papers/sc-fno/sau.json

{
  "paper_id": "sc-fno",
  "paper_title": "Sensitivity-Constrained Fourier Neural Operators (SC-FNO)",
  "D1": [
    {
      "id": "sc-fno-D1-001",
      "claim": "FNO architecture hyperparameters for all six benchmark problems: 4 Fourier layers, width=20 channels, Fourier modes=8 for each applicable dimension (t, x, y), learning rate=0.001, 500 epochs. Learnable parameter counts: 17,921 (ODE1/2), 107,897 (PDE1/2/4), 209,397 (PDE3).",
      "source": "Appendix C, Table C.7"
    },
    {
      "id": "sc-fno-D1-002",
      "claim": "ODE1 (Composite Harmonic Oscillator) discretization and parameter ranges: temporal domain t in [0,1] discretized into N=100 equal time steps; operator maps from M=10 initial steps + parameters p to the next N-M=90 steps. Parameter ranges from uniform distribution: alpha in [1,3], beta in [1,3], gamma in [0,1].",
      "source": "Section 3, Appendix B, Table B.6"
    },
    {
      "id": "sc-fno-D1-003",
      "claim": "ODE2 (Duffing Oscillator) discretization and parameter ranges: temporal domain t in [0,1] with N=100 steps, M=10. Parameters: alpha in [0.02,0.06] (stiffness), beta in [0.01,0.03] (nonlinear stiffness), gamma in [20,60] (driving amplitude), delta in [0.5,1.5] (damping), omega in [0.2,0.6] (frequency), epsilon in [0.0,0.2] (initial position), zeta in [0.0,0.2] (initial velocity).",
      "source": "Section 3, Appendix B, Table B.6"
    },
    {
      "id": "sc-fno-D1-004",
      "claim": "PDE1 (Generalized Nonlinear Damped Wave Equation) discretization and parameter ranges: temporal t in [0,1] with N=30 steps, spatial x in [0,1] with S_x=20 divisions, M=5 initial steps. Parameters: c in [0.0,0.25] (wave speed), alpha in [0.0,0.1] (damping), beta in [0.0,0.25] (stiffness), gamma in [0.0,0.25] (forcing amplitude), omega in [0.0,0.25] (forcing frequency).",
      "source": "Section 3, Appendix B, Table B.6"
    },
    {
      "id": "sc-fno-D1-005",
      "claim": "PDE2 (Forced Burgers' Equation) discretization and parameter ranges: temporal t in [0,pi] with N=30 steps, spatial x in [0,1] with S_x=40 divisions, M=5 initial steps. Parameters: alpha in [0.1,1.0] (advection), gamma in [0.025,0.25] (viscosity), delta in [0.1,0.5] (forcing amplitude), omega in [0.01,0.1] (forcing frequency). Periodic boundary conditions.",
      "source": "Section 3, Appendix B, Table B.6"
    },
    {
      "id": "sc-fno-D1-006",
      "claim": "PDE3 (Navier-Stokes vorticity-stream function) discretization and parameter ranges: spatial x,y in [0,1] with S_x=S_y=64 divisions, temporal t in [0,3], M=1 (map from t=0 to t=3s). Parameters: alpha in [pi,5pi], beta in [pi,5pi]. Reynolds number Re=1000.",
      "source": "Section 3, Appendix B, Table B.6"
    },
    {
      "id": "sc-fno-D1-007",
      "claim": "PDE4 (Allen-Cahn equation) discretization and parameter ranges: temporal t in [0,1] with N=30 steps, spatial x in [0,1] with S_x=40 divisions, M=5 initial steps. Parameters: c in [0.1,0.9] (initial condition scale), alpha in [0.01,1.0] (linear coefficient), beta in [0.01,1.0] (cubic coefficient), omega in [5.0,10.0] (initial condition frequency), epsilon in [0.01,1.0] (diffusion coefficient). Periodic boundary conditions.",
      "source": "Section 3, Appendix B, Table B.6"
    },
    {
      "id": "sc-fno-D1-008",
      "claim": "Dataset specification: training datasets generated using the differentiable numerical solver with 2x10^3 samples for ODE1/2 and PDE1/2, 1x10^3 for PDE3, 500 and 100 samples for PDE4, 100 and 500 samples for zoned PDE2. All parameters sampled from uniform distributions U(a,b). Dataset split: 70% training, 15% validation, 15% testing, with validation/test sets containing parameter values not encountered during training.",
      "source": "Section 3.1, Section 3.4, Appendix B, Table C.8"
    },
    {
      "id": "sc-fno-D1-009",
      "claim": "Batch sizes for training: 16 for ODE1/2, 4 for PDE1/2/3, 1 for PDE4 and the zoned PDE2 (82-parameter) case. Training memory for PDE1: FNO=722MB, SC-FNO=764MB (approximate 6% increase).",
      "source": "Appendix C, Table C.8; Section 3.6"
    },
    {
      "id": "sc-fno-D1-010",
      "claim": "Zoned PDE2 with high-dimensional parameter space: the spatial domain is divided into S=40 segments, each with independent advection alpha_i and forcing amplitude delta_i, plus two global parameters gamma and omega, yielding 2S+2=82 total parameters.",
      "source": "Section 3.4"
    },
    {
      "id": "sc-fno-D1-011",
      "claim": "Dataset generation hardware: a machine equipped with a V100 GPU and four Intel Xeon processors was used for all dataset generation and training experiments.",
      "source": "Appendix D.3"
    },
    {
      "id": "sc-fno-D1-012",
      "claim": "Perturbation-based generalization test: model parameters are perturbed beyond training ranges by lambda=0.4 (40%) to test extrapolation, i.e., testing on range [b, (1+lambda)*b] where b is the upper bound of the training parameter range.",
      "source": "Section 3.2, Table 1"
    }
  ],
  "D2": [
    {
      "id": "sc-fno-D2-001",
      "claim": "SC-FNO forward mapping: the model takes initial conditions u_0, spatial coordinates x, time t, and parameters p as input, and outputs the solution u across all time and space in a single execution. Formulation: u(x,t) = F_SC-FNO(u_0, x, t, p).",
      "source": "Section 2.1"
    },
    {
      "id": "sc-fno-D2-002",
      "claim": "Sensitivity loss L_s: the mean squared error between the predicted Jacobian partial_u_hat/partial_p (computed via automatic differentiation through SC-FNO) and the true Jacobian partial_u/partial_p (pre-computed from differentiable numerical solvers or analytical solutions). Formula: L_s = (1/M) * sum_{j=1}^{M} ||partial_u_hat(x_j, t_j; p)/partial_p - partial_u(x_j, t_j; p)/partial_p||^2, evaluated at M sampled spatiotemporal points.",
      "source": "Section 2.1, Eq (unnumbered)"
    },
    {
      "id": "sc-fno-D2-003",
      "claim": "SC-FNO training objective (Algorithm 2): for each batch, (1) predict state values u_hat = FNO(P), (2) compute data loss L_u = loss(u_hat, u_true), (3) compute predicted Jacobian J_hat = partial_u_hat/partial_P via automatic differentiation, (4) compute sensitivity loss L_s = loss(J_hat, partial_u_true/partial_P), (5) total loss L_total = c_1*L_u + c_2*L_s, (6) backpropagate and update.",
      "source": "Appendix A, Algorithm 2"
    },
    {
      "id": "sc-fno-D2-004",
      "claim": "SC-FNO-PINN training objective (Algorithm 3): extends Algorithm 2 by adding an optional PINN-style PDE residual loss L_eq = residual(u_hat). The total loss becomes L_total = c_1*L_u + c_2*L_s + c_3*L_eq, where L_eq enforces the governing differential equation at collocation points.",
      "source": "Appendix A, Algorithm 3; Section 2.2"
    },
    {
      "id": "sc-fno-D2-005",
      "claim": "Gradient computation Method 1 (AD solver): extends the torchdiffeq ODE-oriented framework to handle PDEs by reformulating them as d{u}/dt = RHS(x), where RHS(x) encapsulates spatial derivative terms. Uses PyTorch automatic differentiation to compute partial_u/partial_p. This method provides the true Jacobians for training SC-FNO.",
      "source": "Section 2.3"
    },
    {
      "id": "sc-fno-D2-006",
      "claim": "Gradient computation Method 2 (finite differences): approximates gradients by solving the PDE multiple times with slightly perturbed parameter values p+Delta_p and computing (u(p+Delta_p) - u(p-Delta_p))/(2*Delta_p) using a fourth-order central finite difference scheme. This is a non-intrusive method applicable to any existing numerical solver.",
      "source": "Section 2.3, Section 3.5"
    },
    {
      "id": "sc-fno-D2-007",
      "claim": "Parameter inversion method: given a trained SC-FNO surrogate and observations u_obs, recover unknown parameters p by gradient-based optimization minimizing ||F_SC-FNO(p) - u_obs||^2 with respect to p via backpropagation. The differentiable surrogate enables efficient gradient computation without repeated PDE solves.",
      "source": "Section 3.1"
    },
    {
      "id": "sc-fno-D2-008",
      "claim": "ODE operator mapping: the neural operator learns to map from the first M time steps of solution u together with parameters p to the next N-M time steps: G: u[0:M] union p -> u[M:N]. Applied to ODE1 and ODE2 with N=100, M=10.",
      "source": "Section 3"
    },
    {
      "id": "sc-fno-D2-009",
      "claim": "PDE operator mapping (time-dependent): learns mapping from first M time steps across spatial domain plus parameters to the next N-M steps: G: u[0:S_x, 0:M] union p -> u[0:S_x, M:N]. Applied to PDE1, PDE2, PDE4.",
      "source": "Section 3"
    },
    {
      "id": "sc-fno-D2-010",
      "claim": "PDE3 operator mapping (direct final-time): learns mapping from initial conditions and parameters directly to the solution at the final time step: G: u[0:S_x, 0:S_y, t=0] union p -> u[0:S_x, 0:S_y, t=3]. No time-stepping is required during forward evaluation.",
      "source": "Section 3"
    },
    {
      "id": "sc-fno-D2-011",
      "claim": "Zoned PDE2 formulation: the spatial domain is divided into S=40 segments, each with independent advection alpha_i and forcing delta_i parameters. The Burgers' equation is solved with piecewise-constant parameters across zones, plus global parameters gamma and omega, yielding 82 total learnable parameters (2S + 2).",
      "source": "Section 3.4"
    },
    {
      "id": "sc-fno-D2-012",
      "claim": "Fundamental distinction: SC-FNO directly supervises time-integrated parameter sensitivities partial_u/partial_p via L_s (computed from forward numerical models), whereas PINNs supervise spatial-temporal derivatives (partial_u/partial_x, partial_u/partial_t) through equation-based loss optimization. The sensitivity partial_u/partial_p is typically absent from PDE formulations, so PINN regularization cannot constrain it.",
      "source": "Section 2.2, Section 3.6"
    },
    {
      "id": "sc-fno-D2-013",
      "claim": "SC-FNO is operator-agnostic: the sensitivity loss L_s can be applied to any neural operator architecture. Demonstrated by adding L_s training to four architectures -- FNO, Wavelet Neural Operator (WNO), Multiwavelet Neural Operator (MWNO), and DeepONet -- with uniform improvements in sensitivity capture and inversion accuracy across all operators.",
      "source": "Section 2.1, Appendix D.1"
    }
  ],
  "D3": [
    {
      "id": "sc-fno-D3-001",
      "claim": "Four-model comparison evaluation: four FNO configurations are compared head-to-head on all benchmark problems -- (1) FNO with L_u only, (2) FNO-PINN with L_u+L_Eq, (3) SC-FNO with L_u+L_s, (4) SC-FNO-PINN with L_u+L_s+L_Eq. All models share identical neural network architectures and inputs; only the loss configuration differs. Loss weighting coefficients c_1, c_2, c_3 control relative contribution of each term. Evaluation covers four dimensions: surrogate solution quality, Jacobian accuracy, inversion accuracy, generalization robustness, and computational cost.",
      "source": "Section 3, Appendix A, Section 2.4"
    },
    {
      "id": "sc-fno-D3-002",
      "claim": "Six benchmark problems of increasing complexity: ODE1 (Composite Harmonic Oscillator, 3 params), ODE2 (Duffing Oscillator, 7 params), PDE1 (Generalized Nonlinear Damped Wave, 5 params), PDE2 (Forced Burgers', 4 params), PDE3 (Navier-Stokes vorticity-stream, 2 params), PDE4 (Allen-Cahn phase transition with bifurcation, 5 params). ODE1 has an analytical solution serving as ground truth for gradient validation.",
      "source": "Section 3, Appendix B"
    },
    {
      "id": "sc-fno-D3-003",
      "claim": "Two-stage parameter inversion experimental design: Stage 1 performs single-parameter inversion -- invert only alpha while treating all other parameters as known, using backpropagation to optimize p by minimizing ||F(p) - u_obs||^2. Stage 2 performs simultaneous multi-parameter inversion -- all parameters of the PDE are inverted jointly. This progressive protocol isolates single-parameter behavior before tackling multi-parameter inversion.",
      "source": "Section 3.1"
    },
    {
      "id": "sc-fno-D3-004",
      "claim": "Perturbation-based generalization test protocol: after training on parameters in range [a, b], the models are tested on perturbed ranges [b, (1+lambda)*b] with systematically increasing lambda (tested at lambda=0.4, i.e., 40% beyond training). This simulates concept drift where the parameter search algorithm ventures into under-sampled regions during inversion. Both solution path u(t) and Jacobian partial_u/partial_p are evaluated.",
      "source": "Section 3.2"
    },
    {
      "id": "sc-fno-D3-005",
      "claim": "Training data volume ablation protocol: models are trained with systematically decreasing sample sizes -- 2000, 1000, 500, 100 samples for PDE1 -- to assess how different loss configurations (L_u only, L_u+L_s, L_u+L_s+L_Eq) affect accuracy and generalization under limited-data scenarios. The remaining portion of the dataset serves as the test set. SC-FNO and SC-FNO-PINN are compared against FNO at each training size.",
      "source": "Section 3.3, Section 3.4"
    },
    {
      "id": "sc-fno-D3-006",
      "claim": "High-dimensional parameter scaling test: PDE2 (Burgers') is modified into a zoned formulation where the spatial domain is divided into S=40 segments with independent advection alpha_i and forcing amplitude delta_i per zone, plus global gamma and omega -- yielding 2S+2=82 total parameters. The same comparison protocol from Section 3.2 is applied at sample sizes N=100 and N=500 to test whether the sensitivity constraint maintains effectiveness in high-dimensional parameter spaces.",
      "source": "Section 3.4"
    },
    {
      "id": "sc-fno-D3-007",
      "claim": "Gradient computation validation protocol: before using AD or FD solvers to generate training data for SC-FNO, both gradient computation methods are first validated against the analytical solution of ODE1 (which has known closed-form sensitivities partial_u/partial_alpha, partial_u/partial_beta, partial_u/partial_gamma). Only after validation are the methods used to generate solution paths and Jacobians for training surrogate models. SC-FNO is trained using the validated gradient data.",
      "source": "Section 3.5, Appendix D.3"
    },
    {
      "id": "sc-fno-D3-008",
      "claim": "Cross-operator validation protocol: the sensitivity constraint L_s is applied to four distinct neural operator architectures -- Fourier Neural Operator (FNO), Wavelet Neural Operator (WNO), Multiwavelet Neural Operator (MWNO), and DeepONet -- to demonstrate operator-agnostic generality. Each operator is evaluated in its standard form and with the sensitivity constraint (SC- variant) on PDE1 and PDE2, each with 2x10^3 training samples.",
      "source": "Section 2.1, Appendix D.1"
    },
    {
      "id": "sc-fno-D3-009",
      "claim": "Dataset generation strategy: all training data is generated synthetically using the differentiable numerical solver (torchdiffeq-based). Parameters are sampled from uniform distributions U(a,b) to ensure balanced coverage across the entire range. True Jacobians partial_u/partial_p are computed and stored once -- a one-time cost per equation. Validation and test sets (15% each) are drawn from parameter values not encountered during training.",
      "source": "Section 3.1, Section 3.2, Section 3.3, Appendix B"
    },
    {
      "id": "sc-fno-D3-010",
      "claim": "Evaluation metrics protocol: two complementary metrics are reported for every experiment -- (1) R^2 (coefficient of determination) measuring explained variance, and (2) Relative L^2 error defined as ||u_hat - u||_2 / ||u||_2. Both metrics are computed for the solution path u(t) and for each individual Jacobian component partial_u/partial_p_i. For multi-parameter inversion, per-parameter R^2 and Relative L^2 are reported to allow per-parameter assessment.",
      "source": "Section 3, Section 3.1, Section 3.2"
    },
    {
      "id": "sc-fno-D3-011",
      "claim": "Inversion evaluation protocol: for parameter inversion tasks, the trained surrogate model is used as a forward simulator within a gradient-based optimization loop. The optimizer (backpropagation through the surrogate) minimizes the discrepancy between surrogate predictions and synthetic observations. Inversion quality is measured by R^2 and Relative L^2 between true and recovered parameters. Experiments use synthetic test data drawn from held-out parameter ranges.",
      "source": "Section 3.1, Appendix D.2"
    },
    {
      "id": "sc-fno-D3-012",
      "claim": "Surrogate quality evaluation under concept drift: models trained on original parameter ranges [a, b] are evaluated on test data from the same distribution AND on perturbed ranges exceeding training. The degradation between in-distribution and out-of-distribution performance quantifies each model's robustness. SC-FNO's perturbation error is compared to FNO's as a ratio to demonstrate sensitivity-constrained models' superior robustness.",
      "source": "Section 3.2"
    },
    {
      "id": "sc-fno-D3-013",
      "claim": "Training cost characterization protocol: for each model configuration (FNO, SC-FNO, FNO-PINN, SC-FNO-PINN) and each benchmark problem, the average training time per epoch is measured. This quantifies the computational overhead of the sensitivity constraint. Additionally, GPU memory consumption is compared between FNO (722MB) and SC-FNO (764MB) on PDE1 to demonstrate that the overhead is modest.",
      "source": "Section 3.6, Appendix C, Table C.8"
    },
    {
      "id": "sc-fno-D3-014",
      "claim": "Experimental control for fair comparison: FNO and SC-FNO share identical neural network architectures, inputs, and hyperparameters (4 Fourier layers, width=20, modes=8, lr=0.001, 500 epochs). Training data, data splits, and evaluation metrics are identical across all model variants. The only difference is the loss function configuration, isolating the effect of the sensitivity constraint from all other confounding factors.",
      "source": "Section 2.4, Section 3, Appendix C"
    }
  ],
  "D4": [
    {
      "id": "sc-fno-D4-001",
      "claim": "Gradient subsampling strategy during training: pre-computed true Jacobians are stored once and reused throughout training. Instead of computing gradients at all N*T spatiotemporal points, each epoch randomly selects a subset of n < N spatial points and t < T time points for gradient comparison. The subsampling varies between epochs to eventually cover the full solution space. Each minibatch requires only one forward pass before applying AD, adding minimal computational overhead.",
      "source": "Section 2.4"
    },
    {
      "id": "sc-fno-D4-002",
      "claim": "The SC-FNO pipeline (end-to-end, 3-stage): (1) pre-compute true solution paths u and Jacobians partial_u/partial_p once using differentiable numerical solvers or finite differences -- a one-time cost per equation; (2) train FNO with combined loss L_total = c_1*L_u + c_2*L_s using stored Jacobians and subsampled gradient evaluation; (3) use the trained surrogate for forward simulation, sensitivity analysis, or gradient-based parameter inversion.",
      "source": "Section 2.4, Section 3.6"
    },
    {
      "id": "sc-fno-D4-003",
      "claim": "Parameter embedding in lifting layer: parameters tau(p) are reshaped and repeated to match the problem's spatiotemporal dimensions, then concatenated with spatial coordinates and initial conditions before being fed into the FNO neural network. FNO and SC-FNO share identical architectures and inputs; they differ only in loss configuration.",
      "source": "Section 2.4"
    }
  ]
}