, technical progress will be necessarily capital-using, the output-capital relation will be declining.
When the output-capital relation is decreasing, we will see that the profit rate will be decreasing.With this argument, we can understand something that appeared irrational: firms to adopt capital-using techniques which eventually will reduce instead of increase its profit rate.This was the question posed by the Okishio theorem (1961Okishio theorem ( , 1967) ) challenging the possibility of a falling tendency for the rate of profit.Yet, as I argued, the business enterprise is acting rationally when it adopts the new technique or machine which is capital-using but efficient.It will have no alternative but to adopt it.Its strategy will be just a defensive one -a strategy to keep it competitive given the fact that the other firms will also the capital-using but cost reducing technique.The fact that, once all business enterprises replaced manpower for a given relatively (to the previous ones) less efficient machine, the resulting output-capital ratio for the whole industry and the average rate of profit will be smaller, is out of the control of each individual firm.This is a perverse but rational effect of mechanization or the adoption of capital using technical progress. 9While capital-using technical progress or mechanization involves the substitution of capital for different activities performed by labor, capital saving technical progress, which also may be called 'modernization', derives from the substitution of new machines for old ones of the same type (i.e., which replace the same kind of labor, or performs the same kind of operation that a previous one performed).It is only the type or model of the machine that changes, since it replaces the same type of labor.The new model, however, is cheaper, or more efficient.In this second case, technical progress besides saving labor saves capital itself, increasing the output-capital relation.While in the case of mechanization the business enterprise had no other alternative than to invest in increasingly less efficient machines, in this case, it again will not have other alternative but in investing in increasingly more productive or less expensive machines -machines that are able to turn out a larger output (with the same quality) per unit of capital.New machines, in this case, are new in relation to other models of machine performing the same operation, while new machines in the previous case are machines performing new operations and thus replacing new types of labor.New machines will only appear in the market as they bring some innovation and lower costs, but there is a major difference between new machines performing new operations which were previously manual, and new machines replacing old machines.In one case we have capital-saving technical progress, in the other, capital-using technical progress.
In the case of neutral technical progress, there is not a specific form of substitution of capital for labor, or the need of reasoning in terms of microfoundations.This sort of technical progress just exists in so far as the two previous processes -mechanization and modernization -compensate one another.At every moment we will have new types of labor being replaced by new types of machines, and old machines being replaced by new models of the same machines ('same' just in so far it replaces the same type of labor).In the first case, technical progress will be capital-using, in the second, capital-saving.If the negative effect of the first is compensated by the positive of the second, technical progress will be neutral.Most growth models concerned with equilibrium (or lack of it) and with the determinants of the rate of growth, as it is the case of the Harrod-Domar and the Solow models, assume neutral technical progress.In the relatively short run (medium run) periods in which such models usually used, such assumption is reasonable and simplifies the model.In the present model, however, principally concerned with distribution in the long run, across the several historical stages or phases of economic development, to renounce such assumption is essential.
The abstract relationships
Given these three forms of technical progress, or the variation of the output-capital ratio, we will have different behaviors of the other central economic variables: the profit rate, the wage rate, and the functional distribution of income.These variables are related among themselves following a simple identity: Let us suppose, first, that the functional distribution of income between profits and wages is constant: R/Y→.In this case, and just having in mind that an increasing capital-output ratio means a decreasing output-capital ratio, it is easy to see, from identity (1), that, if technical progress is capital-using (declining output-capital ratio, Y/K↓), the profit rate will be declining, R/K↓; if technical progress is neutral (constant output-capital ratio, Y/K→), the profit rate will be constant, R/K→; and if technical progress is capital-saving (increasing output-capital ratio, Y/K↑), the profit rate will be increasing, R/K↑.
Thus, we cannot speak of a general tendency of the rate of profit to fall, increase, or remain constant just out of (1).Depending on the type of prevailing technical progress, and given a constant functional distribution of income remains, the rate of profit will correspondingly fall, remain constant, or increase.If, instead, we assume that the constant variable in the long run is the profit rate, as I already argued, and that economic growth is taking place, which will be functional distribution of income and the wage rate for each type of technical progress?
To answer this question, I start by taking the time derivative of equation ( 1), setting it to zero, and substituting R+W for Y.
Computing the derivatives and doing some algebraic manipulations we have the following equation: ( ) The left hand side of the equation ( 2) has the same sign as the rate of change of the functional distribution of income R/W and the right hand side has the opposite sign to the rate of change of Y/K, since: Therefore, equation (2) tells us that if technical progress is capital-using, or Y/K↓, the functional distribution of income will concentrate, so that R/W↑.If technical progress is neutral, the functional distribution of income will remain unchanged, and if technical progress is capital-saving, or Y/K↑, it must be that R/W↓ Now, to understand what happens to the wage rate, let us assume that the population is constant.In a more complete version of the model, population is increasing at a constant rate.
Yet, to simplify the equations and show more clearly the relations between the variables, we assume that population is constant, and then the wage rate, W/L, will depend only on the change of W. We now re-write equation ( 2) to analyze how W varies: Because the profit rate is assumed to be constant, we can substitute the rate of growth of capital for the rate of growth of profits in the above equation, which after some algebraic manipulations yields: From equation (3) we deduce that, if technical progress is neutral or capital-saving, total wages increase because 0 < R/Y < 1 for positive wages and profits, which makes the right hand side of equation ( 3) positive when the rate of growth of output is equal to or higher than the rate of growth of capital.Given the assumption of a constant population, when technical progress is neutral or capital-saving, the wage rate will also increase.
When technical progress is capital-using, on the other hand, equation (3) gives us an ambiguous result.If capital grows at a higher rate than output, the right hand side of equation ( 3) can be either positive or negative, depending on the magnitude of R/Y.Table 1 summarizes these results.In the case of capital-using technical progress or mechanization, which involves a concentrating functional distribution, the wage rate may or may not be decreasing, depending on the rate of growth of income per capita, which influences R/Y.In the summary analysis that I will do of the historical stages of capitalist growth, the fall in the output-capital relation only takes place in the two first stages.In the first stage, the Industrial Revolution, in which mechanization is assumed, this ambiguity will remain.In order to keep the rate of profit constant, the wage rate probably fell, at least in terms of real income and standard of living.In the Competitive Stage, however, stage, the ambiguity will disappear despite mechanization, because I drop the assumption that the rate of profit was constant, because I assume that it was exceptionally high during the industrial revolution, and let it fall for the period.This allows the wage rate to remain approximately constant and the functional distribution of income to concentrate, as probably happened in this period.101960), the 'take-off' takes place.In the Commercial Revolution, primitive accumulation -the initial accumulation of capital through the use of some form of violence -created the conditions for the subsequent generalization of wage labor and the competitive appropriation of surplus through profits (Marx, 1867: I,24).With the Industrial Revolution -a concentrated process of industrialization involving positive externalities or spillovers, and, consequently, high profit rates -, capitalist development becomes selfsustained in so far as the reinvestment of profits to keep pace of technological progress becomes a condition of survival of the business enterprises.
In this search for stylized facts, the Industrial Revolution, which I broadly located for Britain between 1750 and 1815, will be characterized by a high and constant profit rate, while technical progress will be dominantly capital-using.This is consistent with a declining wage 11 I attempted to do that in Bresser-Pereira, 1986.rate, and with the concentration of the functional distribution of income.The wage rate may be declining because it is assumed that workers, immediately before the Industrial Revolution, had a higher standard of living: the first moment of industrialization represented for them 'proletarization' or pauperization.Yet, if income per capita is rising fast, despite the concentration of income, the wage rate may be stagnant or even increasing.Thus, in this phase, we have: The following phase is the Competitive Stage.It is the phase in which economic liberalism or competitive capitalism is dominant.The transition from pre-capitalism was completed.The economy is characterized by a large number of small and medium sized family enterprises.It is essentially competitive, since the gigantic business enterprises are not yet present.The economic system corresponds to the one predicted and describe by the classical liberal economists.Technical progress remains capital-using since mechanization continues intense overcoming the modernization process.Thus, we have decreasing returns.Yet, contrarily to the classical economists' prediction, the wage rate does not fall but probably remains constant.This is possible because the Competitive Stage is the only phase in which the profit rate probably decreases -something that is possible if we assume that it was very high during the Industrial Revolution.This decrease leaves ambiguous the functional distribution of income, which probably continue to concentrate, but much less than in the previous phase, and may even have remained constant, depending on the rate of growth of the income per capita.
Competitive Stage
By the second part of the nineteenth century, around 1870, we have major changes which bring the Classical Stage, the stage the capitalist growth gets fully consolidated: mass production techniques are introduced, the explosion motor replaces the steam motor, and electrical power is dominated and diffused (the Second Industrial Revolution).As a consequence, the economic system turns relatively less competitive, in so far as large business enterprises start dominating the scene, and in so far as workers get organized in unions.Both changes were interdependent: the higher level of workers' organization was only possible in view of the relative oligopolization of markets.From this, follows a major consequence: workers became capable of retaining the productivity gains.Economic theory based on competition assumed that productivity increases would just lead to lower costs which would benefit all, including foreign consumers.The new workers' organization capacity turns possible what, in the late 1940s, the Prebisch's these on the uneven distribution of the productivity gains between developed and developing countries is formulated: while industrial countries which had organized labor were able to conserve productivity gains, disorganized workers producing primary products in developing countries were not, from that deriving the deterioration of the terms of exchange.For our model, only the first aspect of the problem is important.With the Second Industrial Revolution, these characteristics are just enhanced.Markets are increasingly oligopolist, but business enterprises remain competitive enough to keep centrally concerned with the incorporation of technical progress.In so far as mechanization and capital-savings technology compensate one another, the output-capital ratio is basically constant (as growth models usually assume), technical progress is neutral.
From this moment on, workers would be able to augment their wages according to the productivity rate without threatening the profit rate.Capitalism achieves its classical moment.
The great agreement between capitalists and workers, which would assure a relative social peace in developed industrial countries, begins.Technical progress is neutral, the profit rate is constant, and the functional distribution of income, constant; as a consequence, the wage rate increases with productivity.
Classical Stage
Finally, after World War II, we see, first, the definitive rise and affluence of the professional middle class or technobureaucratic class, and, in a second moment, globalization.I refer to this phase as the Knowledge Stage because technical, organizational and communicative knowledge become the strategic factor of production.The system remains capitalist, but the power and income of the individuals possessing one of these three types of knowledge (or, still better, a combination of them) increase in relation to the power of inactive or rentier capitalists.Yet, despite the fact that the managerial elites replace in large extent the capitalist class in controlling the large business organizations, the market logic of capitalism remains unchanged.Or the changes are not big enough to legitimize expressions like post-capitalism.
The economy continues to be essentially market coordinated; the profit motive continues to be central, and capital accumulation with the incorporation of technical progress remains in the core of the growth process.Technical progress remains neutral, although we can already notice a clear tendency for capital-savings technologies to dominate, particularly in the realm of information technology.Yet, three central new facts give rise to the Knowledge Stage: the definitive emergence of the professional or technical middle class, the technology of information revolution, and globalization.These new facts cause a major change in the capitalist system: capital ceases gradually to be the strategic factor of production, as it is replaced by technical and organizational knowledge (Galbraith, 1967;Bresser-Pereira, 1972, 1981).The new middle class, or professional middle class, receiving salaries not wages,12 now share power and income with the capitalist class, to which the most successful managers soon join as they become rich.
Transition Y/K→ R/K↓ W/L↑ R/W↓* Between 1945 and 1970 economic growth is extraordinary.This was the Golden Age of capitalism (Glyn et al., 1988;Marglin, 1990), and, although I am including it in the new phase, it may also be viewed as a transition period.The profit rate keeps high, while wages and particularly salaries are increasing fast.Fast enough to cause, in the 1960s, a profit squeeze, and the fall in the rate of profit (Boddy and Crotty, 1975;Goldenstein, 1999).The neo-liberal ideological wave and the institutional market oriented reforms that began in the following decade are a reaction of the system to restore the satisfying profit rate -something that is achieved in the 1990s (Wolff, 2001;Brenner, 2002;Duménil and Lévy, 2002).Not considering these cyclical variations, technical progress remains basically neutral (although the information technology revolution points out toward a capital-saving type of technical progress),13 the profit rate remains constant, and the distribution between profits and wages also remains basically constant in the Knowledge Stage.Yet, the internal distribution of wages between wages stricto sensu (the remuneration of the industrial and service working class) and salaries (the compensation of the professional middle class) changes decisively in favor of the latter.The personal concentration of income which starts around 1975 all over the world provides evidence of this fact.The figures about income distribution that are normally published by government agencies are figures about the personal (not the functional) distribution of income.The official statistic bodies that do such surveys usually do not clearly distinguish profits from wages lato senso, and definitely do not differentiate wages from salaries, just classifying the population according to percentage groups of income.The changes in the Gini coefficients derived from the tables reflect principally the distribution between salaries and wages, since profits are underestimated in this kind of statistic.In so far as the demand for technical labor increased strongly with the information technology revolution, while the demand for non-technical labor lagged behind, personal concentration of income was inevitable.On the other hand, as the top professional class assumed increasingly the direction of major business enterprises, and capitalists were increasingly reduced to the condition of rentiers or inactive capitalists, they accepted that the profit rate went down a bit (the dent in Figure 1 reflecting this fall), to stabilize at a new constant lower level, the difference being appropriated by salaries, particularly by top executives' high salaries.
After the transition, the basic relations in the Knowledge Stage show the same trends that in the Classical Stage, since technical progress will continue to be defined by a constant outputcapital ratio, while the profit rate and the functional distribution of income remain constant, and the wage rate increases again at the rate of increase in productivity.
Knowledge Stage
Y/K→ R/K→ W/L↑ R/W→ These relations can be seen in Figure 1, where we have the trends of the four variables, or the stylized economic facts for the four phases.Notice that the order of the variables is relatively arbitrary.Technical progress is capital-using in the first and second phase, neutral in the third
The rate of growth
As I remarked in the beginning of the paper, these models do not deal with the factor that cause a higher or a smaller rate of growth, but with the behavior of the profit and the wage rate, i.e., with distribution.This is only partially true.Each of the three forms of technical progress implies a rate of growth, in so far as the output-capital ratio measures the productivity of capital.Given the other variables constant, growth will be higher if technical progress is capital-saving than if it is neutral, and still higher if it is compared with a moment in which predominates capital-using techniques.Yet, the model does not say which will be this rate, because the increase in the labor productivity does not depend only on the type of technical progress, it also depends on the 'intensity' of technical progress: periods of intense technical progress will tend obviously to show higher rates of growth than periods where innovation gets protracted.And, naturally, it is also dependent on the rate of capital accumulation.The intensity of technical progress depends on a large number of variables, as education, entrepreneurial capacity, labor and entrepreneurial motivation, rational allocation of resources, institutions adequate to growth, competent and growth oriented economic policies, etc.These are the microeconomic reforms, in which dynamic capitalist economies are supposed to be permanently involved.
Economic growth depends on a fourth factor, in addition to the rate of capital accumulation, the type, and the intensity of technical progress: macroeconomic stability, which is expressed into stable prices, a moderate interest rate, a competitive exchange rate, and a reasonable degree of full employment.Since economic growth is the long term sum or the integral of short term GDP per capita growth rates, the growth achieved every year is important.In some moments, it is required to sacrifice the short term for the long term, and get involved in economic adjustment, but when a country does this, it expects that the growth rate through time will be higher.Full employment, on its hand, depends on competent macroeconomic policies managing the economic cycle.
What does this long term historical model of growth say in relation to administration of the business cycle and particularly of the long cycles?It says essentially one thing.When an economy faces some king of crisis, this means that the expected profit rate fell down, that investments were reduced.Thus, the solution will necessarily involve the restoration of the profit rate.We can have at least two types of crisis: a slowdown or normal recession, or a major and long term crisis.In the first case, restoring the rate of profit will involve monetary and fiscal policy.However, if the crisis involves a fall in the profit rate, as it happened in the United Sates in the 1970s, or is related to large foreign indebtedness and macroeconomic instability, as has been in most Latin American countries since 1980, the solution will probably involve institutional reforms and wage reduction.Thus, keeping the profit rate constant at a reasonable level is not just a long term sensible assumption; it also may indicate which should be the required long term institutional reforms and the macroeconomic policies that will recover the expected profit rate, stimulate investment, and resume economic growth.
Conclusion
The revised classical model of growth that I just presented is a historical model.
Concomitantly, it is an abstract and general growth model, where the stylized facts about economic growth and distribution appear clearly.It assumes an investment function: capital accumulation depends on the expected rate of profit.It also assumes a production function: growth depends on investments, and on the type and intensity of technical progress.It also depends on full employment, since the production function defines the potential output; the actual product will also depend on effective demand, or on reducing the output hiatus.
The model aims at being simple and general without losing a historical perspective.Thus, it assumes a closed economy, competition, and the existence of only two economic agents: capitalists and workers.The state is present in the model not as an economic agent collecting taxes and providing economic transferences, but only defining the institutions required for markets to operate and the profit rate to be assured at a satisfying level.Given the existence of three types of technological progress (capital using or mechanization, neutral, and capital saving), I show how -in the process of economic growth or increase of labor productivitythe profit rate, the wage rate, and the functional distribution of income between profits and wages vary in relation to these three types of technical progress, which are defined by the technical composition of capital or output-capital relation.In the model, technological progress is defined by the increase of labor productivity (which corresponds to the increase of income per capita, if one assumes as constant the active/inactive labor force relation).
Technological progress will be capital-using if the increase in labor productivity entails the reduction of the output-capital ratio.It will be neutral, if economic growth takes place with a constant output-capital ratio; and it will be capital-saving if this ratio increases.
Marx's theory on the falling tendency of the rate of profit hypothesis is only valid if and while the capital-using technological progress was dominant.If technological progress is assumed to be neutral, the profit rate will remain constant, while the wage rate will increase according to the growth of labor productivity.In the moment that capital-saving technology becomes dominant the wage-rate could increase more than the productivity rate, while the profit rate would remain constant.
The assumption of a constant rate of profit is based on two other assumptions besides the fact that mechanization is dominant just in the early periods of capitalist development: that there is no alternative form of economic organization to capitalism, and that capital accumulation and growth depend on a satisfying profit rate.Thus, the profit rate plays a central role in the model.Whenever appear a tendency to the fall of the rate of profit (as it happen between the late 1960s and the 1980s), the economic and political system reacts in order to restore it.
From this model, and from basic factual knowledge on the history of modern capitalism, it is possible to derive the stylized facts of capitalist growth.Britain and, more generally, the countries that first completed the capitalist revolution are taken for reference.Economic increase in labor productivity is taking place.The model does not discuss this rate.It assumes that it will depend on the rate of capital accumulation, the intensity of technological progress, and the effective use of capital and labor inputs.From the assumption historically verified that economic growth is happening, it looks for the stylized facts involved in this growth process.
In the first stage (the industrial revolution), the only important supposition is that the profit rate is high.Given this assumption, in the second phase (competitive capitalism) the profit may decline without threatening to paralyze the process of capital accumulation.
Technological progress is capital-using, and the functional distribution of income (or surplus value rate) is constant, as the wage rate is reduced to the cost of reproduction of the labor force level, and the rate of profit is declining.In the third period, the Classical Stage, we have a kind of long term steady state.Technical progress becomes neutral, the functional distribution of income between profits and wages is constant, the profit rate constant, and the wage rate increases with productivity.Finally, with technobureaucratic capitalism, technical progress is capital-saving.The functional distribution declines, as the profit rate remains constant, while wages (which now, given the rise of the professional middle class, includes salaries), increase at a higher rate of productivity.To be more precise, the wage rate stricto senso remains constant, but the salary rate increases substantially.On the other hand, given the fact that now the profit rate remunerates principally rentiers or inactive capitalists, and the entrepreneurial activity is also paid with high salaries received by top managers, the capita, Y/N= n, are equal.The variation of y through time is y and fourth.Not considering the 1945-75 transition, only for the Competitive Stage I dropped the assumption of a constant rate of profit, because in the previous phase the rate of profit was above satisfactory level.The functional distribution of income is increasing (concentrating) in the Industrial Revolution, constant in the remaining phases.The wage rate decreases in the Industrial Revolution, turns constant in the Liberal or Marxian phase, and starts to increase with productivity in the Classical Stage.It increases above that level in the transition to the Knowledge Stage, to take into consideration the sharp increase in the professional middle class' salaries, but after that it is supposed to grow with productivity increase.
Figure
Figure 1: Stylized economic facts in four historical stages growth turned out in four phases: the industrial revolution, from late eighteenth century to around 1915; the Competitive Stage, from 1815 to around 1870; the Classical Stage, from 1890 to 1945/70; and the Knowledge Stage, from 1970 till presently.In the four phases,
Table 1 : Types of technical progress, the wage rate and the functional distribution of income, given a constant profit rate.
This is not the moment to review and argue for these historical phases or stages.11TheIndustrial Revolution is a well known process.It is the moment in which the capitalist revolution, which began with the Commercial Revolution, comes to an end.It is the moment when, in Rostow's terms ( | 9,513 | 2018-03-18T00:00:00.000 | [
"Economics"
] |
Solid‐State NMR Spectroscopic Investigation of TiO2 Grown on Silica Nanoparticles by Solution Atomic Layer Deposition
Atomic layer deposition in solution (sALD) is just emerging as a technology for the preparation of thin films. Unlike ALD from the gas phase, it allows for mild reaction conditions in a solvent phase and at room temperature, thus decreasing the energy requirements of the process and widening the range of accessible precursor molecules. In this work, the deposition of thin films of titania on silica is investigated using titanium(IV) isopropoxide (TTIP) and water as precursors, which are alternatingly brought into contact with the support in a home‐built plug flow reactor. The mechanism of covalent grafting of the precursor to the surface, subsequent hydrolysis, and reaction to a layer of titania are investigated in detail using magic angle spinning (MAS) solid‐state nuclear magnetic resonance (NMR) spectroscopy. TTIP preferentially reacts with Q2 groups of condensed silica. 2D solid‐state NMR spectra allow to clearly show the successful grafting of this compound to the support by the appearance of a characteristic signal at −107 ppm, which is tentatively attributed to silicon nuclei in a SiOTi bond, and to reveal the presence of titanol groups on the emerging TiO2 film.
Introduction
Thin film materials with adjustable physical and chemical properties are of high interest for the fabrication of microelectronic DOI: 10.1002/admi.202202131 devices. [1] Among other techniques, thin films have been prepared by atomic layer deposition (ALD) in the gas phase since the 1990s. [2] ALD is achieved by exposing a substrate surface to alternating gaseous precursors. In between the pulses of the precursors, the reaction chamber is flushed with an inert purge gas (usually nitrogen). Thanks to the self-limiting nature of the process, ALD yields atomically controlled films of homogeneous thickness, good conformality, and trenchfill capability. [3,4] However, the technique also requires costly equipment, strict vacuum conditions, and high temperatures, making the process energy-intensive. It is furthermore limited to precursors that are volatile in a suitable temperature range without decomposition. [5] ALD in solution (sALD) can overcome these limitations. In sALD, diluted solutions of the precursors in an inert solvent are brought into contact with the substrate, applying alternating flows of the different solution at room temperature. As in gas-phase ALD, two half-cycles are necessary for one full monolayer deposition. sALD was introduced in 2009 for vanadium-doped TiO 2 and ZrO 2. [6] We have implemented this technology for a wide range of materials which are hardly or not processable by gas-phase ALD, such as TiO 2 , SiO 2 , MgO, [7] and PbS, [8] but also for methylammonium lead iodide perovskites. [8] In sALD, high temperatures are avoided, which opens up a much wider range of possible precursors than gas-phase ALD. At the same time, the advantages of conventional ALD, including the self-limited nature and the homogeneous film thickness are retained.
However, although an idealized model of the surface deposition process in sALD exists, more detailed insights into the processes at the surface are still lacking, in particular with respect to the grafted surface species and the interaction between nongrafted species or solvent molecules with the surface. The goal of sALD is the formation of a homogeneous, closed film after one deposition cycle, but this may not be the case for all substrates and precursors.
TiO 2 is chemically and physically stable and one of the most widely used semiconductors [9] with applications in photovoltaics, [10] sensors, [11] and photocatalysis. [12] We have recently achieved homogeneous deposition of 30 nm thin films of TiO 2 on flat silicon wafers over a SiO 2 layer, using alternating flows of titanium(IV) isopropoxide (TTIP) and water in diethyl www.advancedsciencenews.com www.advmatinterfaces.de ether. [7] An idealized reaction scheme is shown in Figure 1. However, many questions remain open in this process: How do TTIP molecules bind to the silica? Does TTIP during a single half-cycle completely cover substrate? If not, what is the coverage fraction? What is the impact of the pre-treatment of silica in this research model? The detailed elucidation of the reaction will help the optimization of this and other sALD processes.
In this work, we therefore investigate the deposition of titania on silica nanoparticles by solid-state nuclear magnetic resonance spectroscopy under magic angle spinning (MAS NMR), inductively coupled plasma optical emission spectroscopy (ICP-OES), and ellipsometry. By 1D and 2D 1 H and 29 Si MAS NMR techniques, we follow the deposition of the precursor titanium isopropoxide onto silica with different degrees of hydroxylation. The chemical binding of TTIP to the surface, as well as the further reaction toward a TiO 2 layer, can be clearly traced in MAS NMR spectra. We identify a so far unknown silicon species with a chemical shift of −107 ppm, which we attribute to Si-O-Ti bonds. The spectral features of TiOH groups become visible after a full sALD cycle. After 20 cycles, a closed multilayer of TiO 2 has developed on the silica particles.
Experimental Section
All substances were used as received, if not indicated otherwise. Diethyl ether (Sigma-Aldrich, ≥ 99.7%) was recycled by rotational evaporation in-house and stored over molecular sieve. The particle size of silica nanoparticles (Sigma-Aldrich, 99.5%) is in the range of 5-20 nm. The calcination of silica was carried out in a muffle furnace in a flat layer in an open vessel at 500°C (heating ramp 2 K min −1 ) for 8 h, then cooled to 120°C with a rate of 1.5 K min −1 and kept at 120°C. The hot sample was transferred into the glove box and stored under nitrogen.
Pre-Treatment of Silicon Wafers
Silicon wafers (Silicon materials) were cut into small pieces of 0.5 cm 2 and then ultrasonicated for 7 min each first in acetone, then in isopropanol, and then in ethanol, and dried under ambient conditions.
sALD in Batch Mode
In a typical batch sALD process, 0.100 g (1.6 mmol) of silica powder was dispersed in 30 mL anhydrous diethyl ether. 18 μL (0.06 mmol) of TTIP (Acros Organics, ≥99%) were added to the mixture and shaken, to achieve a concentration of 2 mmol L −1 TTIP in the suspension. The suspension was then centrifuged with 6000 rpm (2000 g) for 40 s. After removing the supernatant, the suspension was washed again with anhydrous diethyl ether and centrifuged again. The washing step was repeated ten times. A dry powder was obtained, which corresponds to a "half sALD cycle" sample. To obtain a sample with a complete sALD cycle, the obtained sample was reacted in the same way with a solution of 2 mmol L −1 water in anhydrous diethyl ether, followed by the same washing procedure. All operations are carried under nitrogen atmosphere in a glove box.
sALD Plug Flow Reactor
In a typical preparation using a home-built plug flow reactor, solutions of all precursors as well as the solvent anhydrous diethyl ether were prepared in advance in a nitrogen atmosphere and stored in Schlenk flasks. Water in diethyl ether was ultrasonicated for 20 min prior to use. A plug flow reactor was designed in-house (Figure 2): A silicon tube was assembled with a fitting, followed by a silicon wafer, a layer of cotton, and a filter paper at each end of the tube. Then the tube was filled with ≈180 mg of silica powder, slightly compacted, but not tightly pressed. The increase in thickness of the growing TiO 2 layer was determined on the silicon wafers at both ends of the reactor by ellipsometry. After the reaction, the silica powder was removed from the reactor for analysis. To prevent the precursors from reacting with each other in the pipeline, three branchings were installed in front of the reactor (Figure 2a and Figure S1, Supporting Information). The four channels of the peristaltic pump controlled the entry of the precursors and solvent into the reactor. The inner diameter of the peristaltic tube of the peristaltic pump was fixed to 1.52 mm. The feed duration and flow rate in each channel were controlled using a pump control software by Ismatec. Each cycle consisted of four stages: t 1 , the pulse duration of TTIP feed, t 2 , the purging time after the TTIP flow, t 3 , the pulse time of the water feed, and t 4 , the time of purging after water the feed (for details, see Figure S1, Supporting Information). All subsequent cycles were performed in the same manner.
After the reaction, the tubes were cleaned using subsequently diethyl ether, isopropanol, deionized water, 5% aqueous potassium hydroxide solution, deionized water, and isopropanol, at a rate of 3 mL min −1 for 6 min each.
Solid-State NMR Spectroscopy
MAS NMR spectra were acquired on a 500 MHz (11.7 T) DD2 Agilent wide-bore NMR spectrometer in 1.6 mm zirconia rotors, with the exception of CP build-up experiments, which were recorded on a 400 MHz (9.4 T) Bruker Avance spectrometer in a 4 mm zirconia rotor. Samples were packed into the rotor under an Argon atmosphere in a glove box. For packing of slurries, the rotors were additionally sealed with silicone plugs.
1 H Spectra
1 H spectra were recorded at a MAS rate of 30 kHz using the DEPTH sequence for background suppression, starting with a www.advancedsciencenews.com www.advmatinterfaces.de /2 pulse of 2.5 μs followed by 2 pulses of 5 μs phase-cycled according to a combined "EXORCYCLE" and "CYCLOPS" pulse scheme. [13][14][15] 16 scans were recorded, and the recycles delays were set to five times the 1 H longitudinal relaxation time, determined in saturation recovery experiments. Spectra were referenced to sodium trimethylsilylpropanesulfonate. 13 C cross polarization (CP) spectra were recorded at a MAS rate of 15 kHz. After a /2 pulse of 2.5 μs on 1 H, the RF of 1 H was ramped from 75 to 100 kHz while the RF on 13 C was kept constant at 75 kHz during a contact time of 2 ms (first spinning sideband condition). 100 kHz spinal-64 decoupling was applied on 1 H during acquisition. Recycle delays were set to 5 s, 6000 to 10 000 scans were recorded depending on the sample. For CP experiments on dried and calcined silica, the powders were wetted with diethyl ether prior to measurement. 29 Si CP spectra were recorded at a MAS rate of 15 kHz. After a /2 pulse of 2.5 μs on 1 H, the RF of 1 H was ramped from 55 to 76 kHz while the RF on 29 Si was kept constant at 75 kHz during a contact time of 2 ms (Hartmann-Hahn match). 100 kHz spinal-64 decoupling was applied on 1 H during acquisition. Recycle delays were set to 2 s, 10 000 to 40 000 scans were recorded. Direct excitation spectra were recorded at a MAS rate of 14 kHz, using a /2 pulse of 5.25 μs, an interscan delay of 300 s and around 400 scans depending on the sample. 29 Si heteronuclear correlation (HETCOR) spectra were acquired at a MAS rate of 15 kHz. After /2 pulse of 2.5 μs on 1 H, the RF of 1 H was ramped from 55 to 76 kHz while the RF on 29 Si was kept constant at 75 kHz during a contact time of 2 ms (Hartmann-Hahn match). 700 increments during the evolution period were recorded, while applying 100 kHz FSLG decoupling. 100 kHz spinal-64 decoupling was applied on 1 H during acquisition. 1600 scans were accumulated in every increment with a recycle delay of 2 s.
Ellipsometry
Spectroscopic ellipsometry data were collected on a SENPro spectroscopic ellipsometer from SENTECH with range from 370.0 to 1050.0 nm under a 70°incidence angle. The fit model is based on Cauchy layers for the "TiO 2 -SiO 2 -Si" stack, [16] which is implemented in the data evaluation software Spectra Ray 3.
Energy-Dispersive X-Ray Spectroscopy
Energy-dispersive X-ray (EDX) spectra were acquired on a JSM 6400 by JEOL with an acceleration voltage of 10.00 kV and a magnification of 10 000.
Nitrogen Adsorption-Desorption Isotherms
Textural properties of the calcined silica was assessed by a N 2 adsorption-desorption isotherm at 77 K recorded on an ASAP 2000 from Micrometrics, Inc.: 100 mg of sample were used and outgassed under vacuum at 250°C for 12 h before analysis. The specific surface area of the samples was calculated using the Brunauer-Emmett-Teller (BET) method within the relative pressure range of 0.05-0.2.
Scanning Electron Microscopy
Scanning electron microscopy (SEM) images were recorded with an SE2 detector in the electron microscope ULTRA55 (Carl Zeiss MST AG).
Dynamic light scattering
The dynamic light scattering (DLS) measurements were made with a Zetasizer Nano ZS (Malvern Instruments) with anhydrous diethyl ether as solvent.
Results and Discussion
A silica powder with primary particle size of 5-20 nm was used in this work. To track possible changes in the silica morphology by calcination, the material was characterized after calcination by SEM, N 2 adsorption-desorption isotherms and DLS in anhydrous diethyl ether ( Figures S2-S4, Supporting Information). N 2 adsorption-desorption isotherms show that the calcined silica has a surface area of 630 m 2 g −1 , determined by the BET ( Figure S2, Supporting Information). During calcination, sintering of the primary particles occurred, yielding a size distribution of particle agglomerates centered at 600 nm ( Figure S4, Supporting Information) in DLS measurements.
To achieve sALD of thin layers of titania on powdered silica, two different processes were developed in this work. sALD was carried out in a batch mode in vials, by adding to the dry silica powder a solution of TTIP in anhydrous diethyl ether (called here first half sALD cycle), alternated with a solution of water in anhydrous diethyl ether (second half sALD cycle). The process was carried under a nitrogen atmosphere in a glove box and yielded small amounts of ≈20 mg of silica powder. Losses during washing by centrifugation could not be completely avoided, therefore, this process was only used for experiments in which one half or one full sALD cycle was carried out. By 1 H MAS NMR investigation, ten washing steps and a concentration of 2 mm TTIP and 2 mm water in diethyl ether were found to be optimally suited (see below). For thicker TiO 2 layers produced by multiple sALD cycles, a larger plug flow reactor process was developed, which allowed for sample masses of ≈180 mg of powdered silica. In a plastic tube, the dry powder was packed between filter papers and layers of cotton ( Figure 2a). Solutions of TTIP in anhydrous diethyl ether, water in anhydrous diethyl ether, and pure anhydrous diethyl ether for purging were then flown alternatingly through the packed bed. The operating parameters of the previously constructed experimental pump set-up were based on the existing work by Wu et al. (Table 1). [7] The concentration of precursors was optimized before in batch mode experiments to 2 mm.
The precursor pulse time was optimized so as to achieve a monolayer of TiO 2 in every sALD cycle. Since the thickness of the titania layers on a powdered silica could not directly be obtained, the silicon wafer at the inlet and the outlet of the reactor were measured by ellipsometry, as TTIP readily forms titania on this support (Figure 3). With a pulse duration of about 15 s, a growth rate per cycle (GPC) of 0.3 Å, indicative for the formation of a single TiO 2 layer, was obtained. [7,17] 15 s pulses were therefore retained for further experiments. Larger pulse du- rations lead to a higher GPC, which may be caused by an incomplete removal of precursor. An extension of the purge durations would be therefore necessary for high pulse durations to avoid uncontrolled TiO 2 formation from accumulated precursor. The GPC on the silicon wafer on the inlet were always slightly higher than the GPC on the silicon wafer at outlet of the reactor. A possible explanation is a concentration gradient along the axial direction of the tube reactor, due to the desired reaction of TTIP with the powdered silica. To verify the reproducibility of the experimental setup, the same experiment was repeated 12 times. GPCs were largely reproducible, with a mean value of 0.3 Å per cycle ( Figure S5 and Table S1, Supporting Information).
The thickness increase of the titania layer was also determined as a function of the number of completed sALD cycles (Figure 4). After the first and after ten cycles, the thickness increase differs from the expected values, that is, 0.3 Å after one and 3 Å after ten cycles. After higher cycle numbers, the thickness of the TiO 2 layer is very close to the expected values (6 Å after 20, 12 Å after 40 cycles). This variation in the first few cycles can be attributed to a limited accuracy of the ellipsometry, since the method is close to the detection limit here. To estimate in a semi-quantitative way the amount of titania directly in the powder, the titanium and silicon content of the dried powder were determined by ICP-OES. Note that due to the low TiO 2 content, the amount of Ti is close to the quantification limit of the instrumental setup. However, we observe that for more than ten cycles, the molar ratio of Ti to Si increases linearly with the thickness increase of the titania layers on the silicon wafer. 1 H, 13 C, and 29 Si solid-state NMR spectroscopy under MAS was carried out on samples prepared under various selected conditions. Initially, we studied a series of model samples: the calcined silica, the same support treated with pure TTIP after one half sALD cycle and treated with water saturated diethyl ether after one full sALD cycle. Calcined silica shows a signal around 1.8 ppm, which is known to correspond to isolated silanol groups on the surface, but may also arise from non-accessible silanol groups in ultramicropores (with diameters <1 nm) or occluded in the bulk silica (Figure 5b). [18][19][20] A broad, weak signal around 4-4.5 ppm can be found in some spectra, which is attributed to adsorbed water (Figure 5a). Note that adsorbed water on silica can occur in a range of 1 H chemical shifts between 2 and 4 ppm. Grünberg et al. have explained this phenomenon with binding of adsorbed water to isolated silanol groups and a fast av-erage of protons, which leads to a chemical shift which is weightaveraged according to the proportion of water and silanol groups. At 5 ppm, water is considered to be present in water clusters or as bulk water. [21] Calcined silica was impregnated first with pure TTIP and the slurry was packed into the NMR rotor. The spectrum of this sample is dominated by the two resonances of CH and CH 3 groups of liquid TTIP at 1.1 and 4.5 ppm (Figure 5c). Note that NMR resonances from liquids are usually much narrower than resonances from solids, as the mobility of the liquid phase averages out dipolar couplings between spins, residual chemical shift anisotropies, and chemical shift distributions due to distributions of chemical environment. After washing of excess TTIP and drying, some residual diethyl ether remains on the silica, evidenced by resonances at 1.1 and 3.4 ppm (Figure 5d,e). Note that we did not expose the sample to any low pressure or high temperature treatment after sALD, in order not to alter the surface chemistry, for example, thermal elimination of diisopropyl ether or water from surface species (see below). The resonances of diethyl ether are not sharp like liquid TTIP, but significantly broader, indicating that the sample contains diethyl ether molecules strongly adsorbed on the surface, possibly by hydrogen bonds. It is thus possible to use the signals as an internal chemical shift reference. Unfortunately, the methyl resonance of diethyl ether at 1.1 ppm overlaps with that of TTIP, making it difficult to assess the presence of TTIP in the further experiments by this signal. Interestingly, after one half cycle, two resonances appear at 4.1 and 4.5 ppm, which may be assigned to TTIP and residual water on the surface, but the respective attribution is not unambiguous here (Figure 5d). Both resonances vanish after the treatment of the surface with water second half sALD cycle (Figure 5e), indicating that the hydrolysis of isopropyl groups was successful and no more detectable free or surface bound TTIP remains in the sample. We therefore tentatively assign the resonance at 4.1 ppm to the CH groups of a TTIP molecule that is covalently bound to the surface to form a Si-O-Ti(O i Pr) 3 moiety. Surface-binding may lower the 1 H chemical shift of these hydrogen atoms compared to the free TTIP (CH group at 4.5 ppm, see Figure 5b). 1 H- 29 Si heteronuclear correlation (HETCOR) spectra will give further evidence for such surface-bound species (see below).
The effect of the contact of TTIP and H 2 O in the two half cycles with the silica surface can be followed in 13 C CP MAS NMR spectra. Here too, resonances of ether can be detected in the spectra after each single sALD step (Figure 6, marked in yellow and red). As expected, the presence of TTIP is detected after one half and one and a half sALD cycles by resonances at 76 (CH group, marked in green) and 23 ppm (CH 3 group, marked in blue, Figure 6c). [22] These peaks disappear after the second half sALD cycle and two full sALD cycles (Figure 6b,d), that is, after the hydrolysis of TTIP with water. However, a signal at 29 ppm remains, which may be attributed to residual isopropanol which has been cleaved, but not completely washed away.
We further investigated the impact of calcination of the support on the TiO 2 deposition. In non-calcined silica, a strong, broad resonance between 4 and 8 ppm corresponds to a layer of hydrogen-bonded OH groups, and the resonance at 4.5 ppm to clustered water on the surface (Figure 7a,b). In calcined silica, this broad signal of hydrogen-bonded silanols on the surface is much weaker and has shifted to 6-7 ppm, which is characteristic for smaller amounts of hydrogen-bonded silanol groups (Figure 7c,d), in line with the more complete drying of the material. [19] In samples treated with both 2 mm and with pure TTIP, again, a resonance at 4.1 ppm occurs (Figure 7c,d).
Since the signal reappears in samples after a half cycle TTIP treatment, we tentatively assign this resonance to the CH groups of a TTIP molecule that is covalently bound to the surface to form a Si-O-Ti(O i Pr) 3 moiety. Surface-binding may lower the 1 H chemical shift of these hydrogen atoms compared to the free TTIP (CH group at 4.5 ppm, see Figure 5b). 1 H- 29 Si HETCOR spectra will give further evidence for such surface-bound species (see below).
The optimum concentration of TTIP was then varied using 0.2, 2, and 20 mm TTIP in anhydrous diethyl ether. Signals of TTIP deposited on silica are visible in the 1 H MAS NMR spectra after one half cycle starting from a concentration of 2 mm ( Figure S6, Supporting Information), so this concentration was retained for subsequent experiments both in the batch process and in the plug flow reactor process. Likewise, the number of washing steps was varied from 5 to 20. After ten washing steps, no more residual TTIP remained in 1 H MAS NMR spectra ( Figure S7, Supporting Information), so that this condition was judged to be sufficient for further sample preparations in the batch mode.
Covalent grafting of TTIP to the silica surface will consume silanol groups by the condensation reaction of TTIP with a hydroxyl group and release of a molecule of isopropanol. We have tried in the following to quantify the percentage of surface silanol groups thus consumed, to determine the optimum sALD conditions and in order to probe the existence of preferential reaction sites. The grafting of TTIP was assessed by 29 Si direct excitation (DE) spectra ( Figure S8, Supporting Information). In amorphous silica, Q 2 , Q 3 , and Q 4 species, corresponding to Q n = Si(OH) 4−n (OSi) n , can be readily discerned by their respective chemical shifts around −92, −101, and −110 ppm. DE spectra quantitatively reproduce the relative intensities of Q 2 , Q 3 , and Q 4 sites in a silica material. However, since most of the silica consists of completely condensed Q 4 groups in the bulk material, the relative intensities of Q 2 and Q 3 groups are low, and spectral deconvolution remained ambiguous. As a tendency, the spectra show a decrease of relative intensity of Q 2 and Q 3 sites from pure, calcined silica to silica treated with 2 mm TTIP to silica treated with pure TTIP.
Therefore, the effect of TTIP grafting was further followed by 29 Si CP MAS NMR spectra of the powdered samples before and after the sALD reaction (Figure 8). Note that CP spectra do not reflect the true abundance of each group in the silica. The signal intensity of each silica species depends on the CP buildup time constant T HSi and the 1 H relaxation time constant in the rotating frame T 1 , leading to a characteristic CP build-up curve. [23][24][25] In order to verify if, here, CP spectra are indeed suitable to quantify the amount of consumed Q 2 and Q 3 groups, CP build-up curves were recorded on four different samples: noncalcined silica, calcined silica, calcined silica after one half cycle reaction with 2 mm TTIP, followed by ten washing steps with diethyl ether and after one half cycle reaction with pure TTIP and ten washing steps ( Figures S9 and S10, Supporting Information). Note that for these experiments, the pristine silica powders were wetted with diethyl ether, too, in order to ensure a comparable density of protons on the surface of the reference samples and the samples after half cycle sALD reactions. Within the error, CP build-up curves of Q 2 , Q 3 , and Q 4 groups show very similar slopes up to a contact time of 8 ms ( Figure S10, Supporting Information). This somewhat untypical behavior is certainly due to the high proton density on the surface, ensuring good dipolar interactions between protons and all near-surface silica sites.
Pristine calcined silica shows the highest ratio of Q 3 /Q 4 compared to the samples after sALD (Figure 8a). After one half sALD cycle reaction, samples treated with pure TTIP and with 2 mm TTIP show much lower Q 2 and Q 3 intensities, the material treated with pure TTIP having the lowest remaining content of Q 2 and Q 3 groups. Clearly, pure TTIP leads to a more Adv. Mater. Interfaces 2023, 10, 2202131 Table 2. Relative signal areas of Q 2 , Q 3 , and Q 4 groups after deconvolution of 29 Si CP MAS spectra in Figure 8, and the fraction of reacted Q 2 groups ( Q2 ), Q 3 groups ( Q3 ), and overall fraction of reacted hydroxyl groups ( ) after one half cycle. Two different scenarios are considered: Q 2 groups react with only one hydroxyl group, yielding Q 3 species, or Q 2 groups react with both hydroxyl residues, yielding Q 4 species.
Sample
Relative signal area / % complete coverage of the surface after one half sALD cycle. However, the surface is not fully covered after one half cycle, in which case a nearly complete disappearance of Q 2 and Q 3 signals would be expected. Very small signals of Q 2 and Q 3 sites might remain since silica often contains inaccessible silanol groups in ultramicropores or occluded in the bulk material. [18,19] In non-calcined silica, the decrease of Q 2 and Q 3 groups is less pronounced than in calcined silica (Figure 8b). Here, the presence of water and higher amounts of hydrogen-bonded silanol groups, as observed in the 1 H NMR spectra (Figure 5a), seem to impede TTIP grafting onto the surface, or may lead to intermolecular TTIP homocondensation without any involvement of surface SiOH groups. Q 2 , Q 3 , and Q 4 resonances were deconvoluted and the intensity ratios were analyzed quantitatively. We use here a mathematical model established by Pallister et al. for the grafting of a gallium complex on silica in a gas ALD process. [26] From the peak intensities, the coverage and the fractions of Q 2 and Q 3 silanol groups which have reacted with TTIP were calculated ( Table 2). For details of the calculation, see chapter 1 in the Supporting Information. For the implementation of this model, following Pallister et al., the assumptions were made that Q 2 and Q 3 sites have a similar CP build-up behavior (compare Figure S10, Supporting Information) and that all compounds reacting with the silica will react at the hydroxyl surface sites. [26] The reaction mechanism of Q 2 groups with TTIP can progress via two pathways. In a first scenario, only one of the two silanol groups might react with TTIP (monodentate binding), leaving the other present on the surface, thus Q 2 would transform into Q 3 groups. In a second scenario, both silanol groups on Q 2 species will react with a single TTIP molecule (bidentate binding), cleaving two equivalents of isopropanol, and resulting in a Q 4 group with a fourmembered Si-(O) 2 -Ti ring. Alternatively, on Q 2 groups, the two silanol groups might react with two different TTIP molecules. We therefore consider these two scenarios in the calculation of the fractions of Q 2 and Q 3 groups: 1) Only one hydroxyl group of Q 2 sites reacts with one precursor molecule, yielding Q 3 groups, and 2) a reaction of both hydroxyl groups of Q 2 sites, yielding Q 4 sites. Obviously, in both cases, Q 3 groups are supposed to react to Q 4 groups upon TTIP grafting. Importantly, since the calculation of the coverage of Q 3 groups depends also on the transformation of Q 2 → Q 3 , it is not possible to state which scenario is occurring, or if both cases are competing.
CP build-up curves, discussed above, were recorded on a second set of samples (compare Figures S9 and S10, Supporting Information). Here, likewise, the fractions of reacted Q 2 and Q 3 groups and the overall fraction of reacted silanol groups were calculated individually for spectra with 2, 4, and 8 ms CP contact time (Table S3, Supporting Information). The percentages of reacted groups are in very good agreement when determined from spectra with different contact time. Between the two sets of samples, the percentages of reacted groups are also in good agreement, demonstrating the robustness of the method.
In both calcined and non-calcined samples, Q 2 groups on the surface of silica account for a very small percentage of the total number of hydroxyl groups. However, we find that in both scenarios, Q 2 groups are more prone to react with TTIP (Table 2) than Q 3 groups, since the final fraction of reacted Q 2 groups is significantly higher. This is true for both non-calcined and calcined silica, and both for 2 mm and pure TTIP as a precursor. The most complete coverage is obtained by pure TTIP grafted on calcined silica, where around 80% of Q 2 hydroxyl groups react upon deposition of pure TTIP in scenario 1 (91% in scenario 2), versus 23% of the Q 3 groups (14% in scenario 2). We note that absolute quantification of different silanol species from 29 Si CP spectra is difficult and thus, the values for Q2 , Q3 , and in Table 2 might still be subject to a systematic error. Yet, the trend observed here for the different reactivities of Q 2 and Q 3 groups remains noteworthy. The finding that Q 2 species react more readily is also consistent with the study by Pallister et al. In that study, the authors attributed the higher reactivity of Q 2 groups to a bidentate binding mode and the formation of fourmembered Si-(O) 2 -Ga rings, thus to Q 4 groups, since two ethyl ligands can be cleaved from the Ga complex. Four-membered Si-(O) 2 -Ti rings have been shown to exist in titania-containing zeolites, although with low stability. [27] They may thus explain the preferential reactivity of Q 2 groups toward TTIP. We consider the covalent grafting of two individual TTIP molecules to a single Q 2 site to be less likely, since the first TTIP molecule with its three bulky isopropyl groups would impose a strong steric hindrance for the covalent bonding of a second precursor. Yet, it is possible that while Q 2 species covalently bind to TTIP with one silanol group, the second silanol develops a hydrogen bond to the oxygen atom of another isopropoxyl moiety, thus stabilizing the grafted complex. This may explain the higher reactivity of Q 2 compared to Q 3 groups even in scenario one, considering only monodentate binding.
To obtain a deeper understanding of the silica surface, 2D 1 H- 29 Si HETCOR MAS NMR experiments were carried out. These spectra offer additional information thanks to the correlation of the 1 H and the 29 Si dimension, and increased resolution of the 1 H dimension compared to 1D 1 H spectra due to dedicated pulse schemes which decouple proton dipolar interactions during the acquisition. Calcined silica only shows one correlation of protons in hydrogen bonding around 6 ppm to Q 3 (−101 ppm) and Q 2 (−92 ppm) silica groups (Figure 9a, and schemes in Figure 9f,g). In the samples after one half or one full sALD cycle with 2 mm or pure TTIP, correlations between the methyl protons of diethyl ether with silicon atoms of Q 4 and between protons in CH 2 groups of diethyl ether with Q 3 silica sites can be found. We attribute this to surface-adsorbed diethyl ether, probably hydrogen-bonded to Q 3 sites (schematically shown in Figure 9h). Note that the correlation highlighted in red might also be due to hydrogen atoms of TTIP with Q 4 silica species. Since the chemical shifts of 1 H of methyl groups in both molecules are very close together, these two possibilities cannot be disentangled.
After one half sALD cycle with pure TTIP on calcined silica (Figure 9b), a new correlation appears at ( 1 H) ≈ 4.0 ppm and ( 29 Si) ≈ −107 ppm, which is marked in green. This 29 Si chemical shift is situated between −102 ppm, where Q 3 species are located, and −110 ppm, where Q 4 groups are located. We tentatively assign this signal to a Si-O-Ti species, explaining this unusual 29 Si chemical shift in analogy to Si(1Al) groups in zeolites, in which silicon is bound in a Q 4 surrounding, and in which one silicon neighbor is replaced by aluminium. [28] By replacing one silicon neighbor by titanium, the same effect is observed here. Along these lines, a downfield chemical shift of +7 ppm was also observed between Si(0Ti) and Si(1Ti) sites in a microporous titanogallosilicate by Rocha et al. [29] The resonance is in the following denominated as Q 4 ′. Its 1 H chemical shift is close to the chemical shift of the proton of the CH group of the TTIP molecule, that is, ( 1 H) = 4.0 ppm. The feature can also be seen in the HETCOR spectrum of calcined silica after one half sALD cycle with 2 mm TTIP (Figure 9d). Since the correlation of this peak appears in all samples after one half sALD cycle, we attribute this peak to a surface-bound Si-O-Ti(O i Pr) 3 or bidentate Si-(O) 2 -Ti(O i Pr) 2 species (Figure 9h). This is the first time that a Q 4 -like silicon species bound to one titania atom has been discerned by its 29 Si chemical shift in amorphous silica, an observation made possible by the enhanced resolution of 2D 1 H- 29 Si HETCOR spectra. In the 1D 29 Si CP MAS NMR spectra in this work, as well as in 29 Si (CP) MAS NMR spectra in the work by Pallister et al., [26] this signal was not discernible due to the overlap with the broad neighboring resonances of Q 3 and Q 4 groups.
In the sample after one full sALD cycle, the Q 4 ′ resonance does not appear anymore (Figure 9c). We do however observe a correlation signal at a 29 Si chemical shift of −109 ppm and a 1 H chemical shift of 3.2 ppm (brown). A similar signal at 3.2-3.4 ppm has been observed by Crocker et al. in titania supported on silica, and has been attributed to titanol (TiOH) groups on tetrahedral Ti sites, which are only observed in thin layers of amorphous titania. [30] In contrast, isolated titanol groups on octahedrally coordinated Ti sites, as found in anatase or rutile, typically resonate at 2.0-2.3 ppm, hydrogen bridged titanol groups around 6.7 ppm. [30] This resonance therefore clearly shows here the hydrolysis of TTIP toward TiOH groups, and the correlation between silica Q 4 ′ sites and these titanol groups proves their existence on the surface as Si-O-Ti-OH groups. Note that since lone or geminal titanol groups have the same chemical shift, we cannot discern from the spectra between different (SiO) n Ti(OH) 4−n groups, thus titanium may be grafted to the silica surface via one, two, or three covalent bonds. Moreover, the presence of only the TiOH-Q 4 ′ correlation and no TiOH-Q 3 ′ correlations implies that the scenario 2, that is the reaction of Q 2 into Q 4 ′ groups, occurs preferentially.
The spectrum in Figure 9e was recorded on non-calcined silica instead of calcined silica. Here, the aspect of the spectrum strongly changes, involving several signals which are more difficult to attribute. A new, strong correlation between Q 3 silicon atoms and CH hydrogen atoms of TTIP is found, marked in pink. This correlation reflects the high density of hydroxyl groups on the silica surface in this sample. It may be attributed either to surface-grafted TTIP, in which CH groups are close to neighboring Q 3 hydroxyl species, or to TTIP which is only adsorbed on the surface and not covalently bound (schemes in Figure 9j,k).
The spectrum of non-calcined silica after one full sALD cycle (Figure 10a) again shows hydrogen atoms of diethyl ether correlated for Q 4 and Q 3 species (red and yellow), as well as a broad signal of unreacted Q 3 hydroxyl groups on the silica surface. Importantly, no more signals of TTIP can be found, evidence for the desired complete hydrolytic cleavage and washing of all organic ligands in the second half cycle. The spectrum after 20 sALD cycles (Figure 10b) no longer shows any correlation for Q 4 and Q 4′ species, and the correlation between Q 3 species and ether is also very weak. This is most likely due to the presence of a closed multilayer of TiO 2 on the surface of the silica substrate, which weakens the CP effect, so that the resonance of Q 4 and Q 4′ species disappears. The correlation signal at 4 ppm arises from remaining physisorbed water, which may cover small residual spots of remaining TTIP inaccessible silica surface, or which may be occluded in the TiO 2 film.
Conclusions
We have developed here a first set-up for the ALD from solution of a thin film of titania onto silica. For this study, porous silica particles were used as a model support, their high surface area simplifying further analytical characterization. Furthermore, for the first time, an sALD deposited material was investigated by solidstate NMR spectroscopy. This technique is extremely well suited to track the organic and inorganic surface-deposited species, to discern bound from physisorbed species and to gain deeper insight into the reaction proceeds and its mechanism. A tubular reactor was developed as a simple benchtop set-up to produce around 180 mg of titania-coated silica powder. The addition of silicon wafer at the entrance and at the end of the tubular reactor allowed to verify the thickness increase of the TiO 2 layer in every cycle by ellipsometry. By optimizing the operating parameters of sALD, including the precursor concentration and the pulse duration of precursors, a stable growth rate per cycle of 0.3 Å thickness increase was achieved. By 1 H MAS NMR spectroscopy, the state of the surface before sALD treatment, after one half, and one full sALD cycle, the adsorption of diethyl ether as solvent as well as the effect of TTIP concentration and washing steps could be analyzed, allowing for the optimization of these parameters. Comparing two different supports, calcination of the silica support reduces the amount of hydrogen bonded silanol groups on the surface, leading to a higher proportion of deposited TTIP, this support treatment is thus necessary for future optimizations. 29 Si cross polarization spectra showed the presence of Q 2 , Q 3 , and Q 4 groups on the silica surface and their relative change upon sALD with TTIP. TTIP molecules prefer to react with Q 2 hydroxyl groups compared to Q 3 groups, which we tentatively explain by the formation of Si-(O) 2 -Ti four-membered rings, or to stabilization via hydrogen bonding of a TTIP molecule that is bound in a monodentate way to a Q 2 group. Interestingly, 2D 29 Si HETCOR solid-state NMR spectra further show a unique signal at a 29 Si chemical shift of −107 ppm, which we tentatively attribute to a Si-O-Ti(O i Pr) 3 species. After 20 full sALD cycles, 2D HETCOR spectra show only weak correlations of silica with small amount of residual solvent or water, thus indicating the generation of a closed multilayer of titania on the surface.
In the future, the sALD set-up in the in-house designed plug flow reactor will be further improved, for example, by optimizing the contact between the silica and the fluid. Both the reactor developed here and the analytical insight can be extended to other sALD systems to produce well-controlled thin film on variable supports, while understanding their formation mechanisms on a molecular scale.
Supporting Information
Supporting Information is available from the Wiley Online Library or from the author. | 9,849 | 2023-05-10T00:00:00.000 | [
"Chemistry"
] |
Cognitive Bias in Emissions Trading
: This study investigates whether cognitive biases such as the endowment e ff ect and status quo bias occur in emissions trading. Such cognitive biases can serve as a barrier to trade. This study’s survey-based experiments, which include hypothetical emissions trading scenarios, show that the endowment e ff ect does occur in emissions trading. The status quo bias occurs in only one of the three experiments. This study also investigates whether accumulation of experience can reduce cognitive bias as discovered preference hypothesis expects. The results indicate that practitioners who are supposed to have more experience show no evidence of having less cognitive bias. Contrary to the conventional expectation, the practitioners show significantly higher level of endowment e ff ect than students and only the practitioners show a significant status quo bias. A consignment auction situation, which is used in California’s cap-and-trade program, is also tested; no significant di ff erence between general permission trading and consignment auctions is found.
Introduction
Emissions trading allows entities to achieve emission reduction targets in a cost-effective way through buying and selling emission allowances in emissions trading markets [1,2]. Each trade transaction results in Pareto improvement; the more deals that are made, the more positive the results. However, transactions seem insufficient, especially in the early stage of emissions trading schemes like Korea Emissions Trading Scheme (ETS) and China ETS [3][4][5]. Identifying the reasons for this transaction insufficiency is important.
The literature has attempted to explain why emissions trading is sluggish by discussing the problems faced by the institutions involved and participating companies [3,4,6,7]. Previous studies, however, generally assume these participants have been rational decision makers. The behavioral economics perspective, which can provide new implications by introducing the concept of 'cognitive bias,' has not been adequately considered in the setting of emissions trading. Cognitive bias happens because people depend on intuitive, rapid, and automatic heuristics [8] to 'reduce the complex tasks of assessing probabilities and predicting values to simpler judgmental operations' [9] in their information processing and decision making. Heuristics are very different from rational processing, which is analytic, conscious, and rule-based. Thus, a view that assumes rational decision makers leads to an understanding of the problem and a presentation of solutions that differ widely from those that flow from a view that acknowledges cognitive bias. It is worthwhile trying to understand the problem by using both perspectives.
Some literature discusses the endowment effect and status quo bias to explain the sluggishness of emissions trading [4,6,10,11]. However, the literature has only treated these cognitive biases as a possibility and has not verified their effects using empirical methods in the setting of emissions trading.
Although environmental economists have used laboratory experiments to examine various design choices for emissions trading schemes such as price ceiling, banking, type of auctions, and so on [12,13], they have not directly dealt with cognitive biases. It is important to note that cognitive biases do not always occur. For example, if goods are expected to be resold, the endowment effect tends not to occur [6,14]. Emission allowances do not generate usage value in themselves. The allowances are a kind of permits to emit emissions to the level of the overall limit and have some similar characteristics to financial assets [15]. As Kahneman et al. [14] argues, cognitive biases such as the endowment effect may not occur in an asset of this nature. This study attempts to test whether the endowment effect or status quo bias happen in the setting of emissions trading using a survey-based experiment.
In addition, several researchers have argued for the 'discovered preference hypothesis,' whereby the endowment effect disappears as people become more experienced in a certain decision environment [16][17][18][19][20]. According to this proposition, even if nonprofessional decision makers such as students show an endowment effect or status quo bias, we may assume that the practitioners in charge of emissions trading, who fully understand and have experiences of the system, may show no or less cognitive biases. Therefore, this study also tests whether cognitive bias can be reduced through accumulation of experience by comparing experiment results for two different groups of university students and emission trading practitioners in ETS participating firms.
Another way to reduce cognitive bias may be introduction of consignment auctions as are used in the California's cap-and-trade program, which was launched on January 1, 2012 with an enforceable compliance obligation beginning with 2013 greenhouse gas emissions. California is part of the Western Climate Initiative (WCI) which began in 2007 with an agreement of the governors of Arizona, California, New Mexico, Oregon, and Washington to develop regional greenhouse gas reduction targets and develop a market-based program to meet their targets. In the WCI, California, Quebec, and Ontario are currently running their cap-and-trade programs [21]. The California's cap-and-trade program has adopted the consignment auction for CO 2 allowances from its beginning for ensuring price discovery and directing allowance value to intended uses [11]. The consignment auction combines free allocation and auctioning into one mechanism. Allowances are initially endowed to firms through a free allocation rule. Then, firms consign their initial allowances to the auctioneer. In the auction, firms submit their demand schedules and the overall clearing price is determined. Firms pay the clearing price if the equilibrium allocation of allowances in the auction is more than their original allocation and obtain revenue if the equilibrium allocation is less than their original allocation. Thus, all auction revenue is diverted to the holders of the initial allowance endowments. This novel initial allocation process has been regarded as being both politically appealing and successful in generating clear price discovery [10,11,[22][23][24].
Consignment auctions may reduce the endowment effect, as they require a transaction as soon as emission allowances are initially allocated. In addition, whether participants possess the allocated allowances or not (the status quo) is ambiguous before the consignment auction, which may reduce the status quo bias. Therefore, this study compares experiment results with a consignment auction to those without to investigate whether consignment auctions can reduce cognitive biases.
In sum, the objective of this study is to investigate the existence of the cognitive biases in the setting of emissions trading and the effects of accumulation of experiences and the consignment auction on the biases. To fulfill the objective, this study examines (1) whether the endowment effect and status quo bias occur in emissions trading situations, (2) whether practitioners show less cognitive bias than students, and (3) whether the endowment effect and status quo bias can be reduced in consignment auction situations. Experiments for second and third purpose also serve to verify the endowment effect and status quo bias repetitively. To our knowledge, this study is the first attempt to empirically test cognitive biases in the setting of emissions trading. This paper is organized as follows. In Section 2, we suggest hypotheses and their meanings for cognitive bias in emissions trading. In Section 3, we then introduce the structure of the experiment to test the hypotheses. Section 4 shows the results of the experiment and the analysis. In Section 5, we discuss the implications of the results in real situations. Finally, in Section 6, we provide conclusions and some limitations of this study.
Endowment Effect
When the endowment effect exists, willingness-to-accept (WTA), the minimum amount of money that people are willing to accept to abandon a good, is systematically higher than willingness-to-pay (WTP), the maximum amount people are willing to sacrifice to procure a good [25]. In other words, people ascribe more value to things merely because they own them.
The endowment effect rests on a hypothesis concerning the psychology of preferences associated with prospect theory that predicts loss aversion [19,25]. In loss aversion, losses are perceived to be greater than gains of equal magnitude. When people sell an object, the lost object is perceived as a loss. Thus, the lost value (WTA) is perceived to be higher than the earned value (WTP) when they buy the same object. Knetsch and Sinden [26], Kahneman, Knetsch, and Thaler [27] have shown the existence of the endowment effect through experiments using general consumer goods, such as mugs and chocolate bars. Kahneman et al. [14] have also revealed the effect using public goods that are not traded in the market, such as postal services and park trees.
Further, the concept of endowment effect has been applied in various settings. Frenkel, Heller, and Teper [28] showed that endowment effect offset the winner's curse, which indicates overpayment of buyer in common value auction. Penon and Ortega [29], meanwhile, applied an endowment effect to the problem of firm ownership. In addition, some previous literature has mentioned the endowment effect in emissions trading [4,6,10,11]. However, the literature has only treated these cognitive biases as a possibility and has not verified their effects using empirical methods in an emissions trading situation.
A series of related studies has examined the factors determining endowment effect levels. Novemsly and Kahneman [30] showed that money given up in purchases is not generally subject to loss aversion and thus does not create an endowment effect. Kahneman et al. [14] argued that the endowment effect does not occur if the goods involved are expected to be resold. In their study, no endowment effect was observed for tokens that were to be exchanged for cash after the experiment. Horowitz and McConnell [25] showed that WTA is generally about seven times higher than WTP and that this ratio is lower for commodities traded on the market and greater for non-market goods such as public and environmental goods. Many other studies have found variables that moderate endowment effect levels. Hedonic goods have been found to generate more of an endowment effect than utilitarian goods [31]; high involvement has been found to generate more of an endowment effect [32]; emotions such as disgust and sadness have been found to reduce it [33][34][35]; and low level of empowerment to decision maker incurs more endowment effect [36].
Allowances do not generate utility in themselves; they do so by eliminating fines due to actual emissions or by selling them. Thus, allowances are a kind of claim and have characteristics similar to financial assets [15]. As Novemsly and Kahneman [30], Kahneman et al. [14], and Horowitz and McConnell [25] suggest, allowances may not cause an endowment effect. Kreutzer [6] argues that, even if allowances are not held solely for resale, the existence of reasonably active spot markets for allowances means that they can easily be converted into cash at any time. This easy conversion into cash may reduce the endowment effect. Therefore, this study tests whether the endowment effect occurs in emissions trading using the following hypothesis: Hypothesis 1. In emissions trading, WTA for allowances will be higher than WTP.
List [37] argued that 'prospect theory, which explains endowment effect and status quo bias, adequately organizes behavior among inexperienced consumers, but consumers with intense market experience behave largely in accordance with neoclassical predictions.' Plott and Zeiler [19,20] listed five factors that can eliminate the endowment effect: an explanation of the optimal response, the provision of practice rounds, incentive compatibility with the elicitation mechanism, the mechanism used to elicit valuation, and the gap measurement method. Among those, provision of practice rounds is also related to experience.
This study compares endowment effect levels between students and emissions trading practitioners, who have accumulated trading experience and understand the trading system well and thus are expected to be better decision makers. Given the findings of List [37], Plott, and Zeiler [19,20], practitioners should show either no endowment effect or a smaller effect than is observed in the students. Thus, the following is proposed: Emissions trading practitioners will show less of a gap between WTA and WTP than is observed in students.
Consignment auctions may reduce the endowment effect since they require a consignment to the auctioneer before allocated emission allowances can be held. Knetsch [38] showed that an endowment effect occurred only when the subjects owned the goods for a while before they were told they could trade them. Thus, an endowment effect may not occur before a consignment auction is completed. The following is therefore hypothesized: Hypothesis 3. Consignment auctions generate less of a gap between WTA and WTP than general emissions trading does.
The endowment effect in emissions trading would make firms with excess allowances want to sell at relatively high prices, while firms needing additional allowances would want to purchase at relatively low prices. This could reduce transactions. Hypothesis 1, Hypothesis 2, and Hypothesis 3 assume individual actors rather than organizational ones. If firm decisions differ from those made by individuals, then the hypotheses may not apply to firms. However, this study assumes that firm decisions depend on decisions made by practitioners and executive officers and that firm decisions may not be free of the cognitive biases of individuals if they are held by the majority. This intuition is consistent with the perspective taken by organizational studies. The Echelons theory, proposed by Hambrick [39] and Hambrick and Mason [40] (and widely accepted by organization theorists), argues that executives' experiences, values, and personalities greatly influence their interpretation of the situations they face and, in turn, affect their choices.
The sections below deal with hypotheses on status quo bias. These consider the bias at the individual level, but such bias ultimately affects firm-level decision making.
Status Quo Bias
Kahneman et al. [14] argued that 'one implication of loss aversion is that individuals have a strong tendency to remain at the status quo, because the disadvantages of leaving it loom larger than advantages.' Samuelson and Zeckhauser [41] showed that defining a particular investment alternative as the status quo increases the likelihood of that alternative being selected. Hartman, Doane, and Woo [42] showed the effect of status quo bias by examining electric power services that differed in reliability and price. The concept of status quo bias also has been used to explain why new services or policies are not accepted by people [43,44]. Further, Kreutzer [6] argued that loss aversion and status quo bias can inhibit emissions trading.
The endowment effect and status quo bias both come from loss aversion. Thus, as with the endowment effect, allowances may not cause status quo bias since they are a kind of claim [14,25,30]. This study tests whether status quo bias occurs in emissions trading.
Status quo bias is a situation in which a prior situation (status quo) affects a choice. If status quo bias occurs during a transaction that offers a choice between an excess-allowances situation (EAS) and lack-of-allowances situation (LAS), firms will stay in the EAS when facing an excess-allowances-prior situation (EApS, where the amount of allocated allowances is greater than what is needed); contrariwise, firms will not choose EAS in a lack-of-allowances-prior situation (LApS). Thus, the following is hypothesized: Hypothesis 4. When subjects can choose between an EAS and LAS through a transaction, the EAS will be chosen more often by subjects coming from an EApS than by subjects coming from a LApS.
As List [37], Plott, and Zeiler [19,20] suggested, emissions trading practitioners with a longer experience and better understanding of the trading system may not show status quo bias, as with the endowment effect. Further, some studies explained that people choose status quo bias to avoid a difficult decision [41,45,46]. Then, practitioners who become accustomed to the emission trading context can make more structured decisions with less difficulty. As a result, they may show less status quo bias. Thus, the following is hypothesized: The gap in the ratio of subjects who choose EAS between subjects with EApS and subjects with LapS will be lower when the subjects are emissions trading practitioners than when the subjects are general students.
Furthermore, in consignment auctions, firms are allocated allowances but must consign them to the auctioneer and purchase them again, though they do not pay for the initially allocated amount. Then, the status quo (whether they possess the allocated allowances) is ambiguous in the pre-consignment period. Thus, consignment auctions may induce less status quo bias and the following hypothesis is proposed: The gap in the ratio of subjects who choose EAS between those with EApS and those with a LApS is smaller in consignment auctions than in general emissions trading situations.
Outline of Experiments
This study carried out several survey-based experiments. To test Hypothesis 1 (on the endowment effect), WTA is measured via EAS, and WTP is measured via LAS. Then, the WTA and WTP results are compared. To test Hypothesis 2, two similar experiments involving students and practitioners in a general emissions trade scenario (GETS) are compared. For Hypothesis 3, two different scenarios, one for GETS and another for consignment auctions (CAS), are run with students. Thus, for Hypothesis 1, Hypothesis 2, and Hypothesis 3, the study employed a 2 (WTA vs. WTP) × 3 (students and GETS vs. practitioners and GETS vs. students and CAS) between-subjects design.
To test Hypothesis 4, dealing with status quo bias, the decision of whether to accept a transaction (i.e., to stay at status quo) was tested with EAS (sales situation) and LAS (purchase situation); the two results were then compared. To test Hypothesis 5 and Hypothesis 6, a design similar to that used to test Hypothesis 2 and Hypothesis 3 was employed. Thus, a 2 (EAS and LAS) × 3 (students and GETS vs. students and CAS vs. practitioners and GETS) between-subjects design was used to test the hypotheses on status quo bias.
Scenarios and Questions for Endowment Effect
In each experiment, the subjects were asked to assume they were the CEO of a firm and to answer questions after scenarios and questions were presented. The GETS section for WTA appeared as follows: 'Your firm has been assigned 500,000 tons of allowances from the government for one year from today and your firm estimates 400,000 tons of emissions over the year. You are currently considering selling 100,000 tons of spare allowances on the emissions trading market.
The actual amount of emissions over the year may fluctuate from 400,000 tons. If the amount of allowances (400,000 tons or 500,000 tons) determined by the trade exceeds the total amount of your emissions after one year, you will sell the spare allowances at the market price at that time. If it does not, you will purchase the deficit amount at the market price at that time. The market price at that time is expected to be ₩17,000-27,000/ton.
Question: What is the minimum price at which you are willing to sell your 100,000 tons of spare allowances now? ₩____________/ton.' The WTP section was created by changing the underlined words: '400,000 tons' to '600,000 tons,' 'sell' to 'purchase,' 'spare allowances' to 'deficit allowances,' '400,000 tons or 500,000 tons' to '500,000 tons or 600,000 tons,' and 'minimum' to 'maximum.' These two GETS questionnaires (for WTA and WTP) were distributed to both students and practitioners.
Next, the CAS section for WTA appeared as follows: 'Your firm has been assigned 500,000 tons of allowances from the government for one year from today, and your firm estimates 400,000 tons of emissions over the year.
In this situation, before the firm holds the assigned allowances, the firm has the obligation to entrust the whole amount of allowances to the auctioneer to be auctioned. If you only need 400,000 tons, you will be bidding 400,000 tons and get 400,000 tons with additional revenue (100,000 tons × the final contract price). You are receiving the additional revenue that corresponds to the gap between 500,000 and 400,000 (100,000 tons) because you originally had 500,000 tons of allowances. If you want, you can get 500,000 tons with no additional revenue.
The actual amount of emissions over the year may fluctuate from 400,000 tons. If the amount of allowances (400,000 tons or 500,000 tons) determined by the result of the auction exceeds the total amount of your emissions after one year, you will sell the spare allowances at the market price at that time. If it does not, you will purchase the deficit amount at the market price at that time. The market price at that time is expected to be ₩17,000-27,000/ton. Question: What is the minimum price at which you are willing to sell your 100,000 tons of spare allowances in the consignment auction? ₩__________/ton.' The WTP section was created by making changes similar to those made for the GETS section. The two CAS questionnaires (for WTA and WTP) were distributed only to students.
Scenarios and Questions for Status Quo Bias
The two different prior situations, EApS (sales situation) and LApS (purchase situation), were presented to the subjects after the basic explanation of emissions trading and their decisions on whether to accept an offer of transaction were tested. The EApS section for GETS appeared as follows: 'Your firm has been assigned 600,000 tons of allowances from the government for one year from today and your firm estimates 400,000-600,000 tons of emissions over the year. In this situation, another firm asks you to sell 200,000 tons of allowances. In this deal, there is no transaction cost.
If the amount of your allowances (400,000 tons or 600,000 tons) determined by the trade exceeds the total amount of your emissions after one year, you will sell the spare allowances at the market price at that time. If it does not, you will purchase the deficit amount at the market price at that time. The market price at that time may be different from what it is now.
Question: Will you accept the offer? (Yes or No)' In the scenario for the endowment effect, the question asked the level of price and, to guide the answer, we provided information on expected price levels. For the status quo bias, however, we asked the binary question on whether to accept or not the offer. Thus, information on expected price levels is not necessary and the scenario mentioned only the possibility that the price could be changed.
The LApS scenario was created by making only two changes: changing '600,000 tons' to '400,000 tons' and 'sell' to 'purchase.' The two GETS questionnaires (for EApS and LApS) were distributed to both students and practitioners. Next, to create the consigned auction scenario, a passage in the GETS section ('In this situation, another firm asks you to sell 200,000 tons of allowances. In this deal, there is no transaction cost') was changed to reflect the characteristics of the consignment auction (as in the experiments on the endowment effect). The two CAS questionnaires (for EAS and LAS) were distributed only to students.
Distribution of the Questionnaire
Four sets of questionnaires were used for the students. Each questionnaire had four questions. The first verified whether the subject understood the scenario, since the responses of subjects who did not fully understand the scenarios had to be excluded from the analysis. To check their understanding, the same WTA/WTP scenarios used in experiment for endowment effect were presented; the firm's choice, future total emission amounts, and future price were additionally presented. Finally, the subjects were asked how much they would pay (earn) now and in one year. The second question was about the endowment effect and the WTP/WTA scenarios. The third question was about status quo bias and whether the subjects would accept a transaction. The fourth question was about gender and age. The four sets of questionnaires produce the following: 2 (GETS vs. CAS) × 2 ('WTA and EAS (whether to accept a sales offer or not)' vs. 'WTP and LAS (whether to accept a purchase offer or not)').
For emissions trading practitioners, two sets of questionnaires were used: 'WTA and EAS' vs. 'WTP and LAS' for the same GETS. These also included the three questions described above, except the fourth about gender and age.
Management school students at a national university served as the student respondents. Each subject received one set of questionnaires during a class. For the practitioner group, the sample was collected from around 560 firms that have participated in the Korea ETS for more than three years. The practitioner respondents received questionnaires by e-mail and sent back their responses. They are mostly middle managers in the environment/energy department in the Korea ETS participating companies. Other details on the samples for each experiment are provided below.
Test of Endowment Effect
After questionnaires with incorrect answers to the test questions were excluded, 213 responses remained. Among these, 15 did not include answers to the questions on WTA or WTP. Thus, 198 responses were used to test the endowment effect. Among the 213, one did not include an answer to the question about whether to accept a transaction, so the test on status quo bias analyzed 212 responses. Most wrong answers were given for the CAS question, reflecting the complexity of the scenario. As questionnaires with wrong answers to the test question were excluded, the results of the analyses were not affected by insincere responses. The following analyses were carried out using SPSS ver. 23. Table 1 shows the characteristics of the sample for the test of the endowment effect and the results of the independent sample t-tests. In all three experiments, WTA is higher than WTP, and the differences are significant (Students and GETS, t = 5.801, p = 0.000; Students and CAS, t = 3.783, p = 0.000; Practitioners and GETS, t = 13.624, p = 0.000). Thus, Hypothesis 1, on endowment effect, is supported. Hypothesis 2 concerns the moderating effect of experience on the endowment effect. In the test, a two-way ANOVA was carried out using the results of Students and GETS and Practitioners and GETS. The results are presented in Table 2. The main effect from 'WPA or WTP' is significant (p = 0.000) and also supports Hypothesis 1. The interaction effect from 'WPA or WTP' and 'students or practitioner' is also significant (p = 0.047). However, as Table 1 shows, the gap between WTA and WTP is higher in practitioners than in students. Thus, Hypothesis 2 is not supported. t-test t = 5.801, p = 0.000 ** t = 3.783, p = 0.000 ** t = 13.624, p = 0.000 ** * p < 5%, **p < 1%. + Gender and age were not surveyed for practitioners. Hypothesis 3 concerns the moderating effect of consignment auctions on the endowment effect. In this test, a two-way ANOVA was also carried out using the results of Students and GETS and Students and CAS. As Table 3 shows, the main effect from 'WPA or WTP' is significant (p = 0.000). The interaction effect from 'WPA or WTP' and 'GETS or CAS' is not significant (p = 0.254). Thus, CAS does not show less of an endowment effect, and therefore Hypothesis 3 is not supported.
Test of Status Quo Bias
The test for status quo bias analyzed 212 responses. Table 4 shows the characteristics of the sample and the analysis results. Pearson χ 2 = 40.982, p = 0.000 ** * p < 5%, ** p < 1%. + Gender and age were not surveyed for emissions-trading practitioners.
As Table 4 shows, the result for Students and GETS and Practitioners and GETS is what Hypothesis 4 expected: the ratio of subjects who choose EAS is higher in EApS. In Students and CAS, however, the ratio is slightly higher in LApS. Table 4 also shows the result for a chi-square independence test, wherein only Practitioners and GETS shows a significant difference (Pearson χ 2 = 40.982, p = 0.000). Thus, Hypothesis 4 is supported only for Practitioners and GETS.
Hypothesis 5 expects that practitioners will show less status quo bias. However, only the practitioner experiment (Practitioners and GETS) supports status quo bias significantly; thus, Hypothesis 5 is not supported. This result is consistent with the result on the endowment effect. In tests for both cognitive bias types, practitioners are more sensitive. Hypothesis 6 concerns the moderation effect of consignment auctions. Because status quo bias was not supported in either Students and GETS or Students and CAS, Hypothesis 6 is not supported.
Discussion
This study tested two types of cognitive bias in emissions trading situations: the endowment effect and status quo bias. The endowment effect was supported significantly in all three experiments. The status quo bias was supported significantly only in the practitioner experiment. Overall, this study shows that cognitive bias can occur in emissions trading situations, which may explain the lack of emissions trading transactions.
The finding concerning the WTA/WTP ratio is notable. Horowitz and McConnell [25] showed that the ratio of WTA/WTP is generally about seven. However, in this study, the ratio was 1.149 to 1.329. Allowances are a kind of claim and have characteristics similar to financial assets. This study's finding of a low WTA/WTP ratio is consistent with the finding in previous studies that money, commodities, and goods that are expected to be resold create a lower loss aversion and endowment effect [14,25,30].
In addition, this study tested whether more experience can reduce cognitive bias. Emissions-trading practitioners, who are supposed to have much more experience and understanding of emissions trading, showed no evidence of having cognitive biases lower than those of students. In fact, practitioners showed a significantly higher endowment effect and only the practitioners showed status quo bias. Three possible explanations for these results are discussed below.
First, though practitioners have much experience in emissions trading, they might not have received adequate feedback and might not understand which kind of decisions can lead to cost minimization. The result of an emissions trading transaction depends on many factors, such as the estimate of the amount of their own emissions, the future prices of allowances, and changes in government policy. Thus, they may not perceive that cognitive bias led them to make a wrong choice. For example, even if a high valuation of their allowances came from the endowment effect, they might assume that they estimated future allowances prices incorrectly. Thus, they may not adjust their cognitive biases if their decisions are affected by many factors and are complex. However, this does not explain why practitioners show stronger cognitive biases.
Second, firms' incentive systems for emissions trading may be the cause. If firms provide asymmetric incentives that provide penalties for loss that are higher than the rewards for gains of the same amount, practitioners will feel a level of loss aversion that is higher than what students feel. Loss aversion is the main explanation of the endowment effect and status quo bias [14,25,30]. Even though the questionnaires in this study asked subjects to think of themselves as the CEO, this might not fully exclude the effect of asymmetric incentives. However, this logic does not explain how this asymmetric incentive occurs. If it comes from cognitive biases such as the risk aversion or loss aversion of the organizational members, it forms a circular logic.
Third, firm members have more loss aversion than students, not only in emissions trading, but also in most decision making. Defensive or self-limiting behavior, an important issue in organizational behavior [47,48], relates to risk aversion and can create an endowment effect and status quo bias that is higher than that observed in general students. The higher level of loss aversion of practitioners can be related to the concept of decision avoidance which was discussed in relation to Hypothesis 5 [41,45,46]. If practitioners feel more loss aversion, they will seek more sophisticated decision process and feel more difficulty in decision making even though they have more experience than students. If the higher level of loss aversion occurs because of the general organizational culture, reducing cognitive biases such as the endowment effect and status quo bias may be very difficult. However, firms should be aware of the possibility of cognitive biases and try to reduce them by exercising additional caution and providing education.
In addition, this study tested whether consignment auctions can reduce the endowment effect and status quo bias. In both of the endowment effect and status quo bias experiments, no significant differences between emissions trading with and without consignment auction were observed. Thus, this study has not found any statistically significant evidence supporting that consignment auctions contribute to reducing cognitive biases.
Conclusions and Future Directions
This study tested whether the cognitive biases happen in the setting of emissions trading and showed that the cognitive biases occur in some situations. We also tested the effect of more experience and consignment auction on the cognitive biases. Contrary to conventional expectations, practitioner groups who were supposed to have more experience on emissions trading showed higher cognitive biases than the student group. In addition, despite the intuition that consignment auction may reduce the feeling of possession and make the status quo ambiguous, consignment auction also did not significantly mitigate the cognitive biases.
The cognitive bias may help partly explain the lack of transactions in emissions trading that is observed in the early stage of emissions trading schemes including Korea ETS and China ETS. Endowment effect would make companies have a stronger tendency to hold more of their acquired emission allowances than to sell them at proper prices. Status quo bias also make companies reluctant to do transactions. Then, the emissions trading market is subject to lack of supply and trading. The price of the allowance surges near the time to submit the emission permits. This is burdensome for companies to meet their targets by purchasing allowances and an authority makes some changes to the rules of emissions trading schemes in order to appease the companies complaining the high price and the lack of supply. Actually, this phenomenon occurred several times in Korea, where emissions trading was introduced in 2015.
The resulting lack of trading could hinder cost-effective greenhouse gas mitigation in general. Thus, it is important to know how to mitigate the cognitive bias in order to invigorate emissions trading. Because accumulation of experience and consignment auction did not significantly lower cognitive biases in this study, other methods to reduce cognitive biases should be developed. It is possible to think that decision-making structures and reward systems in an organization are important to make people in charge of emissions trading behave rationally. With respect to decision-making structures, the result of Chan and Saqib [36] should be noticed. They showed that the endowment effect occurs when people feel low power and argued that feeling powerful can reduce the endowment effect. This idea is worth testing in emissions trading situations. Further, Dean, Kıbrıs, and Masatlioglu [45], Gerasimou [49], and Tversky and Shafir [50] suggested decision avoidance concept to explain status quo bias. Decision avoidance means that if practitioners feel less difficulty in decision making, they will show less status quo bias. Some methods to reduce difficulty in decision making should be developed; such as more information about the market situation and government policy and education about emission trading. More studies about these issues are worth doing.
Finally, we mention a few limitations of this study and those will be helpful to make future research more valid. First, this study adopted a simple survey-based experiment rather than experiment using incentive-compatible mechanism such as auction or BDM (Becker-DeGroot-Marschak) method, in which all participants can achieve the best outcome to themselves just by acting according to their true preferences, to make the experiment mimic real-world trading. Even though designing incentive-compatible mechanisms which mimic real emission trading is difficult and costly, it will increase the validity of the results.
Second, another thing to be noticed is that the number of sample in each experiment is small. Even though most of survey-based experiments in the literature used 30~40 subjects in each experiment as like this study, a higher number of subjects will increase the robustness of the results.
Third, the scenarios in the experiments are more or less abstract and did not fully reflect the real situation. It was to make the scenario easy to understand where subjects could focus on the main questions, specifically for the student group. However, development of more detailed scenario with the incentive compatible mechanism can be more realistic experiment. Further, the scenario of consignment auction is more complex than that of general emissions trading, which makes some inconsistency in comparing the results of emissions trading between the cases with and without consignment auction.
Fourth, this study showed that practitioners were more sensitive to the endowment effect and status quo bias, but the discussion about possible causes relied upon inferences rather than evidence. It also showed that status quo bias was significant only in the experiment with practitioners. These issues also need additional research. | 8,478.4 | 2019-03-05T00:00:00.000 | [
"Environmental Science",
"Economics"
] |
Convergence structures and Hausdorff uo-Lebesgue topologies on vector lattice algebras of operators
A vector sublattice of the order bounded operators on a Dedekind complete vector lattice can be supplied with the convergence structures of order convergence, strong order convergence, unbounded order convergence, strong unbounded order convergence, and, when applicable, convergence with respect to a Hausdorff uo-Lebesgue topology and strong convergence with respect to such a topology. We determine the general validity of the implications between these six convergences on the order bounded operator and on the orthomorphisms. Furthermore, the continuity of left and right multiplications with respect to these convergence structures on the order bounded operators, on the order continuous operators, and on the orthomorphisms is investigated, as is their simultaneous continuity. A number of results are included on the equality of adherences of vector sublattices of the order bounded operators and of the orthomorphisms with respect to these convergence structures. These are consequences of more general results for vector sublattices of arbitrary Dedekind complete vector lattices. The special attention that is paid to vector sublattices of the orthomorphisms is motivated by explaining their relevance for representation theory on vector lattices.
INTRODUCTION AND OVERVIEW
In an earlier paper [15] the authors studied aspects of locally solid linear topologies on vector lattices of order bounded linear operators between vector lattices. Particular attention was paid to the possibility of introducing a Hausdorff uo-Lebesgue topology on such vector lattices.
Such vector lattices of operators carry two, and possibly three, natural convergence structures (order convergence, unbounded order convergence, and, when applicable, convergence with respect to the Hausdorff uo-Lebesgue topology), as they can be defined for arbitrary vector lattices. For vector lattices of operators, however, besides these 'uniform' convergence structures, there are also two, or possibly three, corresponding 'strong' counterparts that can be defined in the obvious way. Several relations between the resulting six convergence structures on vector lattices of operators were also investigated in [15]. In view of their relevance for representation theory in vector lattices, special emphasis was put on the orthomorphisms on a Dedekind complete vector lattice. In this case, implications between convergences hold that do not hold for more general vector lattices of operators. Furthermore, it was shown that the orthomorphisms are not only order continuous, but also continuous with respect to unbounded order convergence on the vector lattice and with respect to a possible Hausdorff uo-Lebesgue topology on it.
Apart from their intrinsic interest, the results in [15] can be viewed as a part of the groundwork that has to be done in order to facilitate further developments of aspects of the theory of vector lattices of operators. The questions that are asked are natural and basic, but even so the answers are often more easily formulated than proved.
In the present paper, we take this one step further and study these six convergence structures in the context of vector lattice algebras of order bounded linear operators on a Dedekind complete vector lattice. Also here there are many natural questions of a basic nature that need to be answered before one can expect to get much further with the theory of such vector lattice algebras and with representation theory on vector lattices. For example, is the left multiplication by a fixed element continuous on the order bounded linear operators with respect to unbounded order convergence? Is the multiplication on the order continuous linear operators simultaneously continuous with respect to a possible Hausdorff uo-Lebesgue topology on it? Given a vector lattice subalgebra of the order continuous linear operators, is the closure (we shall actually prefer to speak of the 'adherence') in the order bounded linear operators with respect to strong unbounded order convergence again a vector lattice subalgebra? Is there a condition, sufficiently lenient to be of practical relevance, under which the order adherence of a vector lattice subalgebra of the orthomorphisms coincides with its closure in a possible Hausdorff uo-Lebesgue topology? Building on [15], we shall answer these questions in the present paper, together with many more similar ones. As indicated, we hope and expect that, apart from their intrinsic interest, this may serve as a stockpile of basic, but non-elementary, results that will facilitate a further development of the theory of vector lattice algebras of operators and of representation theory in vector lattices. This paper is organised as follows.
Section 2 contains the necessary notation, definitions, and conventions, as well as a few preparatory results that are of interest in their own right. Proposition 2.6 shows that, in many cases of practical interest, a unital positive linear representation of a unital f-algebra on a vector lattice is always an action by orthomorphisms. Its consequence Corollary 2.7 specialises this to the case of left and right multiplications of order bounded operators by orthomorphisms.
In Section 3, we study the validity of each of the 36 possible implications between the 6 convergences that we consider on vector lattice algebras of order bounded linear operators on a Dedekind complete vector lattice. We do this for the order bounded linear operators as well as for the orthomorphisms.
The results that are already in [15] and a few additional ones are sufficient to complete the tables 3.1 and 3.2.
Section 4 contains our results on the continuity of the left and right multiplications by a fixed element with respect to each of the six convergence structures on the order bounded linear operators. For this, we distinguish between the multiplication by an arbitrary order bounded linear operator, by an order continuous one, and by an orthomorphism. By giving (counter) examples, we show that our results are sharp in the sense that, whenever we state that continuity holds for multiplication by, e.g., an orthomorphism, it is no longer generally true for an arbitrary order continuous linear operator, i.e., for an operator in the 'next best class'. We also consider these questions for the orthomorphisms. The results are contained in tables 4.14 to 4. 16.
In Section 5, we investigate the simultaneous continuity of the multiplication with respect to each of the six convergence structures. When there is simultaneous continuity, the adherence of a subalgebra is, of course, again a subalgebra. With only one exception (see Corollary 5.6 and Example 5.7), we give (counter) examples to show that our conditions for the adherence of an algebra to be a subalgebra again are 'sharp' in the sense as indicated above for Section 4. Section 6 is dedicated to the equality of various adherences of vector sublattices and vector lattice subalgebras. It is also indicated there how representation theory in vector lattices leads quite naturally to the study of vector lattice subalgebras of the orthomorphisms (see the Theorems 6.1 and 6.2), thus motivating in more detail the special attention that is paid to the orthomorphisms in [15] and in the present paper.
PRELIMINARIES
In this section, we give the notation, conventions, and definitions used in the sequel. We also include a few preliminary results.
All vector spaces are over the real numbers and all vector lattices are supposed to be Archimedean. We let E + denote the positive cone of a vector lattice E. The identity operator on a vector lattice E will be denoted by I, or by I E when the context requires this. The characteristic function of a set S is denoted by χ S .
Let E be a vector lattice, and let x ∈ E. We say that a net (x α ) α∈A in E is order convergent to x ∈ E (denoted by x α o − → x) when there exists a net (x β ) β∈B in E such that y β ↓ 0 and with the property that, for every β 0 ∈ B, there exists an α 0 ∈ A such that |x − x α | ≤ y β 0 whenever α in A is such that α ≥ α 0 . Note that, in this definition, the index sets A and B need not be equal.
A net (x α ) α∈A in a vector lattice E is said to be unbounded order convergent Order convergence implies unbounded order convergence to the same limit. For order bounded nets, the two notions coincide.
Let E and F be vector lattices. The order bounded linear operators from E into F will be denoted by ob (E, F); this is a Dedekind complete vector lattice when F is. We write E for ob (E, ). A linear operator T : E → F between two vector lattices E and F is order continuous when, for every net (x α ) α∈A in An order continuous linear operator between two vector lattices is automatically order bounded; see [4, Lemma 1.54], for example. The order continuous linear operators from E into F will be denoted by oc (E, F). We write E oc for oc (E, ).
Let F be a vector sublattice of a vector lattice E. Then F is a regular vector sublattice of E when the inclusion map from F into E is order continuous. Ideals are regular vector sublattices. For a net in a regular vector sublattice F of E, its unbounded order convergence in F and in E are equivalent; see [18,Theorem 3.2].
An orthomorphism on a vector lattice E is a band preserving order bounded linear operator. We let Orth(E) denote the orthomorphisms on E. Orthomorphisms are automatically order continuous; see [4,Theorem 2.44]. An overview of some basic properties of the orthomorphisms that we shall use can be found in the first part of [15,Section 6], with detailed references included.
A topology τ on a vector lattice E is a uo-Lebesgue topology when it is a (not necessarily Hausdorff) locally solid linear topology on E such that, for a net For the general theory of locally solid linear topologies on vector lattices we refer to [3]. A vector lattice need not admit a uo-Lebesgue topology, and it admits at most one Hausdorff uo-Lebesgue topology; see [8, Propositions 3.2, 3.4, and 6.2] or [26, Theorems 5.5 and 5.9]). In this case, this unique Hausdorff uo-Lebesgue topology is denoted by τ E .
Proof. The equivalence of the parts (1) and (2) Let X be a non-empty set. As in [15], we define a convergence structure on X to be a non-empty collection of pairs ((x α ) α∈A , x), where (x α ) α∈A is a net in X and x ∈ X , such that: Replacing nets by sequences and subnets by subsequences gives the usual sequential convergence structure, as in [5, Definition 1.7.1].
Remark 2.2.
Our definition of a convergence structures is actually a possible definition of a so-called net convergence structure. The theory of their counterparts, the so-called filter convergence structures, has been canonised in [5]. It is only recently that a definition of a net convergence structure (more sophisticated than ours) has been given that can be shown to yield a natural bijection between the net convergence structures and the filter convergence structures on a set; see [24]. In this definition, not all index sets for the nets in the structure are admitted, the admissible index are allowed to be merely pre-ordered, and property (1) in the above definition is replaced with two others. We refer to [24] for further details, and content ourselves with our definition above that is sufficient for our merely descriptive purposes.
Suppose that is a convergence structure on a non-empty set X . For a nonempty subset S ⊆ X , we define the -adherence of S in X as We set a ( ) := . A subset S of X is said to be -closed when a (S) = S. It is evident how define the adherence of a subset in the case of a sequential convergence structure. The following result, which was mentioned in [15, Section 1] without proof (see also [16,Section 8] for special cases), was already established in a context of order convergence on partially ordered vector spaces as [27,Theorem 3.1]. Its sequential version is proved using a similar argument. Lemma 2.3. Let X be a non-empty set, and let be a convergence structure on X . Then the -closed subsets of X are the closed sets of a topology τ on X .
Proof. It is trivial that and X are -closed, and it is immediate that an arbitrary intersection of -closed subsets of X is -closed. We claim that a (S 1 ∪ S 2 ) = a (S 1 ) ∪ a (S 2 ) for arbitrary S 1 , S 2 ⊆ X , which implies that finite unions of -closed subsets are again -closed. Since it is obvious that a (S 1 ∪ S 2 ) ⊇ a (S 1 ) ∪ a (S 2 ), we need to show only the reverse inclusion. We may suppose that S 1 , S 2 = . Take an x ∈ a (S 1 ∪ S 2 ). Then there exists a net (x α ) α∈A in S 1 ∪ S 2 such that ((x α ) α∈A , x) ∈ . If there is a tail (x α ) α≥α 0 that is contained in S 1 , then it follows from property (1) of that x ∈ a (S 1 ). If no tail of (x α ) α∈A is contained in S 1 , then the set B := {β ∈ A : x β ∈ S 2 } is a co-final subset of A. Hence (x β ) β∈B can canonically be viewed as a subnet of (x α ) α∈A that is contained in S 2 , and property (1) implies that x ∈ a (S 2 ). In both cases, It is not generally true that a (S) is τ -closed. We have the following result, the final statement of which was already mentioned without proof for special cases in [16,Section 8]. Its sequential version is valid by essentially the same proof.
Lemma 2.4. Let X be a non-empty set, let be a convergence structure on X , and let S ⊆ X . Then Proof. Property (2) of implies that S ⊆ a (S), and the obvious monotony of the adherence set map implies that S τ = a (S τ ) ⊇ a (S). The remaining two statements follow easily from the chain of inclusions.
On a vector lattice E, the set of all pairs of order convergent nets and their order limits forms a convergence structure o on E. Likewise, there is a convergence structure uo on E and, when applicable, a topological convergence structure τ E . For a subset S of E, we shall write a o (S) for a o (S), a uo (S) for a uo (S), and, when applicable, S τ E for a τ E (S). The corresponding sequential convergence structures are denoted by σo , σuo , and, when applicable, σ τ E , respectively. There are self-explanatory notations a σo (S), a σuo (S), and, when applicable, a σ τ E (S). We shall also speak of the order adherence (or oadherence) of a subset, rather than of its o -adherence; etc. Note that the order adherence a o (S) of S is what is called the 'order closure' of S in other sources. Since this 'order closure' need not be closed in the τ o -topology on E, we shall not use this terminology that is prone to mistakes.
Let E and F be vector lattices, where F is Dedekind complete. Suppose that is a vector sublattice of ob (E, F). As for general vector lattices, we have the convergence structures o ( ), uo ( ) and, when applicable, a convergence structure τ on . In addition to these 'uniform' convergence structures, there are in this case also 'strong' ones that we shall now define. Let (T α ) α∈A be a net in , and let T ∈ . Then we shall say that (T α ) α∈A is strongly order convergent The set of all pairs of strongly order convergent nets in and their limits forms a convergence structure SO on . Likewise, the net is strongly unbounded order convergent to T (denoted by T α SUO −−→ T ) when it is pointwise unbounded order convergent to T , resulting in a convergence structure SUO on . When E admits a Hausdorff uo-Lebesgue topology τ E , then a net is strongly convergent with respect to τ E to T (denoted by T α S τ E − − → T ) when it is pointwise τ E convergent to T , yielding to a convergence structure S τ on . As for the three convergence structures on general vector lattices, we shall simply write a SUO ( ) for the SUO -adherence a SUO ( ) of a subset of ; etc. We shall use a similar simplified notation for adherences corresponding to the sequential strong convergence structures that are defined in the obvious way.
The adherence of a set in a convergence structure obviously depends on the superset, since this determines the available possible limits of nets. In an ordered context, there can be additional complications because, for example, the notion of order convergence of a net itself depends on the vector lattice that the net is considered to be a subset of. It is for this reason that, although we have not included the superset in the notation for adherences, we shall always indicate it in words. Let X be a convergence structure on a non-empty set X , and let Y be a convergence structure on a non-empty set Y . A map Φ : X → Y is said to be X -Y continuous when, for every pair ((x α ) α∈A , x) in X , the pair ((Φ(x α )) α∈A , Φ(x)) is an element of Y . We shall speak of S τ E -o continuity rather than of S τ Eo continuity; etc.
Let E be a vector lattice. For T ∈ ob (E), we define ρ T , λ T : ob (E) → ob (E) by setting ρ T (S) := ST and λ T (S) := T S for S ∈ ob (E). We shall use the same notations for the maps that result in other contexts when compositions with linear operators map one set of linear operators into another.
For later use in this paper, we establish a few preparatory results that are of some interest in their own right.
Lemma 2.5.
Let be an f-algebra with an identity element e, and let E be a vector lattice with the principal projection property. Let a ∈ + , and suppose that is a positive linear map such that π(e) = I. Then π(a) ∈ Orth(E).
Proof. It is obvious that π(a) ∈ ob (E), so it remains to be shown that π(a) is band preserving on E. We know from [4, Theorem 2.57] that for n ≥ 1. Take x ∈ E + . Then we have for n ≥ 1. Let B x be the band generated by x in E, and let P x ∈ ob (E) be the order projection onto B x . Using that π(a)x ≥ 0 and equation (2.1), we have for all n ≥ 1. Hence (I − P x )[π(a)x] = 0, so that π(a)x ∈ B x . Since x was arbitrary, this shows that π(a) is band preserving.
Proposition 2.6.
Let be an f-algebra with an identity element e, and let E be a vector lattice with the principal projection property. Suppose that π : → ob (E) is a positive linear map such that π(e) = I. Then π( ) ⊆ Orth(E). If, in addition, π preserves the multiplication, then π is a unital vector lattice algebra homomorphism from into Orth(E).
Proof. It is clear from Lemma 2.5 that π maps into Orth(E). Suppose that π also preserves the multiplication. In this case, we note that Orth(E) is a unital Archimedean f-algebra (see [4,Theorem 2.59]), so that it is semiprime by [11,Corollary 10.4]. Since is likewise semiprime, it follows from [11, p. 96] (see also [11, part (i) of Theorem 3.7]) that π is a vector lattice homomorphism.
The following is immediate from Proposition 2.6 and-for its first part-the commutativity of Orth(E) (see [ show that the (then coinciding) maps ρ and λ even provide a vector lattice algebra isomorphisms between the Orth(E) and Orth(Orth(E). This is also true when E is not Dedekind complete.
IMPLICATIONS BETWEEN CONVERGENCES ON VECTOR LATTICES OF OPERATORS
In this section, we investigate the implications between the six convergences on the order bounded linear operators and on the orthomorphisms on a Dedekind complete vector lattice. Without further ado, let us simply state the answers and explain how they are obtained.
For a general net of order bounded linear operators (resp. orthomorphisms) on a general Dedekind complete vector lattice, the implications between order convergence, unbounded order convergence, convergence in a possible Hausdorff uo-Lebesgue topology, strong order convergence, strong unbounded order convergence, and strong convergence with respect to a possible Hausdorff uo-Lebesgue topology, are given in Table 3.1 (resp. Table 3.2).
In Orth(E), uo and SUO convergence of nets coincide, as do a possible τ Orth (E) and S τ E convergence.
In these tables, the value in a cell indicates whether the convergence of a net in the sense that labels the row of that cell does (value 1) or does not (value 0) in general imply its convergence (to the same limit) in the sense that labels the column of that cell. For example, the value 0 in the cell (uo, S τ E ) in Table 3.1 indicates that there exists a net of order bounded linear operators on a Dedekind complete vector lattice E that admits a Hausdorff uo-Lebesgue topology τ E , such that this net is unbounded order convergent to zero in ob (E), but not strongly convergent to zero with respect to τ E . The value 1 in the cell (uo, S τ E ) in Table 3.2, however, indicates that every net of orthomorphisms on an arbitrary Dedekind complete vector lattice E that admits a Hausdorff uo-Lebesgue topology τ E , such that this net is unbounded order convergent to zero, is strongly convergent to zero with respect to τ E .
We shall now explain how these tables can be obtained. Obviously, the order convergence of a net of operators implies its unbounded order convergence, which implies its convergence in a possible Hausdorff uo-Lebesgue topology. There are similar implications for the three associated strong convergences. Furthermore, an implication that fails for orthomorphisms also fails in the general case. Using these basic facts, it is a logical exercise to complete the tables from a few 'starting values' that we now validate.
For Table 3.1, we have the following 'starting values': It is easily checked that the above information suffices to complete both tables.
Remark 3.3.
(1) Every order bounded net of orthomorphisms on an arbitrary Dedekind complete vector lattice E that is strongly order convergent to zero, is order convergent to zero in Orth(E); see [15, Theorem 9.4]; (2) Every sequence of orthomorphisms on a Dedekind complete Banach lattice E that is strongly order convergent to zero, is order convergent to zero in Orth(E); see [15, Theorem 9.5]; (3) The validity of all zeroes in Table 3.1 (resp. Table 3.2) follows from the existence of a net of order bounded linear operators (resp. orthomorphisms) on a Dedekind complete Banach lattice for which the implication in question does not hold. With the cell (SO, o) in Table 3.2 as the only exception, such a net of operators on a Banach lattice can even be taken to be a sequence. This follows from an inspection of the (counter) examples referred to above when validating the 'starting' zeroes in the tables.
CONTINUITY OF LEFT AND RIGHT MULTIPLICATIONS
In this section, we study continuity properties of left and right multiplication operators. For example, take an arbitrary T ∈ ob (E), where E is an arbitrary Dedekind complete vector lattice that admits a Hausdorff uo-Lebesgue topology τ ob (E) . Is it then true that λ T : ob (E) → ob (E) maps unbounded order convergent nets in ob (E) to τ ob (E) convergent nets (with corresponding limits)? If not, is this then true when we suppose that T ∈ oc (E)? If not, is this true when we suppose that T ∈ Orth(E)? One can ask a similar combination of questions, specifying to classes of increasingly well-behaved operators, for each of the 6 · 6 = 36 combinations of convergences of nets in ob (E) under consideration in this paper. There are also 36 combinations to be considered for right multiplication operators. This section provides the answers in all 72 cases; the results are contained in the Tables 4.14 and 4.15. For the example that we gave, the answer is negative and remains so for arbitrary T ∈ oc (E), but it becomes positive for arbitrary T ∈ Orth(E). For Orth(E), there are similar questions to be asked for its left and right regular representation, but their number is smaller. Firstly, we see no obvious better-behaved subclass of Orth(E) that we should also consider. Secondly, since Orth(E) is commutative, there is only one type of multiplication involved. Thirdly, as in Table 3.2, there are two pairs of coinciding convergences. All in all, there are only 4 x 4 = 16 possible combinations that actually have to be considered for the regular representation of Orth(E). Also in this case, all answers can be given; the results are contained in Table 4. 16. As it turns out, Table 4.16 is identical to Table 3.2. There appears to be no a priori reason for this fact; it is simply the outcome.
We shall now set out to validate the Tables 4.14, 4.15, and 4.16. Fortunately, we do not need individual results for every cell. Upon considering the multiplications by the orthomorphism that is the identity operator, the zeroes in the Tables 3.1 and 3.2 already determine the values in many cells. For the remaining ones, the combination of the 'standard' implications that were already used for the Tables 3.1 and 3.2 and a limited number of results and (counter) examples already suffices. We shall now start to collect these.
We start with o-o and SO-SO continuity.
Proof. We prove the parts (1) and (2). Take T ∈ ob (E), and let (S α ) α∈A ⊆ ob (E) be a net such that S α o − → 0 in ob (E). By passing to a tail, we may assume that (|S α |) α∈A is order bounded in ob (E).
The parts (3) and (4) are immediate consequences of the definitions.
We now show that the condition in the parts (2) and (4) of Proposition 4.1 that T ∈ oc (E) cannot be relaxed to T ∈ ob (E).
n=1 be the sequence of standard unit vectors in E, and let c denote the sublattice of E consisting of the convergent sequences. We define a positive linear functional f c on c by setting . We define S n ∈ oc (E) for n ≥ 1 by setting The sequence (S n ) ∞ n=1 , being order convergent to S, is also strongly unbounded order convergent to S in ob (E).
Remark 4.3.
Examples 4.2 also shows that, already for a Banach lattice E, λ T need not even be sequentially o-τ ob (E) continuous, sequentially o-S τ E continuous, sequentially SO-τ ob (E) continuous, or sequentially SO-S τ E continuous for arbitrary T ∈ ob (E).
Remark 4.4.
The o-o continuity (appropriately defined) of left and right multiplications on ordered algebras is studied in [2]. It is established on [2, p. 542-543] that, for a Dedekind complete vector lattice E, the right and left multiplication by an element T of the ordered algebra L(E) of all (!) linear operators on E are both order continuous on L(E) in the sense of [2] if and only if the left multiplication is, which is the case if and only if T ∈ oc (E). The proof refers to [1, Example 2.9 (a)], which is concerned with multiplications by a positive operator T on the ordered Banach algebra L(E) of all (!) bounded linear operators on a Dedekind complete Banach lattice E. It is established in that example that the simultaneous order continuity of the right and left multiplication by T on L(E) in the sense of [1] is equivalent to T being order continuous. On [1, p. 151] it is mentioned that this criterion for the order continuity of an operator can also be presented for an arbitrary Dedekind complete vector lattice. Although it is not stated as such, and although a proof as such is not given, the author may have meant to state, and have known to be true, that, for a Dedekind complete vector lattice E and T ∈ ob (E), λ T and ρ T are both o-o continuous on ob (E) in the sense of the present paper if and only if λ T is, which is the case if and only if T ∈ oc (E). Using arguments as on [1, p. 151] and [2, p. 542-543], the authors of the present paper have verified that-this is the hard part-for T ∈ ob (E), the o-o continuity of λ T on ob (E) in the sense of the present paper does imply that T ∈ oc (E). Hence the three properties of T ∈ ob (E) mentioned above are, indeed, equivalent; a result that is to be attributed to the late Egor Alekhno.
We use the opportunity to establish the following side result, which follows easily from combining each of [25, Satz 3.1] and [7, Proposition 2.2] with the parts (1) and (2) of Proposition 4.1.
Proposition 4.5. Let E be a Dedekind complete vector lattice. Then:
(1) the map T → ρ T defines an order continuous lattice homomorphism ρ : (2) the map T → λ T defines an order continuous lattice homomorphism λ : . Remark 4.6. In [28, Problem 1], it was asked, among others, whether, for a Dedekind complete vector lattice E, the left regular representation of ob (E) is a lattice homomorphism from ob (E) into ob ( ob (E)). In [10,Theorem 11.19], it was observed that the affirmative answer is, in fact, provided by [25, Satz 3.1]. Part (2) of Proposition 4.5 gives still more precise information.
Part (1), which relies on [7, Proposition 2.2], implies that the right regular representation of oc (E) is an order continuous lattice homomorphism from oc (E) into ob ( oc (E)), with an image that is, in fact, contained in oc ( oc (E)). After this brief digression, we continue with the main line of this section, and consider uo-uo and SUO-SUO continuity of left and right multiplication operators.
We prove part (4). We now show that the condition in the parts (1), (2), and (4) of Proposition 4.7 that T ∈ Orth(E) cannot be relaxed to T ∈ oc (E).
Examples 4.9.
(1) We first give an example showing that λ T and ρ T need not be uo-uo continuous on ob (E) for T ∈ oc (E).
for f ∈ E. For n ≥ 1, we define the positive operator S n on E by setting We claim that (S n ) ∞ n=1 is a disjoint sequence in ob (E). Let m, n ≥ 1 with m > n. Take a k ≥ 1 such that 1/k < 1/n − 1/m. For every f ∈ E + , [4, Theorem 1.51] then implies that and S n ( f · χ [(i−1)/k,i/k] ) are disjoint for i = 1, . . . , k. Hence S m ∧ S n = 0, as claimed.
By [18,Corollary 3.6], the disjoint sequence (S n ) ∞ n=1 is unbounded order convergent to zero in ob (E). On the other hand, it is easy to see that ρ T (S n ) = λ T (S n ) = T = 0 for all n ≥ 1. Hence neither of (ρ T (S n )) ∞ n=1 and (λ T (S n )) ∞ n=1 is unbounded order convergent to zero in ob (E). This shows that neither ρ T nor λ T is uo-uo continuous on ob (E).
(2) We now give an example showing that λ T need not be SUO-SUO continuous on ob (E) for T ∈ oc (E).
For n ≥ 1, define the positive operator S n on E by setting S n f : n=1 is strongly unbounded order convergent to zero. On the other hand, it is easily seen that λ T (S n )χ [0,1] = χ [0,1] = 0 for n ≥ 1. This implies that (λ T (S n )) ∞ n=1 is not strongly unbounded order convergent to zero, so that λ T is not SUO-SUO continuous on ob (E).
Remark 4.10.
Examples 4.9 also shows that ρ T and λ T need not be uo-τ ob (E) continuous and that λ T need not be SUO-S τ E continuous on ob (E) for arbitrary T ∈ oc (E) = ob (E). In fact, the sequential versions of these continuity properties can already fail to hold, even in cases where E is a Banach lattice with an order continuous norm.
We now turn to the Hausdorff uo-Lebesgue topologies, where we shall make use of Theorem 2.1. Proposition 4.11. Let E be a Dedekind complete vector lattice that admits a (necessarily unique) Hausdorff uo-Lebesgue topology τ E , so that ob (E) also admits a (necessarily unique) Hausdorff uo-Lebesgue topology τ ob (E) . Then: Proof. We know from Corollary 2.7 that ρ T , λ T ∈ Orth( ob (E)) when T ∈ Orth(E), and then the parts (1) and (2) We now show that the condition in the parts (1), (2), and (4) of Proposition 4.11 that T ∈ Orth(E) cannot be relaxed to T ∈ oc (E).
Examples 4.12.
(1) We first give an example showing that λ T and ρ T need not be τ ob (E)τ ob (E) continuous on ob (E) for T ∈ oc (E). For this, we resort to the context and notation of part (1) of Examples 4.9. In that example, we know that S n uo − → 0 in ob (E), and then certainly S n τ ob (E) − −−− → 0. Since ρ T (S n ) = λ T (S n ) = T = 0 for all n ≥ 1, we see that neither ρ S nor λ S is τ ob (E) -τ ob (E) continuous on ob (E).
(2) We give an example showing that λ T need not be S τ E -S τ E continuous on ob (E) for T ∈ oc (E). For this, we resort to the context and notation of part (2) of Examples 4.9. In that example, we know that Remark 4.13. Examples 4.12 also shows that ρ T and λ T need not even be τ ob (E) -τ ob (E) continuous and that λ T need not even be S τ ob (E) -S τ ob (E) continuous on ob (E) for arbitrary T ∈ oc (E) = ob (E). In fact, the sequential versions of these continuity properties can already fail to hold, even in cases where E is a Banach lattice with an order continuous norm.
We now have sufficient material at our disposal to determine the tables mentioned at the beginning of this section.
For right multiplications on ob (E), the results are in Table 4.14. The value in a cell with a row label indicating a convergence structure 1 and a column label indicating a convergence structure 2 is to be interpreted as follows: (1) A value {0} (resp. Orth(E), resp. oc (E)) means that ρ T is 1 -2 continuous on ob (E) for every Dedekind complete vector lattice E and for every T ∈ {0} (resp. T ∈ Orth(E), resp. T ∈ oc (E)), but there exist a Dedekind complete vector lattice E and a T ∈ Orth(E) (resp. T ∈ oc (E), resp. T ∈ ob (E)) for which this is not the case; (2) A value ob (E) means that ρ T is 1 -2 continuous on ob (E) for every Dedekind complete vector lattice E and for every T ∈ ob (E).
As mentioned in the beginning of this section, a zero in Table 3.1 gives {0} in Table 4.14. It is easily verified that the remaining values can be determined using that order convergence implies unbounded order convergence, which implies τ E convergence when applicable; that analogous implications hold for their strong versions; that order convergence implies strong order convergence; combined with Proposition 4.1, Proposition 4.7, Remark 4.10, Proposition 4.11, and Remark 4.13.
For left multiplications on ob (E), the results are in Table 4.15, with a similar interpretation of the values in the cells as for Table 4.14.
For Table 4.15, the values of the cells can be determined using the zeroes in Table 3.1, the 'standard implications' as listed for Table 4.14, combined with For multiplications on Orth(E), the continuity properties are given by Table 4.16. In that table, a value 1 in a cell with a row label indicating a convergence structure 1 and a column label indicating a convergence structure 2 means that the maps ρ T = λ T : Orth(E) → Orth(E) is 1 -2 continuous for all T ∈ Orth(E). A value 0 means that there exists a Dedekind complete vector lattice E and a T ∈ Orth(E) for which this is not the case.
In Orth(E), uo and SUO convergence of nets coincide, as do a possible τ Orth (E) and S τ E convergence.
The values in the cells of Table 4.16 can be determined using the zeroes in Table 3.2, the 'standard implications' as listed for Table 4.14; the fact that Orth(E) is a regular vector sublattice of ob (E); the facts that unbounded order convergence and strong unbounded order convergence coincide on Orth(E), as do a possible τ Orth(E) and S τ E convergence; combined with Proposition 4.1, Proposition 4.7, and Proposition 4.11.
SIMULTANEOUS CONTINUITY OF MULTIPLICATIONS AND ADHERENCES OF SUBALGEBRAS OF ob (E)
In this section, we study the simultaneous continuity of the multiplications in subalgebras of ob (E) (where E is a Dedekind complete vector lattice) with respect to the six convergence structures under consideration in this paper. This is motivated by questions of the following type. Suppose that E admits a Hausdorff uo-Lebesgue topology. Take a subalgebra (not necessarily a vector lattice subalgebra) of ob (E). Is its adherence a S τ E ( ) in ob (E) with respect to strong τ E convergence again a subalgebra of ob (E)? This is not always the case, not even when ⊆ oc (E); see Example 5.13. When ⊆ Orth(E), however, the answer is affirmative; see Corollary 5.12.
As is easily verified, it follows already from the continuity of the left and right multiplications with respect to strong τ E convergence (see Proposition 4.11) that a S τ E ( ) · a S τ E ( ) ⊆ a S τ E (a S τ E ( )) when ⊆ Orth(E), but that is not sufficient to show that a S τ E ( ) is a subalgebra. The simultaneous continuity of the multiplication in Orth(E) with respect to strong τ E convergence in Orth(E) would be sufficient to conclude this, and this can indeed be established; see Proposition 5.11.
For each of the remaining five convergence structures, we follow the same pattern. We establish (this also relies on the single variable results in Section 4) the simultaneous continuity of the multiplication with respect to the convergence structure under consideration, and then conclude that the pertinent adherence of a subalgebra is again a subalgebra. For the latter result it is-as the above example already indicates-essential to impose an extra condition on the subalgebra . This condition depends on the convergence structure under consideration. Natural extra conditions are that be a subalgebra of Orth(E) or of oc (E), and we do indeed obtain positive results under such conditions. We also have fairly complete results showing that the relaxation of the pertinent condition to the 'natural' next lenient one does, in fact, render the statement that the adherence is a subalgebra again invalid. This also implies that multiplication is then not simultaneously continuous.
In the cases where the lattice operations are known to be simultaneously continuous with respect to the convergence structure under consideration, it obviously also follows that the pertinent adherence of a vector lattice subalgebra is a vector lattice subalgebra again.
We shall now embark on this programme. We start with order convergence, which is the easiest case. For this, we have the following result on the simultaneous continuity of multiplication.
Proposition 5.1. Let E be a Dedekind complete vector lattice. Suppose that
Proof. It is clear that S ∈ oc (E). By passing to a tail, we may suppose that (|T β |) β∈B is bounded above by some R ∈ ob (E) + . Using the parts (1) and (2) of Proposition 4.1 for the final order convergence, we have that The following is now clear from Proposition 5.1 and the simultaneous order continuity of the lattice operations.
Corollary 5.2. Let E be a Dedekind complete vector lattice. Suppose that is a subalgebra of oc (E). Then the adherence a o ( ) in ob (E) is also a subalgebra of oc (E). When is a vector lattice subalgebra of oc (E), then so is a o ( ).
We now show that the condition in Corollary 5.2 that ⊆ oc (E) cannot be relaxed to ⊆ ob (E).
Example 5.3.
Take E = ℓ ∞ and let (e n ) ∞ n=1 be the standard sequence of unit vectors in E. We define T ∈ ob (E) as in Examples 4.2. For n ≥ 1, we now define S ′ n ∈ oc (E) by setting and S ′ ∈ oc (E) by setting It is easily verified that T 2 = 0, that S ′ n S ′ m = 0 for m, n ≥ 1, and that S ′ n T = T S ′ n = 0 for n ≥ 1. Hence := Span{T, S ′ n : n ≥ 1} is a subalgebra of ob (E). As S ′ n ↑ S ′ in ob (E), both S ′ and T are elements of a o ( ).
However, T S ′ / ∈ a o ( ). In fact, T S ′ is not even an element of a SO ( ) ⊇ a o ( ). To see this, we observe that T S ′ e 2 = e 1 = 0, and that, as is easily verified, T S ′ e 2 ⊥ Re 2 for all R ∈ . Hence there cannot exist a net Now we turn to the strong order adherences of subalgebras of ob (E). We start by showing that Orth(E) is closed in ob (E) under the convergences under consideration in this paper. We recall from Theorem 2.1 that either all of E, Orth(E), and ob (E) admit a Hausdorff uo-Lebesgue topology, or none does.
Lemma 5.4. Let E be a Dedekind complete vector lattice. Then Orth(E) is closed in ob (E) under order convergence, unbounded order convergence, strong order convergence, and strong unbounded order convergence. Suppose that E admits a (necessarily unique) Hausdorff uo-Lebesgue topology. Then Orth(E) is closed in
Proof. A band in a (not necessarily Dedekind complete) vector lattice is not only closed under order convergence, but also closed under unbounded order convergence (see [18,Proposition 3.15]) and under convergence in a Hausdorff locally solid linear topology on the lattice (see [3,Theorem 2.21(d)]. This implies that Orth(E) is closed in ob (E) under order convergence, unbounded order convergence, and convergence in a possible Hausdorff uo-Lebesgue topology on ob (E). It also implies that, for each of the three strong convergences in ob (E) under consideration, a limit in ob (E) of a net of orthomorphisms, i.e., of order bounded band preserving operators, is again an order bounded band preserving operator, i.e., an orthomorphism.
Proposition 5.5. Let E be a Dedekind complete vector lattice. Suppose that (S α ) α∈A
is a net in Orth(E) such that S α SO − → S in ob (E) for some S ∈ ob (E) and that Proof. Lemma 5.4 shows that S ∈ Orth(E). Take x ∈ E. By passing to a tail, we may suppose that (|T β x|) β∈B is bounded above by some y ∈ E + . By applying [4, Theorem 2.43] and the order continuity of |S| for the final convergence, we see that The following is now clear from Proposition 5.5. Corollary 5.6. Let E be a Dedekind complete vector lattice. Suppose that is a subalgebra of Orth(E). Then the adherence a SO ( ) in ob (E) is also a subalgebra of Orth(E).
We now show that the condition in Corollary 5.6 that ⊆ Orth(E) cannot be relaxed to ⊆ ob (E). At the time of writing, the authors do not know whether it might be relaxed to ⊆ oc (E).
Example 5.7. We resort to the context and notation of Example 5.3. In that example, we had operators T, S ′ ∈ a o ( ) such that T S ′ / ∈ a SO ( ). Since a o ( ) ⊆ a SO ( ), this example also provides an example as currently needed.
We turn to unbounded order adherences and strong unbounded order adherences.
Proposition 5.8. Let E be a Dedekind complete vector lattice. Suppose that
Then S, T ∈ Orth(E), and S α T β uo − → ST in ob (E). Seven similar statements hold that are obtained by, for each of the three occurrences of unbounded order convergence, either keeping it or replacing it with strong unbounded order convergence.
Proof. We start with the statement for three occurrences of unbounded order convergence. For this, we first suppose that S = T = 0.
For α ∈ A, let P α be the order projection in Orth(E) onto the band B α in Orth(E) that is generated by (|S α | − I) + . Then 0 ≤ P α I ≤ P α |S α | ≤ |S α | by [15,Lemma 6.9]. Hence P α I uo − → 0 in ob (E), so that also P α I uo − → 0 in the regular vector sublattice Orth(E) of ob (E) by [18,Theorem 3.2]. Since the net (P α I) α∈A is order bounded in Orth(E), we see that Combining the fact that |S α | ≤ I + P α |S α | by [15, Proposition 6.10(2)] with equation (5.2), we have, for α ∈ A, β ∈ B, For the case of general S and T , we first note that S, T ∈ Orth(E) as a consequence of Lemma 5.4. On writing We now show that, neither for a uo ( ) to be a subalgebra of ob (E), nor for a SUO ( ) to be a subalgebra of ob (E), the condition in Corollary 5.9 that ⊆ Orth(E) can be relaxed to ⊆ oc (E). for n ≥ 1, and we define T ∈ oc (E) by setting x i e i ∈ E. Set S n := S 1,2 − S 1,n+3 for n ≥ 1. It is not hard to check that T 2 = T , that S n T = T S n = 0 for n ≥ 1, and that S m S n = 0 for m, n ≥ 1. Hence := Span{T, S n : n ≥ 1} is a subalgebra of oc (E). Using [4, Theorem 1.51], it is easy to see that (S 1,n+3 ) ∞ n=1 is a disjoint sequence in ob (E), so that S 1,n+3 uo − → 0 in ob (E) by [18, Corollary 3.6]. Hence S n uo − → S 1,2 in ob (E), showing that S 1,2 ∈ a uo ( ). Obviously, T ∈ a uo ( ). We claim that, however, T S 1,2 is not even an element of τ ob (E) ⊇ a uo ( ). In order to see this, we observe that T S 1,2 = S 1,3 and, using [4, Theorem 1.51], that S 1,3 ⊥ T and S 1,3 ⊥ S n for n ≥ 1. Hence T S 1,2 ⊥ , which implies that x i e i ∈ ℓ 1 , we have S 1,n+3 x = x 1 e n+3 for n ≥ 1. This implies that S 1,n+3 SUO −−→ 0 in ob (E), showing that S n SUO −−→ S 1,2 in ob (E). Hence S 1,2 ∈ a SUO ( ). Obviously, T ∈ a SUO ( ). We claim that, however, T S 1,2 is not even an element of a S τ E ( ) ⊇ a SUO ( ). In order to see this, it is sufficient to observe that T S 1,2 e 1 = e 3 = 0 and that T S 1,2 e 1 ⊥ Re 1 for all R ∈ . This implies that We turn to closures in a Hausdorff uo-Lebesgue topology and strong closures with respect to a Hausdorff uo-Lebesgue topology. We recall once more from Theorem 2.1 that either all of E, Orth(E), and ob (E) admit a Hausdorff uo-Lebesgue topology, or none does. If they do, then, by general principles (see [26,Proposition 5.12]), τ Orth(E) is the restriction of τ ob (E) to Orth(E).
Then S, T ∈ Orth(E), and S
. Seven similar statements hold that are obtained by, for each of the three occurrences of τ ob (E) convergence, either keeping it or replacing it with strong τ E convergence.
Proof. We start with the statement for three occurrences of τ ob (E) convergence. For this, we first suppose that S = T = 0.
We can use parts of the proof of Proposition 5.8 here. In that proof, it was established that, for α ∈ A, there exists a band projection P α in Orth(E) such that . For the case of general S and T , we first note that S, T ∈ Orth(E) as a consequence of Lemma 5.4. On writing The following is now clear from Proposition 5.11 and the simultaneous continuity of the lattice operations with respect to the τ ob (E) topology.
Corollary 5.12. Let E be a Dedekind complete vector lattice that admits a (necessarily unique) Hausdorff uo-Lebesgue topology τ E . Suppose that is a subalgebra
of Orth(E). Then the closure τ ob (E) in ob (E) and the adherence a S τ E ( ) in ob (E) are equal, and are subalgebras of Orth(E). When is a vector lattice subalgebra of Orth(E), then so is We now show that, neither for τ ob (E) to be a subalgebra of ob (E), nor for a S τ E ( ) to be a subalgebra of ob (E), the condition in Corollary 5.12 that ⊆ Orth(E) can be relaxed to ⊆ oc (E).
Example 5.13. We return to the context and notation of Example 5.10. In that example, we saw that S n uo − → S 1,2 in ob (E). Then certainly S n τ ob −−→ S 1,2 in ob (E), so that both T and S 1,2 are elements of τ ob (E) . We saw in Example 5.10, however, that T S 1,2 / ∈ τ ob (E) .
It was also observed that S n SUO −−→ S 1,2 in ob (E). Since E is atomic, the unbounded order convergence of a net in E and its convergence in the Hausdorff uo-Lebesgue topology on E are known to coincide (see [9, Lemma 3.1] and [26,Lemma 7.4]). Thus also S n S τ E − − → S 1,2 , so that both T and S 1,2 are elements of a S τ E ( ). We saw in Example 5.10, however, that T S 1,2 / ∈ a S τ E ( ).
EQUALITY OF ADHERENCES OF VECTOR SUBLATTICES
In this section, we establish the equality of various adherences of vector sublattices with respect to convergence structures under consideration in this paper. We pay special attention to vector sublattices of the orthomorphisms on a Dedekind complete vector lattice. Apart from the intrinsic interest of the results, our research in this direction is also motivated by representation theory. We shall now explain this.
Suppose that E is a vector lattice, and that is a non-empty set of order bounded linear operators on E. Typical examples to keep in mind are those where is a group of order automorphisms of E (this arises naturally when considering positive representations of groups on vector lattices), or where is a (vector lattice) algebra of order bounded linear operators (this arises naturally when considering positive representations of (vector lattice) algebras on vector lattices). One of the main issues in general representation theory is to investigate the possible decompositions of a module into submodules. In our case, this is asking for decompositions E = F 1 ⊕ F 2 as an order direct sum of vector sublattices F 1 and F 2 that are both invariant under . It is well known (see [29,Theorem 11.3] for an even stronger result) that F 1 and F 2 are then projection bands that are each other's disjoint complements. Their respective order projections then commute with all elements of . Conversely, when an order projection has this property, then E is the order direct sum of its range and its kernel, and both are invariant under . All in all, the decomposition question for the action of on E is the same as asking for the order projections on E that commute with . This makes it natural to ask for the commutant of in Orth(E), where these order projections reside. This commutant is obviously an associative subalgebra of Orth(E). Somewhat surprisingly, it is quite often also a vector sublattice of Orth(E). For example, this is always true for Banach lattices, in which case the operators in need not even be regular. Being bounded is enough, as is shown by the following result, for which the Banach lattice need not even be Dedekind complete. Proof. It is obvious that ′ Orth is an associative subalgebra of Orth(E) that contains I and that is closed with respect to the coinciding operator norm and order unit norm · I . An appeal to [15, Theorem 6.1] then finishes the proof.
For Dedekind complete vector lattices, we have the following. Proof. For T ∈ Orth(E), we let λ T (resp. ρ T ) denote the corresponding left (resp. right) multiplication operator on ob (E). Since Corollary 2.7 implies that λ T − ρ T is an orthomorphism on ob (E), its kernel is a band in ob (E) by [4,Theorem 2.48]. It follows from this that the commutants of and B in Orth(E) are equal.
We shall now show that this common commutant in Orth(E) is a vector sublattice of Orth(E). In view of what we have already established, we may suppose that consists of one positive operator S on E. We shall show that for T 1 , T 2 ∈ Orth(E), T 1 ∨ T 2 commutes with S whenever T 1 and T 2 do. Obviously, (T 1 ∨ T 2 )S = λ T 1 ∨T 2 (S) which, by part (2) of Corollary 2.7 equals (λ T 1 ∨λ T 2 )(S). Using once more from part (2) of Corollary 2.7 that left multiplications by elements of Orth(E) are orthomorphisms on ob (E), [4,Theorem 2.43 As a consequence of the assumption, this equals (ST 1 ) ∨ (ST 2 ). By a reasoning similar to the one just given, but now using part (1) of Corollary 2.7, this equals S(T 1 ∨ T 2 ). Hence ′ Orth is a vector sublattice of Orth(E).
It is clear that ′ Orth is an associative subalgebra of Orth(E) containing I and that I, which is a weak order unit of Orth(E), is also one of ′ Orth . We turn to the remaining statements when S ⊆ oc (E). Suppose that (T α ) α∈A is a net in ′ Orth , that T ∈ ob (E), and that T α o − → T in ob (E). Then certainly T ∈ Orth(E). Using that ⊆ oc (E), it follows from Proposition 4.1 that T commutes with all elements of . Hence T ∈ ′ Orth , and we conclude that ′ Orth is an order closed vector sublattice of ob (E). Obviously, it is then also order closed in every regular vector sublattice of ob (E) containing it. We have thus established part (1). In Theorem 6.2, when ⊆ oc (E), then the vector lattice ′ Orth is a Dedekind complete vector lattice with the identity operator I as a weak order unit. The unbounded version of Freudenthal's spectral theorem (see [23,Theorem 40.3], for example) then shows that an arbitrary element T ∈ ′ Orth is an order limit of a sequence of linear combinations of the components of I in ′ Orth . Since the latter are precisely the order projections that commute with we see that, in this case, ′ Orth does not only contain all information about the collection of bands in E that reduce , but that it is also completely determined by this collection.
On a later occasion, we shall report more elaborately on the procedures of taking commutants and also of taking bicommutants in the context of operators on vector lattices and Banach lattices, as well as on their relations with reducing projection bands for sets of operators; see [13,14]. For the moment, we content ourselves with the general observation that the study of vector lattice subalgebras of the orthomorphisms is relevant for representation theory on vector lattices.
We shall now set out to study one particular aspect of this, namely, the equality of the adherences of vector sublattices of the orthomorphisms with respect to several of the convergence structures under consideration in this paper. Although from a representation theoretical point of view it would be natural to require that they also be associative subalgebras, this does, so far, not appear to be relevant for these issues. Such results on equal adherences can then also be obtained for associative subalgebras of the orthomorphisms on a Banach lattice, as a consequence of the fact that their norm closures in the orthomorphisms are. in fact, vector sublattices to which the previous results can be applied.
Regarding the results below that are given for vector sublattices of the orthomorphisms, we recall that, for a Dedekind complete vector lattice, several adherences coincide for subsets of the orthomorphisms. Indeed, since, for nets of orthomorphisms, unbounded order convergence coincides with strong unbounded order convergence, and since the convergence in a possible Hausdorff uo-Lebesgue topology coincides with the corresponding strong convergence, the corresponding adherences of subsets of the orthomorphisms are also equal. The same holds for sequential adherences. For reasons of brevity, we have refrained from including these 'obviously also equal' adherences in the statements.
Although our motivation leads us to study vector sublattice of the orthomorphisms, the results as we shall derive them for these are actually consequences of more general statements for vector lattices that need not even consist of operators. These are of interest in their own right. Other such results are [ in .
Before giving the proof, we mention the following fact that is easily verified. Suppose that X is a topological space that is supplied with two topologies τ 1 and τ 2 , where τ 2 is weaker than τ 1 . Then S Proof. It follows from [15, Theorem 6.1] that · e is a Banach f-subalgebra of . Being a vector sublattice of , Proposition 6.4 shows that the sets in the second line of equation (6.1) are equal. Since the convergence of a net in the order unit norm · e implies its order convergence to the same limit (and then also its convergence in τ to the same limit), we are done by an appeal to the remark preceding the proof.
The following is now clear from Theorem 6.6 and the argument in the proof of Corollary 6.5. in Orth(E); here · denotes the coinciding operator norm, order unit norm with respect to the identity operator, and regular norm on Orth(E).
We shall now continue by establishing results showing that the closures of vector sublattices (or associative subalgebras) in a possible Hausdorff uo-Lebesgue topology coincide with their adherences with respect to various convergence structures on the enveloping vector lattices (or vector lattice algebras) under consideration in this paper.
Needless to say, under appropriate conditions, 'topological' results as obtained above may apply at the same time as 'adherence' results to be obtained below. For reasons of brevity, we have refrained from formulating such 'combined' results.
Let us also notice at this point that the results below imply that the adherences of vector sublattices that occur in the statements are closed with respect to the pertinent convergence structures. Indeed, these adherences are set maps that map vector sublattices to vector sublattices. When they agree on vector sublattices with the topological closure operator that is supplied by the Hausdorff uo-Lebesgue topology, then they, too, are idempotent. For example, the unbounded order adherence of the vector sublattice F in Proposition 6.8, is unbounded order closed. For reasons of brevity, we have refrained from including such consequences in the results.
We start by considering two cases where the enveloping vector lattices have weak order units.
On combining Theorem 2.1, Proposition 6.8, and [15, Proposition 6.5], the following is easily obtained. We recall that a subset of a vector lattice is said to be an order basis when the band that it generates is the whole vector lattice. We continue by considering cases where the enveloping vector lattice (or vector lattice algebra) has a strong order unit.
It is known that the o-adherence of a vector sublattice of a Dedekind complete Banach lattice E with a strong order unit can be a proper sublattice of its uo-adherence; see [17, Lemma 2.6] for details. When the vector sublattice contains a strong order unit of E, however, then this cannot occur, not even in general vector lattices. This is shown by the following preparatory result. Lemma 6.10. Let E be a vector lattice with a strong order unit. Suppose that F is a vector sublattice of E that contains a strong order unit of E. Then a o (F) = a uo (F) and a σo (F) = a σuo (F) in E.
Proof. We prove that a o (F) = a uo (F). It is clear that a o (F) ⊆ a uo (F). For the reverse inclusion, we choose a positive strong order unit e of E such that e ∈ F. Take x ∈ a uo (F), and let (x α ) α∈A be a net in F such that x α uo − → x in E. There exists a λ ∈ ≥0 such that |x| ≤ λe. For α ∈ A, set y α := (−λe ∨ x α ) ∧ λe. Clearly, ( y α ) α ⊆ F and y α uo − → (−λe ∨ x) ∧ λe = x. Since the net ( y α ) α∈A is order bounded in E, we have that y α o − → x in E. Hence x ∈ a o (F). We conclude that a uo (F) ⊆ a o (F).
The proof for the sequential adherences is a special case of the above one.
Remark 6.11. For comparison, we recall that, for a regular vector sublattice F of a vector lattice E, it is always the case that a o (F) = a uo (F) in E, and that these coinciding subsets are order closed subsets of E; see [17,Theorem 2.13]. For this to hold, no assumptions on E are necessary.
The following is immediate from Proposition 6.8 and Lemma 6.10. The following result follows from the combination of Theorem 2.1, Theorem 6.12, and [15,Proposition 6.5]. In view of [15, Proposition 6.5], the natural condition to include is that E have an at most countably infinite order basis, but it is easily verified fact that, for a Banach lattice, this property is equivalent to having a weak order unit. We now turn to closures and adherences of associative subalgebras of a class of f-algebras with strong order units. For this, we need the following preparatory result. Lemma 6.14. Let E be a Banach lattice, and let A be a subset of E. Then a σo (A) = a σo (A) in E, where A denotes the norm closure of A.
Proof. We need to prove only that a σo (A) ⊆ a σo (A). For this, we may suppose that A = . Take x ∈ a σo (A) and a sequence (x n ) ∞ n=1 in A such that x n σo − → x in E. For n ≥ 1, take an y n ∈ A such that y n − x n ≤ 2 −n . For n ≥ 1, define z n by setting z n := ∞ m=n | y m − x m |, which is meaningful since the series is absolutely convergent. It is clear that z n ↓. Since z n ≤ 2 −n+1 , we have z n ↓ 0 in E. The fact that | y n − x n | ≤ z n for n ≥ 1 then shows that | y n − x n | σo − → 0 in E. From 0 ≤ | y n − x| ≤ | y n − x n | + |x n − x| σo − → 0, we then see that y n σo − → x in E. Hence x ∈ a σo (A), as desired. Theorem 6.15. Let be a Dedekind complete unital f-algebra with the countable sup property, such that its identity element e is also a strong order unit of , and such that it is complete in the submultiplicative order unit norm · e on . Suppose that admits a (necessarily unique) Hausdorff uo-Lebesgue topology τ . Let be an associative subalgebra of such that · e contains a strong order unit of . Then τ = a σo ( ) = a o ( ) = a σuo ( ) = a uo ( ) = · e τ = a σo ( · e ) = a o ( · e ) = a σuo ( · e ) = a uo ( · e ) (6.2) in .
Proof. We know from [15, Theorem 6.1] that · e is a Banach f-subalgebra of . Then Theorem 6.12 shows that all equalities in the second line of equation (6.2) hold. Furthermore, it is obvious that and that a σo ( ) ⊆ a σuo ( ) ⊆ τ .
Using that a σo ( ) = a σo ( · e ) by Lemma 6.14 and that-see the proof of Theorem 6.6-we also know that τ = · e τ , it then follows that all sets in equation (6.2) are equal.
The following is now clear from Theorem 2.1, Theorem 6.15, and [15, Proposition 6.5]. in Orth(E); here · denotes the coinciding operator norm, order unit norm with respect to the identity operator, and regular norm on Orth(E).
Acknowledgements. During this research, the first author was supported by a grant of China Scholarship Council (CSC). The authors thank Ben de Pagter for pointing out that a (not necessarily unital) positive algebra homomorphism between two unital Archimedean f-algebras is automatically a vector lattice homomorphism, which has led to Proposition 2.6, Corollary 2.7, and Theorem 6.2 being in a stronger form than in the original manuscript. They also thank the anonymous referee for their careful reading of the manuscript and the many detailed suggestions for improvement of the presentation. | 16,067.8 | 2020-11-07T00:00:00.000 | [
"Mathematics"
] |
Impaired resolution of blood transcriptomes through tuberculosis treatment with diabetes comorbidity
Abstract Background People with diabetes are more likely to develop tuberculosis (TB) and to have poor TB‐treatment outcomes than those without. We previously showed that blood transcriptomes in people with TB‐diabetes (TB‐DM) co‐morbidity have excessive inflammatory and reduced interferon responses at diagnosis. It is unknown whether this persists through treatment and contributes to the adverse outcomes. Methods Pulmonary TB patients recruited in South Africa, Indonesia and Romania were classified as having TB‐DM, TB with prediabetes, TB‐related hyperglycaemia or TB‐only, based on glycated haemoglobin concentration at TB diagnosis and after 6 months of TB treatment. Gene expression in blood at diagnosis and intervals throughout treatment was measured by unbiased RNA‐Seq and targeted Multiplex Ligation‐dependent Probe Amplification. Transcriptomic data were analysed by longitudinal mixed‐model regression to identify whether genes were differentially expressed between clinical groups through time. Predictive models of TB‐treatment response across groups were developed and cross‐tested. Results Gene expression differed between TB and TB‐DM patients at diagnosis and was modulated by TB treatment in all clinical groups but to different extents, such that differences remained in TB‐DM relative to TB‐only throughout. Expression of some genes increased through TB treatment, whereas others decreased: some were persistently more highly expressed in TB‐DM and others in TB‐only patients. Genes involved in innate immune responses, anti‐microbial immunity and inflammation were significantly upregulated in people with TB‐DM throughout treatment. The overall pattern of change was similar across clinical groups irrespective of diabetes status, permitting models predictive of TB treatment to be developed. Conclusions Exacerbated transcriptome changes in TB‐DM take longer to resolve during TB treatment, meaning they remain different from those in uncomplicated TB after treatment completion. This may indicate a prolonged inflammatory response in TB‐DM, requiring prolonged treatment or host‐directed therapy for complete cure. Development of transcriptome‐based biomarker signatures of TB‐treatment response should include people with diabetes for use across populations.
INTRODUCTION
Diabetes mellitus (DM) negatively impacts on tuberculosis (TB) control efforts by increasing the risk of Mycobacterium tuberculosis infection 1 and of progression to active TB disease three-fold. 2,3The growing prevalence of DM, particularly in countries with high burdens of TB, means DM now underlies around 15% of TB cases globally, 4 accounting for 10% of TB deaths in HIV-negative people.
Concomitant DM negatively affects TB-treatment outcomes and is associated with increased risks of delayed sputum conversion, relapse, treatment failure and death: the relative risk for each poor outcome is ∼2 to ∼5 in metaanalyses. 5,6It is unknown whether extending standard TB treatment would improve the outcome for TB-DM comorbid patients, or whether alternative treatment is required, such as host-directed therapy.Improvement of diabetes management in people with TB-DM comorbidity may also improve TB outcomes.A pragmatic clinical study 7 linked to this one showed that structured DM monitoring and intervention improved glycaemic control in TB-DM patients, but was underpowered to determine the effect on TB treatment outcome.
The worldwide DM prevalence is ∼463 million people and is estimated to rise to 700 million by 2045. 8The majority of people have type-2 DM, caused by a reduction in the response to insulin thereby reducing its ability to control target cell metabolism, which triggers an increase in insulin production leading to pancreatic damage through exhaustion, and impaired glucose tolerance.There is a spectrum from normal through to overt DM via intermediate hyperglycaemia (IH), and people with IH are more likely to develop DM in the future. 9As well as measures such as impaired fasting glucose and the impaired glucose tolerance test, the HbA1c concentration can indicate an individual's position on this spectrum. 9Infectious diseases, including TB, can cause temporary stress hyperglycaemia, which carries a higher risk of adverse events than longer-term pre-diabetes. 10TB-induced stress hyperglycaemia also makes DM diagnosis difficult: some people with apparent newly diagnosed DM at TB diagnosis no longer reach DM diagnostic criteria after TB treatment. 11B incidence and TB-DM treatment outcomes are worse in people with poorly controlled DM with higher HbA1c concentrations.12 People with TB-DM comorbidity have altered immunity compared with people with uncomplicated TB, with both innate and adaptive immune responses affected.13 In plasma, various inflammatory cytokines such as IL-1β, IL-17A, interferon (IFN)γ and TNFα are more elevated in people with TB-DM 14,15 and TB-pre-diabetes 16 than in people with uncomplicated TB.People with TB-DM have more circulating Th1 and Th17 cells and fewer Tregs.In uncomplicated TB, peripheral immune responses typically resolve to normal levels during successful TB treatment. 15n contrast, the excessive inflammatory plasma cytokine responses in TB-DM are still evident after treatment completion, 17 and dendritic cell, monocyte 18 and T cell differentiation 19 aberrations are still present at 2 months, although resolved by 6 months, indicating a delayed response to TB treatment in TB-DM patients.
1][22][23] With successful TB treatment, this transcriptomic signature is rapidly down-regulated, has largely diminished after 2 months of treatment and mostly disappears by 12 months, 21,23,24 mirroring clinical resolution and chest X-ray improvement; however, transcriptomes do not fully resolve with poor TB treatment outcome, 25 including in people with TB-DM comorbidity. 26We recently showed 27 that DM comorbidity, as well as IH, significantly affects the TB diagnosis biosignature, causing an enhanced inflammatory but reduced type 1 IFN response.This is in concordance with reduced IFNβ responses to Toll-like receptor stimulation in people with DM. 28 Differences in the changes in blood transcriptomes through TB treatment between people with TB-DM comorbidity and those with uncomplicated TB have not been described.The main aim of this study was to determine whether transcriptomic biosignatures resolve normally in people with TB-DM co-morbidity, or whether changes during TB treatment are kinetically or qualitatively different to those observed in people with uncomplicated TB alone.Additionally, differences in TB patients with pre-diabetes or IH compared with uncomplicated were identified.The characterisation of any such differences between people with TB-DM and TB-only may indicate the underlying mechanisms for worse TB treatment outcomes and may indicate promising avenues for the development of new therapies.
Patient recruitment and classification
Newly diagnosed patients with bacteriologically confirmed pulmonary TB, with or without concomitant DM, were recruited in three locations: Bandung, Indonesia (UNPAD), Cape Town, South Africa (SUN) and Craiova, Romania (UMFCV), as part of the TANDEM project. 29Exclusion criteria were multi-drug-resistant TB, HIV positivity, pregnancy, other serious co-morbidity or corticosteroid use.In South Africa, healthy controls (HCs) without TB, diabetes or hyperglycaemia were also enrolled: all had laboratory HbA1c < 5.7% were sputum smear and culture negative and had normal chest X-rays.Samples were not available from HCs in the Indonesian cohort.All participants gave written informed consent.The study was approved by the LSHTM Observational Research Ethics Committee (6449/July2013), the UNPAD Health Research Ethics Committee (377/UN6.C2.1.2/KEPK/PN/2012), the SUN Health Research Ethics Committee (N13/05/064/July2013) and the UMFCV Committee of Ethics and Academic and Scientific Deontology (94/September2013).
All TB patients underwent first line TB treatment according to WHO guidelines.Most patients diagnosed with DM received the local standard of care treatment, and the medication taken was noted.A TB-DM subgroup within the Indonesian cohort had intensive HbA1c monitoring as part of a pragmatic randomised control trial, with DM medication changed accordingly. 7Participants were classified by DM/glycaemia status at TB diagnosis and after 6 months of TB treatment (Figure 1 and Table S1).The 'TB-DM' group included patients with both pre-existing and newly diagnosed DM (Table S2).People with newly diagnosed TB-DM had laboratory HbA1c test ≥6.5% with confirmatory HbA1c test ≥6.5% or fasting blood glucose ≥7 mmol/L at TB diagnosis, 29,30 followed by a further HbA1c test ≥6.5% after 6 months of TB treatment.Infection, including TB, can drive impairments in glucose control leading to elevated HbA1c, which can then resolve when the infection is cleared.In order to distinguish between people with TB-induced, transiently elevated HbA1c from people who had pre-diabetes irrespective of their TB, we further sub-classified people who had HbA1c ≥5.7% at TB diagnosis.TB patients whose HbA1c test results were ≥5.7% and <6.5% at both TB diagnosis and at 6 months were deemed to have pre-diabetes ('TB-preDM').Patients whose HbA1c result was ≥5.7% at TB diagnosis but <5.7% at 6 months were deemed to have TB-related IH at TB diagnosis ('TBrel-IH').
Sample collection and RNA extraction
Venous blood samples (2.5 mL) were collected into PAXgene Blood RNA Tubes (PreAnalytiX) from TB patients prior to TB treatment initiation (W0) and at intervals through treatment (W2, 4,8,16,26) up to 12 months post diagnosis (W52) and stored at −80 • C prior to analysis.Total RNA was extracted using RNeasy spin columns (Qiagen) and quantified by Nanodrop (Agilent).
2.3
Unbiased whole genome RNA-Seq RNA samples that were processed for RNA-Seq analysis were quality-assessed using the LabChip GX HiSens RNA system (PerkinElmer).Total RNA samples were processed using the poly-A tail Bioscientific NEXTflex-Rapid-Directional mRNA-seq method with the Caliper SciClone to generate libraries, which were single-end sequenced using the NextSeq500 High Output kit V2 (Illumina) for 75 cycles.Data are deposited in the NCBI-GEO database, accession number GSE193978.STAR (v2.5.1b) 31 was used to align the sequence data from FASTQ files to the Human g1kv37 reference genome, and quality control was performed with FastQC. 32Downstream data analysis was performed in R. 33 HTseq-count (v0.61) was used for transcript quantification, 34 and lowly abundant transcripts were removed.Data were normalised using the DESeq2 (1.30.0) 35 R package, which included a correction for sex.
For the MaSigPro 36 analysis, due to the number of timepoints, a quadratic regression model (degrees of freedom = 2) was executed.MaSigPro uses a two-step regression-based approach which finds genes with temporal differences and also differences between groups.It is similar to a two-way ANOVA but for longitudinal RNA-Seq data, accounting for similarity between samples from the same individual.This includes an initial least-squares technique and then stepwise regression.The coefficients obtained then undergo hierarchical clustering to group the genes together that behave similarly.False discovery correction was done using the Benjamini-Hochberg method with an adjusted p value < .05deemed to be significant.From the genes that were found to be differentially expressed between clinical groups, the R package tmod 37 and its HGtest function were used for modular analysis, with all genes used as the background.Modules with an adjusted p value < .05were deemed significant.Modular activity was calculated by summing the differential expression of genes in the TB-DM group relative to the TB-only group within a module and then dividing by the number of genes within that module.Molecular degree of perturbation (MDP) analysis was performed using the R package mdp. 38The g:profiler webtool 39 was used for gene ontology and pathway analyses of gene lists.
Targeted gene expression profiling
Reverse-Transcriptase Multiplex Ligation-dependent Probe Amplification (dcRT-MLPA) was performed using the SALSA MLPA kit (MRC-Holland) as described elsewhere. 40RT primers and half-probes were designed by Leiden University Medical Centre (LUMC, Leiden, the Netherlands) 41,42 and included sequences for four housekeeping genes and 144 selected key immune-related genes to profile specific compartments of the human immune response (Table S3): (1) dcRT-MLPA data were analysed to identify differentially expressed genes (DEGs) between groups at diagnosis by the non-parametric Mann-Whitney U-test with Benjamini-Hochberg correction for multiple testing.Ingenuity pathway analysis (IPA-60467501) (QIAGEN) was used to explore interactive networks between the DEGs.MDP analysis (mdp R package), 38 partial least squaresdiscriminant analysis (PLS-DA) (mixOmics R package) 43 and Pearson correlations of gene expression (log2 FC) versus HC were performed in R version 4.0.2.Longitudinal changes in gene expression levels from diagnosis (baseline) to 6 months (Indonesian cohort) or 12 months (South African cohort) were assessed by means of linear mixed models for repeated measurements over time.Models were fitted to Log 2 -transformed measurements in the lme4 R package using the lmer function. 44Grouptime interactions were included as fixed effects and the patient identifier-time interactions were included as random effects.For the South African cohort, we forced a b-spline at 6 months, which enabled us to identify altered gene expressions during treatment (0-6 months) as well as altered gene expression after treatment (6-12 months).Time was coded as 0 for the first timepoint (diagnosis) and as a continuous variable for the time difference between the two time points.p Values were adjusted for multiple testing using the false discovery rate method of Benjamini-Hochberg. 45An adjusted p value < .05 and a log2-fold change (FC) <− .6 and > .6 were set as thresholds for the identification of DEGs.Genes that were below the detection limit in >90% of the samples per cohort were excluded from the analysis.Signatures with the best discriminatory capability were identified using logistic regression with lasso regularisation (glmnet R package). 46Leave-one-out cross validation and train-test split were used to assess the performance of the trained regression models.The classifying performance of the models were assessed by evaluating the sensitivity, specificity, receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) with 95% confidence interval (CI), and box-andwhiskers-plots representing the predicted probability for each class were used to evaluate the classifying performance of the models.Ingenuity pathway analysis was used to identify top canonical pathways, upstream regulators and causal networks in the biosignature model and MaSig-Pro RNASeq DEG lists.Therapeutic drug inhibitors were identified using the Therapeutic Target Database 47 and the GeneCards Database. 48
Study population
TB patients were recruited in South Africa, Indonesia and Romania, as a nested sub-study within the TANDEM project 29 (Figure 1 and Table S1).Sixty-eight patients had TB and DM comorbidity, of whom 49 had pre-existing diabetes and 19 were diagnosed upon study recruitment (Table S2).Forty-two TB patients without diagnosed DM were classified as having uncomplicated TB-only, whereas 53 had TBrel-IH at the time of TB diagnosis which resolved by the end of treatment.Thirty-two patients had TB-preDM, with persistent moderately elevated glycaemia.TB patients were all followed up for 18 months, with blood samples collected during TB treatment.In South Africa, HCs (n = 27) were also recruited, with blood samples collected at one time.Their RNASeq profiles have been published previously, 27 and they are included here in the targeted gene dcMLPA analysis for reference.All groups were evenly sex balanced, except for male predominance sub-study if they met the inclusion and exclusion criteria.Study participants were followed up at time points shown, with blood samples taken for gene expression analyses.The primary aim was to compare people with TB and with TB-DM comorbidity through TB treatment.
Secondarily, we analysed gene expression in TB patients with stable (TB-preDM) or transient (TBrel-IH) elevated glycaemia, as we discovered that this also impacted gene expression.
in the TB-PreDM group.Age ranges were similar across clinical groups.In the Indonesian TB-DM group, there was a highly significant decrease in HbA1c through TB treatment, which was likely due to the intensive DM follow-up in Bandung (Figure S1 and Table S1); this was not evident in South Africa or Romania.
Global longitudinal transcriptomes in TB-DM
Gene expression was determined in venous blood by RNA-Seq in a subgroup of study participants from the four TB patient clinical groups (TB-DM: n = 34; TB-PreDM: n = 10; TBrel-IH: n = 20; TB-only: n = 16; Table S1 and Figure 1).The MDP of gene expression in individual samples from patients with TB-only or TB-DM over time was calculated relative to the mean gene expression at diagnosis in people with TB-only (Figure 2).We have previously shown that the gene expression in TB and TB-DM patients at TB diagnosis is perturbed relative to HCs. 27 The overall gene expression was different between TB and TB-DM patients at diagnosis.As expected, there were changes in gene expression during TB treatment in the TB-only group, which were evident by week 2 and continued throughout treatment.The global gene expression change in the TB-DM group through treatment was of lower magnitude, indicating less impact of TB treatment: global gene expression in the TB-DM group remained different to the TB-only group at all time points (Figure 2).
The MaSigPro package performs a two-step regression analysis.This novel statistical approach identifies genes that change significantly across groups and through time.Traditional methods rely on pairwise comparisons which would be ineffective at capturing the dynamic nature of longitudinal data. 36This analysis identified 167 genes with significantly different changed expression between TB-DM and TB-only groups through TB treatment, in the combined dataset from South Africa, Indonesia and Romania.Hierarchical clustering of these genes based on similar expression patterns yielded nine clusters (Figure 3 and Table S4).Clusters which were more highly expressed in TB-DM patients throughout treatment (clusters 1, 2, 4 and 8) were enriched for genes involved in the innate immune response, IL-4 signalling, protein dimerisation and neutrophil chemotaxis, determined using the DAVID Functional Annotation Tool 49 (Table 1).Cluster 6 exhibited divergence between TB and TB-DM patients only at week 8 of treatment: this cluster was enriched for genes involved in anti-viral and IFN signalling responses.Clusters more highly expressed in TB-only patients (clusters 3, 5, 7 and 9) were smaller and enriched for alternative splice variants and immunoglobulins.
Aberrant longitudinal transcriptomes in TB patients with intermediate hyperglycaemia
Previously 27 we showed that gene expression in TBrel-IH is more similar to people with diagnosed DM and TB than with TB-only at TB diagnosis.We repeated the MaSigPro analyses separately for South Africa and Indonesia, combining patients with pre-DM and intermediate glycaemia, to determine how transcriptomes changed through TB treatment in intermediate groups (Figure S2).In South Africa, the analysis resulted in 1179 transcripts separated into three hierarchical clusters, which changed through treatment differently across clinical groups (Figure S2A and Table S5), with the combined intermediate group behaving more similarly to TB-DM.Similar results were obtained with the Indonesian cohort, with 2354 tran-F I G U R E 3 MaSigPro analysis of change in gene expression through TB treatment in blood samples from patients in all three populations combined (South Africa, Indonesia and Romania).MaSigPro identified genes that behave similarly between patient groups using hierarchical clustering.Results are shown for log-transformed normalised count for the TB-only group or TB-DM.Bars show mean ± 1 SEM.Data were filtered to remove lowly abundant transcripts prior to analysis.scripts across four hierarchical clusters behaving differently between clinical groups (Figure S2B and Table S6).
A core list of 102 genes overlapped between MaSigPro analyses for the combined cohort from Romania, South Africa and Indonesia, and from the latter two populations separately (Figure S3 and Table S7).Gene ontology and pathway analyses of this core list using the g:profiler webtool revealed functional enrichment of genes involved in the immune response, in the response to biotic stimuli, and gene products localising to intracellular vesicles (Figure S4).We hypothesised there would be differences between the TB-preDM and TBrel-IH groups, as the initially elevated HbA1c in the TBrel-IH group was directly ascribed to TB and resolved during TB treatment, whereas there was persistence of hyperglycaemia in the TB-preDM group through TB treatment; however, there was no evidence to support this postulate, as gene expression changes in both TB-preDM and TBrel-IH patients were largely similar to each other, and the similarity to TB-DM or TB-only patient groups varied by gene cluster (Figure 4).Longitudinal mixed effects model analysis of mean expression within the core gene list clusters showed highly significant changes across all four clinical groups throughout treatment, with differences between the groups in larger gene clusters (Table S8).Importantly, there was no interaction between clinical group and time, showing there was resolution of expression in all groups through treatment, albeit from different starting points and at different rates.
Modular analysis of DEGs
DEGs identified in the MaSigPro analyses were used in modular analyses to understand biological differences between clinical groups in South Africa and Indonesia.The DEGs were used as a foreground signal against all genes (Tables S9 and S10 respectively).Immune activation, monocytes and neutrophils were the most statistically significant differentially expressed modules.The most statistically significant modules were investigated further by calculating their modular activity in TB-DM relative to TBonly through time.The top module in both populations was immune activation, which was upregulated in TB-DM compared with TB-only throughout treatment.In both populations, different modules fluctuated and behaved inversely to one another between TB-DM and TB-only (Figure 5).
Impact of DM on the TB treatment response using targeted gene expression profiling
We performed targeted profiling of TB-relevant immune gene expression in an expanded cohort from South Africa with more intensive sampling, using dcRT-MLPA (TB-DM: n = 19; TB-PreDM: n = 28; TBrel-IH: n = 32; TB-only: n = 17; HC: n = 27; Table S11).At baseline, the overall gene expression perturbation, including genes from modules previously highlighted, was similar in all study groups and significantly increased compared with HCs (Figure S5A).PLS-DA displayed a clear although partial separation of all the TB groups irrespective of DM or glycaemia from HCs, suggesting distinct genes are perturbed (Figure S5; B27).
Gene expression was strongly correlated between TB-only and TB-DM, TB-preDM or TBrel-IH, but with some outlier genes which were affected by glycaemic status (Figure 6A).The number of DEGs relative to HCs was higher in the TB-DM (n = 11 DEGs), TB-preDM (n = 7 DEGs) and TBrel-IH (n = 14 DEGs) groups than TB-only (n = 3 DEGs) at TB diagnosis (Figure 6B).In these groups, the number of DEGs progressively reduced over time, indicating a resolution of expression through treatment.In particular, normalisation of expression of genes such as GNLY and GBP1 occurred by 2 weeks in the TB-only group but was delayed in TB-DM, TB-preDM and TBrel-IH.
Longitudinal S7) was summed for those genes within each MaSigPro gene cluster (Figure 3) for individual patients (log 2 scale).Only MaSigPro clusters with >3 genes in the core gene list are shown.Points show the mean ± SEM for each of the four clinical groups at each timepoint.perturbation over time (Figures 7A and S6A).Gene expression changes through treatment, identified by linear mixed models, showed some consistency across TB groups, with the South African cohort exhibiting downregulation of GBP5, GBP1 and IFITM3 (Figure 7B) and the Indonesian cohort showing downregulation of GBP5 and IFITM3 and upregulation of GNLY (Figure S6B) from diagnosis to 6 months.Importantly, the number of upregulated DEGs in response to TB treatment increased with rising glycaemia in both cohorts (South Africa: TB-only: 6 DEGs, TB-preDM: 10 DEGs, TBrel-IH: 12 DEGs, TB-DM: 14 DEGs; Indonesia: TB-only: 9 DEGs, TBrel-IH: 13 DEGs, TB-DM: 22 DEGs).We did not find any evidence that the change in the expression of the most significantly DEGs correlated with the change in the glycaemic control in the TB-DM group, tested in the Indonesian cohort (Figure S7), suggesting this is an independent measure of TB disease resolution.Notably, no DEGs were detected between 6 and 12 months in the South African cohort, except for GBP5 (p < 1e−10) in patients with TBrel-IH (Figure S8).
F I G U R E 5
Modular activity of the most significant modules in TB-DM relative to TB-only in (A) South Africa and (B) Indonesia.Modular analysis was performed between TB-DM and TB-only patients and the most statistically significant were chosen (p value < .05).Modular activity calculated by summing the expression of genes within a module and dividing by the number of genes within that module.
Ingenuity pathway analysis showed the majority of treatment-response DEGs in TB-only and TB-preDM were IFN-signalling genes (ISGs) (Figures 8C and S6C).In contrast, in TBrel-IH and TB-DM patients, although downregulation of ISGs through treatment was observed, the major change was upregulation of genes associated with adaptive immunity (T-cell subset markers, Th1-associated genes, Treg-associated genes, cytotoxicity markers).Overall, the dcRT-MLPA confirmed that although TB-associated gene profiles showed similar patterns and rate of change in TB patients and people with TB-DM, the magnitude was different.
Identification of a signature for TB treatment-response
As TB transcriptomic signatures were altered in people with DM or IH, we identified signatures with the highest classifying power to discriminate between patients at diagnosis and end of TB treatment irrespective of diabetes/glycaemia by pooling all TB patients, using logistic regression with lasso regularisation.Initially, signatures were developed in the South African and Indonesian cohorts separately (Tables 2 and S12).The classifying capability of each signature against the training (AUC range: 0.73-1.0)and validation (AUC range: .69-.92) cohorts for each clinical group was reasonably good (Figures 8A and B and S9A and B).To improve the classification performance and reduce cohort dependency, the datasets of both cohorts were pooled, and a combined two cohort 15-gene signature developed.This showed enhanced classification performance across the cohorts, with ROC analysis showing AUCs of .88 for TB-only, .96for TBrel-IH and .85 for TB-DM, with excellent classification retained in individual cohorts (Figures 8C and S9C).The kinetic profiles of six representative genes are shown in Figure S10.
Ingenuity pathway analysis revealed the top network of genes included in the model were centred on TNF/NF-κB/MAPK (Figure S11A), with the top upstream regulators including proteins such as natural cytotoxicity triggering receptor and UL16 binding protein, which are involved in NK cell mediated killing, as well as the binding partners solute carrier family 15 member 4 and TLR adaptor interacting with endolysosomal SLC15A4, which are involved in regulation of TLR7 and TLR8 signalling (Table S13).An online search revealed limited drugs currently available or under development to target these regulators.This analysis was extended to include the genes from the RNASeq global MaSigPro which were upregulated in TB, downregulated by TB treatment but consistently more highly expressed in TB-DM, that is, genes in Clusters 1, 2, 6 and 8 from Figure 3.The top network identified was an inflammatory network centred on NF-κB and MAPK, alongside TA B L E 2 Gene expression signature predicting month 6 versus diagnosis, obtained by pooling the study groups and cohorts (South Africa + Indonesia). the IFN response (Figure S11B).The top canonical pathways identified included inflammasomes and cytokine signalling.Various compounds exist which target some of the upstream regulators identified, such as Emapalumab for IFNγ, anakinra for IL-1α, or H-151 which is under development for STING1 antagonism (Table S13).
DISCUSSION
In this longitudinal analysis of blood transcriptomes, excessive gene expression perturbation previously described at TB diagnosis 14,27 continued throughout six months of TB treatment in pulmonary TB patients with diabetes co-morbidity.However, qualitatively and kinetically similar changes occurred in patients with or without diabetes, suggesting prolonged TB treatment might be sufficient to restore normal transcriptomes and potentially improve TB treatment outcomes: this would need to be tested in a clinical trial.Whilst DM itself causes altered blood transcriptomes, we have previously shown these are qualitatively different to the changes seen in TB-DM, 27 making it unlikely the remaining TB-related transcriptomic signature in TB-DM patients is caused by DM directly.TB patients with either pre-diabetes or TB-related IH also exhibited greater magnitudes of gene expression perturbation throughout treatment, similar to patients with diagnosed diabetes.There was no clear difference between the pre-diabetes and TB-related IH groups through treatment, indicating that aberrant glycaemic control in TB and early in TB treatment is sufficient to cause prolonged excessive gene expression abnormalities despite resolution of glycaemic control in the latter group.Our data published here and in our previous cross-sectional study 27 are not fully consistent with results reported by Prada-Medina et al 13 in an Indian TB-DM cohort, whereby they concluded that the differences in the TB-DM transcriptome compared with TB-only were largely driven by diabetes-related signatures: the difference may be related to the classification of participants as we have only included people with no evidence of any hyperglycaemia in our clinical group.The patient classification utilised in this study was primarily based on measures of hyperglycaemia.People with type 2 diabetes have complex metabolic and lipid disturbances, including elevated triglycerides, reduced high-density lipoproteins and increased low-density lipoproteins in blood.Such changes have also been observed in TB-DM co-morbidity, 50 with measurements of altered carbohydrate, amino acid and lipid metabolism able to clearly discriminate between TB and TB-DM. 51It is likely not the hyperglycaemia per se which has caused the alterations in the gene expression in people with TB-DM in the current study: rather this would be the overall result of the complex metabolic disturbance.Here, the overall consistency in change of gene expression through treatment, irrespective of diabetes status, enabled derivation of accurate predictive models of TB treatment response, which could be used effectively in populations with or without diabetes.
Diabetes has a negative effect on TB treatment outcomes, 5,6 for unclear reasons.One explanation could be a qualitatively different immune response in diabetes, leaving people persistently susceptible to bacterial replication and disease reactivation.An alternative explanation is that excessive inflammation and immune activation at diagnosis in TB-DM means patients require longer or, more likely, different treatment to reach the same endpoint as people with uncomplicated TB, so that they are not left susceptible to TB recurrence.Our data support the latter model, as all gene clusters differentially expressed between clinical groups exhibited similar changes, but of different magnitude.Prolonged enhanced concentrations of the pro-inflammatory cytokines IL-1β and TNFα in blood, and reduced anti-inflammatory IL-10 in sputum, have been observed in patients with TB and diabetes comorbidity, 52 in accordance with our results.Bronchial spread often persists beyond treatment initiation, with new or expanding cavities appearing on PET-CT scans 4 weeks into treatment in one-fifth of pulmonary TB patients. 53Plausibly, increased ongoing bacterial spread in patients with diabetes co-morbidity causes persistent pro-inflammatory responses: the peripheral transcriptome correlates with lung inflammatory activity in TB patients. 54Restoration of normal transcriptomes, and presumably improved lung resolution, could potentially also be achieved by co-administration of host-directed therapy alongside standard treatment.Therapy which dampens pro-inflammatory responses, such as corticosteroids or matrix metalloproteinase inhibitors, 55 would have added benefit by reducing lung damage, which often persists after microbiological cure. 56Our data suggest upstream regulators of inflammatory responses, such as inhibition of IL-1α or STING1, might provide tractable drug targets for TB-DM comorbidity.The role of specific DEG or of upstream regulators could be tested using mouse models of diabetes coupled with CRISPR technology, to determine the impact on TB disease pathology and treatment response following infection with M. tuberculosis.Antihyperglycaemic therapy, such as metformin, leads to more balanced, less inflammatory responses to M. tuberculosis, 57 and has been suggested as adjunctive therapy for TB, particularly in patients with diabetes. 58Our transcriptomic data suggest that patients with either pre-diabetes or TB-related IH would also benefit from prolonged or adjunctive host-directed therapy, in alignment with the observed worse TB treatment outcomes in people with transient hyperglycaemia. 59Improved DM management per se might also improve TB treatment outcomes, but there is currently weak evidence in this field and larger clinical trials are warranted worldwide. 60In one recent study in the UK, TB patients with well-controlled DM appear to have had normal TB treatment outcomes, 61 whereas we have previously found that many TB patients with DM have been very poorly managed across four high TB burden countries. 29he ability to monitor TB treatment and predict outcome would be beneficial for clinical management.We show that transcriptomic models can be derived from host blood which reflect TB treatment-response irrespective of glycaemia.These models worked well across geographically and ethnically diverse populations, enhancing their utility for drug development.There were however substantial differences between the three populations in Indonesia, South Africa and Romania, potentially reflecting different genetics of both host and pathogen, alongside other parameters such as social determinants including smoking and alcohol consumption, and exposure to other microbes.The best models include genes involved in IFN signalling, known to be suppressed at TB diagnosis in TB-DM patients, 27 which we found were enhanced mid-way through treatment but did eventually resolve by 6 months.People protected against TB development display balanced prostaglandin E2 and lipoxin expression in lungs, preventing TB disease progression following infection. 62Drugs which target 5-lipoxygenase restrict lung pathology and reduce bacterial replication in murine models, by lowering the type 1 IFN response 63 ; the increases through treatment in the TB-DM cohort may plausibly relate to sustained infection and accompanying inflammation.In TB-DM patients, the inflammation-related genes resolved more linearly through TB treatment, but remained elevated to the end of TB treatment, persisting until 12 months post-diagnosis in the South African cohort.In future studies, it would be important to test whether prolonged treatment with standard therapy impacts blood transcriptomes beyond the 6 month time point.Increased doses of anti-TB drugs might also lead to better treatment outcomes in TB-DM.In a complementary paper, 26 transcriptomic signatures indicative of treatment outcome have been derived that can be used in patients with either DM or IH.Together, these papers show that signatures related to poor TB outcome are distinct from the excessive and prolonged inflammation observed in TB-DM.A strength of our study is the inclusion of several timepoints through TB treatment, particularly in the South African cohort, allowing a detailed kinetic analysis of how gene expression resolves in people with TB-DM.Ahead of the study, we did not know whether gene expression resolution would be similar between people with TB-DM and TB but of a different magnitude, whether there would be delayed kinetic of expression resolution, or whether there would be qualitatively different changes in gene expression -all of these scenarios would have been in keeping with the increased poor treatment outcomes experienced by people with TB-DM.Theoretically, transcriptomic differences between people with TB-DM and people with TB-only might be caused by diabetes medications; however, we consider this to be unlikely as we have shown that people with TB-IH, who are not taking any diabetes medicines, have similar profiles to those with TB-DM who are. 27Also, the administration of the diabetes drug metformin, which is widely used by people with TB-DM in this study, has no effect on ex vivo blood transcriptomes when administered to healthy individuals. 57he strengths of our study include the detailed clinical and temporal characterisation of TB patients with or without diabetes from three cohorts from three continents, with varied genetic and social backgrounds, with the derivation of a biosignature of TB treatment which applies across all groups.The limitations of the current study include the modest sample size, and that not all samples were analysed in depth by RNASeq to quantify the entire transcriptome: our data support the undertaking of a large scale prospective clinical study of biosignatures for the prediction of delayed immune/inflammatory resolution in TB and diabetes comorbid patients.Such a study should also include HCs and people with diabetes only in all cohorts, with samples collected longitudinally from these groups: this was another limitation in the current study, as samples were only available from these groups at one time point and only in South Africa and Romania.In this study, we used HbA1c to characterise people with TBrel-IH and pre-diabetes.A more comprehensive approach would have included impaired fasting glucose and impaired oral glucose tolerance test; however, these markers overlap and all are associated with future risk of diabetes. 64People with diabetes often have a range of clinical complications, such as heart disease, kidney disease, nerve damage and problems with other infections.Our study was not powered to investigate the impact of diabetic complications on gene expression as only two TB-DM participants in South Africa and two in Indonesia had experienced significant clinical complications.A future biomarker large scale study should include assessment of the impact of these potential confounders on the TB-DM transcriptome, which was beyond the scope of the current study.It would also be valuable to follow up people with pre-diabetes and who are 'latently' infected with M. tuberculosis, to determine how these conditions drive each other long-term.
These findings further illustrate how comorbidity with diabetes affects the host response to M. tuberculosis infection and to antibiotic treatment, and how a better understanding of these interactions could be exploited to reduce poor TB treatment outcomes associated with TB and diabetes comorbidity.
A C K N O W L E D G E M E N T S
We are grateful to all the study participants for blood and data donations.We acknowledge Bahram Sanjabi, Desiree Brandenburg-Weening, and Pieter van der Vlies for assistance with the RNA-Seq, Evelien Temminck for providing technical assistance with dcRT-MLPA experiments, and J Erni Durdevic for providing statistical and machine learning advice.
F I G U R E 1
Recruitment of participants with TB into the TANDEM study, and selection of participants for inclusion in the gene expression analyses.The TANDEM study was a multi-centre, multidisciplinary project investigating various factors in TB and diabetes co-morbidity.This bioprofiling study was nested within the TANDEM Master study, in which 2185 TB patients were recruited to undergo screening for diabetes.They were initially classified into those with diabetes or without diabetes and were recruited into the bioprofiling for 15 min using Moloney Murine Leukemia Virus reverse transcriptase (Promega) with gene-specific RT-primers (Sigma-Aldrich), followed by inactivation of the enzyme by heating at 98 • C for 2 min.The left-and right-hand half probes were hybridised to the cDNA at 60 • C overnight, followed by ligation at 54 • C for 15 min using ligase-65 (MRC-Holland), and inactivation by heating at 98 • C for 5 min.Ligated probes were amplified by PCR (33 cycles at 95 • C for 30 s, 58 • C for 30 s and 72 • C for 60 s, followed by one cycle at 72 • C for 20 min).PCR products were 1:10 diluted in Highly deionised (Hi-Di) formamide (Ther-moFisher) containing the 400HD Rhodamine X (ROX) fluorophore size standard (ThermoFisher).PCR products were denatured at 95 • C for 5 min, stored immediately at 4 • C and analysed on an Applied Biosystems 3730 capillary sequencer in GeneScan mode (BaseClear).Trace data were analysed using GeneMapper software 5 (Applied Biosystems).The areas of each assigned peak (arbitrary units) were exported for analysis in R (version 3.6.3).Data were normalised to the housekeeping gene glyceraldehyde 3-phosphate dehydrogenase (GAPDH) and signals below the threshold value for noise cutoff in GeneMapper (log2 transformed peak area 7.64) were assigned the threshold value for noise cutoff.
F I G U R E 2
Molecular degree of perturbation plots representing change in global gene expression in blood relative to patients with TB-only at TB diagnosis.Gene expression was determined by RNA-Seq of whole venous blood from pulmonary TB patients from all three clinical locations with (TB-DM: n = 34) or without (TB-only: n = 18) concomitant diabetes, at TB diagnosis and during TB treatment.The bars show the median and 1.5*inter-quartile range.
F I G U R E 4
Gene expression through treatment in TB patients with pre-diabetes or TB-related intermediate hyperglycaemia, relative to TB-DM and TB-only patients.The expression of genes in the Core 102 genelist (Table
F I G U R E 6
Gene expression profiles in TBrel-IH and TB-DM are not completely normalised to healthy control profiles at the end of TB treatment.(A) Scatter plots representing Pearson correlations between expression of all genes in targeted dcRT-MLPA panel in TB patients relative to healthy controls (y-axes) versus the other study groups relative to healthy controls (x-axes), plotted as log 2 FC.Red line corresponds to line of best fit and shaded bands indicate confidence intervals.Genes regulated log 2 FC < -.6 or > .6 are annotated.(B) Differential Expression Analysis was performed on GAPDH-normalised log 2 -transformed targeted gene expression data of the South African cohort.Volcano plots representing DEGs at diagnosis and at different timepoints post TB treatment initiation of TB patients categorised based on their diabetes/glycaemia status compared with the healthy controls.The y-axis scales of all plots are harmonised per study group.p Values, -log 10 -transformed for better visualisation, are plotted against log 2 FC.Genes with p < .05 and log 2 FC < -.6 or > .6 were labelled as DEGs.
F I G U R E 7
TB treatment response in TB patients is dependent on diabetes/glycaemia status.MDP and differential expression analyses were performed on GAPDH-normalised log 2 -transformed targeted gene expression data of the South African cohort.(A) MDP analyses of the different study groups showing the impact of TB treatment on the overall gene perturbation over time.Samples of patients at diagnosis were used as baseline controls.(B) Volcano plots representing DEGs regulated during TB treatment of TB patients categorised based on their diabetes/glycaemia status.The y-axis scales of all plots are harmonised per study group.p Values, -log 10 -transformed for better visualisation, are plotted against log 2 FC.Genes with p < .05 and log 2 FC < -.6 or > .6 were labelled as DEGs.(C) IPA interactive network analyses of DEGs regulated during TB treatment.The various shapes of the nodes represent the functional classes of the gene products.Gene modules are indicated by distinctive colours.
F I G U R E 8
Identification of common host biomarker signatures associated with TB treatment response irrespective of population heterogeneity and diabetes/glycaemia severity.South African, Indonesian or pooled cohort transcriptomic datasets of TB patients independent of their diabetes/glycaemia status were used to train the models.Receiver operating characteristic (ROC) curves (sensitivity plotted against 1-specificity) and area under the curve (AUC) with 95% confidence intervals (CI) show the classifying performance of the trained models.(A) The model trained on 70% of the South African dataset was tested in the remaining 30% of the South African dataset split into the different TB study groups (left panel) and validated using the complete dataset of the Indonesian cohort split into the different TB study groups (right panel).(B) The model trained on 70% of the Indonesian dataset was tested in the remaining 30% of the Indonesian dataset split into the different TB study groups (left panel) and validated using the complete dataset of the South African cohort split into the different TB study groups (right panel).(C) The model trained on 70% of the pooled (South African and Indonesian) dataset was tested in the remaining 30% of the pooled dataset split into the different TB study groups that both cohorts have in common (left panel) and validated using the complete dataset of the South African cohort split into the different TB study groups (middle panel) or the complete dataset of the Indonesian cohort split into the different TB study groups (right panel).
N T G. W. holds patents about methods of tuberculosis diagnosis and tuberculosis biomarkers which are unrelated to the current study.No other authors have any declared conflicts of interest.F U N D I N G I N F O R M AT I O NThe research leading to these results, as part of the TANDEM Consortium, has received funding from the European Community's Seventh Framework Programme (FP7/2007-2013) under Grant Agreement no.305279.THMO also received funding from the Netherlands Organization for Scientific Research (NWO-TOP Grant Agreement No. 91214038).D ATA AVA I L A B I L I T Y S TAT E M E N TThe data that support the RNA-Seq findings of this study are openly available in NCBI-GEO at https://www.ncbi.nlm.nih.gov/geo/,accession number GSE193978.The data that support the dcRT-MLPA findings of this study are available in the supplementary material of this article.O R C I DJacqueline M. Cliff https://orcid.org/0000-0002-5653-1818RE F E R E N C E S
function Cluster Overall pattern Number of transcripts Protein coding Processed transcript Pseudo-gene Regulatory RNAs b Top non-redundant functions from DAVID a
Clusters of genes differentially expressed between TB-DM and TB-only patients in MaSigPro analysis of the combined RNA-Seq dataset from South Africa, Indonesia and Romania. | 9,919 | 2022-02-08T00:00:00.000 | [
"Medicine",
"Biology"
] |
Ensuring/insuring resilient energy system infrastructure
Natural disasters significantly impact energy systems and dependent critical infrastructures, causing severe human and economic losses in modern society. Given the increasing effects of climate change on both the frequency and the severity of extreme weather events, energy systems must adapt to cope with this new and evolving risk environment. In this perspective, we argue that re/insurers have an interest in supporting resilient infrastructure as well as the know-how to do so. Specifically, insurers can support resilient infrastructure by offering resilient-oriented insurance products, such as parametric insurance. Integrating resilience into re/insurance requires integrating existing assessment methods, including risk assessment, to develop innovative insurance products that help clients cope with climate change. Developing insurance products alongside industrial, academic, and government partners is key to making both effective and cost-attractive policies. While our argument is tailored towards energy infrastructure and climate change-related threats, resilience-based insurance would also be useful in mitigating the losses caused by other extreme and hybrid threats across interdependent critical infrastructure networks.
Introduction
Natural disasters significantly impact energy systems and dependent critical infrastructures (CIs) (European Union 2008; The White House 2013), causing severe human and economic losses in modern society (Radu 2021).Given the increasing effects of climate change on both the frequency and the severity of extreme weather events (Intergovernmental Panel on Climate Change 2021), energy systems must adapt to cope with this new and evolving risk environment (International Energy Agency 2021).Energy systems have traditionally been designed under a cost-and risk-minimisation approach in order to ensure supply can meet demand, whereby system reliability is held up as a key objective (Čepin 2011;Zio 2013) and enforced by regulations (e.g., European Union 2017; Government of South Australia 2023; North American Electric Reliability Corporation 2023).Climate change challenges this design paradigm by exposing infrastructure to events (e.g., extreme storms, coastal and river flooding, heatwaves, droughts) for which it was not designed and by introducing greater uncertainty into what conditions infrastructure may face in the future, considering also the amplification effects of interrelated human activities (European Commission 2021a).Recent unprecedented weather conditions have caused major energy system failures (Table 1), with the potential for damage only increasing as the frequency and severity increases (International Energy Agency 2021).
Coping with the impacts of climate change comes at a cost.Climate-related losses have multiplied over recent decades by an order of magnitude (United States Department of Energy 2013; Centre for Research on the Epidemiology of Disasters 2023), with the global losses in 2022 worth approximately 284 billion USD (Swiss Re 2023a, b).A large portion of these losses are uninsured (European Environment Agency 2022; Swiss Re 2023a, b), reflecting both regional re/insurance gaps as well as the difficulty of having private insurers cover the losses of extreme threats (XTs), which have the potential to inflict such vast and costly damages.Despite the undesirability of climate change risks, re/ insurers also face pressure to adapt their product offerings: insurance products are typically designed following a riskmanagement approach based on actuarial sciences, which is increasingly inappropriate given the evolving climate risk environment.
The challenges facing energy infrastructure calls to shift from a risk-based paradigm to one that is resilience-oriented (Linkov et al. 2014).The concept of resilience focuses on minimizing total loss of performance (Fig. 1) and facilitates engineering design without requiring the exact nature of the hazard/threat or probability of occurrence to be defined, unlike the risk-centric approach.While developing resilient infrastructure is a clear policy goal (The White House 2013; European Union 2022), the scale of energy system vulnerability to effects of climate change suggests that achieving energy infrastructure resilience requires going beyond studying technical and organizational factors-the focus of most resilience studies to date-and asking how increasing access to financial resources can contribute to infrastructure resilience.Understanding the links between access to finance and energy system resilience is especially important given that increasing resilience is associated with higher system design costs (Jin et al. 2019).An overarching question in the direction of a more disaster-resilient society then arises: how to cover the cost of increased system resilience?
In our perspective, re/insurers are part of the solution.Here, we argue that re/insurers have business interests in supporting resilient infrastructure as well as the knowhow to do so.Specifically, we ask: (1) why resiliencebased insurance would support more effective insurance policies for energy infrastructure; and (2) how to integrate resilience assessment into current insurance practices.Ultimately, integrating resilience into insurance practices requires adopting multi-disciplinary assessment methods, developing new types of insurance products, and joint efforts from industrial, academic, and governmental partners to ensure that the proposed products are effective and complementary to other resilience-enhancing measures.While some calls have previously been issued for re/insurers to contribute to systems resilience (Swiss Re 2019;Radu 2021), our perspective is novel in that it provides an actionable framework towards developing new resilience-oriented insurance products.We also provide a comprehensive overview of insurance practices in the energy sector, which has so far been lacking but is essential for aligning industrial, government, and research efforts towards energy system resilience.
The remainder of the perspective is presented as follows.We first introduce current insurance practices, trends, and challenges in Section 2, before arguing why resilience-based insurance would support more effective insurance policies in Section 3. In Section 4, we introduce how insurers could develop resilience-based policies.Section 5 concludes.
A B
Governments have an interest in supporting energy systems since the systems are critical in supporting a functioning society (Rinaldi 2001).This position differs from that of private insurers, who are profit-motivated and unable or unwilling to carry the costs of major disasters.Overall, the combination of self-insurance, mutual insurance pools, and governmental assistance has historically left little room for private insurers in the energy sector.
The insurance policies that are available in the energy sector are often targeted towards site-specific packages, such as for production facilities, transformer stations, and associated buildings, like offices.These sites are insurable because hazards can be defined for a single site (they constitute "spot risks") and insurance packages can include standard products, like business interruption, property damage, and construction risk (Personal communication 2023;Munich Re 2023).Insurers also offer technology-specific packages, such as for hydropower dams and run-of-river facilities (Allianz 2023).Novel renewables-focused insurance packages can also be used to hedge the financial risks associated with the technologies.For example, some products cover losses due to severe resource shortages, such as a lack of wind, sun, and water for wind, solar, and hydroelectric power production, respectively (Swiss Re 2017;Munich Re 2023).
On the other hand, transmission and distribution system lines are normally excluded from insurance policies (Swiss Re 2022).Part of the reason for this is that appropriate insurance policies can be difficult to design, given that transmission systems may cross multiple jurisdictions (Frye & Emmons 2005), are difficult to characterise given their vast geographic scope (Personal communication 2023; Gangcuangco 2023), and that distribution lines suffer frequent outages (Eto et al. 2019).Some products for electricity and gas pipelines do exist (Allianz 2023;Chubb 2023) but are much less common.If coverage is provided, it may be limited in geographic scope, e.g., the length of transmission line covered or in maximum payout value (Swiss Re 2023a, b).
Unlike the rest of the energy sector, nuclear power plant producers have well-established insurance conventions.Insurance arrangements to cover the costs of a nuclear accident have been in place since the mid-twentieth century, with a mutual insurance pool created in the United Kingdom in 1956 (Faure and Vanden Borre 2008), insurance requirements introduced in the United States in 1957 through the Price-Anderson act (Nuclear Regulatory Commission 2022a), and international protocols from the International Atomic Energy Agency and the Organization for Economic Cooperation and Development signed in the 1960s to establish global payout standard (International Atomic Energy Agency 2023; Nuclear Energy Agency 2023).More recently, the Convention on Supplementary Compensation created a mutual insurance pool between nations (International Atomic Energy Agency 1998).Some nations additionally require some private insurance coverage, such as the United States and Switzerland (Nuclear Regulatory Commission 2022b; Swiss Federal Office of Energy 2023).Generally, the need to establish insurance protocols for nuclear power lay in the scale of potential damages, which far exceed that of other power generation units.We defer the interested reader to the Nuclear Energy Agency (Nuclear Energy Agency 2023) and International Atomic energy agency for more details (International Atomic Energy Agency 2023).
Challenges to pricing climate change-related risks
It is neither clear how to price risks associated with climate change-related XTs, nor what sort of role private insurers will play in mitigating the associated damages (Radu 2021).Generally, this uncertainty stems from the difficulties in designing appropriate climate change insurance products and the obstacles insurers face in designing new insurance products more generally.
The most problematic issue in developing insurance policies to mitigate climate change risks involves the difficulty in calculating the probability of occurrence.Classic risk analysis hinges upon being able to identify hazards, consequences of the hazard occurring, and the probability of occurrence in the form of a Probabilistic Risk Assessment (PRA) (Kaplan and Garrick 1981).However, the uncertainty surrounding climate change prohibits the systematic characterisation of future natural threats.Even when qualitative estimates surrounding likelihood are produced and classes of risks can be defined (Table 2), it is challenging to estimate the full, energy-related losses associated with a changing climate.Energy services are firstly valuable for the services they provide and, while multiple methods exist to estimate the current value of lost energy services (Electric Power Research Institute 1996; National Renewable Energy Laboratory 2022), estimating the value of future lost services must consider the evolution of energy technologies and the services themselves (United States Department of Energy 2013).Energy systems are highly interconnected with one another and with dependent CIs-like water distribution, communication, and transportation networks-as well as increasingly digitalised (European Commission 2022a): calculating energy system losses must therefore also consider how other infrastructures and digital systems can mitigate or worsen such losses.
Besides the issues of relying on traditional PRA methods for developing insurance products, many other factors exclude climate change-related risks from being an ideal insurable risk (Rejda and McNamara 2014).First, there is relatively little experience with climate change risks in comparison to other insurable assets that have been subject to a particular threat, e.g., cars hit by another vehicle driver.
For example, though there are many electricity interconnection systems, few in warm regions have experienced catastrophic cold snaps as the Texan power grid did in 2021 (CBS News 2021; The Texas Tribune 2021).In addition, the effects of climate change often manifest as disaster scenarios, the full costs of which insurers may struggle to carry in absence of uneconomic premiums.Unlike the ideal case, climate change is not a chance occurrence (Intergovernmental Panel on Climate Change 2021).Insuring against conscious, human-initiated events is complicated given that probabilities can be difficult to predict and the losses can be arbitrarily large.This is particularly true for "hybrid threats", when multiple human-made crises occur simultaneously, for example, in the form of energy shortages during a geopolitical crisis (Sengupta and Eddy 2022), or massive flooding in areas facing an ongoing economic crisis (The World Bank 2022; Lederer 2023).
Developing insurance products for emerging threats is a generally challenging endeavour.The absence of data, general lack of threat awareness, and risk of product price volatility all complicate the design and selling of novel insurance products (United States Department of Energy 2013; KPMG 2019).However, single events can spur rapid industry change.For example, prior to the September 11 attack on the World Trade Center, terrorism risks were often covered by private insurance (European Central Bank 2007).High-impact, low-probability events like terrorist attacks are unattractive to insurers because of the potentially massive liabilities associated with an event.We imagine it is likely that governments must continue to act as insurers of last resort to respond to such devastating emerging threats.However, as the impacts of climate change become more known and the frequency of climate change incidents rises, it is worth asking if private insurers can carry at least some of the more reasonably foreseeable losses (Radu 2021).
Why re/insurer resilience
The continued failure of energy systems in face of "extraordinary-yet-unsurprising" climate-related events (Seneviratne et al. 2012) is evidence that the dominant, probabilistic hazard-based risk-management strategy is insufficient on its own.Moving forward, we envision a role for insurers in supporting the development of more resilient energy infrastructure due to their extensive knowledge and history in managing risk.Although taking on this role as resilience coordinators would require insurers to reimagine parts of their services (Section 3.1), we can identify two main reasons for them to do so, namely, to seize a new business opportunity (Section 3.2) and to adapt their existing coverage areas to new realities (Section 3.3).Transmission and distribution lines their ensuing losses; this knowledge would be hugely meaningful in providing clients with advice on how to improve their resilience most effectively.This guidance would be particularly useful for energy infrastructure players given that information regarding failures can be difficult to obtain.The difficulty stems from three sources.First, there is relatively little experience with catastrophic energy system failures, which are high-impact, low-probability events.Modelling failures, or creating synthetic data, can also be very challenging given the complexity of energy systems and knock-on or "cascading" effects a single failure might entail (Kirschen 2002;Gjorgiev and Sansavini 2022).Second, even if the data exists, confidentiality requirements limit the ability of individual actors to retrieve the information needed to conduct their own analyses.Third, even if data can be collected, a detailed data collection process is resource intensive, particularly if data is collected from a variety of jurisdictions, where taxonomies, reporting procedures, and level of data granularity may all differ.Insurers are well stationed to address these challenges given their comprehensive database of loss-of-service data in terms of both hazard and resulting impact.Indeed, insurers have sufficient knowledge to derive the relationship between intensity of an event, like a particular weather event and the ensuing damages.
Seizing new business opportunities
The need for resilient energy systems presents new business opportunities for insurers.Climate change increases the challenge of indemnifying losses within the energy sector and, while self-insurance and governmental assistance may have previously been sufficient to cover the costs of major events, the increasing frequency and severity of climate change events signals a new opportunity for insurers to provide support.This support would be particularly welcome in developing nations, where there are: • Larger gaps in insurance coverage (Ernst & Young 2023); • Fewer possibilities for governments to act as insurers of the last resort; It is not strictly necessary for insurers to provide complete coverage to the energy sector-self-insurance and government assistance can continue to play a role-but they can potentially support a fuller coverage.
Although insurers have typically had little-to-no risk appetite for insuring energy assets, proactively supporting resilient energy infrastructure would benefit other business areas.First, more resilient energy infrastructure can reduce loss payouts to non-energy clients.Energy systems are valuable for the services they provide: loss of power reduces the ability to conduct business and manage a transportation system, and introduces health risks (Rinaldi et al. 2001;Rutter and Keirstead 2012).Second, supporting resilient infrastructure construction would create greater opportunities for insurance investments in the infrastructure sector.To cover potential loss payouts, insurers manage vast amounts of capital in investment divisions (Zurich 2019).The way this capital is managed is highly regulated but, generally, insurers invest in low-risk opportunities with steady, longterm returns.Infrastructure investments, including energy, can meet these criteria and additionally provide a natural hedge against local currency risk (United Nations Development Programme 2020).Currently, only around 2% of insurers capital stock is allocated to infrastructure investments (Shindo and Stewart 2021; United Nations Development Programme 2020), but the share varies regionally alongside investment regulations (Shindo and Stewart 2021).There are efforts to increase the favorability of infrastructure investment conditions, most notably from the European Union who aim to unlock private investment for renewable energy (European Commission 2015; European Union 2015).However, the viability of such investments depends upon the resilience of the underlying investment, e.g., against suffering construction delays and unforeseen outages.
Adapting to new risk environment
Offering resilience-oriented policies would help insurers better address the threats facing their current client base.Risks have been assessed on a probabilistic basis, where risks are defined as triplets of hazards, the probability of occurrence, and consequences of the hazard occurring (Kaplan and Garrick 1981).However, this approach is no longer entirely suitable given the evolving nature of the climate change-related risks facing energy systems.It is increasingly difficult to characterise hazards and assign probabilities of occurrence given the dynamic ways in which Earth's natural patterns are evolving and the unknown ways that energy infrastructure will perform in climate conditions for which it was not designed (Seneviratne et al. 2012;Panteli and Mancarella 2015).These challenges extend to the dependent CI, like hospitals and transportation systems, and to other threats, like pandemics and cyber-attacks.The increasingly interconnectedness of energy systems, particularly with information and communication technologies, makes assessing the damage associated with a given disturbance challenging to evaluate.
Insurers need to adapt their existing products to help clients cope with this new risk environment.Here, developing resilience-oriented policies would help insurers provide meaningful cover in the face of a deeply uncertain future (Cox 2012).Specifically, incorporating a resilience approach would allow insurers to provide coverage to the class of threats which cannot be so clearly defined as is required in traditional risk-management approaches.In practice, a resilience-centric approach would emphasise system performance in the face of a given challenge (Park et al. 2013) rather than specifically focusing on what led to the disruption.It lends naturally to a "what-if?"type of scenario-based analysis, where particular system configurations can be explored.For example, recovery strategies can be developed for a power production plant that has limited staff available, restricted outside communication, and safety-critical systems whose functions must be ensured-irrespective of if the situation was induced by a flood, fire, windstorm, etc.This type of scenario-based analysis would be particularly useful in the context of assessing hybrid threats, where the data needed to conduct probabilistic assessments is sparse-to-unavailable.
Offering resilience-based insurance products would provide a direct financial benefit to policyholders to increase their own adaptive capacities, which are crucial for navigating unforeseen challenges.Understanding how policyholders would respond to a given challenge facilitates developing customer-appropriate resilience strategies, especially in consideration to their particular financial, organisational, and technological capacities.Instead, it would allow insurers to define policies to encompass multiple dimensions of resilience and target the most effective measures for the case at hand.For example, it may be impossible to design an energy system or system to be robust against all possible events, but efficient recovery strategies may exist.The move towards customised insurance policies is also aligned with wider trends in the insurance industry (McKinsey and Company 2021) and the good practice of seeking project-specific insurance policies (Munich Re 2016).
Acknowledging that not all threats are appropriately modelled from a risk perspective would free insurers from the burden of identifying a probability to every (potentially unimaginable) loss scenario and invite new solutions for handling emerging threats.Notably, a resilience-based infrastructure approach does not exclude risk mitigation.On the contrary, the approaches are complementary to one another (Aven 2019).Risk analysis is key in supporting insurers design damage-based losses, particularly for predictable risks and potential in situations where parametric policies (Section 4.2) cannot be struck.Resilience approaches, however, are better suited to situations of deep uncertainty and that involve many stakeholders (Linkov et al. 2014;Zio 2016).Together, risk-informed and resilience-oriented policies operated in tandem would offer more comprehensive coverage to all nature of threats.
How to integrate resilience into insurance policies
Integrating energy infrastructure resilience into insurance practices requires developing appropriate products.This can be achieved by taking three key steps: (i) by incorporating existing resilience assessment methods and tools into XT pricing frameworks, (ii) by targeted product classes, and, most importantly, (iii) by working in conjunction with partners to ensure cohesive, effective, and desirable insurance projects.
Integrating multi-disciplinary methods and tools for comprehensive resilience assessments
Developing effective pricing strategies for resilient energy infrastructure in the face of climate change and other XTs requires identifying the most vulnerable energy system states.This latter goal is critical for developing resilience improvement strategies irrespective of the probability of and exact reasons for arriving in the vulnerable state.In particular, more detailed understanding of vulnerabilities facilitates more effective strategy making: for example, identifying whether investing in renovations or additional emergency preparedness training would better minimise the total loss of system performance.To identify the most vulnerable states, insurers must integrate multi-disciplinary approaches to assess the vulnerabilities of energy infrastructure exposed to XTs in a more detailed fashion and, therefore, to obtain a more comprehensive picture of climate change damages.Specifically, comprehensive resilience assessments would integrate the insights from natural, engineering, and management sciences.As a start, integrating climate and weather models (the "hazard" information) into energy infrastructure models characterising both the spatial (e.g., topological, structural) and functional (e.g., operations, connected services and endusers) exposure of the infrastructure to the spectrum of XTs can help identify the most vulnerable system configurations on a level to support practical decision-making.Incorporating global weather forecast models like ICON (COSMO 2022) and ECMWF (Owens and Hewson 2018) into natural hazard maps allows for a location-specific representation of the intensity of a given weather event (e.g., maximum wind gust speed for storms, water level for floods) at different time scales, taking into account different emission scenarios and the inherently stochastic nature of weather processes (via ensemble forecasts, as in Röösli et al. 2021).The adoption of regional climate models also enables modelling combinations of climate and weather-related drivers that can potentially lead to compound events with a significant impact on energy infrastructure assets, capturing their intrinsic interactions and temporal/spatial dependencies via bottom-up approach (e.g., as in Zscheischler et al. 2018;Culley et al. 2016).These models can then be used in combination with asset fragility curves from reliability engineering (e.g., for power systems components as in Panteli et al. 2016;Dunn et al. 2018;Fang et al. 2019;Ma et al. 2021), cascading failure simulators (e.g., Gjorgiev and Sansavini 2022;Mühlhofer et al. 2023), and stress-test methodologies and tools (e.g., Lo Sardo et al. 2019;Esposito et al. 2020;European Commission 2016) to evaluate how single or multiple asset disruptions can propagate within the energy system and across interdependent CIs.Some government agencies, like the American Federal Emergency Management Agency (2011) and the Swiss Seismological services (2023), even already provide standardised tools and data for identifying high-risk areas for a large set of natural hazards (e.g., earthquakes, floods, tsunamis, and hurricanes) and estimating associate physical, economic, and social impacts.Altogether, these tools are useful in the preparation, mitigation, response, and recovery from natural disasters.
In addition to the existing tools, academia continues to propose new methods to support emergency management of natural hazards for a wide spectrum of weather-/climaterelated XTs and for different types of CIs (Aznar-Siguan and Bresch 2019; Merz et al. 2020;National Renewable Energy Laboratory 2023).Policymakers are equally interested in developing comprehensive resilience approaches, as evidenced from established administrative divisions and issued calls for proposal for increasing societal resilience on a holistic basis (United Nations Habitat 2023; European Commission 2022b, 2022c).Including academic and government partners within the scope of building comprehensive resilience assessment plans can only support the ongoing efforts within re/insurance companies to develop resilience consulting services (Milliman 2023;Zurich 2021) and policy to close the gap in insurance coverage for developing nations (InsuResilience Global Partnership 2022; Insurance Development Forum 2023).
The approaches we describe necessarily require some metrics to be used to quantify system resilience.The use of system metrics in the study of resilience is somewhat controversial, with some authors arguing that resilience should only be defined with relation to specific threats (Haimes 2009) or that resilience is should be understood as biologically enabled adaptive capacity (Woods and Hollnagel 2006;Woods 2015), and thereby making the use of technical system metrics like "energy not served" unsuitable.
Our perspective aligns with the corpus who argue that resilience metrics are required in order to support system design and develop practical, comparable strategies for improving resilience.Resilience metrics may be defined for resilience capabilities-namely, the prediction, absorption, recovery, and adaptation (see Amini et al. 2023 and Fig. 1)-and many authors have already proposed such metrics.For instance, Panteli et al. (2017) proposed a set of four indicators to comprehensively characterize the operational and infrastructural resilience performance of an electric power system exposed to XTs, namely: (1) how fast and (2) how low system performance resilience can drop after the manifestation of a given XT, (3) how extensive is the duration of the post-XT degraded state, and (4) how promptly the system reaches its pre-event state.Although the metrics in Panteli et al. (2017) cannot address all elements of resilience-building, i.e., adaptive capacity or system operator's ability to cope with surprise, we nonetheless see the use of technically derived metrics critical instruments for conducting a cost-benefit analysis of resilience measures, particularly for investment planning activities.
Parametric insurance as a resilience-oriented insurance product
Unlike traditional insurance policies, parametric insurance indemnifies the causes of damage rather than the ensuing losses (SwissRe Corporate Solutions 2018; Marsh & McLennan 2018).Parametric insurance is characterised by (Swiss Re Corporate Solutions 2018): • A triggering event, which can be objectively deemed to have occurred; and • A payout mechanism, agreed upon prior to the triggering event and paid irrespective of the actual damages ensuing from the event.
There are three main benefits of parametric insurance in the context of fostering resilient energy infrastructure.First, because parametric insurance insures the causes of damage rather than the resulting losses, it serves as a proactive motivation for policyholders to avoid damages beyond a specific cost threshold.Second, since the payout is defined in advance, policyholders are motivated to limit the scope of damages resulting from a specific event: indeed, losses exceeding the payout threshold will not be covered.Third, parametric insurance eliminates the need to conduct a detailed loss assessment, thereby shortening the time until payout is received, and the recovery process can begin (Horton 2018).Shortening the time to payout and recovery is particularly beneficial in developing nations to avoid secondary disasters (Clyde & Co 2018), like a public health crisis caused by inoperable electric water pumps.These three elements affect all parts of the resilience curve (Fig. 1), with the former two aspects both also supporting the goal of risk minimization.
Though parametric insurance is already available to a limited extent in the energy sector (Munich Re 2023), increasing its usage requires further work on several fronts.Triggering events may be defined with existing metrics, like cyclone strength as measured by the Australian Bureau of Meteorology and earthquake severity as measured by the United States Geological survey (Swiss Re Corporate Solutions 2018); however, more complicated weather indices might be needed to reflect the variety of stresses that can lead to losses in the energy sector.For instance, the electricity grid may fail due a local problem that propagates or "cascades" to the rest of the system, or due to a widespread heatwave that reduces capacities throughout the system.These events would, respectively, require local and distributed indices, like 24-h rainfall at a specific weather station, or the average temperature increase above normal across a range of weather stations.Index acceptability would be subject to negotiation and may be difficult to define due to a lack of data availability and information reliability.The independence of index providers is also required to preserve the inherently transparent of parametric insurance policies (Horton 2018).In addition, the process of defining appropriate payout mechanisms will require learning on the part of insurers.Some of this learning can be fostered through experience ("trial-and-error") but can also be supported by methodological advancements.
In the future, other insurance policies could also benefit system resilience in similar ways to parametric insurance by targeting a fast payout and policy accessibility.We focus here on parametric insurance given its application in current practice and the uncertainty of whether new, alternative products could meet local re/insurance regulatory requirements (e.g., Solvency II requirements in the European Union; European Union 2009).One of the most valuable actions insurers could take in supporting such collaborative efforts would be to provide policyholders with clearer information on potential natural hazards using simple tools, like online dashboards for visualising location-specific climate threats and performing system-level XT impact assessment.These tools facilitate the identification and prioritisation of disaster management strategies, allowing for a timely and transparent XT-induced disaster response.Using storyline planning to foster understanding and situational awareness of climate change and other XT-related risks would also generally facilitate communication between clients, government, and academia, thereby clarifying the roles and responsibilities of each actor in a disaster situation.
A collaborative effort towards a resilience-based climate change-related risk pricing
In turn, researchers should support insurance companies and policyholders better understand how climate change-related and hybrid risks can impact energy systems and the dependent CIs, as well as the affected communities relying on them.Researchers can do so by providing indications to regulators on cost-effective risk mitigation and resilience enhancement strategies against XT, and by translating their resilience assessments into policy-relevant socio-technical impact measures.This aspect is also relevant to the envisaged paradigm shift towards parametric insurance solutions for all XTs that require guidelines for identifying and using the parameter-based set of resilience metrics and indicators.
Last, establishing a structured dialogue between all actors involved in the XT pricing decision-making process would help align their respective disaster risk-management structures and be able to respond more quickly to disasters.Private insurance could help close protection gaps in existing coverage and clarify what mechanisms exist to efficiently disperse recovery funds.In the European Union, for example, existing legal provisions only cover disaster-financing aspects to a limited extent (Radu 2021).Transferring some risk away from government by subscribing to disaster insurance policies can significantly reduce a disaster's impact on public finances and could, for example, be enforced through regulation (Radu 2021;Frye and Emmons 2005;Nuclear Regulatory Commission 2022b).However, transferring risk first requires a better alignment between insurance needs and practices and EUlevel disaster risk-management-related policies and adaptation strategies (European Commission 2021b).
Conclusions
Increasing the resilience of energy systems and dependent CIs is urgently needed to mitigate the losses caused by climate change.Despite the limited historical role of private re/insurers in the energy sector and the challenges of insuring climate-related extreme threats (XTs), re/insurers also have vested interests in promoting resilient infrastructure as means to support their business.In this context, the paper provides an overview of current practices, trends, and challenges in insuring energy system infrastructure, envisages a future role for insurers as resilience coordinators in the development of more XT-resilient energy infrastructure, and proposes an actionable framework to integrate resiliencebased policies into current insurance practice.
Re/insurers can promote resilient energy infrastructure through their policies, as they have historically promoted risk-minimisation.Practically, this could be achieved by conducting multi-disciplinary resilience assessments, by offering parametric insurance policies, and by working together with clients, academia, and government to ensure product desirability and effectiveness.As with other types of insurance, we would expect resilience-oriented insurance policies to develop as re/insurers gain practical experience designing the policies (Young et al. 2016) and in combining them with existing products, like disaster bonds (Polacek 2018).Figure 2 summarises our perspective.
Finally, many of the arguments in this perspective could equally apply to CIs in general, like information and communications technology, and other XTs, like cyber-attacks.It is likely that resilience-oriented insurance policies would also mitigate the damages in those contexts, especially given the interconnectedness of CIs and the damages associated with hybrid threats.Although private re/insurers cannot be expected to carry the full losses resulting from extreme events, they have the incentive and the know-how to help society become more resilient to such events.
Fig
Fig. 2 Graphical summary of the perspective
Table 1
(Fink et al. 2009f climate-and weather-induced energy system failures Extreme wind from winter storm Kyrill led to two million homes without power and four to seven billion Euros worth of insured losses(Fink et al. 2009).Slovenia 2014 Extreme snowfall led to power system failure in which over 100 generating stations were affected (German Federal Agency for Technical Relief 2014).Tropical cyclone Idai left an estimated one million one people without electricity service.In response to downed transmission lines between South Africa and Mozambique, South Africa implemented rolling blackouts (Nasa Earth Observatory 2019; Hill 2019).Barbados, St. Lucia 2021 Hurricane Elsa caused power outages for the entire island of Barbados and 90% of St. Lucian homes (National Hurricane Center 2022).United States 2021 Record cold resulted in power failure and water crisis due to pipes bursting in the state of Texas.Millions were left without power and roughly half the state's population experienced issues accessing clean water (CBS News 2021; The Texas Tribune 2021; Oxner 2021).Belgium, Germany, Netherlands 2021 Extreme precipitation and flooding led to damages in electricity and gas networks.At the peak of the event, over 240,000 people lost power.Over 130 km of natural gas pipelines were damaged, with full restoration occurring as late as five months after the initial disruption (Koks et al. 2022).China 2022 Typhoon Chaba results in over 230,000 people losing power (South China Morning Post 2022).
(Interfax 2023)Fire damaged electricity transmission lines and caused a nuclear power plant to be taken offline, resulting in twenty million consumers losing power amidst a heatwave and drought (Grant and Davies 2023).Kyrgyzstan 2023 Severe cold caused emergency power rationing, leading to daily blackouts two to three hours in length in conditions as cold as −52 °C(Interfax 2023).2Insuring
energy systems 2.1 Background on insurance in the energy sector
FortisBC 2022).However, energy systems continue to struggle recovering from natural disasters and climate change events (Table1).Insurers hold a uniquely deep understanding of how energy systems might fail and
Table 2
Selected climate-induced risks for the energy sector (International Energy Agency Climate Risks 2021) | 7,784 | 2023-08-21T00:00:00.000 | [
"Engineering",
"Environmental Science"
] |