Add files using upload-large-folder tool

.gitattributes

@@ -178,3 +178,25 @@ fc0904622d1a9197d5f70db4aedec8e101d5b84440d49d5b4d038ce254db87d1/preprint/images
 3b16c1a1cb0da25846c83bf4e800968ae2ecd44d56416ed9d713b378cf69173d/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text
 3b16c1a1cb0da25846c83bf4e800968ae2ecd44d56416ed9d713b378cf69173d/preprint/images/Figure_6.tif filter=lfs diff=lfs merge=lfs -text
 3b16c1a1cb0da25846c83bf4e800968ae2ecd44d56416ed9d713b378cf69173d/preprint/images/Figure_2.tif filter=lfs diff=lfs merge=lfs -text
+f9ec1a4f49fef715e40c47572e436a1c5bcdece7eb718c358a88d32efaaf1ddd/preprint/images/Figure_3.tif filter=lfs diff=lfs merge=lfs -text
+f9ec1a4f49fef715e40c47572e436a1c5bcdece7eb718c358a88d32efaaf1ddd/preprint/images/Figure_5.tif filter=lfs diff=lfs merge=lfs -text
+f9ec1a4f49fef715e40c47572e436a1c5bcdece7eb718c358a88d32efaaf1ddd/preprint/images/Figure_7.tif filter=lfs diff=lfs merge=lfs -text
+f9ec1a4f49fef715e40c47572e436a1c5bcdece7eb718c358a88d32efaaf1ddd/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text
+f9ec1a4f49fef715e40c47572e436a1c5bcdece7eb718c358a88d32efaaf1ddd/preprint/images/Figure_1.tif filter=lfs diff=lfs merge=lfs -text
+f9ec1a4f49fef715e40c47572e436a1c5bcdece7eb718c358a88d32efaaf1ddd/preprint/images/Figure_6.tif filter=lfs diff=lfs merge=lfs -text
+f9ec1a4f49fef715e40c47572e436a1c5bcdece7eb718c358a88d32efaaf1ddd/preprint/images/Figure_2.tif filter=lfs diff=lfs merge=lfs -text
+af9e4c13933752b5df110fee3914a2e6592236fdccb7636eb6156bc6b5244794/preprint/images/Figure_5.tif filter=lfs diff=lfs merge=lfs -text
+af9e4c13933752b5df110fee3914a2e6592236fdccb7636eb6156bc6b5244794/preprint/images/Figure_3.tif filter=lfs diff=lfs merge=lfs -text
+af9e4c13933752b5df110fee3914a2e6592236fdccb7636eb6156bc6b5244794/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text
+af9e4c13933752b5df110fee3914a2e6592236fdccb7636eb6156bc6b5244794/preprint/images/Figure_2.tif filter=lfs diff=lfs merge=lfs -text
+af9e4c13933752b5df110fee3914a2e6592236fdccb7636eb6156bc6b5244794/preprint/images/Figure_1.tif filter=lfs diff=lfs merge=lfs -text
+53b312065a46a22c9efb5be896a397d500703899f16c61630ebab85e62a61df4/preprint/images/Figure_3.tif filter=lfs diff=lfs merge=lfs -text
+53b312065a46a22c9efb5be896a397d500703899f16c61630ebab85e62a61df4/preprint/images/Figure_5.tif filter=lfs diff=lfs merge=lfs -text
+53b312065a46a22c9efb5be896a397d500703899f16c61630ebab85e62a61df4/preprint/images/Figure_2.tif filter=lfs diff=lfs merge=lfs -text
+53b312065a46a22c9efb5be896a397d500703899f16c61630ebab85e62a61df4/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text
+53b312065a46a22c9efb5be896a397d500703899f16c61630ebab85e62a61df4/preprint/images/Figure_1.tif filter=lfs diff=lfs merge=lfs -text
+9c5f3c6c3c9cc918fbfea512fdd5d12030a7b89970df30ad941da48534a996ab/preprint/images/Figure_5.tif filter=lfs diff=lfs merge=lfs -text
+9c5f3c6c3c9cc918fbfea512fdd5d12030a7b89970df30ad941da48534a996ab/preprint/images/Figure_2.tif filter=lfs diff=lfs merge=lfs -text
+9c5f3c6c3c9cc918fbfea512fdd5d12030a7b89970df30ad941da48534a996ab/preprint/images/Figure_3.tif filter=lfs diff=lfs merge=lfs -text
+9c5f3c6c3c9cc918fbfea512fdd5d12030a7b89970df30ad941da48534a996ab/preprint/images/Figure_4.tif filter=lfs diff=lfs merge=lfs -text
+9c5f3c6c3c9cc918fbfea512fdd5d12030a7b89970df30ad941da48534a996ab/preprint/images/Figure_1.tif filter=lfs diff=lfs merge=lfs -text

03e2bbf832d9683ec5ce0abc6b992fd7d8bbafa9f7e03faa1162f30a7e8ab680/preprint/images_list.json

@@ -0,0 +1,42 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/Figure_1.png",
+        "caption": "Silent speech recognition system using strain sensors and deep learning. a, Expanded view of single crystalline silicon nanomembrane (SiNM)-based stretchable strain sensor; the thickness of transfer-printed SiNM is 300 nm. Metal interconnect comprises a thin layer of Au(250 nm)/Cr(5 nm) deposited by thermal evaporation. Both substrate and encapsulation layers are made of spin-coated polyimide double-layer (thickness: 3.4 \u00b5m). Inset: optical microscopic image of a unit cell of strain gauge. b, 3-dimensional modeling of a human face wearing four devices. Each device is integrated with two strain gauges, a vertical gauge, and a horizontal gauge. Inset: photographs of the deformation of each sensor during silent speech of vowel \u201ca\u201d; scale bars, 2 mm. c, Waveform and heatmaps of the resistance changes from eight-channel strain gauges with respect to time during the pronunciation of the word, \u201cabsolutely.\u201d d, Overall flowchart of silent speech recognition system, including strain DAQ, data preprocessing, feature extraction, and word classification. The 100 words used in this study are randomly selected from LRW-1000, which is generally considered a benchmark in word recognition.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_2.png",
+        "caption": "Characterizations of SiNM-based biaxial strain gauge. a, Magnified optical image of biaxial strain gauges comprising a horizontal gauge and a vertical gauge; scale bar, 300 \u00b5m. b,c, Finite element analysis of applied local strain to the biaxial strain gauge on an elastomer substrate with 30% stretching along x (b) and y (c) directions. d,e, Photographs of the biaxial strain gauges before and after 30% stretching test (left) and resistance change of both horizontal and vertical gauges during in vitro test at 10%, 20%, and 30% stretching (right) in x (d) and y (e) directions, showing independent sensing properties of the biaxial strain gauges where the stretching in parallel and perpendicular directions to the gauge is prone to apply dominant and minor strain, respectively; scale bars, 1 mm. Inset: enlarged photos of a biaxial strain gauge under an applied strain. f, Relative change in electrical resistance during 50000 cycles of 30% stretching along y direction under 10 mm/s. Each cycle has a start delay and end delay of 1 s. g, Comparison of sensitivity to strain between SiNM- and metal-based strain gauge through 10%, 20%, and 30% cyclic stretching test. h, Waveforms of corresponding in vivo test of both SiNM- and metal-based strain gauges during silent speech of the word \u201cwithout\u201d (h, top), and magnified plots of channels 2 (red highlight) and 7 (blue highlight) (h, bottom). i, Normalized waveforms of h in the training phase.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_3.png",
+        "caption": "Method and validation results of silent word recognition. a, Pipeline of our deep learning model architecture comprising mainly 3D convolutional layers. b, Procedure of evaluating the proposed silent speech recognition system. 100 datasets, comprising 100 words each, are randomly divided into five folds and cross-validated; 58 and 42 datasets out of 100 datasets are acquired from different subjects: A and B, respectively. c, Comparison of the recognition performance of two different classifier models, SVM and our deep learning model, as the number of trained data increases. Each accuracy rate is the average value of five independent validations where \u201cFold 5\u201d in b is fixed as a test dataset, and n datasets randomly selected from the other four folds are trained in our deep learning model. d, Word recognition rates in the number of sensor channels. Each accuracy of n channels out of eight channels is the arithmetic mean of the accuracies from all the 8Cn combinations. e, Confusion matrices of word prediction results from three different classifier models, including correlation (left), SVM (middle), and 3D convolution (right), with the average accuracy rates of 10.26%, 76.30%, and 87.53%, respectively.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_4.png",
+        "caption": "Analysis and verification of results. a, T-distributed stochastic neighbor embedding (t-SNE) of 100 words, visualizing the result of validation 1 in Fig. 3b. All 100 words are allocated in points of different colors. The denser the cluster of same-colored points is plotted, the more the model classifies them as similar data. b, t-SNE of the most confused 10 words out of 100 words. c, Two pairs of normalized waveforms of words exhibiting similar facial movement during pronunciation (\u201cFAMILY\u201d/\u201cFamilies\u201d and \u201cincrease\u201d/\u201cdegrees\u201d). d, Relevance-CAM (R-CAM) that explains our model by highlighting which region of the entire waveform is dominant to classify each word, demonstrating that the model focuses on characteristic signal parts where the variance of the resistance is large.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_5.png",
+        "caption": "Control experiment of silent word recognition using EMG. a\u2013c, Schematics (top) and photographs (bottom) of three epidermal sEMG electrodes with different electrode dimensions. The exposed contact areas of small-sized (a), medium-sized (b), and large-sized (c) electrodes are ~0.1, ~5.5, and ~22.3 mm2, respectively. Other than the exposed contact area is encapsulated with a polyimide layer. Inset in (a, bottom): magnified optical image of small-sized electrode; scale bars, 1 mm (a,b) and 1.5 mm (c). d\u2013f, Raw EMG signals of three sEMG electrodes while the subjects clench their jaw tightly with electrodes attached to the buccinators. Insets: magnified views of noise part. g, Confusion matrix of the result of silent 100 words recognition using our small-sized EMG sensor (four channels with eight electrodes on buccinators, levator anguli oris, depressor anguli oris, and anterior belly of digastric) with the recognition rate of 35%. A total of 100 datasets are acquired, 46 of which are from Subject A and 54 are from Subject B. h, t-SNE of the total 100-word classification.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    }
+]

03e2bbf832d9683ec5ce0abc6b992fd7d8bbafa9f7e03faa1162f30a7e8ab680/preprint/preprint.md

@@ -0,0 +1,162 @@
+# Abstract
+
+A wearable silent speech interface (SSI) is a promising platform that enables verbal communication without vocalization. The most widely studied methodology for SSI focuses on surface electromyography (sEMG). However, sEMG suffers from low scalability because of signal quality-related issues, including signal-to-noise ratio and interelectrode interference. Hence, in this study, we present a novel SSI by utilizing crystalline-silicon-based strain sensors combined with a 3D convolutional deep learning algorithm. Two perpendicularly placed strain gauges with minimized cell dimension (< 0.1 mm<sup><span citationid="CR2" class="CitationRef">2</span></sup>) could capture the biaxial strain information with high reliability. We attached four strain sensors near the subject’s mouths and collected strain data of unprecedently large wordsets (100 words), which our SSI can classify at a high accuracy rate (87.53%). Several analysis methods were demonstrated to verify the system’s reliability, as well as the performance comparison with another SSI using sEMG electrodes with the same dimension, which exhibited an accuracy rate of 35.00%.
+
+# Full Text
+
+The lack of clinical treatment for speech impediments caused by aphasia or dysarthria has been promoting various studies toward improving nonacoustic communication efficiency<sup>1–5</sup>. Silent speech recognition is one of the most promising approaches for addressing the above problems, in which facial movements are tracked by visual monitoring<sup>6–10</sup> or nonvisual capturing of various biosignals<sup>11–14</sup>. Visual monitoring, a well-known vision recognition, is the most direct method to map speech-related movements and has the highest spatial resolution<sup>15,16</sup>. Nevertheless, the continuous shooting of the face in a static environment is indispensable to avoid an accuracy drop by body motion or light-induced artifacts, which leads to user inconvenience for daily communication routines. Furthermore, human–machine interfaces that exploit wearable electronics<sup>1–5,12–14,17–23</sup> for biosignal recording are used in a relatively dynamic environment. Electrophysiological signals, such as electroencephalography (EEG)<sup>11,24–26</sup>, electrocorticography (ECoG)<sup>27–29</sup>, and surface electromyography (sEMG)<sup>12,30,31</sup>, have been extensively studied for SSI. Neural signals, including EEG and ECoG, contain an enormous amount of information regarding brain activity in specific local regions that are activated during speech. However, EEG suffers from signal attenuation due to the skull and scalp<sup>32</sup>, thereby impeding the differentiation of a large number of words driven by complex electrical activities<sup>33</sup>. By contrast, ECoG exhibits a much higher signal-to-noise ratio (SNR), but it has limitation in clinical use because it is an invasive approach involving craniotomy. The sEMG, which measures electrical activities from facial muscles, can be extracted noninvasively and has relatively less complexity. Nonetheless, the low spatial resolution regarding SNR<sup>34</sup> and interelectrode correlation<sup>35,36</sup> hinders its application in the classification of a larger number of words. Furthermore, external issues, including signal degradation mainly due to body wastes, such as sweat and sebum alongside skin irritation, preclude long-term monitoring in real life<sup>37</sup>.
+
+Facial strain mapping using epidermal sensors provides another prospective platform to achieve a silent speech interface (SSI) with outstanding performance. Various studies have explored the robustness of strain gauges in diverse facial movement detection applications<sup>13,38,39</sup>, such as facial expression recognition and silent speech recognition. However, the large deformation of facial skin generated during expression or speech mostly relies on stretchable organic material-based strain sensors fabricated in a bottom–up approach<sup>13,38,40</sup>. These devices can make conformal contact with the skin and endure tensile stress in severe deformation environments but suffer from their intrinsic device-to-device variation and poor long-term stability. These properties are critical drawbacks regarding deep learning-assisted classification because their repeatability is directly related to system accuracy. By contrast, inorganic materials, such as metals and semiconductors, are representative materials for fabricating strain gauges with high reliability and fast response times. The resistance of a conventional metal-based strain sensor varies according to the geometrical changes under the applied strain, resulting in a relatively low gauge factor (~2). However, for a semiconductor-based strain gauge, the piezoresistive effect is a dominant factor regarding the resistance change<sup>41–43</sup>. Under applied strain, the shift in bandgap induces carrier redistribution, thereby changing the mobility and effective mass of semiconductor materials. Because the resistance change caused by the piezoresistive effect has a few orders higher magnitude than that caused by the geometrical effect, the semiconductor-based strain gauge has an incomparable gauge factor (~100) to the metal-based strain gauge.
+
+In this study, we propose a novel SSI using a strain sensor based on single crystalline silicon with a 3D convolution deep learning algorithm to overcome the shortcomings of the existing SSI. The silicon gauge factor can be calculated using the equation:  
+$G=(\varDelta R/R)/(\varDelta L/L)=1+2\nu +\pi {\rm E}$, where ν and π are the Poisson’s ratio and piezoresistive coefficient, respectively. Boron doping with a concentration of 5 × 10<sup>18</sup> cm<sup>−3</sup> was adopted to minimize resistance change due to external temperatures<sup>44</sup> while maintaining its relatively high piezoresistive coefficient (~80% of its value)<sup>45</sup>. High Young’s modulus (Ε) of Si contributes to the fast response time as well as sensitivity according to the equation:  
+${\rm T}=\eta /{\rm E}$, where Τ is the relaxation time and η is the viscous behavior term. However, since single crystalline silicon exhibits inherent rigidity with a high Young’s modulus, stretchability must be achieved by modifying its structure into a fractal serpentine design<sup>17,19,46</sup>. Our epidermal strain sensor was fabricated in a self-standing ultrathin (<8 µm) mesh and serpentine structure without requiring an additional elastomeric layer, thereby providing enhanced air and sweat permeability<sup>47</sup> and comfort when attached. Additionally, we devised a biaxial strain sensor that can measure directions and magnitudes two-dimensionally by placing two extremely small-sized (<0.1 mm<sup>2</sup>) strain gauges in the horizontal and vertical directions, respectively. Based on a heuristic area feature study, four biaxial strain sensors were attached to the part where the skin changes the most during silent speech. Because direct electrical contact is not required for strain measurements, our devices can leverage double-sided encapsulation, which minimizes signal degradation caused by external factors. Strain data of 100 words randomly selected from Lip Reading in the Wild (LRW)<sup>48</sup>, each with 100 repetitions from two participants, were collected and used for deep learning model training. Our model with a 3D-convolution algorithm produced 87.53% recognition accuracy, which is an unprecedented high performance for this number of words compared with the existing SSIs using a strain gauge. Analysis of data measured over multiple days from two subjects suggested that our system captured each word characteristic rather than the individual user’s characteristics or precise attachment locations. We believe that this result is comparable with the state-of-the-art result of the SSI using the sEMG dry electrode, whose dimension is approximately two orders of magnitude exceeding our strain gauge<sup>12</sup>. We also fabricated an sEMG sensor with identical dimensions as our strain gauge, which exhibited 35.00% accuracy. This comparison verifies the advantage of our system’s high scalability, facilitating extended word classification.
+
+## Overview of SSI with a strain sensor
+
+Figure 1a shows the stacked structure of our stretchable sensor embedding two silicon nanomembrane (SiNM)-based strain gauges (thickness~300 nm) located perpendicular to each other in flexible polymer layers. The total thickness of the fabricated device was less than 8 µm, enabling the conformal attachment to the skin when a water-soluble tape was used as a carrier of temporary tattoo. During silent speech, muscle movements around the mouth induce skin deformation, which can be precisely monitored using perpendicularly placed strain gauges. Highly sensitive SiNM-based strain gauges and flexible polyimide film have relatively high Young’s moduli of approximately 130 and 1 GPa, respectively, making them inappropriate candidates for stretchable devices. Therefore, the whole components of our sensor are patterned into mesh and serpentine structures to achieve stretchability and long-term stability for this application<sup>17,19,46</sup>.
+
+Each part of the face skin differs in stretching degree and direction when speaking silently, depending on the targeted words. Accordingly, determining proper sensor locations significantly contributed to SSI performance. An auxiliary vision recognition experiment that extracts the area features of the face was conducted for this purpose (Supplementary Fig. 1a). Among the randomly partitioned 24 compartments around the mouth, the sections with larger areal changes during silent speech were assumed to involve more strain gradients. Relevance-weighted class activation map (R-CAM) analysis revealed significant changes in the areas just below the lower lip (Sections 1, 4, 5, and 9 in Supplementary Fig. 1b)<sup>49</sup>. Additionally, the ablation study revealed that significant differences were unobserved in the recognition accuracy between acquiring data from the half side and both sides of the face, as the facial skin moved almost symmetrically during silent speech (Supplementary Fig. 2a). Considering the ease of attachment and diversity of signals, the four sites were determined as S1(A), S2(B), S3(C), and S4(D) (Fig. 1b), matching Sections 15, 16, 20, and 24, respectively, in Supplementary Fig. 2b.
+
+Figure 1c shows that the four strain sensors, each incorporating two SiNM gauges, captured the resistance change in the time domain through eight independent channels when a word such as “ABSOLUTELY” was silently pronounced. Different resistance changes were monitored at each channel according to the varying shape of the subject’s mouth by mapping the normalized resistance changes at each channel over time into a 2 × 4 heatmap. Concatenating these matrices according to the time sequences, the targeted words can be digitized into a three-dimensional (3D) matrix containing each designated position and time information.
+
+Figure 1d shows the overall flow of the hardware and software processes of our SSI. In this study, when an enunciator silently uttered a random word out of the 100 words, strain information from the eight channels was recorded by a data acquisition (DAQ) system (Supplementary Figs. 3 and 4). Considering the positional correlation between the biaxial gauges, the 1D signal data were processed as sequential 2D images as input data. We adopted a 3D convolutional neural network (CNN) to encode spatiotemporal features from an SiNM strain gauge signal. We trained our network with five-fold cross-validation and analyzed how it makes decisions based on explainable artificial intelligence.
+
+## Hardware characterization of the biaxial strain sensor
+
+The facial skin expands and contracts in all directions based on a specific point when a person speaks. Therefore, the degree of skin extension and information on the direction are necessary for accurate tracking of facial skin movement. Here, we designed a biaxial strain sensor that independently quantified the strain in two mutually orthogonal directions by integrating a pair of strain gauges positioned in the horizontal and vertical directions, respectively (Fig. 2a).
+
+To characterize the electrical properties of the SiNM-based strain gauge, uniaxial tensile stress was applied on the x- and y-axis up to 30%, considering the elastic limit of the facial skin during silent speech<sup>50</sup>. Finite element analysis (FEA) of the strain distribution demonstrates that a horizontal gauge experiences much higher strain with 30% x-axis stretching compared to vertical gauge, and vice versa with 30% y-axis stretching (Fig. 2b–c). This result corresponds with the actual uniaxial stretching test. Supplementary Note 1 and Supplementary Table 1 detail the Piezoresistive Multiphysics model used in FEA. Figure 2d–e shows the relative resistance change of the horizontal and vertical gauges, respectively, showing a stepwise increase regarding the increment in applied strain. When a collateral force was applied to the strain gauge, it induced a dominant resistance change, whereas the orthogonal force induced a relatively small resistance change. Along with high sensitivity, reliable DAQ is important for SSI applications. To confirm our sensor repeatability, a cyclic stretching test was also conducted by attaching the device to an elastomer with a modulus comparable with that of human skin. Even after 50000 repetitions of 30% stretching, our strain sensor showed negligible change in its resistance, confirming its high reliability.
+
+A metal-based strain gauge with an identical structure to an SiNM-based strain gauge was also fabricated to check the feasibility of this application. Figure 2g shows the comparison of relative resistance changes between SiNM-based and metal-based gauges while stretching up to 30%. The result showed that the SiNM-based gauge was approximately 42.7, 28.9, and 20.8 times more sensitive for 10%, 20%, and 30% stretching, respectively, than those of the metal gauge. Figure 2h shows the captured relative resistance change of two gauges for eight channels while silently pronouncing the same word “WITHOUT”. Through its high gauge factor, the SiNM-based gauge exhibited a remarkable waveform, whereas the metal-based gauge showed almost indistinguishable changes. For a normalized waveform, which is an input form for feature extraction, the SiNM-based gauge exhibited a distinct resistance change between 0.5 and 1.5 s, whereas no conspicuous change was monitored because of the similar level of noise for metal-based gauge.
+
+## Three-dimensional CNN for SiNM strain gauge signal analysis
+
+Our goal was to classify the 100 words from the SiNM strain gauge signals with a time length of 2 s measured at 300 frames per second. To utilize both spatial and temporal information, we used a 3D CNN model for the classification task. Figure 3a illustrates the detailed architecture of the model. Our model comprised seven 3D convolution layers and three fully connected (FC) layers. We used the kernel size of (3,3,3), padding (1,1,1), and stride (1,1,1) except the Conv3 layer where we used the kernel size of (3,1,3), padding (1,0,1), and strides (2,1,2) for downsampling. For each layer, we used instance normalization and ReLU activation. The pooling layer was not used to preserve localized spatial information. We flattened the output features of the last convolution layer (Conv7), then it was connected to several FC layers for classification. We used cross-entropy loss and the Adam optimizer<sup>51</sup> to train our 3D CNN. More details are provided in Supplementary Table 2.
+
+We utilized the cross-entropy loss and Adam optimizer to train our 3D CNN. More details are provided in Supplementary Table 2.
+
+## Results of silent speech recognition
+
+We performed a word classification task with our SSI system to 100 datasets per 100 words recorded by two subjects (See Supplementary Table 3 for details). To provide an insight on the generalized performance of our proposed system to an independent dataset, we performed five-fold cross-validation tests with randomly mixed datasets. Figure 3b shows the results. The accuracy of the five-fold cross-validation test ranged from 80.1–91.55%, and the average was 87.53%. We also evaluated word classification accuracy by varying the number of trained data, and compared the results with a conventional support vector machine (SVM)-based classification model (Fig. 3c)<sup>52</sup>. Not surprisingly, the accuracy of the “Fold 5” validation set test improved as the trained data increased, from 23.70% with 10 cases to 87.50% with 80 cases. Our model showed at least 15% higher accuracy than the SVM model when the number of trained dataset is larger than or equal to 20. We also investigated performance variation depending on the number of sensors used. As shown in Fig. 3d, word accuracy improved from 49.87–87.53% as the number of sensors increased from 2 to 8. We obtained these results by averaging all the feasible combinations of horizontal and vertical channels.
+
+To evaluate our SSI performance, we compared it with the following classifier models: correlation and SVM. Figure 3e shows the confusion matrix of the recognition results for 20 datasets (Fold 3) using these classifier models. The correlation model used one target dataset as a reference, and the results were calculated using the cosine similarity of each word between the reference and other datasets. This experiment predicted words with the highest similarity scores. We repeated this operation by changing the reference dataset. Figure 3e shows the results obtained by averaging all the cases. Our proposed method’s accuracy reached 91.55% for the “FOLD 3” validation set, significantly exceeding those of the correlation and SVM (average accuracy: 10.26% and 76.30%, respectively). Supplementary Tables 4 and 5 present the accuracy per word of our SSI. Furthermore, we evaluated the performance variation of our model to unseen data, of which accuracy may drop due to the mismatch of sensor location and subject dependency. Although the unseen datasets taken from the completely different domain from the test datasets were used, however, the classification accuracy could gradually be improved if we adapted the model using a transfer learning, which increased the accuracy sharply even up to 88%. This demonstrated that our sensors extracted meaningful values even if the attached points could be slightly misplaced. Supplementary Table 6 details the accuracy of the results.
+
+## Visualization
+
+To visualize the high-dimensional features learned from deep learning models, we utilized t-distributed stochastic neighbor embedding (t-SNE)<sup>53</sup>, which is commonly used to map high-dimensional features into two- or three-dimensional planes. We visualized high-dimensional feature outputs of the 3D convolutional deep learning model in two dimensions (Fig. 4a and Supplementary Fig. 5). The t-SNE results for the 100 classes of the test dataset showed that each class was well grouped together. We selected 10 specific classes out of the 100 words and visualized them for further analysis (Fig. 4b). Observably, the words with similar pronunciations were mapped to be close to each other (“INCREASE” vs “DEGREES” and “FAMILY” vs “FAMILIES”). Supplementary Tables 4 and 5 summarize quantitative results obtained by these confusing words. The raw signal waveform of these similar words resembled each other (Fig. 4c). Therefore, it is inevitably difficult for the word-based classification model to distinguish between these similar pronounced words. Notably, our model can provide correct classification by detecting changes in the muscles around the mouth.
+
+We analyzed the characteristic of our deep learning-based classification model through R-CAM<sup>49</sup>, a method for visualizing how much each region is affected by a classification task. Figure 4d illustrates the R-CAM results to the words, “ABSOLUTELY” and “AFTERNOON”. For both words, our model focused on the part in which the S2 sensor signal (third- and fourth-row signals) showed dominant characteristic movements. Regarding the word “ABSOLUTELY,” our model focused on the downward and upward convexities of sensor S2 at the time of 0.6 s. Concerning “AFTERNOON,” similarly, our model focused on the downward convex point in both cases, which is at around 1 s for “AFTERNOON(i)” and at around 0.7 s for “AFTERNOON(ii).” The results demonstrated that our model was not overfitted to signal data but focused on characteristic signal parts where the resistance variance was large.
+
+## Comparison of word recognition performance with sEMG
+
+As shown in Fig. 5a–c, three types of epidermal sEMG electrodes with various dimensions were also fabricated to determine the dependence of electrode size to acquired signal quality. The surface area of the small-sized electrode almost resembled the unit cell of our strain gauge so that we could fairly compare the scalability of the two systems, whereas those in medium- and large-sized electrodes were comparable with the conventional epidermal sEMG electrodes for other SSIs<sup>12,31</sup>. A pair of two-channel sEMG electrodes and one commercial EMG reference electrode were attached to the buccinators and near the posterior mastoid, respectively, and the sEMG signal was obtained at a sampling frequency of 1000 Hz when the subject’s jaw was clenched. The raw sEMG signal was preprocessed with a commercial EMG module comprising three filters and an amplifier before being transmitted to a DAQ module (Supplementary Fig. 6). The calculated SNR (1.517, 5.964, and 8,378 for small-, medium-, and large-sized electrodes, respectively) increased as the electrode dimension increased because of the lowered surface impedance (see Fig. 5a–c bottom), revealing the limitation in improving the spatial resolution of sEMG data.
+
+To compare the word classification accuracy of sEMG based model with that of our stran gauges-based system, four pairs of small-sized sEMG electrodes were attached to the facial muscles, which are generally selected for SSI, including buccinators, levator anguli oris, depressor anguli oris, and the anterior belly of digastric (Supplementary Fig. 7)<sup>12</sup>. As in the case of DAQ using our strain gauge, 100 datasets of sEMG signals were obtained from the two subjects when silently speaking 100 words, followed by hardware and software signal processing (Supplementary Fig. 6). The preprocessed datasets were randomly partitioned into five folds, and each fold feature was extracted and cross-validated using the same method (see the flowchart in Fig. 3b). Figure 5d shows the confusion matrix of classification results, where the average recognition accuracy was 35.00%. Although the state-of-the-art performance with high accuracy (~92%) was demonstrated, the system electrode size was two orders of magnitude larger than that of this work<sup>12</sup>. With the sEMG waveforms from 100 words data as inputs, the feature embeddings output by the deep learning model are shown in Fig. 5e. The 2D t-SNE mapping showed that the points with the same color were scattered rather than clustered at a specific location, indicating the difficulty in learning the representation of the scattered raw data information. Supplementary Fig. 8 shows the magnified t-SNE plot with labeling of 100 words. This result, probably due to the diminished SNR, symbolized the impeding factor of sEMG for extended word recognition because more data with high spatial resolution induce a higher classification accuracy of extended wordsets.
+
+# Conclusion
+
+In summary, a single crystalline-silicon-based strain gauge with a mesh and serpentine structure could be a promising candidate for silent speech communication with high scalability. Controlled doped single crystalline silicon, having the advantages of high gauge factor and stability as an inorganic material, establishes a more accurate system for SSI through deep learning model training with high reliability and repeatability. The FEA simulation and automatic stretching test results demonstrated that two adjacent gauges positioned perpendicular to each other are suitable for measuring the two-dimensional movement of the skin. Additionally, we demonstrated that the silicon-based strain gauge provides superior sensitivity compared to the metal-based one with the same structure under a strain of 30%. Coupled with a novel 3D convolution deep learning model, we achieved a word recognition accuracy of 87.53% to 100 words with eight strain gauges, whereas eight EMG electrodes with the same dimensions as ours only yielded an accuracy of approximately 35.00%. These results suggest a new platform by scaling the number of channels of the sensor system for SSI with a high spatiotemporal resolution, thereby providing a phoneme unit recognition capability that was previously impossible with any other systems.
+
+# Method
+
+## Materials
+
+SOITEC supplied SOI wafers (300 nm Si/1000 nm SiO₂), and KAYAKU Advanced Materials supplied 495 PMMA A8. Polyamic acid solution (12.8 wt%; 80% NMP/20% aromatic hydrocarbon) was purchased from Sigma-Aldrich. Two positive photoresists used for the photolithography process, MICROPOSIT S1805 and AZ 5214E, were from DOW and MicroChemicals, respectively. A photoresist developer and AZ 300mif from MicroChemicals were used for both photoresists. All the materials for cleaning (HF solution, buffered oxide etchant 6:1, sulfuric acid, and hydrogen peroxide) were purchased from REAGENTS DUKSAN. The PDMS base and curing agent, Sylgard 184, were purchased from DOW. Cr etchant (CT-1200S) and Au etchant (AT-409LB) were purchased from JEONYOUNG. Cu etchant (CE-100) was purchased from the Transene Company.
+
+## SiNM transfer process
+
+First, the SOI wafer was deep cleaned using piranha solution (H₂SO₄:H₂O₂ = 3:1) at 100°C for 15 min and buffered oxide etchant for 5 s, followed by boron doping (high energy implantation, Axcelis) at a dose of 5e14 cm⁻², followed by rapid thermal annealing (RTA200H-SP1, NYMTECH) at 1050°C for 90 s. The above cleaning process was repeated once more after the doping process. Second, microholes with 3 µm diameter and 50 µm pitch were defined throughout the device layer of the SOI wafer via UV–lithography (MDA-400S, Midas System) and reactive ion etching (Q190620-M01, Young Hi-Tech). MICROPOSIT S1805 was used as a positive photoresist for better adhesion with an elastomer stamp due to its high surface uniformity. Hole-patterned SOI wafers were then immersed in the HF solution for 25 min to dissolve the BOX layer and to release the device layer from the handle substrate. After rinsing with DI water, the released SiNM was transferred to the elastomer stamp (PDMS base: curing agent = 4:1) with moderate pressure, and the stamp was then pressed onto a PI layer soft-baked at 110°C for 1 min. After baking at 150°C for 3 min, the transfer printing process was completed by removing the stamp and photoresist.
+
+## Biaxial strain sensor fabrication
+
+The fabrication process started with preparing the two substrates on a silicon thermal oxide wafer cleaned using a piranha solution. A 500-nm-thick PMMA was spin coated and baked at 180°C for 3 min as a sacrificial substrate to release completed devices after the whole fabrication process. Subsequently, a thin film of a PI double-layer (~ 3.4 µm) was formed as a supporting substrate by spin coating of liquid polyamic acid solution on the PMMA sacrificial layer. The PI substrate was then fully baked in a vacuum oven at 210°C for 2 h after the transfer printing process of the SiNM layer. Two gauges perpendicular to each other were defined by photolithography and RIE. For better electrical contact with metallization, SiNM-based gauges were cleaned with buffered oxide etchant to remove the native oxide layer. Thermal evaporators (KVE-T2000, Korea Vacuum Tech) were used to deposit the metal layer of Au(250 nm)/Cr(5 nm) followed by UV–lithography with AZ-5214e positive photoresist to avoid overetching and then wet etching. After spin coating and curing the additional PI double-layer (~ 3.4 µm) for the encapsulation layer, 150 nm of the Cu mask layer was deposited and patterned to define the mesh and serpentine design. The whole structure of the device was then dry etched according to the etch mask, and the Cu mask was then wet etched. Afterward, the PMMA sacrificial layer was dissolved by immersing it in an acetone bath, and the released device was transferred to the water-soluble tape. Supplementary Fig. 9 illustrates the schematics of the fabrication process. For the metal-based strain sensor, the whole sequence was the same, except that the SiNM transfer and metal wet etching processes were substituted for deposition of Au(50 nm)/Cr(5 nm) and liftoff, respectively. Step-by-step fabrication process is detailed in Supplementary Note 2.
+
+## sEMG electrode fabrication
+
+As with the strain sensor above, the fabrication process of sEMG electrodes started with a coating of the PMMA sacrificial layer and a subsequent thin film of the PI layer (1.7 µm) on a cleaned silicon thermal oxide wafer. An electrode layer of Au(160 nm)/Cr(5 nm) was deposited by thermal evaporation and then patterned via UV–lithography and wet etching. An additional layer of PI (1.7 µm) for passivation was spin coated followed by thermal deposition of a Cu(100 nm) mask layer, as mentioned before. The unmasked area was dry etched using RIE, providing a mesh and serpentine design. Through this process, the designated active area was simultaneously exposed to direct contact with the skin (Fig. 5a–c). Finally, the device was immersed in an acetone bath to remove the underlying PMMA layer, resulting in device detachment from the handle substrate and subsequent transfer to a water-soluble tape. Supplementary Fig. 10 illustrates the schematics of the fabrication process, and step-by-step fabrication process is detailed in Supplementary Note 3.
+
+## Experimental process of strain DAQ
+
+Before attaching the strain sensor, the targeted skin was cleaned with ethanol and water. A skin-safe pressure-sensitive adhesive (Derma-tac from Smooth-On) was applied to the designated position, which was selected through the preliminary study presented in Supplementary Figs. 2 and 3. Water-soluble tapes with our strain sensor transferred on were attached to the position with moderate pressure, and DI water was then gently sprayed using a dispenser for 1 min to dissolve the PVA film. Residues of water-soluble tape were carefully peeled up using a tweezer. The strain sensors were connected to the breadboard comprising voltage divider circuit components by a presoldered ACF cable and jumper wires. The voltage divider provided V<sub>in</sub> from a 3 V common supply voltage generated by a voltage output DAQ module (PXIe-6738 from NI). V<sub>in</sub> was then measured with a voltage input DAQ module (PXIe-6365 from NI) with a 300 Hz sampling frequency. The flowchart and experimental setting image of the strain DAQ system with the voltage divider circuit are shown in Supplementary Figs. 3 and 4.
+
+## Experimental process of sEMG DAQ
+
+For an unbiased comparison of the two SSIs, the DAQ of sEMG was performed following previous literature with state-of-the-art performance. A pair of sEMG electrodes on a water-soluble tape was transferred to a 3 M Tegaderm with 2 cm spacing, followed by the removal of the water-soluble tape in a temporary tattoo-like manner. After cleaning the allocated locations with ethanol and water, four pairs of electrodes were attached to the adhesive of Tegaderm (Supplementary Fig. 10). The reference electrode was attached near a posterior mastoid, which was electrically neutral from the sEMG measurement sites. All the electrodes were connected to commercial sEMG modules (PSL-iEMG2 from PhysioLab), incorporating three filters and an amplifier. The obtained sEMG signals were then carried to the voltage input DAQ module (PXIe-6365, NI) with a 1000 Hz sampling frequency. Supplementary Fig. 6 shows the details of the DAQ process.
+
+## Software environment and the SSI process
+
+The environment was based on Ubuntu 18.04. CUDA 11.2, anaconda3, and python 3.8 were installed. Adjacent location values were located sequentially to reflect the geometric characteristics in which they were correlated with each other. Signals with a video as input are X∈ℝ¹×H×W×T, where 1 is the number of videos, and T is the number of frames. H × W is the size of the frame where H is 2 with the paired horizontal and vertical axes, and W is 4 with the number of the location of strain gauge sensors. To extract the properties of resistance changes, we used min–max normalization for each signal and applied a Savitzky–Golay filter to reduce noise. The preprocessed signal data were fed into the 3D convolution-based model to consider tempospatial information.
+
+# References
+
+1. Zhou, Z. et al. Sign-to-speech translation using machine-learning-assisted stretchable sensor arrays. *Nature Electronics* **3**, 571–578 (2020).
+2. Moin, A. et al. A wearable biosensing system with in-sensor adaptive machine learning for hand gesture recognition. *Nature Electronics* **4**, 54–63 (2021).
+3. Wen, F., Zhang, Z., He, T. & Lee, C. AI enabled sign language recognition and VR space bidirectional communication using triboelectric smart glove. *Nat Commun* **12**, 5378 (2021).
+4. Lu, Y. et al. Decoding lip language using triboelectric sensors with deep learning. *Nat Commun* **13**, 1401 (2022).
+5. Zhao, J. et al. Passive and Space-Discriminative Ionic Sensors Based on Durable Nanocomposite Electrodes toward Sign Language Recognition. *ACS Nano* **11**, 8590–8599 (2017).
+6. Chung, J.S., Senior, A., Vinyals, O. & Zisserman, A. in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 3444–3453 (2017).
+7. Martinez, B., Ma, P., Petridis, S. & Pantic, M. in ICASSP 2020–2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 6319–6323 (2020).
+8. Zhang, X., Cheng, F. & Wang, S. in Proceedings of the IEEE/CVF International Conference on Computer Vision 713–722 (2019).
+9. Ma, P., Martinez, B., Petridis, S. & Pantic, M. in ICASSP 2021–2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 7608–7612 (IEEE, 2021).
+10. Ma, P., Wang, Y., Shen, J., Petridis, S. & Pantic, M. in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision 2857–2866 (2021).
+11. Anumanchipalli, G.K., Chartier, J. & Chang, E.F. Speech synthesis from neural decoding of spoken sentences. *Nature* **568**, 493–498 (2019).
+12. Wang, Y. et al. All-weather, natural silent speech recognition via machine-learning-assisted tattoo-like electronics. *npj Flexible Electronics* **5**, 1–9 (2021).
+13. Wang, Y. et al. A durable nanomesh on-skin strain gauge for natural skin motion monitoring with minimum mechanical constraints. *Science Advances* **6**, eabb7043 (2020).
+14. Wagner, C. et al. Silent speech command word recognition using stepped frequency continuous wave radar. *Scientific Reports* **12**, 1–12 (2022).
+15. Ren, S., Du, Y., Lv, J., Han, G. & He, S. in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 13325��13333 (2021).
+16. Afouras, T., Chung, J.S. & Zisserman, A. in ICASSP 2020–2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2143–2147 (IEEE, 2020).
+17. Kim, D.-H. et al. Epidermal electronics. *science* **333**, 838–843 (2011).
+18. Gao, W. et al. Fully integrated wearable sensor arrays for multiplexed in situ perspiration analysis. *Nature* **529**, 509–514 (2016).
+19. Sim, K. et al. Metal oxide semiconductor nanomembrane–based soft unnoticeable multifunctional electronics for wearable human-machine interfaces. *Science advances* **5**, eaav9653 (2019).
+20. Kwon, Y.-T. et al. All-printed nanomembrane wireless bioelectronics using a biocompatible solderable graphene for multimodal human-machine interfaces. *Nature communications* **11**, 1–11 (2020).
+21. Zhu, M. et al. Haptic-feedback smart glove as a creative human-machine interface (HMI) for virtual/augmented reality applications. *Science Advances* **6**, eaaz8693 (2020).
+22. Miyamoto, A. et al. Inflammation-free, gas-permeable, lightweight, stretchable on-skin electronics with nanomeshes. *Nature nanotechnology* **12**, 907–913 (2017).
+23. Wang, S. et al. Skin electronics from scalable fabrication of an intrinsically stretchable transistor array. *Nature* **555**, 83–88 (2018).
+24. Herff, C. et al. Brain-to-text: decoding spoken phrases from phone representations in the brain. *Frontiers in neuroscience* **9**, 217 (2015).
+25. Nguyen, C.H., Karavas, G.K. & Artemiadis, P. Inferring imagined speech using EEG signals: a new approach using Riemannian manifold features. *Journal of neural engineering* **15**, 016002 (2017).
+26. Proix, T. et al. Imagined speech can be decoded from low-and cross-frequency intracranial EEG features. *Nature communications* **13**, 1–14 (2022).
+27. Martin, S. et al. Word pair classification during imagined speech using direct brain recordings. *Scientific reports* **6**, 1–12 (2016).
+28. Angrick, M. et al. Speech synthesis from ECoG using densely connected 3D convolutional neural networks. *Journal of neural engineering* **16**, 036019 (2019).
+29. Pei, X., Barbour, D.L., Leuthardt, E.C. & Schalk, G. Decoding vowels and consonants in spoken and imagined words using electrocorticographic signals in humans. *Journal of neural engineering* **8**, 046028 (2011).
+30. Meltzner, G.S. et al. Development of sEMG sensors and algorithms for silent speech recognition. *Journal of neural engineering* **15**, 046031 (2018).
+31. Liu, H. et al. An epidermal sEMG tattoo-like patch as a new human–machine interface for patients with loss of voice. *Microsystems & nanoengineering* **6**, 1–13 (2020).
+32. Mahmood, M. et al. Fully portable and wireless universal brain–machine interfaces enabled by flexible scalp electronics and deep learning algorithm. *Nature Machine Intelligence* **1**, 412–422 (2019).
+33. Guenther, F.H. et al. A wireless brain-machine interface for real-time speech synthesis. *PloS one* **4**, e8218 (2009).
+34. Huigen, E., Peper, A. & Grimbergen, C. Investigation into the origin of the noise of surface electrodes. *Medical and biological engineering and computing* **40**, 332–338 (2002).
+35. De Luca, C.J., Kuznetsov, M., Gilmore, L.D. & Roy, S.H. Inter-electrode spacing of surface EMG sensors: reduction of crosstalk contamination during voluntary contractions. *Journal of biomechanics* **45**, 555–561 (2012).
+36. Rodriguez-Falces, J., Neyroud, D. & Place, N. Influence of inter-electrode distance, contraction type, and muscle on the relationship between the sEMG power spectrum and contraction force. *European journal of applied physiology* **115**, 627–638 (2015).
+37. Abdoli-Eramaki, M., Damecour, C., Christenson, J. & Stevenson, J. The effect of perspiration on the sEMG amplitude and power spectrum. *Journal of Electromyography and Kinesiology* **22**, 908–913 (2012).
+38. Han, S. et al. Multiscale nanowire-microfluidic hybrid strain sensors with high sensitivity and stretchability. *npj Flexible Electronics* **2**, 1–10 (2018).
+39. Ravenscroft, D., Prattis, I., Kandukuri, T., Samad, Y.A. & Occhipinti, L.G. in 2021 IEEE International Conference on Flexible and Printable Sensors and Systems (FLEPS) 1–4 (IEEE, 2021).
+40. Wang, H. et al. High-Performance Foam-Shaped Strain Sensor Based on Carbon Nanotubes and Ti3C2T x MXene for the Monitoring of Human Activities. *ACS nano* **15**, 9690–9700 (2021).
+41. Iqra, M., Anwar, F., Jan, R. & Mohammad, M.A. A flexible piezoresistive strain sensor based on laser scribed graphene oxide on polydimethylsiloxane. *Scientific reports* **12**, 1–11 (2022).
+42. Takamatsu, S. et al. Plastic-scale-model assembly of ultrathin film MEMS piezoresistive strain sensor with conventional vacuum-suction chip mounter. *Scientific Reports* **9**, 1–8 (2019).
+43. Kim, J. et al. Stretchable silicon nanoribbon electronics for skin prosthesis. *Nature communications* **5**, 1–11 (2014).
+44. Norton, P. & Brandt, J. Temperature coefficient of resistance for p-and n-type silicon. *Solid-state electronics* **21**, 969–974 (1978).
+45. Won, S.M. et al. Piezoresistive strain sensors and multiplexed arrays using assemblies of single-crystalline silicon nanoribbons on plastic substrates. *IEEE Transactions on Electron Devices* **58**, 4074–4078 (2011).
+46. Webb, R.C. et al. Ultrathin conformal devices for precise and continuous thermal characterization of human skin. *Nature materials* **12**, 938–944 (2013).
+47. Wang, Y. et al. Low-cost, µm-thick, tape-free electronic tattoo sensors with minimized motion and sweat artifacts. *npj Flexible Electronics* **2**, 1–7 (2018).
+48. Yang, S. et al. in 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019) 1–8 (IEEE, 2019).
+49. Lee, J.R., Kim, S., Park, I., Eo, T. & Hwang, D. in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 14944–14953 (2021).
+50. Silver, F.H., Siperko, L.M. & Seehra, G.P. Mechanobiology of force transduction in dermal tissue. *Skin Research and Technology* **9**, 3–23 (2003).
+51. Kingma, D.P. & Ba, J. Adam: A method for stochastic optimization. *arXiv preprint arXiv:1412.6980* (2014).
+52. Tang, Y. Deep learning using linear support vector machines. *arXiv preprint arXiv:1306.0239* (2013).
+53. Van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. *Journal of machine learning research* **9** (2008).
+
+# Supplementary Files
+
+- [Supplementary.docx](https://assets-eu.researchsquare.com/files/rs-1657508/v1/20d9df0e4408fe5dc9e8753f.docx)  
+  Dataset 1
+
+- [RealTimeSubjectA.mp4](https://assets-eu.researchsquare.com/files/rs-1657508/v1/98439e60c1a33681cee59350.mp4)  
+  Supplementary Video 1
+
+- [RealTimeSubjectA.zip](https://assets-eu.researchsquare.com/files/rs-1657508/v1/53d5a44823e67c3b3317b18e.zip)  
+  Supplementary Video 1
+
+- [RealTimeSubjectB.mp4](https://assets-eu.researchsquare.com/files/rs-1657508/v1/597c307906fa344643662c25.mp4)  
+  Supplementary Video 2
+
+- [RealTimeSubjectB.zip](https://assets-eu.researchsquare.com/files/rs-1657508/v1/765bcfa359e4bccc5dac50bb.zip)  
+  Supplementary Video 2

072b9ce21b7943bdfb95e67483287fad20b8c8258dec744bf664f859e7a97f73/preprint/images_list.json

@@ -0,0 +1,66 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/Figure_1.png",
+        "caption": "Variation in discharge from C7 in the Krycklan catchment showing summer low flow days <0.1 mm day -1 (a), and the distribution of discharge over the years with 0.1 mm day -1 thresholds (b). The jitter dots are actual discharge values for each year from which the ridgeline distribution curve was estimated in panel B. ",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_2.jpeg",
+        "caption": "Changes in DOC concentrations across 13 nested sub-catchments over 17 years in relation to summer low flows showing (a) Drought (r2 range 0.29-0.65, p<0.05) and (b) Post-drought (r2 range 0.28-0.67, p<0.05). The points represent the changes in concentration of each summer fitted with regression lines for individual catchments. The red horizontal line indicates zero change while the vertical grey line indicates average low flow days (18 days).  Note differences in y-axis scales.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_3.jpeg",
+        "caption": "Monthly variation in DOC changes relative to the long term means from 2003-2019 in the Krycklan sub-catchments showing the long-term DOC mean (black line) and the years with the high number of low flow days (above the 90th percentile) (blue line), years with the lowest number of low flow days (orange) and the years with average low flow days (grey dots) in Krycklan. The loess regression curves show the average DOC change in years with the high number of low flow days (blue) and low number of low flow days (orange) ",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_4.jpeg",
+        "caption": "The Drought and Post-drought effects of LMW DOC (a: where r2 range 0.20-0.54, p<0.05 and, b: where r2 range 0.21-0.64, p<0.05 respectively) and CN ratio (c: where r2 range 0.27-0.47, c: where r2 range 0.2-0.59, p<0.05) of 13 boreal sub-catchments stream chemistry. Data for the CN ratio for the driest summer (2006) is missing because sampling started in 2007. Additionally, for some of the larger sites (C10, C12, C14, C15), sampling stopped in 2017 hence data for 2018 and 2019 were unavailable.  ",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_5.jpeg",
+        "caption": "Drought and post-drought effects on groundwater DOC (a, b), LMW DOC (c, d), and CN ratio (e, f). Riparian wells values were obtained from averages of samples at depths of 0.1-0.65 m while values from the mire wells were obtained from averages of sampling at depths 2-2.5, which is consistent with the dominant flow paths for the two catchments (Laudon et al. 2013).",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_6.png",
+        "caption": "DOC slope responses to summer low flows, modeled using the best predictor (catchment size) during (a) Drought and (b) Post-drought. Note only significant values were used in the models  ",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_7.jpeg",
+        "caption": "Conceptual responses of DOC, LMW DOC, and CN ratio to summer low flow conditions at different time intervals representing our predictions for (a) the Drought effects during summer, and (b) the Post-drought effects after the first rewetting in relation to long-term averages normalized to seasons. ",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/[IMAGE_METHODS_1].png",
+        "caption": "",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    }
+]

072b9ce21b7943bdfb95e67483287fad20b8c8258dec744bf664f859e7a97f73/preprint/preprint.md

@@ -0,0 +1,159 @@
+# Abstract
+
+One likely consequence of global climate change is an increased frequency and intensity of droughts at high latitudes. We use a 17-year record from 13 nested boreal streams to examine the direct and lagged effects of summer drought on the quantity and quality of dissolved organic carbon (DOC) inputs from catchment soils. Protracted periods of drought reduced DOC concentrations in all catchments but also led to large pulses of DOC inputs upon rewetting in autumn. Concurrent changes in DOC optical properties and chemical character suggest that seasonal drying and rewetting triggers soil processes that alter the forms of carbon supplied to streams. Contrary to common belief, the clearest drought effects were observed in larger watersheds, whereas responses were most muted in smaller catchments. Collectively, our results reveal an emerging shift in the seasonal distribution of DOC concentrations and character, with potentially far-reaching consequences for northern aquatic ecosystems.
+
+# Introduction
+
+One major effect of ongoing climate change is a greater frequency and severity of hydrological drought<sup>1</sup>. An increased likelihood of such events is notably significant for northern landscapes, which are covered by vast areas of aquatic and wetland ecosystems<sup>2</sup>, and serve as an important source and conduit for nutrients, metals, and carbon to downstream recipients. Yet, much of our understanding of drought effects on streams and catchments come from research in biomes where such events historically have been common<sup>3</sup>. By comparison, the consequences of intensified drying/rewetting cycles at high latitudes for aquatic ecosystems and biogeochemical processes remain poorly investigated [Gomez et al. 2020]. This knowledge gap, together with the large pool of soil organic matter that can be mobilized in northern landscapes<sup>4</sup>, may lead to dramatic and unexpected impacts on aquatic ecosystems in a future dominated by more extreme weather events.
+
+Northern streams and lakes are often typified by high concentrations of dissolved organic carbon (DOC), which plays a wide range of geochemical<em><sup>5</sup></em>, biogeochemical<em><sup>6</sup></em>, and ecological roles<sup>7</sup> and is thus an important indicator of water quality. DOC supply from soils influences the transport and bioavailability of heavy metals and anthropogenic organic compounds<sup>8</sup>, represents a main energy source for northern aquatic food webs<sup>6</sup>, and promotes the production of harmful byproducts of chlorine disinfection during drinking water sanitization<sup>9</sup>. Further, variation in DOC ‘quality’, as represented by shifts in the relative composition of organic compounds and their degree of biological reactivity, can further shape biogeochemical processes in aquatic systems, including rates of microbial metabolism<sup>10</sup> and nutrient transformations<sup>11</sup>. Given these important roles, environmental changes that result in an alteration in the amount and characteristics of DOC supplied to aquatic systems can result in widespread consequences for northern aquatic ecosystems.
+
+DOC production and mobility in landscapes are driven by the combination of soil biogeochemical processes and the strength and timing of hydrological connections between terrestrial source areas, deeper groundwater systems, and stream channels<sup>12</sup>. It is generally recognized that elevated flow conditions promote DOC supply by strengthening the connections between these terrestrial sources and streams<sup>13</sup>. In this way, the timing of high flows, together with temperature-driven changes in soil processes, can shape the overall seasonality of DOC supply to streams<sup>14</sup>. However, much less is known about how the amount, timing, and chemical character of DOC are altered by seasonal drought episodes, which reduce lateral connectivity, but also set the stage for biogeochemical and microbial processes in dry and disconnected soils<sup>15</sup>. Such biogeochemical and microbial changes could alter the pool of organic matter that becomes mobilized when flow conditions resume.
+
+Here we ask how the severity of summer drought episodes drives the seasonal patterns of DOC quantity and quality in a boreal stream network. To answer this, we investigated how stream DOC concentrations, the 254/365 nm absorbance ratio, used here as a proxy for low molecular DOC (LMW DOC)<sup>16-18</sup>, and the carbon/nitrogen (CN ratio) respond to summer low flow conditions over 17-years in 13 nested catchments that differ in size and land cover. Given the importance of hydrology as a transport vector for DOC, we predicted that droughts would disconnect organic-rich soil layers in upper horizons from lateral flow paths, resulting in lower stream DOC, LMW DOC, and CN ratio, depending on the severity of these events. During the rewetting phase, we tested whether the DOC quantity and quality simply return to pre-drought conditions or whether prolonged dry periods alter the amount and composition of DOC that is mobilized to streams. Finally, we assessed how the variation in drying/rewetting shaped the DOC response across the drainage network depending on catchment properties, specifically variation in land cover and catchment size.
+
+# Results And Discussion
+
+## Mechanisms underlying drought responses
+
+Over the 17 year record, summer low flow hydrology varied considerably with mean daily minimum discharge ranging several orders of magnitude between the driest and wettest summers (0.0003 to 0.13 mm day⁻¹). The most pronounced summer low flows occurred in 2006 and 2018 with 62 and 41 days where daily discharge was below the 0.1 mm day⁻¹ threshold (Figure 1). This inter-annual variability in hydrology had clear consequences for DOC concentrations, which declined at each site as the severity of summer drought conditions increased. For all catchments, summers with more than 40 days of low flow conditions had the largest percentage decrease in DOC concentrations from the 17-year average (20-55% in small to larger catchments, p<0.05), respectively; Figure 2a). These concentration declines as drought severity increased are consistent with reduced hydrological connectivity to more surficial, organic-rich riparian soils as the groundwater table drops and intersect deeper horizons that store less organic matter¹⁹,²⁰. As further support for these mechanisms, DOC concentrations measured in near stream wells also declined as the drought severity increased (Figure 5). Here, DOC concentrations in riparian wells were reduced by 75% compared to pre-drought values (r²=0.72, p<0.05), whereas concentrations in nearby mire wells showed only minor changes (r²=0.11, p>0.1). By comparison, the wettest years in the record were associated with elevated DOC concentrations in streams, increasing from 10-20% relative to the long-term mean. Collectively, these inter-annual differences highlight the overwhelming role of hydrological connectivity between streams and near-surface soils (Supplementary Figure S1) that supply DOC to aquatic systems²¹. Importantly, while warmer temperatures have the potential to increase DOC supply from riparian soils²² and peat¹⁴,²³, our results suggest that potential changes in hydrology, including greater drought frequency, could fundamentally shift the seasonality of DOC in boreal aquatic ecosystems (Figure 3).
+
+Drought episodes also directly influenced DOC's character, yet these effects were more variable and in some cases subtle. For example, the CN ratio declined at all sites as drought severity increased (Figure 4c, Table 1, r²>0.17, p<0.05 for 12 out of 13 sites). We attribute the decrease in this ratio to a shift in the contribution to stream DOM from organic-rich, near-surface soils to deeper strata, where soils are more biologically processed and tend to have lower CN ratios²⁴,²⁵. At greater soil depths, there are also higher levels of reduced inorganic nitrogen (ammonium) which can contribute to the lower CN ratio as calculated here²⁶,²⁷. By contrast, the LMW DOC, changed only marginally (p<0.05 for nine out of thirteen sites, Table 1), with bidirectional responses to drought across catchments (Figure 4a), suggesting that the direct effects of drying are less systematic for the character of DOC as represented by this index. Regardless, the decreasing trends for both indexes were mirrored by observations in the riparian and mire wells, with the largest declines observed in years with the longest summer droughts (Figure 5 c, e, respectively, riparian wells r²=0.82, mire wells r²=0.81, p<0.1). Lower groundwater LMW DOC and CN ratio were synchronous with lower quantities of DOC observed in the streams during drought periods compared to pre-drought conditions. Thus, droughts limit the mobility of organic carbon across landscape types, reduce stream DOC concentrations, and alter key characteristics of DOC across the stream network. Such changes to stream chemistry could have cascading consequences for downstream aquatic ecosystems in the future, should extreme drought events increase in frequency and severity.
+
+## Post-drought recovery of stream biogeochemistry to summer droughts
+
+Periods of low flow conditions, which cause longitudinal fragmentation and lateral flow disconnections can also lead to increased production of DOC in upper soil horizons that may be mobilized when dry periods are terminated²⁸. Consistent with this, the rewetting periods following summer drought were associated with high DOC concentrations in streams across the Krycklan network, with the largest increases observed during years where rewetting followed the most severe dry periods. Indeed, the driest summers resulted in 100-150% increases in DOC concentration after rewetting across all catchments (Figure 2b, p<0.05, Table 1). As above, these episodes of DOC flushing in streams were also mirrored in observations from riparian (50% increase, r²=0.31, p<0.05) and mire wells (150% increase r²=0.79, p<0.05) (Supplementary Figure S1 b). In addition to the effects on concentration, DOC properties were also influenced by the severity of the preceding drought during rewetting periods. Specifically, the LMW DOC increased linearly in 12 of the 13 catchments (p<0.05), while the CN ratio increased in 11 out of 13 sites, from pre-drought conditions in the study streams (Figure 5b, d, Table 1). Patterns in groundwater LMW DOC and CN ratio during these rewetting periods were also similar to stream observations where both showed increasing trends in riparian and mire wells (p<0.1) (Figure 5 d, f). Thus, the changes in both DOC quantity and quality parameters suggest that seasonal low flow periods mobilize accumulated solutes during the following rewetting period to an extent that is proportional to the severity of the drought across all landscapes.
+
+Elevated DOC concentrations following the drought periods are in line with the reconnection of near-surface organic-rich soil horizons. Yet, the increases in the DOC concentration, as well as changes in DOM properties, as summer droughts became more severe suggests that biogeochemical processes strongly influenced the soil DOC pools during dry periods. Several mechanisms have been linked to small-scale increases in soil organic matter mineralization and DOC production in response to drying/rewetting cycles. For instance, droughts have been found to decrease phenolic microbial inhibiter compounds in wetlands resulting in increased organic matter decomposition and an increase in carbon loss in peats²⁹. Additionally, droughts increase the temperature and degree of aeration of soils that are normally inundated, upregulating organic matter decomposition³⁰,³¹. Finally, rewetting these exposed soils also trigger the physical and microbial processes that promote organic matter mineralization [Borken and Matzner, 2008]. While we cannot resolve amongst these mechanisms, the observed patterns suggest that processes in seasonally-exposed organic soils support a large pulse of DOM upon rewetting, including a greater fraction of LWM forms that are more bioavailable for aquatic organisms³². In fact, based on prior studies in the Krycklan catchment³², the changes in LMW DOC we observe in response to rewetting (ca.0.2- 0.5 units) in both surface and groundwaters could correspond to as much as a 50% increase in microbial growth efficiency in streams. In addition to these implications for energy mobilization by aquatic microbes, increased autumn pulses of DOC may also have implications for the fate and transport of a variety of hazardous pollutants such as mercury⁸, and, when combined with longer-term browning trends, contribute to poorer water quality and higher water treatment cost³³.
+
+## Network Scale responses
+
+Responses to drought at different time intervals were notably variable across the river network, reflecting differences in the sensitivity to low flow disturbance. Yet, the importance of the catchment size as a mediator of these responses differed depending on the time frame considered. For example, larger catchments showed both the greatest decline in DOC concentration in response to drought (Figure 6a) and the largest increases in DOC upon rewetting during the post-drought period (Figure 6b). Strong drought responses in the larger catchment likely relate to their greater distance to near-surface organic DOC sources that feed headwaters (Figure 6, Table 1). Isolation from these sources is exacerbated by the increasing influence of deeper and DOC-poor groundwater as catchment size increases¹². As a result, even small losses in connectivity to more DOC-rich headwaters during drought may cause the chemistry of larger rivers to shift abruptly towards the character of deeper groundwater sources. Upon rewetting, these larger streams and rivers have such low DOC concentrations that the sudden reconnection to upstream headwaters creates a strong chemical response (Figure 6b, p<0.05 for 10 of the 13 catchments, Table 1). By comparison, headwaters are seldom supported by these deeper groundwater sources³⁴, and hence their responses to drying and rewetting events are more attenuated. In this sense, larger river systems may be less prone to complete water loss during drought than headwaters, but nonetheless, show stronger biogeochemical responses to such events. Thus, while larger rivers may sustain aquatic communities during low flow periods, dramatic increases in DOC after rewetting may lead to a host of ecological and societal challenges³³ including greater stress for organisms in cases of rapid pH declines⁶ and ultimately higher cost for the production of potable water³⁵.
+
+While catchment size clearly plays an overarching role in regulating stream chemistry following prolonged dry periods, it is not possible to exclude the influence of land cover effects. For instance, several small catchments are dominated by peat-forming mires and display the lowest response to droughts (Figure 6b). Contrastingly, the stronger response by larger, forest-dominated catchments could also be an indication of the degree to which they dry during more severe droughts, as reduced mire cover is linked to lower water storage capacity, and greater evaporative losses, and thus potentially weakened ecohydrological resilience to drought³⁶. Conversely, catchments with greater peat coverage, showed the weakest post-drought response, suggesting these landscape elements confer resilience to such events, likely by acting as important water storage zones³⁷ with the potential to dampen effects of long-term environmental change³⁸.
+
+## Conceptualizing drought impacts on stream chemistry
+
+Integrating long-term monitoring data of biogeochemistry and hydrology with modeling techniques provide a more comprehensive understanding of how climate extremes feedback on the mobilization and biogeochemical cycling of soil organic carbon across temporal and spatial scales in boreal catchments. Observed seasonal variation in amplitude of DOC, LMW DOC, and CN ratio demonstrate differential responses in catchment biogeochemical processes to droughts, such that stream water quality is not only affected by reduced soil water supply, but also by declines in the leaching and availability of organic resources important for aquatic microbial processes. Further, the combined responses of the three variables suggest that, while drought effects on stream water chemistry are direct and immediately result in lower than average DOC responses (Figure 7a), the indirect, lagged effects are magnified and extend beyond the duration of the disturbance itself (Figure 7b). Overall, increases in the intensity of seasonal drying/rewetting cycles have the potential to shift the seasonality of DOC in boreal streams by reducing summer peaks in concentration while causing anomalously high concentrations during periods of hydrological reconnection later in the autumn (Figure 7b). Therefore, how recipient aquatic ecosystems cope with such changes during and following droughts, remains a key question.
+
+Almost two decades of monitoring data show that inter-annual variation in summer low flows shape the seasonality in DOC quantity and quality in boreal streams more than is currently appreciated. While much emphasis is currently placed on the direct effects of climate warming at high latitudes³⁹, our study indicates that potential hydrological changes will likely be another important driver of carbon mobilization and water chemistry change. In light of the current climate projections of an increase in drought frequency in Scandinavia⁴⁰, our results suggest the immediate declines in the quality and quantity of organic carbon in streams during summer followed by lagged increases during rewetting. Despite the variations in landscape properties, all catchments showed similar responses to droughts, however, the magnitude of these responses was more pronounced in the largest catchments. Changes to both the quantity and quality of carbon across the stream network can potentially have vital implications for the aquatic ecosystems that rely on the seasonal balance of DOC production and mobilization from catchment soils⁴¹. Overall, these results highlight the importance of integrating responses at multiple temporal and spatial scales and present a step forward in establishing a unifying theory of drought impacts in boreal biogeochemistry.
+
+# Methods
+
+The Krycklan Catchment Study (KCS) (64.23° N, 19.77° E) is located in Northern Sweden and consists of 13 long-term monitoring streams. The sub-catchments vary in size from small headwaters (0.12 km²) to the large outlet (67 km²; <sup>42</sup>). Land cover is dominated by forest till soils (47% to 100% among monitoring sites), lakes (0 to 6%), and peatlands (referred to as mires) (0% to 51% areal coverage). Fluvial sediments dominate the lower parts of the catchment below the highest postglacial coastline (Supplementary Table S1.). The bedrock consists of 94% metasediments/metagraywacke, 4% acid and intermediate metavolcanic rocks, and 3% basic metavolcanic rocks. Soil mineralogy is dominated by quartz (31–43%), plagioclase (20–25%), K-feldspar (16–33%), amphiboles (7–21%), muscovite (2–16%), and chlorite (1–4%) <sup>43</sup>. Forests are predominantly Norway spruce (Picea abies, 25%) and Scots pine (Pinus sylvestris, 66%), with 9% deciduous forest. Mean annual precipitation recorded between 1991–2010 was 610 mm from which 35% was classified as snow during winter (December–April) <sup>44</sup>. In January, the average air temperature is -9.5 ±4.1°C while July temperatures are 14.5±1.7°C <sup>44</sup>. During the spring, (mid-April), snowmelt accounts for approximately 40% of the annual runoff. There are low impacts from land use with 2% agricultural lands, less than 100 inhabitants and 0.63% of the catchment was subject to final felling, annually (1999-2010). The hydrological regime, landscapes, and land uses of Krycklan Catchment are considered representative of much of the boreal landscape [Laudon et al., 2013].
+
+## Discharge data
+
+Discharge measurements used to classify summer low flow conditions were based on a small, centrally located headwater sub-catchment (C7, 0.45 km²) for which we have the longest, most detailed record in the Krycklan Catchment. The C7 sub-catchment drain a mix of forest (81%) and mire (19%) land cover, has a mean specific runoff that falls near the average for all streams in the Krycklan Catchment, and hence provides a reasonable proxy for discharge including drought condition in the area <sup>45</sup>. Using this standardized runoff also made it possible to compare the responses between catchments to the same drought period. Stage height was determined from a 90⁰V-notch weir in a heated house with a pressure transducer connected to a Campbell Scientific data logger <sup>46</sup>. Daily discharge was calculated from measurements of stage height and an established rating curve based on more than 1000 manual salt dilution and bucket-method measurements <sup>47</sup>.
+
+## Surface and groundwater sampling
+
+During summer, surface water samples were collected every other week from each site in acid-washed, high-density polyethylene bottles, which were kept in cold storage until analysis. Samples were collected monthly in winter. Sampling was synchronous across all sub-catchments occurring mostly on the same day, which provided 190 monthly DOC averages during the investigated period (2003-2019) for each of the 13 catchments (Unity Svartberget Data, <a href="https://franklin.vfp.slu.se/">https://franklin.vfp.slu.se/</a>). In addition, groundwater was sampled from a nest of suction lysimeters that collect riparian soil water from 0.10 to 0.65 m at ~0.10 m intervals. Finally, mire groundwater samples were collected from wells installed from 2 to 2.5 m. Both sampling of forest riparian soil solution and mire groundwater occurred seasonally; here we used samples collected at the onset of summer (end of May-June) and mid-summer/early autumn (Jul-Sep). The averages of all depths were used for the analysis for both the riparian and mire wells. Annual sampling in the riparian zone began in 2003. Similar sampling in the mire started later (2009) with sufficient seasonal samples for our purposes, collected on average every second year.
+
+In the laboratory, surface water (total 7060 samples from all sites) and groundwater samples (827 samples from both riparian and mire wells) were analyzed for DOC and total nitrogen (TN) concentrations using a Shimadzu TOC-VCPH analyzer after acidification to remove inorganic compounds <sup>38</sup>. The CN ratio was calculated by dividing DOC by TN. We did not subtract the dissolved inorganic carbon (NH₄, NO₂, NO₃) from TN estimates due to the limited data in the larger sites, however, since the proportion of inorganic N is relatively small (9% of the TN during the summer period) compared to total organic nitrogen, we do not expect that this would affect our results. Filtered water samples were also analyzed for absorbance (243 surface samples and 72 groundwater) using wavelength ranging from 200 to 600 nm, at 1 nm intervals at a scan speed of 240 nm min⁻¹ and a slit width of 2 nm using a Lambda 40 UV-visible spectrophotometer (Perkin Elmer, Waltham, MA, USA). A 1 cm quartz cuvette with Milli-Q water as the blank was used for measuring the samples. We used the absorbance wavelengths A₂₅₄ and A₃₆₅ nm for the ratio (Abs ratio A₂₅₄/A₃₆₅) in this analysis to trace the effects of droughts on the low molecular weight DOC compounds (LMW DOC). The 254/365 absorption ratio is positively correlated to bacterial production in natural waters <sup>16</sup> and negatively to the molecular weight of DOC <sup>17,18</sup> thus can be used as a qualitative measure of low molecular weight DOC.
+
+## Hydrological droughts
+
+We represented inter-annual variation in the extent of summer drought over the 17 years study period using the occurrence of low flow conditions as a proxy. Within this study period, a threshold of 0.1 mm day⁻¹ was used to represent low flow conditions, which corresponds to daily discharge less than the 10th percentile value based on summer observations over the last 30 years. For this estimate, the summer season was defined based on air temperatures (July-Sept) <sup>48</sup>. The ggridges density distribution function from the ggplot 2 package in R was then used to display the proportion of discharge below the thresholds in each year and to visualize changes in the distribution over space and time. Ridge line plots calculate density estimates from actual data (jitter point, Figure 1 b) and plot those using ridgeline visualization. From the density distribution of all the years, we can observe that the majority of days during the driest summers (2006, 2018) had discharge levels below the 10th percentile (0.1 mm day⁻¹) (Figure 1). Years with no low flow days below the threshold were 2012, 2016, and 2017 which showed an almost even distribution of discharge over the summer season (Figure 1). The years with the average number of low flow days were 2005 and 2015 (15 and 16 days respectively, Figure 1).
+
+## Drought impacts on DOC, LMW DOC, and CN
+
+The impacts of low flow conditions on DOC concentrations were investigated at two temporal scales to explore the direct drought effects as well as the subsequent responses during the rewetting stage. The direct response captures changes in DOC, LMW DOC, and CN that occur during summer low flow periods. The post-drought effects investigated the change in these same response variables after the first rewetting. We similarly investigated drought effects on groundwater DOC, LMW DOC, and CN ratio in forested riparian and mire soils during summer low flows and post-drought after the first rewetting.
+
+The effects of prolonged summer low flow on both surface water and groundwater DOC were determined by first averaging the DOC concentrations during the summer period (Jul and Sept) in each catchment for each year. We then calculated the difference between individual summer averages and the long-term summer average of the 17 years and expressed as percentage change, following:
+
+[IMAGE_METHODS_1]
+
+where the drought effect (DOC<sub>d</sub>) was determined as the percentage difference in DOC each summer (DOC<sub>a</sub>) compared to the long-term summer average (DOC<sub>b</sub>). We used linear regression with the number of low flow days in the summer period to test the prediction that average DOC concentrations would decline with drying severity (Figure 2). Finally, we used a similar regression approach to test whether DOC in forest and mire groundwater were also affected by drought severity by comparing values measured in the summer to pre-drought values (end of May-June average).
+
+The effect of prolonged summer low flows on the initial flush of DOC (DOC<sub>d</sub>) upon rewetting was estimated as the difference between DOC measured after the longest low flow period (when there was a rain event that caused an increase in runoff) (DOC<sub>a</sub>) and DOC measured before the onset of low flows each summer (DOC<sub>b</sub>) (equation. 1). Here, we expected that DOC produced in soils during the dry periods would be flushed out in the first rewetting event, reflecting processes hindering DOC consumption or favoring DOC production during drought. We used linear regression with the number of low flows in each summer to test whether changes in DOC were related to the duration of the low flow periods. For groundwater analysis, similar calculations were used to show the differences between the post-summer (September values) and pre-summer (-June) values and to ask whether there were any post-drought effects on groundwater in either the forest soils or mires (Fig. S2). All differences were expressed as percentage changes from the pre-drought concentrations. As above, we used the linear regression with the number of flow days below the 0.1 mm day⁻¹ threshold to test whether the magnitude of these (lagged) effects was related to the severity of summer drying.
+
+## Drought and Post-drought effects on LMW DOC and CN ratio
+
+The drought and post-drought effects analysis for LMW DOC and CN ratio followed the same procedure as used in the DOC modeling for stream water and groundwater data. With these two independent variables of carbon character, we expected that changes that occur as a result of the prolonged dry summer periods would be reflected in the quality of the carbon (LMW DOC and CN ratio) when soils are flushed during rewetting. In these analyses, we expected that if the drought and post-drought effects on DOC are purely hydrological, there would be no change in the LMW DOC and CN ratio indicating that the quality is unaffected by prolonged dry periods. Here, an increasing trend in the absorption ratio (254/365 nm) signifies a shift to carbon with low molecular weight and higher bacterial productivity <sup>16,49</sup> while increasing CN ratio may indicate increasing biodegradability of DOC <sup>50</sup>. Conversely, decreasing absorption ratios indicates a shift to more aromatic compounds with higher molecular weight, while a lower CN ratio indicates DOC supply from more strongly processed soils at lower depths <sup>51</sup>.
+
+## Slope relationships with land cover and catchment area
+
+To better understand the drivers of drought response, we tested how the slope of the regression relating DOC change to drought duration at each site varied with catchment features. The slope of this relationship represents the rate of change in DOC concentration as drought magnitude increases at each stream, thus providing an integrative assessment of drought sensitivity. We used stepwise multiple linear regression (in Minitab 18.1) to test whether this response varied among streams as a function of subcatchment size, as well as the percentage of peat soils, forest, lake, and sedimentary soil cover in each subcatchment.
+
+# References
+
+1. Buntgen, U. et al. Recent European drought extremes beyond Common Era background variability. *Nat Geosci*, doi: 10.1038/s41561-021-00698-0 (2021).
+2. Vonk, J. E. et al. Reviews and syntheses: Effects of permafrost thaw on Arctic aquatic ecosystems. *Biogeosciences* **12**, 7129–7167, doi: 10.5194/bg-12-7129-2015 (2015).
+3. Granados, V. et al. The interruption of longitudinal hydrological connectivity causes delayed responses in dissolved organic matter. *Sci Total Environ* **713**, doi: 10.1016/j.scitotenv.2020.136619 (2020).
+4. Bradshaw, C. J. A. & Warkentin, I. G. Global estimates of boreal forest carbon stocks and flux. *Global Planet Change* **128**, 24–30, doi: 10.1016/j.gloplacha.2015.02.004 (2015).
+5. Aiken, G. R., Hsu-Kim, H. & Ryan, J. N. Influence of Dissolved Organic Matter on the Environmental Fate of Metals, Nanoparticles, and Colloids. *Environ Sci Technol* **45**, 3196–3201, doi: 10.1021/es103992s (2011).
+6. Creed, I. F. et al. Global change-driven effects on dissolved organic matter composition: Implications for food webs of northern lakes. *Global Change Biol* **24**, 3692–3714, doi: 10.1111/gcb.14129 (2018).
+7. Karlsson, J. et al. Light limitation of nutrient-poor lake ecosystems. *Nature* **460**, 506-U580, doi: 10.1038/nature08179 (2009).
+8. Dittman, J. A. et al. Mercury dynamics in relation to dissolved organic carbon concentration and quality during high flow events in three northeastern US streams. *Water Resour Res* **46**, doi: 10.1029/2009wr008351 (2010).
+9. Ritson, J. P. et al. The effect of drought on dissolved organic carbon (DOC) release from peatland soil and vegetation sources. *Biogeosciences* **14**, 2891–2902, doi: 10.5194/bg-14-2891-2017 (2017).
+10. Brailsford, F. L. et al. Microbial uptake kinetics of dissolved organic carbon (DOC) compound groups from river water and sediments. *Sci Rep-Uk* **9**, doi: 10.1038/s41598-019-47749-6 (2019).
+11. Rodriguez-Cardona, B., Wymore, A. S. & McDowell, W. H. DOC:NO3- ratios and NO3- uptake in forested headwater streams. *J Geophys Res-Biogeo* **121**, 205–217, doi: 10.1002/2015jg003146 (2016).
+12. Tiwari, T., Laudon, H., Beven, K. & Agren, A. M. Downstream changes in DOC: Inferring contributions in the face of model uncertainties. *Water Resour Res* **50**, 514–525, doi: 10.1002/2013wr014275 (2014).
+13. Zarnetske, J. P., Bouda, M., Abbott, B. W., Saiers, J. & Raymond, P. A. Generality of Hydrologic Transport Limitation of Watershed Organic Carbon Flux Across Ecoregions of the United States. *Geophys Res Lett* **45**, 11702–11711, doi: 10.1029/2018gl080005 (2018).
+14. Christ, M. J. & David, M. B. Temperature and moisture effects on the production of dissolved organic carbon in a Spodosol. *Soil Biol Biochem* **28**, 1191–1199, doi: 10.1016/0038-0717(96)00120-4 (1996).
+15. Roth, V. N. et al. Persistence of dissolved organic matter explained by molecular changes during its passage through soil. *Nat Geosci* **12**, 755-+, doi: 10.1038/s41561-019-0417-4 (2019).
+16. Berggren, M., Laudon, H. & Jansson, M. Landscape regulation of bacterial growth efficiency in boreal freshwaters. *Global Biogeochem Cy* **21**, doi: 10.1029/2006gb002844 (2007).
+17. Dehaan, H. Solar Uv-Light Penetration and Photodegradation of Humic Substances in Peaty Lake Water. *Limnol Oceanogr* **38**, 1072–1076, doi: 10.4319/lo.1993.38.5.1072 (1993).
+18. Dahlen, J., Bertilsson, S. & Pettersson, C. Effects of UV-A irradiation on dissolved organic matter in humic surface waters. *Environ Int* **22**, 501–506, doi: 10.1016/0160-4120(96)00038-4 (1996).
+19. Ledesma, J. L. J. et al. Towards an Improved Conceptualization of Riparian Zones in Boreal Forest Headwaters. *Ecosystems* **21**, 297–315, doi: 10.1007/s10021-017-0149-5 (2018).
+20. Gomez-Gener, L., Lupon, A., Laudon, H. & Sponseller, R. A. Drought alters the biogeochemistry of boreal stream networks. *Nat Commun* **11**, 1795, doi: 10.1038/s41467-020-15496-2 (2020).
+21. Laudon, H. et al. Patterns and Dynamics of Dissolved Organic Carbon (DOC) in Boreal Streams: The Role of Processes, Connectivity, and Scaling. *Ecosystems* **14**, 880–893, doi: 10.1007/s10021-011-9452-8 (2011).
+22. Winterdahl, M. et al. Riparian soil temperature modification of the relationship between flow and dissolved organic carbon concentration in a boreal stream. *Water Resour Res* **47**, doi: 10.1029/2010wr010235 (2011).
+23. Pastor, J. et al. Global warming and the export of dissolved organic carbon from boreal peatlands. *Oikos* **100**, 380–386, doi: 10.1034/j.1600-0706.2003.11774.x (2003).
+24. Gundersen, P. et al. Do indicators of nitrogen retention and leaching differ between coniferous and broadleaved forests in Denmark? *Forest Ecol Manag* **258**, 1137–1146, doi: 10.1016/j.foreco.2009.06.007 (2009).
+25. Rumpel, C. & Kogel-Knabner, I. Deep soil organic matter-a key but poorly understood component of terrestrial C cycle. *Plant Soil* **338**, 143–158, doi: 10.1007/s11104-010-0391-5 (2011).
+26. Callesen, I., Raulund-Rasmussen, K., Westman, C. J. & Tau-Strand, L. Nitrogen pools and C: N ratios in well-drained Nordic forest soils related to climate and soil texture. *Boreal Environ Res* **12**, 681–692 (2007).
+27. Marty, C., Houle, D., Gagnon, C. & Courchesne, F. The relationships of soil total nitrogen concentrations, pools and C:N ratios with climate, vegetation types and nitrate deposition in temperate and boreal forests of eastern Canada. *Catena* **152**, 163–172, doi: 10.1016/j.catena.2017.01.014 (2017).
+28. Werner, B. J. et al. High-frequency measurements explain quantity and quality of dissolved organic carbon mobilization in a headwater catchment. *Biogeosciences* **16**, 4497–4516, doi: 10.5194/bg-16-4497-2019 (2019).
+29. Fenner, N. & Freeman, C. Drought-induced carbon loss in peatlands. *Nat Geosci* **4**, 895–900, doi: 10.1038/Ngeo1323 (2011).
+30. Bragazza, L. et al. Persistent high temperature and low precipitation reduce peat carbon accumulation. *Global Change Biol* **22**, 4114–4123, doi: 10.1111/gcb.13319 (2016).
+31. Stirling, E., Fitzpatrick, R. W. & Mosley, L. M. Drought effects on wet soils in inland wetlands and peatlands. *Earth-Sci Rev* **210**, doi: 10.1016/j.earscirev.2020.103387 (2020).
+32. Berggren, M., Laudon, H. & Jansson, M. Bacterial utilization of imported organic material in three small nested humic lakes. *Int Ver Theor Angew* **30**, 1393-+ (2010).
+33. Kritzberg, E. S. et al. Browning of freshwaters: Consequences to ecosystem services, underlying drivers, and potential mitigation measures. *Ambio* **49**, 375–390, doi: 10.1007/s13280-019-01227-5 (2020).
+34. Peralta-Tapia, A. et al. Scale-dependent groundwater contributions influence patterns of winter baseflow stream chemistry in boreal catchments. *J Geophys Res-Biogeo* **120**, 847–858, doi: 10.1002/2014jg002878 (2015).
+35. Lavonen, E. E., Gonsior, M., Tranvik, L. J., Schmitt-Kopplin, P. & Kohler, S. J. Selective Chlorination of Natural Organic Matter: Identification of Previously Unknown Disinfection Byproducts. *Environ Sci Technol* **47**, 2264–2271, doi: 10.1021/es304669p (2013).
+36. Geris, J., Tetzlaff, D. & Soulsby, C. Resistance and resilience to droughts: hydropedological controls on catchment storage and run-off response. *Hydrological Processes* **29**, 4579–4593, doi: 10.1002/hyp.10480 (2015).
+37. Karlsen, R. H. et al. The role of landscape properties, storage and evapotranspiration on variability in streamflow recessions in a boreal catchment. *J Hydrol* **570**, 315–328, doi: 10.1016/j.jhydrol.2018.12.065 (2019).
+38. Fork, M. L., Sponseller, R. A. & Laudon, H. Changing Source-Transport Dynamics Drive Differential Browning Trends in a Boreal Stream Network. *Water Resour Res* **56**, doi: 10.1029/2019WR026336 (2020).
+39. Crowther, T. W. et al. Quantifying global soil carbon losses in response to warming. *Nature* **540**, 104-+, doi: 10.1038/nature20150 (2016).
+40. Spinoni, J., Vogt, J. V., Naumann, G., Barbosa, P. & Dosio, A. Will drought events become more frequent and severe in Europe? *Int J Climatol* **38**, 1718–1736, doi: 10.1002/joc.5291 (2018).
+41. Kaiser, K., Canedo-Oropeza, M., McMahon, R. & Amon, R. M. W. Origins and transformations of dissolved organic matter in large Arctic rivers. *Sci Rep-Uk* **7**, doi: 10.1038/s41598-017-12729-1 (2017).
+42. Laudon, H., Sponseller, R. A. & Bishop, K. From legacy effects of acid deposition in boreal streams to future environmental threats. *Environ Res Lett* **16**, doi: 10.1088/1748-9326/abd064 (2021).
+43. Ledesma, J. L. J. et al. Riparian zone control on base cation concentration in boreal streams. *Biogeosciences* **10**, 3849–3868, doi: 10.5194/bg-10-3849-2013 (2013).
+44. Oni, S. K. et al. Long-term patterns in dissolved organic carbon, major elements and trace metals in boreal headwater catchments: trends, mechanisms and heterogeneity. *Biogeosciences* **10**, 2315–2330, doi: 10.5194/bg-10-2315-2013 (2013).
+45. Karlsen, R. H. et al. Landscape controls on spatiotemporal discharge variability in a boreal catchment. *Water Resour Res* **52**, 6541–6556, doi: 10.1002/2016wr019186 (2016).
+46. Laudon, H. et al. The Krycklan Catchment Study-A flagship infrastructure for hydrology, biogeochemistry, and climate research in the boreal landscape. *Water Resour Res* **49**, 7154–7158, doi: 10.1002/Wrcr.20520 (2013).
+47. Laudon, H., Kohler, S. & Buffam, I. Seasonal TOC export from seven boreal catchments in northern Sweden. *Aquat Sci* **66**, 223–230, doi: 10.1007/s00027-004-0700-2 (2004).
+48. Tiwari, T., Sponseller, R. A. & Laudon, H. Extreme Climate Effects on Dissolved Organic Carbon Concentrations During Snowmelt. *J Geophys Res-Biogeo* **123**, 1277–1288, doi: 10.1002/2017jg004272 (2018).
+49. Fellman, J. B., D'Amore, D. V., Hood, E. & Boone, R. D. Fluorescence characteristics and biodegradability of dissolved organic matter in forest and wetland soils from coastal temperate watersheds in southeast Alaska. *Biogeochemistry* **88**, 169–184, doi: 10.1007/s10533-008-9203-x (2008).
+50. Hunt, A. P., Parry, J. D. & Hamilton-Taylor, J. Further evidence of elemental composition as an indicator of the bioavailability of humic substances to bacteria. *Limnol Oceanogr* **45**, 237–241, doi: 10.4319/lo.2000.45.1.0237 (2000).
+51. Kindler, R. et al. Dissolved carbon leaching from soil is a crucial component of the net ecosystem carbon balance. *Global Change Biol* **17**, 1167–1185, doi: 10.1111/j.1365-2486.2010.02282.x (2011).
+
+# Tables
+
+**Table 1. The regression coefficient of Drought and Post-drought responses of DOC, LMW DOC, and CN ratio in the sub-catchments used in this study**
+
+| Sites | Drought |  |  | Post-drought |  |  |
+|-------|---------|----|----|--------------|----|----|
+|       | DOC     | LMW DOC | CN ratio | DOC | LMW DOC | CN ratio |
+| C1    | 0.42    | 0.20    | 0.25     | 0.33    | 0.24    | 0.14*  |
+| C2    | 0.29    | 0.37    | 0.27     | 0.39    | 0.56    | 0.50   |
+| C4    | 0.41    | 0.45    | 0.43     | 0.65    | 0.64    | 0.44   |
+| C5    | 0.33    | 0.28    | 0.47     | 0.07*   | 0.36    | 0.58   |
+| C6    | 0.36    | 0.02*   | 0.31     | 0.42    | 0.36    | 0.20   |
+| C7    | 0.51    | 0.14*   | 0.39     | 0.67    | 0.27    | 0.40   |
+| C9    | 0.43    | 0.20    | 0.27     | 0.58    | 0.17*   | 0.36   |
+| C10   | 0.52    | 0.04*   | 0.20     | 0.60    | 0.21    | 0.43   |
+| C12   | 0.65    | 0.02*   | 0.37     | 0.45    | 0.45    | 0.59   |
+| C13   | 0.30    | 0.01*   | 0.29     | 0.12*   | 0.59    | 0.14*  |
+| C14   | 0.60    | 0.40    | 0.41     | 0.28    | 0.31    | 0.46   |
+| C15   | 0.52    | 0.54    | 0.13*    | 0.20*   | 0.35    | 0.33   |
+| C16   | 0.39    | 0.50    | 0.30     | 0.31    | 0.39    | 0.31   |
+
+All values significant at p<0.05 except *
+
+# Supplementary Files
+
+- [Supplementary.docx](https://assets-eu.researchsquare.com/files/rs-1137926/v1/ec4859fae1cc04b1b31b21ab.docx)  
+  Supplementary information

0881102f92b71ddf7f1d973c580e2e057a1c1d457488b491ccfa9270b62ba04a/preprint/images_list.json

@@ -0,0 +1,50 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/Figure_1.png",
+        "caption": "Characterizing different promoters to drive ATOH1 expression and hair cell regeneration in adult mouse utricle ex vivo. A) Utricles from 4-10-week-old mice were treated with 1mM gentamicin followed by AAV8-CMV-ATOH1-H2B-EGFP, AAV8-GFAP-ATOH1-H2B-EGFPor AAV8-RLBP1-ATOH1-H2B-EGFP, then examined at day 12 and day 20. B-G) Utricles treated with CMV-ATOH1, GFAP-ATOH1or RLBP1-ATOH1for 12 and 20 days. Insets showing the GFP signals. H-I) GFP intensity in hair cells and supporting cells after treatment with CMV, GFAP and RLBP1 promoters. J) Quantification shows hair cell counts at day 12 and 20 after treatment with CMV-ATOH1,GFAP-ATOH1and RLBP1-ATOH1. Over the time, hair cell number significantly increased with CMV-ATOH1 and GFAP-ATOH1, but not RLBP1-ATOH1treatment. Data shown as mean\u00b1S.D., compared using one-way ANOVA. *p<0.05, **p<0.01, ***p<0.001. Scale bars: B-G 100mm.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_2.png",
+        "caption": "Single-cell transcriptomic analyses of utricles treated with AAV8-CMV-, GFAP- or RLBP1-ATOH1-H2B-EGFP. A) Utricles from 4-10-week-old mice were treated with 1mM gentamicin followed by AAV8-CMV-ATOH1-H2B-EGFP, AAV8-GFAP-ATOH1-H2B-EGFP or AAV8-RLBP1-ATOH1-H2B-EGFP (4x1011 vg/mL), and processed for single-cell RNA sequencing at day 7, day 12 and day 20. B) UMAP plot of integrated dataset of 3 different virus treatment from all timepoints showing 6,901 single cells following all quality control steps including only hair cells (HC) and supporting cells (SC). Six cell clusters including, converting SC 1, converting SC 2, immature HC, mature HCa and mature HCb were identified. B\u2019) Schematic showing supporting cell-to-hair cell conversion. C) UMAP plot colored by CMV-ATOH1, GFAP-ATOH1 or RLBP1-ATOH1 treatment. D) Heatmap showing differentially expressed genes among the 6 cell clusters. The top differentially expressed genes of each cluster are shown on the right. A full list of these genes is found in Table S1. The heatmap is colored by relative expression from -2 (magenta) to 2 (yellow). E) UMAP plot showing the Monocle3 pseudotime value of each cell, progressing from gray to purple. F) The fraction of cells from different treatment groups that were assigned to each cell type. G) The fraction of cells from different timepoints that were assigned to each cell type.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_3.png",
+        "caption": "Supporting-cell-specific promoters drive regenerated hair cell maturation ex vivo. A) Violin plot depicting ATOH1-EGFP-transgene expression levels in 6 cell clusters. B) UMAP plot of ATOH1-EGFP-transgene expression level. C-D) Schematic showing persistently and transiently upregulated ATOH1-EGFP-transgene in supporting cells and regenerated hair cells after CMV-ATOH1 treatment, and GFAP-ATOH1 or RLBP1-ATOH1 treatment, respectively. E-H) Ridge plots of hair cell maturity, immaturity score, type II hair cell and type I hair cell scores of 6 different cell clusters. A full list of these genes is found in Table S3. I) Dot plots showing top 18 highly enriched mature hair cell genes from each cell cluster. J) Dot plots showing top 10 highly enriched immature hair cell genes from each cell cluster.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_4.png",
+        "caption": "AAV8-GFAP-ATOH1 induced hair cell regeneration in the adult mouse utricle in vivo. A) 4\u201310-week-old mice were treated with IDPN (4 or 5 mg/g) at day 1, AAV8-GFAP-ATOH1-H2B-EGFP (1.71x1013 vg/mL) were injected through posterior semicircular canal at day 2 or day 15. Organs were examined at 2 weeks and 1 month after injection. B-C\u2019\u2019) Treated utricles had more Myosin7a+ (red) hair cells compared to contralateral ear at both 2 weeks and 1 month. D-D\u2019) Representative high magnification images of injected and contralateral ears at 2 weeks (from B and B\u2019\u2019). GFP+ hair cells were observed in utricle treated with GFAP-ATOH1 (arrowhead). E-F) GFAP-ATOH1 significantly increased Myosin7a+ (red) hair cells in both the extrastriolar and striolar regions at 2 weeks and 1 month compared to contralateral ear. Red dots represent utricles with surgery at day 2 and black dots represent surgery at day 15. G-H\u2019\u2019) Crista from injected ear has more Myosin7a+ (red) hair cells compared to contralateral ear at both 2 weeks and 1 month. I-J\u2019) Utricles from injected ear had more bundles (phalloidin, green) than contralateral ear. I\u2019 and J\u2019 shown the high magnification pictures in boxes. K) The number of phalloidin-labeled bundles increased significantly in utricles from injected ear compared to contralateral controls. L-N) Expression of Ctbp2+ puncta (gray) on the basolateral surfaces of utricular hair cells from naive, injected ear and contralateral ear. Insets are the 3D view of hair cells at 1 month. O) Quantification of Ctbp2+ puncta in Myosin7a+ (red) hair cells. Number of Ctbp2+ puncta per hair cell increased significantly in contralateral ear compared to naive and injected ear. Data shown as mean\u00b1S.D, compared using paired Student\u2019s t-tests and Ordinary one-way ANOVA. *p<0.05, **p<0.01, ***p<0.001. Scale bars: B-H\u2019\u2019, I-J 100mm, D-D\u2019, P-R 20mm, I\u2019-J\u2019 10mm, L and N 3mm, M and O 2mm.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_5.png",
+        "caption": "Single-cell transcriptomic analyses of IDPN-damaged utricles treated with GFAP-ATOH1 in vivo. A) 4\u201310-week-old mice were treated with IDPN (5 mg/g) at day 1, AAV8-GFAP-ATOH1-H2B-EGFP (1.71x1013 vg/mL) at day 15. Single-cell RNA sequencing was performed at 2 weeks and 1 month after injection. B) UMAP plot of integrated dataset of 3 different treatment groups (control, IDPN, IDPN and GFAP-ATOH1) from both timepoints. There were 8,399 single cells following all quality control steps including only hair cells and supporting cells. Six cell clusters including supporting cells, stage 1, stage 2, stage 3 regenerated hair cells, type II and type I hair cells were identified. C) UMAP plot colored by control, IDPN or IDPN and GFAP-ATOH1 treatment. D) Heatmap showing differentially expressed genes among the 6 cell clusters. The top differentially expressed genes of each cluster are shown on the right. A full list of these genes is found in Table S4. The heatmap is colored by relative expression from -2 (magenta) to 2 (yellow). E) UMAP plot showing the Monocle 3 pseudotime value of each cell, progressing from gray to purple. F) The fraction of cells from different treatment groups assigned to each cell cluster. G) The fraction of cells from different timepoints assigned to each cell cluster. H) The fraction of cells from different timepoints with GFAP-ATOH1 treatment assigned to each cell cluster.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_6.png",
+        "caption": "Supporting-cell-specific promoters drives maturation of regenerated hair cells in vivo. A) Violin plot of ATOH1-EGFP-transgene expression level in 6 cell clusters. B) UMAP plot of ATOH1-EGFP-transgene expression levels. C-F) Ridge plots of hair cell maturity, immature hair cell, type II hair cell and type I hair cell scores of 6 different cell clusters. A full list of these genes is found in Table S3. G) Dot plots showing top 18 highly enriched mature hair cell genes from each cell cluster. J) Dot plots showing top 10 highly enriched immature hair cell genes from each cell cluster.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    }
+]

0881102f92b71ddf7f1d973c580e2e057a1c1d457488b491ccfa9270b62ba04a/preprint/preprint.md

@@ -0,0 +1,245 @@
+# Abstract
+
+Vestibular hair cells are mechanoreceptors critical for detecting head position and motion. In mammals, hair cell loss causes vestibular dysfunction as spontaneous regeneration is nearly absent. Constitutive expression of exogenous ATOH1, a hair cell transcription factor, increases regeneration of hair cells, but these cells fail to mature. To mimic native hair cells which downregulate endogenous ATOH1 as they mature, we engineered viral vectors carrying the supporting cell promoters GFAP and RLBP1. In utricles damaged ex vivo, both CMV-ATOH1 and GFAP-ATOH1 increased regeneration more effectively than RLBP1-ATOH1, while GFAP-ATOH1 and RLBP1-ATOH1 induced hair cells exhibiting more mature transcriptomes. In utricles damaged in vivo, GFAP-ATOH1 induced regeneration of hair cells expressing genes representing maturing type II hair cells, and more hair cells with bundles and synapses than untreated organs. Together our results demonstrate the efficacy of spatiotemporal control of ATOH1 overexpression in inner ear regeneration.
+
+**Biological sciences/Neuroscience/Regeneration and repair in the nervous system**  
+**Biological sciences/Biotechnology/Gene therapy**  
+**Biological sciences/Neuroscience/Auditory system/Hair cell**  
+**hair cells**  
+**utricle**  
+**regeneration**  
+**vestibular**  
+**mouse**
+
+# INTRODUCTION
+
+Loss of vestibular hair cells renders the inner ear balance organs dysfunctional. A variety of insults, including ototoxic drugs, infection, genetic mutations, and ageing can damage vestibular hair cells<sup>1,2</sup>, causing vestibular deficits that affect more than 35% of the US population aged 40 and above<sup>3</sup>. Since hair cells are the primary mechanoreceptors and are highly susceptible to damage, hair cell degeneration is deemed the main pathology underlying human vestibulopathy<sup>4</sup>. As we currently lack effective treatments for vestibulopathy, there is a pressing need for a biological treatment—therapies that regenerate vestibular hair cells can serve this unmet need.
+
+In non-mammalian species, hair cells are terminally differentiated, whereas supporting cells, accessory epithelial cells that interdigitate with hair cells within the sensory epithelium, have a remarkable ability to regenerate lost hair cells allowing recovery of auditory and vestibular functions<sup>5–11</sup>. While the capacity to regenerate hair cells is substantially less in the mammalian inner ear, vestibular supporting cells in mammals have retained some latent potential for hair cell regeneration, even at mature ages<sup>12–15</sup>. This contrasts sharply with adult cochlear supporting cells, which show little to no regenerative capacity, likely because their chromatin structure is less accessible around hair cell genes compared to their vestibular counterparts<sup>14</sup>.
+
+While mice can regenerate up to 30% of lost vestibular hair cells after damage, regenerated hair cells appear immature, and it remains unclear if vestibular function can be fully restored<sup>12,16,17</sup>. Nevertheless, multiple reports have demonstrated that this spontaneous regenerative response can be augmented with exogenous factors, with some reporting modest recovery of function<sup>13,18</sup>. Thus, the underlying regenerative machinery in the mammalian vestibular system appears to be intact and manipulatable.
+
+One promising therapeutic target is the basic helix-loop-helix transcription factor, ATOH1, which is both necessary and sufficient for inducing the hair cell fate during normal development<sup>19</sup>. Multiple studies in rodents and human vestibular organ have demonstrated that transgenic or viral overexpression of ATOH1 forces the direct conversion of vestibular supporting cells into new hair cells at adult ages<sup>13,18,20–24</sup>. However, the reported degrees of differentiation of the regenerated hair cells were mixed, with their transcriptomes undetermined. However, improvements in rotarod performance, vestibulo-ocular reflex (VOR) gain, and vestibular evoked potential (VsEP) thresholds have been reported<sup>13,18,22,23</sup>, suggesting some functional recovery.
+
+During hair cell development, ATOH1 is transiently expressed in progenitor cells and immature hair cells and becomes downregulated or silenced as hair cells differentiate<sup>25–27</sup>. In the cochlea, persistent overexpression of ATOH1 in differentiating hair cells delays development and can be cytotoxic<sup>28,29</sup>. ATOH1 is also a potent determinant of cell fate in the cerebellum, spinal cord, and skin<sup>30–32</sup>. Thus, the level, temporal control, and cell type-specific distribution of ATOH1 expression are important.
+
+With the aim of developing an optimized AAV-based ATOH1 gene therapy, we used single-cell RNA-Seq and explored the impact of defined promoters on ATOH1 transgene expression level, timing, effects in distinct cell types, and the maturity of regenerated vestibular hair cells<em>ex vivo</em> and<em>in vivo</em>. We found that the level of ATOH1 overexpression in supporting cells strongly correlates with new hair cell formation, with high levels of ATOH1 expression needed for efficient cell fate conversion. Furthermore, we provide single cell transcriptomic data demonstrating that silencing ATOH transgene enables greater maturation of regenerated hair cells, and that ATOH1-mediated transdifferentation of supporting cells induced cells that resemble endogenous type II hair cells. Finally, we show that off-target ATOH1 expression in cells other than supporting cells can alter their transcriptional identity without hair cell differentiation. Collectively, our results underscore the importance of spatiotemporal transcriptional control of ATOH1 in hair cell regeneration and maturation and should guide future design of ATOH1-based gene therapies.
+
+# RESULTS
+
+The level and duration of ATOH1 transgene expression determine supporting-cell-to-hair-cell conversion in the adult mouse utricle ex vivo
+
+Adenoviral-based overexpression of mouse or human ATOH1 converts supporting cells into new hair cells in cultured utricles from adult mice, rats, and humans<sup>14, 21, 33, 34</sup>. Early work comparing promoters delivered via adenoviral vectors concluded that lower levels of ATOH1 expression resulted in more regenerated hair cells compared to higher levels<sup>34</sup>. To begin to determine how spatiotemporal expression of ATOH1 affects hair cell regeneration, we cultured adult mouse utricle, lesioned hair cells with the aminoglycoside gentamicin, and subsequently transduced them with escalating titers of AAV-CMV-ATOH1 (Fig. S1, S2A-F). We observed a strong, dose-dependent conversion of supporting cells to Pou4f3 + hair cells (Fig. S2 G-H), consistent with previous results<sup>20</sup>.
+
+Higher titers of AAV increased the total number of supporting cells transduced (Fig. S2 G-I), and may contribute to the efficiency of AAV-CMV-ATOH1 in inducing hair cell conversion. We reasoned that gene regulatory elements (e.g., promoters) that exhibit differing strengths and are specific to supporting cells may allow titration of the level of ATOH1 transgene expression at equivalent AAV doses. To identify such regulatory elements, candidate sequences were identified from single-cell genomics data<sup>35</sup>, cloned into AAV constructs driving an H2B-EGFP transgene (Fig. S1), and tested in adult utricle explants at similar doses (AAV8-CMV-H2B-EGFP (8.8x10<sup>11</sup> vg/ml), AAV8-GFAP-H2B-EGFP (1.52x10<sup>13</sup> vg/ml), or AAV8-RLBP1 -H2B-EGFP (8.6x10<sup>12</sup> vg/mL)). We selected the AAV8 serotype as the vector because it has been shown to transduce utricular hair cells and supporting cells at high rates in vivo<sup>20, 36</sup>. We transduced utricle explants from adult mice with AAV8 carrying the regulatory sequences associated with cytomegalovirus (CMV), glial fibrillary acidic protein (GFAP), and retinaldehyde binding protein 1 (RLBP1) for 4 days and found decreasing levels of EGFP intensity within supporting cells among these promoters (Fig. S3 A-D’’’, S5A). When ATOH1-GFP under different promoters was delivered, the CMV promoter induced relatively higher ATOH1-GFP intensity than GFAP and RLBP1 (1.5 and 6.6-fold), and GFAP was 4.4-fold more intense than RLBP1 at 12 days ex vivo. These differences increased to 2.2, 12.1, and 4.4-fold, respectively, at 20 days ex vivo (Fig. 1 I). Furthermore, the CMV promoter drove robust EGFP expression in both hair cells and supporting cells, while with GFAP and RLBP1 expression was mostly restricted to supporting cells, suggesting that these regulatory elements are not active in hair cells (Fig. S3 B-D’’’’).
+
+Having identified a set of regulatory elements driving varying degrees of EGFP expression, we next tested the relationship between levels of ATOH1 transgene expression and hair cell regeneration. Gentamicin-damaged utricle explants<sup>37, 38</sup> were transduced with AAV8-CMV-ATOH1-H2B-EGFP (5.7x10<sup>12</sup> vg/ml), AAV8-GFAP-ATOH1-H2B-EGFP (7.2x10<sup>12</sup> vg/ml), or AAV8-RLBP1-ATOH1-H2B-EGFP (1.4x10<sup>13</sup> vg/mL) and then cultured for an additional 4–13 days to allow time for hair cell regeneration to occur (Fig. 1 A-G, S3E-H’’). Regeneration of hair cells was observed in all three treatment groups, and the level of regeneration increased in the order (CMV > GFAP > RLBP1) similar to their corresponding EGFP intensity (Fig. 1 H-J, S3E-I). The number of hair cells per utricle at the end of the culture period were 1278.8 ± 355.1 for AAV8-CMV-ATOH1, 715.4 ± 437.7 for AAV8-GFAP-ATOH1, and 230.5 ± 113.3 for AAV8-RLBP1-ATOH1 (p < 0.05, one-way ANOVA; Fig. S2 E, 3F-I).
+
+To compare the extent of regeneration over time with each regulatory element, we separately assessed the number of hair cells in utricle explants at five and 13 days after AAV transduction. The mean number of hair cells per utricle increased significantly between the two timepoints for CMV- and GFAP-ATOH1, but not RLBP1-ATOH1 treatment (Fig. 1 B-G, J), indicating that new hair cell generation was still occurring between the 2 timepoints. Furthermore, the number of hair cells remained higher for CMV than GFAP and RLBP1 at both timepoints (Fig. 1 H-J). Taken together, these data suggest that CMV is a stronger promoter than both GFAP and RLBP1 and induced more hair cell regeneration in damaged mouse utricles ex vivo.
+
+## scRNA-seq reveals transcriptome-wide changes during AAV-ATOH1-mediated supporting cell-to-hair cell conversion
+
+To map the transcriptomic changes during supporting cell-to-hair cell conversion in utricles treated with the ATOH1 transgene, gentamicin-damaged utricle explants were transduced with AAV8-CMV-ATOH1, AAV8-GFAP-ATOH1, or AAV8-RLBP1-ATOH1 and then were maintained in culture for 7, 12, or 20 days before being processed for single-cell RNA-Seq (scRNA-Seq; Fig. 2 A). A total of 34,574, 30,428, and 21,891 high-quality cells were collected for each regulatory element, respectively (Fig. S4 A-D). After merging these datasets and using marker gene expression to enrich supporting cells and regenerating hair cells (Fig. S4 A-I), the final dataset consisted of 6,901 cells. UMAP and heatmap visualization of the merged dataset from cells taken at all timepoints revealed six distinct cell clusters that were separated based on gene expression into distinct phases of supporting cell-to-hair cell conversion (Fig. 2 B-B’, D). The six clusters organized by differentiation status along an unbiased pseudotime gradient (Fig. 2 E), with maturation of regenerated hair cells progressing over time in culture (Fig. 2 G). We also found 1 clusters of immature and 2 clusters of mature regenerated hair cells, which we designed as hair cell a and β, suggesting a divergent of cell fates. Compared to 7 days in vitro (DIV) when only immature hair cells were present, more mature hair cells were observed at 12 DIV and 20 DIV (61.3% and 70.2% of HCs, respectively; Fig. 2 G). While these six clusters were observed after transduction with all three regulatory elements, higher proportions of converting supporting cells and mature hair cells were found after treatment with ATOH1 driven by CMV and GFAP than by RLBP1 (Fig. 2 F).
+
+Supporting cells that responded to ATOH1 overexpression (stage 1 converting SC and stage 2 converting SC) initially downregulated known supporting cell genes (Otol1, Otog, Tspan8, and Gsn), and upregulated Notch pathway genes (Hey1, Hes5, Jag1, and Notch1) prior to the onset of expression of early hair cell genes (Pou4f3, Lhx3, Hes6, and Dlk2) (Fig. 3 J, S5E-F and Table S1). A “Notch Activity Score” based on expression of 57 genes associated with the canonical Notch pathway (see Methods, Table S3) showed that Notch pathway activity is the highest in the stage 1 and 2 converting supporting cell clusters (Fig. S5 E-F). As Notch signaling suppresses supporting cells from acquiring a hair cell fate<sup>39–41</sup>, these findings suggest that Notch gene expression becomes upregulated in converting supporting cells possibly as a result of exogenous Atoh1. At 20 DIV, there were substantially higher percentages of stage 1 and 2 converting supporting cells (Fig. 2 G, S5B). Interestingly, many supporting cells did not show any signs of conversion despite displaying levels of ATOH1 transgene expression comparable to converting SCs, with 31.5% of SCs at 20 DIV remaining in the supporting cell cluster (Fig. 2 G, Fig. 3 A, Table S1). These ATOH1-transgene+, non-converting SCs exhibited lower Notch activity scores than stage 1 and 2 converting supporting cells (Fig. S5 E-F).
+
+Consistent with EGFP intensity corresponding to each regulatory element (Fig. 1 B-F), the levels of ATOH1 transgene in supporting cells was similarly ordered (average across all cells, CMV = 18.6 normalized read counts; GFAP = 3.3; RLBP1 = 1.2; Fig. S5 A). Moreover, the percentage of regenerated hair cells (immature hair cells, mature hair cell a and b) increased with regulatory elements associated with higher ATOH1 transgene expression level (CMV = 28.7%; GFAP = 24.9%; RLBP1 = 12.2%; Fig. 2 F).
+
+## ATOH1 transgene expression under control of a supporting-cell-specific regulatory element becomes silenced in regenerated hair cells
+
+Two of the regulatory elements used to drive ATOH1 expression, GFAP and RLBP1, led to limited or no EGFP expression in hair cells (Fig. 1 B-F, S3F-H’’). Based on these results, ATOH1 transgene expression was expected to become downregulated as supporting cells convert into hair cells (Fig. 3 A-D). ATOH1 transgene expression was detected in several clusters (supporting cells, stage 1 and 2 converting supporting cells, immature hair cells, and mature hair cells α) for all three regulatory elements utilized, indicating that they are active during the early stages of supporting cell conversion and hair cell specification (Fig. 3 A). In contrast, the mature hair cell β cluster segregated into distinct groups visible on the UMAP based on regulatory element and ATOH1 transgene level (Fig. 2 B-C, Fig. 3 B). Hair cells from the CMV group almost exclusively comprised the mature hair cell α cluster (99.8% of cells), whereas the GFAP and RLBP1 groups comprised the mature hair cell β cluster (92.9% of cells; Fig. 2 B-C). As predicted, the ATOH1 transgene was expressed at high levels in the mature hair cell α cluster, but it was significantly reduced or absent in mature hair cell β (Fig. 3 A-B), suggested that it is downregulated as regenerated hair cell mature.
+
+ATOH1 autoregulates its own gene expression by binding to regulatory regions in the 3’ enhancer of the endogenous gene<sup>42, 43</sup>. We distinguished Atoh1 transcripts originating from the endogenous locus from vector transcripts via the WPRE element at the 3’ end of the ATOH1 transgene. Interestingly, expression of endogenous Atoh1 was detected in the immature hair cell, mature hair cell α, and mature hair cell β clusters, but not supporting cell or either converting supporting cell clusters (Fig. S5 C-D). This finding suggests that ATOH1 protein produced from the vector transgene induced expression of endogenous Atoh1, and endogenous Atoh1 expression persisted in the mature hair cell β cluster after Atoh1 transgene expression from the GFAP and RLBP1 regulatory elements became silenced.
+
+## Silencing of AAV-Atoh1 transgene from supporting cell-specific regulatory elements promotes maturation of regenerated vestibular hair cells ex vivo
+
+Atoh1 is expressed transiently in hair cells during development, and misexpression of Atoh1 in developing hair cells delays maturation<sup>44, 45</sup>. We compared the transcriptomes of the mature hair cell α and β clusters to assess whether silencing of Atoh1 transgene expression driven by supporting cell-specific regulatory elements promotes maturation. As a measure of maturation, “hair cell maturity indices” were created based on the top 200 mature hair cell and 100 immature hair cell genes that are up- or down-regulated as hair cell differentiate during development (see Methods, Table S3) and were used to score cell clusters along the supporting cell-hair cell axis. The mature hair cell β cluster scored higher on the “Mature hair cell” and lower on the “Immature hair cell” indexes compared to mature hair cell α cluster (Fig. 3 E-F), confirming that the former is more mature at the transcriptome level. The expression of select index genes is shown in Fig. 3 I-J, highlighting the differentially expressed genes in the mature hair cell β cluster including those critical for synaptogenesis (Otof, Slc17a8), stereocilia formation (Espn, Fscn2), and mechanotransduction (Clrn1) (Fig. S5 G-H).
+
+Mammalian vestibular organs contain two types of hair cells, type I and type II, that are found throughout the organ, along with a central striolar region and a peripheral extrastriolar zone<sup>46, 47</sup>. The functional significance of hair cell subtypes and zonal distribution have yet to be fully elucidated, but type I hair cells in the striolar region are associated with fast, phasic responses and type II hair cells with more static, tonic responses<sup>48, 49</sup>. To assess the extent regenerated hair cells differentiate into specialized hair cell subtypes, we developed maturation indices for type I and II hair cells based on the expression of genes which are upregulated in each subtype in the adult mouse utricle (see Methods, Table S3). We found that the mature hair cell β cluster scored higher on the type II hair cell index compared to the mature hair cell α cluster (Fig. 3 G-H). Cells from both the mature HCα and mature HCβ clusters scored relatively low on type I hair cell index, indicating regenerated HCs were not adopting a type I fate ex vivo (Fig. 3 H). These data indicate that the expression of ATOH1 transgene under control of supporting cell-specific regulatory elements becomes silenced in regenerating hair cells as they differentiate, and this reduction in transgene expression promotes further maturation towards a more mature type II HC fate.
+
+## AAV8 displays robust tropism for vestibular supporting cells after IDPN damage in adult mice in vivo
+
+So far, the ex vivo findings indicated that AAV-ATOH1 under control a supporting cell-specific regulatory element enhances maturation of regenerated hair cells, thereby serving as a promising gene therapy candidate for vestibular hair cell regeneration in vivo. We next assessed the in vivo tropism of various capsids in an established model of hair cell damage.
+
+3,3’-iminodipropionitrile (IDPN) is a well-described vestibulotoxin that induces dose-dependent HC death after a single systemic administration to adult mice in vivo<sup>13, 50</sup>. In 4–10-week-old mice, IDPN administration resulted in significant loss of HCs in utricular maculae (67% loss in extrastriola, 84% loss in striola, respectively, p < 0.001) and cristae (61% loss in peripheral zone, 93% loss in central zone, respectively, p < 0.01) within one week (Fig. S6 A-G). Hair cell loss was comparable between the striolar and extrastriolar regions of utricles, and the lesions across both ears were highly symmetric in each mouse (Fig. S6 A-D). Immunolabeling with antibodies to the type II HC marker Sox2 demonstrated that IDPN preferentially killed type I HCs as 96% of the surviving HCs were Sox2+ (Fig. S6 E).
+
+To identify a capsid with optimal supporting cell tropism in IDPN-damaged vestibular organs in vivo, AAV1-, AAV8-, or AAV9-CAG-H2B-EGFP was administered via the posterior semicircular canal (PSC) one day after treatment with saline or IDPN. Two weeks later, vestibular organs were harvested and examined (Fig. S7 A). Analysis of the resulting EGFP expression revealed that supporting cell transduction with AAV8 and AAV9 in utricular maculae and cristae was more robust than AAV1 in both the saline control and IDPN conditions (Fig. S7 B-D, L-N). Furthermore, the transduction with AAV8 and AAV9 in utricular supporting cells and hair cells appeared higher after IDPN damage, reaching > 78% transduction rates for both cell types (Fig. S7 F-G, K-L). On average, AAV8 transduced slightly more supporting cells than AAV9 in the IDPN-treated utricle, with > 90% of supporting cells expressing EGFP in the extrastriola and striola. Of note, no EGFP expression in the contralateral ear was observed in any condition (Fig. S7 E’, H’, O’, and R’), indicating that the contralateral ear could serve as untreated controls for ATOH1 gene therapy.
+
+## AAV8-GFAP-ATOH1 drives supporting cell-specific transgene expression in adult mouse vestibular organs in vivo
+
+Amongst the three regulatory elements tested ex vivo, GFAP was deemed optimal because it led to strong and targeted ATOH1 expression in supporting cells. To further assess the potential utility of GFAP as a regulatory element for ATOH1 gene therapy, its activity profile was assessed in adult mouse utricle and cristae in vivo. Based on the above results, AAV8 was selected as the capsid and AAV8-GFAP-H2B-EGFP was administered locally at two doses (1x10<sup>10</sup> and 3x10<sup>10</sup> vg/mL per ear) via the PSC of adult mice (Fig. S8 A). Two weeks later utricles showed robust EGFP in supporting cells, with little to no detectable EGFP expression in hair cells (Fig. S8 B). No statistically significant difference was observed between the two doses (Fig. S8 C-G). These results indicate that GFAP drove robust EGFP selectively in supporting cells in vivo.
+
+## AAV8-GFAP-ATOH1 induces regeneration of vestibular HCs with stereocilia bundles and presynaptic ribbons in vivo
+
+To assess whether AAV8-GFAP-ATOH1-H2B-EGFP promotes regeneration of vestibular hair cells in IDPN-treated mice in vivo, the virus was administered unilaterally one or 14 days after IDPN treatment. Vestibular organs were harvested and examined 2–4 weeks after AAV administration (Fig. 4 A). Virus injection resulted in robust expression of EGFP in both utricles and cristae, whereas no EGFP was detected in uninjected contralateral ears (Fig. 4 B-D’, G-H, insets). Utricles and cristae from injected ears also displayed more Myosin7a + HCs in comparison to contralateral controls (Fig. 4 B-D’, G-H’’, S10A-C, E-H). Quantification showed significantly more Myosin7a + HCs in both striolar and extrastriolar regions of utricles from injected ears compared to contralateral controls at 2 weeks (70.6 ± 51.5 vs. 26.9 ± 28.0 and 109.7 ± 62.3 vs. 68.6 ± 43.8 cells per 10,000 µm<sup>2</sup> in striola and extrastriola, respectively) and 1 month (142.7 ± 83.5 vs. 47.0 ± 36.7 and 176.4 ± 50.8 vs. 93.7 ± 36.2 cells per 10,000 µm<sup>2</sup> in striola and extrastriola, respectively) (Fig. 4 B-C’’, E and F). Similarly, significantly more regenerated hair cells were observed in the peripheral and central regions of cristae 2 weeks (22.0 ± 8.0 vs. 7.6 ± 3.6 and 16.4 ± 3.6 vs. 1.6 ± 1.7 cells per 10,000 µm<sup>2</sup> in peripheral and central regions, respectively) and 1 month after AAV administration (26.8 ± 7.1 vs. 12.0 ± 7.6 and 18.0 ± 6.4 vs. 2.8 ± 1.5 cells per 10,000 µm<sup>2</sup> in peripheral and central regions, respectively) (Fig. 4 G-H, S10G-I), suggesting that AAV8-GFAP-ATOH1-H2B-EGFP promoted regeneration of vestibular hair cells. Quantification of Myosin7a + hair cells 2 weeks after AAV injection confirmed a significant, dose-dependent, increase in the number of hair cells in both striolar and extrastriolar regions of injected ears compared to contralateral control ears (Fig. S9 A-F). At the high dose, 912.4 ± 144.7 new HCs were added to the utricle (Fig. S9 G), suggesting that about 57.9% of the degenerated HCs had been replaced. In addition, the numbers of Myosin7a + HCs increased with time in both the striolar (70.6 ± 51.5 vs. 142.7 ± 83.5 cells per 10,000 µm<sup>2</sup> at 2 weeks and 1 month, respectively) and extrastriolar (109.7 ± 62.3 vs. 176.4 ± 50.8 cells 10,000 µm<sup>2</sup> at 2 weeks and 1 month, respectively) regions of the utricle (Fig. 4 E-F, S10D).
+
+Hair cells require stereociliary bundles for mechanoreception and synapses to relay information centrally via afferent neurons<sup>51, 52</sup>. One month after AAV8-GFAP-ATOH1-H2B-EGFP administration to IDPN-treated mice, the density of phalloidin-labeled bundles throughout the utricle of injected ears were 1.8-fold higher than contralateral control ears (p < 0.05; Fig. 4 I-K).
+
+Ribbon synapses form along the basolateral surface of vestibular hair cells and are identifiable with the presynaptic marker Ctbp2<sup>53</sup>. Surprisingly, utricular HCs in uninjected ears from IDPN-treated mice showed a significant increase in the density of Ctbp2-labeled puncta relative to HCs from naive, undamaged mice (Fig. 4 L-O). This phenomenon has been reported in surviving hair cells before<sup>54</sup>, but whether it resulted from damage-induced fragmentation of pre-existing synapses, an additive process, or remodeling was unclear. In contrast, utricular HCs from ears administered AAV8-GFAP-ATOH1-H2B-EGFP showed comparable Ctpb2 puncta density relative to naive HCs (Fig. 4 O). Among 144 hair cells examined in the virus-treated utricle (n = 6), all of them displayed Ctbp2 + puncta, suggesting that they expressed components needed for synaptic transmission.
+
+## AAV8-GFAP-ATOH1 induces regeneration of mature type II HCs in vivo
+
+We further characterized in vivo vestibular hair regeneration after AAV8-GFAP-ATOH1-H2B-EGFP using scRNA-seq. Utricles and cristae were collected from IDPN-damaged mice treated with AAV8-GFAP-ATOH1-H2B-EGFP (IDPN and GFAP-ATOH1), IDPN-damaged mice (IDPN), or naive mice (Control; Fig. 5 A, S13A) at 2 weeks and 4 weeks post-treatment (4 weeks and 6 weeks after IDPN administration). The organs were pooled by group, single cells were isolated and processed for scRNA-seq. A total of 12,130; 11,111; and 15,146 cells were collected for the control, IDPN, and IDPN and GFAP-ATOH1 treated utricles, respectively, as well as 18,102; 27,072; and 32,927 cells for the cristae (from 6 mice). Marker gene identification in unbiased clusters of the merged groups revealed rich datasets that included all the various cell types that constitute the utricle and crista: hair cells, supporting cells, transitional epithelial cells (TECs), dark cells, melanocytes, mesenchymal cells, glia, endothelial cells, smooth muscle cells, and immune cells (Fig. S11 E-I and S14E-I). In the GFAP-ATOH1-treated utricle and crista, ATOH1 transgene was detected in mesenchymal cells, TECs, supporting cells, and supporting cells presumably converting into hair cells (Fig. S11 J and S14J). GFAP regulatory element activity in mesenchyme was expected as GFAP expression has been reported<sup>55</sup>.
+
+To examine supporting cells converting into hair cells, supporting cell and hair cell transcriptomes from all groups were examined and re-clustered (8,399 cells from the utricles and 7,855 cells from the cristae). Unbiased clustering identified six clusters in both the utricle and cristae (Fig. 5 B, D and S13B, D). Marker gene expression revealed three of the clusters were supporting cells, mature type I and II hair cells. Supporting cells and type II hair cells were present in similar proportions for all three treatment groups (Fig. 5 F, S13F). In contrast, type I hair cells were scant in the IDPN only and IDPN and GFAP-ATOH1 groups, consistent with the finding that IDPN primarily ablates type I hair cells (Fig. 5 F, S6E and S13F)<sup>56</sup>.
+
+The other three clusters in the utricle primarily comprised cells from the IDPN and IDPN + GFAP-ATOH1 groups. Gene expression analysis indicated that these groups contained cells representing various stages of supporting cell-to-hair cell conversion, thereby revealing the transcriptional map along the supporting cell-to-hair cell axis after in vivo treatment with GFAP-ATOH1. The cells were then ordered along a pseudotime trajectory and were designated stage 1–3 regenerating hair cells, with stage 1 being the earliest stage of conversion and stage 3 being the most mature state (Fig. 5 B and E). Out of Stage 1–3 regenerated hair cells, 86.2% came from utricles treated with IDPN + GFAP-ATOH1 (Fig. 5 F). These clusters were characterized by hair cell genes known to be expressed during native development and were relatively more mature than converting supporting cells previously described ex vivo (Fig. 5 D, Table S4). Compared to undamaged controls, there was an increase in the proportion of stage 1–3 hair cells in the IDPN only group, consistent with a low level of spontaneous regeneration after IDPN damage alone<sup>13</sup>. Further, there were substantially more regenerating hair cells with ATOH1 overexpression, than after IDPN alone (Fig. 5 F). Similar to the ex vivo results, ATOH1 transgene expression levels were highest in supporting cells and stage 1 regenerated hair cells, and then were gradually downregulated in stage 2 and 3 regenerated hair cells (Fig. 6 A-B). Thus, like ex vivo, the GFAP regulatory element also became silenced as supporting cells convert into hair cells in vivo (Fig. 6 A-B). As the ATOH1 transgene became downregulated, the transcriptomes of the stage 1–3 regenerated HCs scored higher in maturity indices, suggesting a more mature phenotype (Fig. 6 C-D, G-H). In addition, based on type I and type II HC indices, regenerating hair cells primarily adopted a type II fate as they progressed from stage 1 to 3 (Fig. 6 E-F). No expression of type I hair cell marker genes or increase in type I index score was detected as regenerating HCs differentiated, indicating that ATOH1 overexpression mainly induces a type II hair cell fate in vivo (Fig. 6 E-F).
+
+Analysis of the 7,855 supporting cells and hair cells isolated from untreated, IDPN, and IDPN and ATOH1-treated cristae revealed six clusters that are comparable to those found in the utricle (Fig. S13 and S14). In cristae, ATOH1 transgene expression decreased along the trajectory from supporting cells to hair cells (Fig. S15 A-B), and hair cells maturity score increased (Figures S15 C-D, G-H). At the transcriptome level, newly regenerated hair cells in the cristae also resemble type II hair cells more than type I (Fig. S15 E-F).
+
+In both the utricle and crista, the expression of Notch pathway genes is highest in supporting cells and stage 1 regenerated HCs and is decreased as maturity increases (Fig. S12 C-D and S16C-D). This agrees with the ex vivo results, where Notch activity decreases as hair cells mature.
+
+In summary, these data demonstrate that use of the supporting cell-specific promoter GFAP to drive transient ATOH1 overexpression is effective in enhancing maturation of the regenerated type II-like HCs.
+
+# DISCUSSION
+
+Vestibular hair cells are required for detecting head acceleration and rotation. Their loss causes vestibular hypofunction. Because there is minimal spontaneous regeneration in the mature, mammalian vestibular organs, vestibular hypofunction is deemed irreversible. Independent studies have found that ATOH1 overexpression either via transgenic approaches or viral transduction increase hair cell regeneration in the damaged mature mouse utricle<sup>13, 18, 20</sup>. However, regenerated hair cells failed to fully mature, in part because these approaches have employed constitutive overexpression of ATOH1. As developing hair cells naturally downregulate ATOH1<sup>25–27</sup> and ATOH1 overexpression has been reported to impede maturation of cochlear hair cells<sup>28, 29</sup>, we have engineered an ATOH1 transgene to be driven by supporting cell promoters to test whether transient overexpression of ATOH1 can induce both hair cell regeneration and subsequent maturation. In damaged utricles *ex vivo* and *in vivo*, we found that constitutive and ubiquitous overexpression of ATOH1 using AAV8-CMV-ATOH1 induced more regenerated hair cells than AAV8-GFAP-ATOH1 and AAV8-RLBP1-ATOH1. In addition, AAV8-GFAP-ATOH1 and AAV8-RLBP1-ATOH1 achieved transient overexpression of ATOH1, resulting in supporting cells that downregulate Notch signaling, upregulate endogenous ATOH1, and convert into hair cells that display transcriptome consistent with maturing type II hair cells. Together, our results suggest that transient ATOH1 expression is a promising approach to initiate mature vestibular hair cell regeneration. Our results are also consistent with a similar approach that transiently overexpresses ATOH1 in the cochlea<sup>57</sup>.
+
+## Responsiveness of supporting cells to ATOH1 overexpression
+
+Relative to cochlear supporting cells, vestibular supporting cells in mature mice are more responsive to singular overexpression of the hair cell transcription factor ATOH1<sup>14,58</sup>. In our experiments with AAV8-GFAP-ATOH1, we observed different stages of converting supporting cells and maturing hair cells that expressed the ATOH1 transgene, clearly indicating that supporting cells responded to AAV8-GFAP-ATOH1 and acquired a hair cell fate. One interpretation that there are different stages of converting supporting cells and maturing hair cells is that supporting cell-to-hair cell conversion was asynchronized. Alternatively, supporting cell-to-hair cell conversion may occur in parallel but at different rates. Both possibilities are supported by the findings that the hair cell number was higher at the later than earlier time points *in vivo*.
+
+While it is expected that mesenchymal cells do not convert and become hair cells despite expression of the ATOH1 transgene, some supporting cells also failed to convert despite expression of the ATOH1 transgene. Considering Notch signaling is known to govern lateral inhibition in the mature utricle<sup>40, 59</sup>, one explanation is that there is active Notch signaling in supporting cells as a result of surviving hair cells, which are present in both the gentamicin and IDPN damage paradigms. One may speculate that regions with higher density of surviving hair cells will contain supporting cells expressing greater levels of Notch target genes, which we have observed in a subset of supporting cells in our single-cell dataset. We postulate that only supporting cells with high levels of exogenous and endogenous *ATOH1* can overcome the inhibitory notch activity and drive transdifferentiation of supporting cells to mature hair cells, whereas supporting cells with lower *ATOH1* levels, would be unable to overcome Notch signaling and as such would remain in a supporting cell state.
+
+## Differentiation of regenerated hair cells
+
+The mature mammalian vestibular organs harbor type I and II hair cells which feature distinct cell body and bundle morphology, transcriptomes, basolateral potassium currents, and innervation patterns<sup>46, 47</sup>. While there are reports claiming relative importance of each hair cell subtype, both are likely important for the overall organ function. After damage of the mouse utricle, spontaneously regenerated hair cells demonstrated features of immature type I and II hair cells<sup>13, 16, 17</sup>. Our study builds on the previous study<sup>60</sup> and provides a comprehensive transcriptomic atlas of sensory and non-sensory cell types of the mature mouse utricle, and a new transcriptomic atlas of the mature mouse cristae. Using transcriptomes of native mature type I and II hair cells and supporting cells as benchmarks, we found that AAV8-GFAP-ATOH1 induces robust supporting cell-to-hair cell conversion. Remarkably, regenerated hair cells mature as they downregulate the ATOH1 transgene and, transiently upregulate endogenous ATOH1, and express mature hair cell genes including those representing the mature type II hair cells. These results are consistent with the previous report that ATOH1 overexpression resulted in primarily regenerated type II hair cells<sup>13</sup>. Moreover, AAV8-GFAP-ATOH1 treatment resulted in an increase of bundle-bearing hair cells and regenerated hair cells displaying a subset of bundle genes. Lastly, our study shows that regenerated hair cells after AAV8-GFAP-ATOH1 treatment express some synapse genes and treated organs containing hair cells with synapse counts comparable to native hair cells. These results suggest that the regenerated hair cells can mature partially and may display some features necessary for mechantransduction and synaptic transmission. Since damaged hair cells have been reported to self-repair<sup>54</sup>, these results should be further validated with fate-mapping experiments.
+
+In summary, our study describes the effectiveness and limitation of gene therapy using supporting cell promoters to drive transient overexpression of ATOH1 in vestibular hair cell regeneration. Our results provide a high-resolution transcriptomic atlas of the mature mouse vestibular organs and a promising approach to regenerate mature type II hair cells.
+
+# References
+
+1. Tsuji, K. et al. Temporal bone studies of the human peripheral vestibular system. Aminoglycoside ototoxicity. Ann Otol Rhinol Laryngol Suppl 181, 20–25, doi: 10.1177/00034894001090s504 (2000).
+2. Rauch, S. D., Velazquez-Villasenor, L., Dimitri, P. S. & Merchant, S. N. Decreasing hair cell counts in aging humans. Ann N Y Acad Sci 942, 220–227, doi: 10.1111/j.1749-6632.2001.tb03748.x (2001).
+3. Agrawal, Y., Carey, J. P., Della Santina, C. C., Schubert, M. C. & Minor, L. B. Disorders of balance and vestibular function in US adults: data from the National Health and Nutrition Examination Survey, 2001–2004. Arch Intern Med 169, 938–944, doi: 10.1001/archinternmed.2009.66 (2009).
+4. Sayyid, Z. N., Kim, G. S. & Cheng, A. G. Molecular therapy for genetic and degenerative vestibular disorders. Curr Opin Otolaryngol Head Neck Surg 26, 307–311, doi: 10.1097/MOO.0000000000000477 (2018).
+5. Corwin, J. T. Postembryonic production and aging in inner ear hair cells in sharks. J Comp Neurol 201, 541–553, doi: 10.1002/cne.902010406 (1981).
+6. Popper, A. N. & Hoxter, B. Growth of a fish ear: 1. Quantitative analysis of hair cell and ganglion cell proliferation. Hear Res 15, 133–142, doi: 10.1016/0378-5955(84)90044-3 (1984).
+7. Jorgensen, J. M. & Mathiesen, C. The avian inner ear. Continuous production of hair cells in vestibular sensory organs, but not in the auditory papilla. Naturwissenschaften 75, 319–320, doi: 10.1007/BF00367330 (1988).
+8. Corwin, J. T. & Cotanche, D. A. Regeneration of sensory hair cells after acoustic trauma. Science 240, 1772–1774, doi: 10.1126/science.3381100 (1988).
+9. Ryals, B. M. & Rubel, E. W. Hair cell regeneration after acoustic trauma in adult Coturnix quail. Science 240, 1774–1776, doi: 10.1126/science.3381101 (1988).
+10. Dooling, R. J., Ryals, B. M. & Manabe, K. Recovery of hearing and vocal behavior after hair-cell regeneration. Proc Natl Acad Sci U S A 94, 14206–14210, doi: 10.1073/pnas.94.25.14206 (1997).
+11. Smolders, J. W. Functional recovery in the avian ear after hair cell regeneration. Audiol Neurootol 4, 286–302, doi: 10.1159/000013853 (1999).
+12. Golub, J. S. et al. Hair cell replacement in adult mouse utricles after targeted ablation of hair cells with diphtheria toxin. J Neurosci 32, 15093–15105, doi: 10.1523/JNEUROSCI.1709-12.2012 (2012).
+13. Sayyid, Z. N., Wang, T., Chen, L., Jones, S. M. & Cheng, A. G. Atoh1 Directs Regeneration and Functional Recovery of the Mature Mouse Vestibular System. Cell Rep 28, 312–324 e314, doi: 10.1016/j.celrep.2019.06.028 (2019).
+14. Jen, H. I. et al. Transcriptomic and epigenetic regulation of hair cell regeneration in the mouse utricle and its potentiation by Atoh1. Elife 8, doi: 10.7554/eLife.44328 (2019).
+15. Kawamoto, K., Izumikawa, M., Beyer, L. A., Atkin, G. M. & Raphael, Y. Spontaneous hair cell regeneration in the mouse utricle following gentamicin ototoxicity. Hear Res 247, 17–26, doi: 10.1016/j.heares.2008.08.010 (2009).
+16. Gonzalez-Garrido, A. et al. The Differentiation Status of Hair Cells That Regenerate Naturally in the Vestibular Inner Ear of the Adult Mouse. J Neurosci 41, 7779–7796, doi: 10.1523/JNEUROSCI.3127-20.2021 (2021).
+17. Wang, T. et al. Uncoordinated maturation of developing and regenerating postnatal mammalian vestibular hair cells. PLoS Biol 17, e3000326, doi: 10.1371/journal.pbio.3000326 (2019).
+18. Schlecker, C. et al. Selective atonal gene delivery improves balance function in a mouse model of vestibular disease. Gene Ther 18, 884–890, doi: 10.1038/gt.2011.33 (2011).
+19. Atkinson, P. J., Huarcaya Najarro, E., Sayyid, Z. N. & Cheng, A. G. Sensory hair cell development and regeneration: similarities and differences. Development 142, 1561–1571, doi: 10.1242/dev.114926 (2015).
+20. Guo, J. Y. et al. AAV8-mediated Atoh1 overexpression induces dose-dependent regeneration of vestibular hair cells in adult mice. Neurosci Lett 747, 135679, doi: 10.1016/j.neulet.2021.135679 (2021).
+21. Shou, J., Zheng, J. L. & Gao, W. Q. Robust generation of new hair cells in the mature mammalian inner ear by adenoviral expression of Hath1. Mol Cell Neurosci 23, 169–179, doi: 10.1016/s1044-7431(03)00066-6 (2003).
+22. Staecker, H., Praetorius, M., Baker, K. & Brough, D. E. Vestibular hair cell regeneration and restoration of balance function induced by math1 gene transfer. Otol Neurotol 28, 223–231, doi: 10.1097/MAO.0b013e31802b3225 (2007).
+23. Staecker, H. et al. Optimizing atoh1-induced vestibular hair cell regeneration. Laryngoscope 124 Suppl 5, S1-S12, doi: 10.1002/lary.24775 (2014).
+24. Taylor, R. R. et al. Characterizing human vestibular sensory epithelia for experimental studies: new hair bundles on old tissue and implications for therapeutic interventions in ageing. Neurobiol Aging 36, 2068–2084, doi: 10.1016/j.neurobiolaging.2015.02.013 (2015).
+25. Yang, H., Xie, X., Deng, M., Chen, X. & Gan, L. Generation and characterization of Atoh1-Cre knock-in mouse line. Genesis 48, 407–413, doi: 10.1002/dvg.20633 (2010).
+26. Pan, N. et al. A novel Atoh1 "self-terminating" mouse model reveals the necessity of proper Atoh1 level and duration for hair cell differentiation and viability. PLoS One 7, e30358, doi: 10.1371/journal.pone.0030358 (2012).
+27. Liu, Z. et al. Age-dependent in vivo conversion of mouse cochlear pillar and Deiters' cells to immature hair cells by Atoh1 ectopic expression. J Neurosci 32, 6600–6610, doi: 10.1523/JNEUROSCI.0818-12.2012 (2012).
+28. Zheng, J. L. & Gao, W. Q. Overexpression of Math1 induces robust production of extra hair cells in postnatal rat inner ears. Nat Neurosci 3, 580–586, doi: 10.1038/75753 (2000).
+29. Gubbels, S. P., Woessner, D. W., Mitchell, J. C., Ricci, A. J. & Brigande, J. V. Functional auditory hair cells produced in the mammalian cochlea by in utero gene transfer. Nature 455, 537–541, doi: 10.1038/nature07265 (2008).
+30. Ben-Arie, N. et al. Math1 is essential for genesis of cerebellar granule neurons. Nature 390, 169–172, doi: 10.1038/36579 (1997).
+31. Miesegaes, G. R. et al. Identification and subclassification of new Atoh1 derived cell populations during mouse spinal cord development. Dev Biol 327, 339–351, doi: 10.1016/j.ydbio.2008.12.016 (2009).
+32. Wright, M. C. et al. Unipotent, Atoh1 + progenitors maintain the Merkel cell population in embryonic and adult mice. J Cell Biol 208, 367–379, doi: 10.1083/jcb.201407101 (2015).
+33. Taylor, R. R. et al. Regenerating hair cells in vestibular sensory epithelia from humans. Elife 7, doi: 10.7554/eLife.34817 (2018).
+34. Praetorius, M. et al. Adenovector-mediated hair cell regeneration is affected by promoter type. Acta Otolaryngol 130, 215–222, doi: 10.3109/00016480903019251 (2010).
+35. Jan, T. A. et al. Spatiotemporal dynamics of inner ear sensory and non-sensory cells revealed by single-cell transcriptomics. Cell Rep 36, 109358, doi: 10.1016/j.celrep.2021.109358 (2021).
+36. Wang, G. P. et al. Adeno-associated virus-mediated gene transfer targeting normal and traumatized mouse utricle. Gene Ther 21, 958–966, doi: 10.1038/gt.2014.73 (2014).
+37. Cunningham, L. L., Cheng, A. G. & Rubel, E. W. Caspase activation in hair cells of the mouse utricle exposed to neomycin. J Neurosci 22, 8532–8540, doi: 10.1523/JNEUROSCI.22-19-08532.2002 (2002).
+38. Forge, A., Li, L. & Nevill, G. Hair cell recovery in the vestibular sensory epithelia of mature guinea pigs. J Comp Neurol 397, 69–88 (1998).
+39. Collado, M. S. et al. The postnatal accumulation of junctional E-cadherin is inversely correlated with the capacity for supporting cells to convert directly into sensory hair cells in mammalian balance organs. J Neurosci 31, 11855–11866, doi: 10.1523/JNEUROSCI.2525-11.2011 (2011).
+40. Lin, V. et al. Inhibition of Notch activity promotes nonmitotic regeneration of hair cells in the adult mouse utricles. J Neurosci 31, 15329–15339, doi: 10.1523/JNEUROSCI.2057-11.2011 (2011).
+41. Wang, G. P. et al. Notch signaling and Atoh1 expression during hair cell regeneration in the mouse utricle. Hear Res 267, 61–70, doi: 10.1016/j.heares.2010.03.085 (2010).
+42. Helms, A. W., Abney, A. L., Ben-Arie, N., Zoghbi, H. Y. & Johnson, J. E. Autoregulation and multiple enhancers control Math1 expression in the developing nervous system. Development 127, 1185–1196, doi: 10.1242/dev.127.6.1185 (2000).
+43. Shi, F., Cheng, Y. F., Wang, X. L. & Edge, A. S. Beta-catenin up-regulates Atoh1 expression in neural progenitor cells by interaction with an Atoh1 3' enhancer. J Biol Chem 285, 392–400, doi: 10.1074/jbc.M109.059055 (2010).
+44. Kelly, M. C., Chang, Q., Pan, A., Lin, X. & Chen, P. Atoh1 directs the formation of sensory mosaics and induces cell proliferation in the postnatal mammalian cochlea in vivo. J Neurosci 32, 6699–6710, doi: 10.1523/JNEUROSCI.5420-11.2012 (2012).
+45. Cai, T. et al. Characterization of the transcriptome of nascent hair cells and identification of direct targets of the Atoh1 transcription factor. J Neurosci 35, 5870–5883, doi: 10.1523/JNEUROSCI.5083-14.2015 (2015).
+46. Desai, S. S., Zeh, C. & Lysakowski, A. Comparative morphology of rodent vestibular periphery. I. Saccular and utricular maculae. J Neurophysiol 93, 251–266, doi: 10.1152/jn.00746.2003 (2005).
+47. Desai, S. S., Ali, H. & Lysakowski, A. Comparative morphology of rodent vestibular periphery. II. Cristae ampullares. J Neurophysiol 93, 267–280, doi: 10.1152/jn.00747.2003 (2005).
+48. Eatock, R. A. & Songer, J. E. Vestibular hair cells and afferents: two channels for head motion signals. Annu Rev Neurosci 34, 501–534, doi: 10.1146/annurev-neuro-061010-113710 (2011).
+49. Contini, D. et al. Intercellular K(+) accumulation depolarizes Type I vestibular hair cells and their associated afferent nerve calyx. Neuroscience 227, 232–246, doi: 10.1016/j.neuroscience.2012.09.051 (2012).
+50. Soler-Martin, C., Diez-Padrisa, N., Boadas-Vaello, P. & Llorens, J. Behavioral disturbances and hair cell loss in the inner ear following nitrile exposure in mice, guinea pigs, and frogs. Toxicol Sci 96, 123–132, doi: 10.1093/toxsci/kfl186 (2007).
+51. Pacentine, I., Chatterjee, P. & Barr-Gillespie, P. G. Stereocilia Rootlets: Actin-Based Structures That Are Essential for Structural Stability of the Hair Bundle. Int J Mol Sci 21, doi: 10.3390/ijms21010324 (2020).
+52. Nouvian, R., Beutner, D., Parsons, T. D. & Moser, T. Structure and function of the hair cell ribbon synapse. J Membr Biol 209, 153–165, doi: 10.1007/s00232-005-0854-4 (2006).
+53. Schug, N. et al. Differential expression of otoferlin in brain, vestibular system, immature and mature cochlea of the rat. Eur J Neurosci 24, 3372–3380, doi: 10.1111/j.1460-9568.2006.05225.x (2006).
+54. Kim, G. S. et al. Repair of surviving hair cells in the damaged mouse utricle. Proc Natl Acad Sci U S A 119, e2116973119, doi: 10.1073/pnas.2116973119 (2022).
+55. Rio, C., Dikkes, P., Liberman, M. C. & Corfas, G. Glial fibrillary acidic protein expression and promoter activity in the inner ear of developing and adult mice. J Comp Neurol 442, 156–162, doi: 10.1002/cne.10085 (2002).
+56. Llorens, J. & Dememes, D. Hair cell degeneration resulting from 3,3'-iminodipropionitrile toxicity in the rat vestibular epithelia. Hear Res 76, 78–86, doi: 10.1016/0378-5955(94)90090-6 (1994).
+57. Bi, Z. et al. Development and transdifferentiation into inner hair cells require Tbx2. Natl Sci Rev 9, nwac156, doi: 10.1093/nsr/nwac156 (2022).
+58. Atkinson, P. J., Kim, G. S. & Cheng, A. G. Direct cellular reprogramming and inner ear regeneration. Expert Opin Biol Ther 19, 129–139, doi: 10.1080/14712598.2019.1564035 (2019).
+59. Collado, M. S., Burns, J. C., Hu, Z. & Corwin, J. T. Recent advances in hair cell regeneration research. Curr Opin Otolaryngol Head Neck Surg 16, 465–471, doi: 10.1097/MOO.0b013e32830f4ab5 (2008).
+60. McInturff, S., Burns, J. C. & Kelley, M. W. Characterization of spatial and temporal development of Type I and Type II hair cells in the mouse utricle using new cell-type-specific markers. Biol Open 7, doi: 10.1242/bio.038083 (2018).
+61. Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat Methods 16, 1289–1296, doi: 10.1038/s41592-019-0619-0 (2019).
+62. Johnson, S. L., Safieddine, S., Mustapha, M. & Marcotti, W. Hair Cell Afferent Synapses: Function and Dysfunction. Cold Spring Harb Perspect Med 9, doi: 10.1101/cshperspect.a033175 (2019).
+63. Stuart, T. et al. Comprehensive Integration of Single-Cell Data. Cell 177, 1888–1902 e1821, doi: 10.1016/j.cell.2019.05.031 (2019).
+
+# Supplementary Files
+
+- [SuppFigure1vectormapV201.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/cc6353f795c5071e45d0c1f7.png)  
+  Supplementary figure 1. Vector map. A) Schematic showing the vector map of all used virus. AAV carrying mouse Atoh1 was used for *ex vivo* experiments, and AAV carrying human ATOH1 for *in vivo* experiments.
+
+- [Suppfigure2relatedtofigure1V701.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/80b62136328dd5ad431555af.png)  
+  Supplementary figure 2. AAV-CMV-ATOH1 induced regeneration of hair cells in a dose-dependent manner. A) Gentamicin ablated hair cells in a dose-dependent manner in adult mouse utricle *ex vivo*. Utricles were cultured with two different doses (0.1 and 1mM) of gentamicin for 1 day, then examined 3 days after. B-D) Fewer hair cells (Myosin7a, magenta and Pou4f3, gray) in damaged utricles compared to control. Few hair cells survived after 1mM gentamicin treatment. E) Pou4f3+ hair cells decreased in a dose-dependent manner, r²=0.98. F) Utricles were treated with 1mM gentamicin for 1 day, followed by 3 days of different doses of AAV1-CMV-ATOH1-H2B-EGFP treatment. G) More hair cells (Pou4f3, gray) were observed with increasing dose of CMV-ATOH1. GFP signal (inserts in G) also increased. H) Pou4f3+ hair cells increased in a dose-dependent manner, r²=0.92. I) No significant change in Sall2+ supporting cells in different viral titers, r²=0.28. Nonlinear regression (curve fit) was used. Scale bar: B-D, G 100mm.
+
+- [Suppfigure3relatedtofigure1V201.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/6a117de390a293699770b6b9.png)  
+  Supplementary figure 3. Supporting-cell-specific promoters driving ATOH1 regenerated hair cells in adult mouse utricle *ex vivo*. A) Utricles from 4-10-week-old mice were first cultured with 4x10¹¹ vg/mL AAV8-CMV-H2B-EGFP, AAV8-GFAP-H2B-EGFP or AAV8-RLBP1-H2B-EGFP for 3 days then in virus free medium for additional 4 days and examined at day 7 after treatment. B-D) GFP signals after each treatment. GFP signal is more intense in CMV promoter treated utricles compared to supporting-cell-specific-promoters GFAP and RLBP1. Inserts are the gain-adjusted GFP signals. B’-D’’’) High magnification pictures of hair cell (Pou4f3, magenta) layer and supporting cell (Sall2, red) layer in box of B-D. E) Utricles from 4-10-week-old-mice were treated with 1mM gentamicin followed by CMV-ATOH1, GFAP-ATOH1 or RLBP1-ATOH1 treatment, then in virus free medium for additional 7 days and examined at day 9. F-H’) There were more hair cells (Pou4f3, magenta) in utricles treated with CMV-ATOH1 and GFAP-ATOH1 than RLBP1-ATOH1. GFP signals were more intense in CMV-ATOH1 treated utricle. Insets shown the GFP signals separately. F’’-H’’) Shown gain adjusted GFP signal. I) Quantification showing that hair cell number increased significantly with CMV-ATOH1 compared to RLBP1-ATOH1 treatment. Data shown as mean±S.D., compared using one-way ANOVA. ***p<0.001. Scale bars: B-D, F-H’’ 100mm, B’-D’’’’ 20mm.
+
+- [SuppFigure4scRNAinvitrorelatedtofigure2V701.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/7e2d5a3a87ea256b76d6a6a3.png)  
+  Supplementary figure 4. Single-cell transcriptomes of regenerating utricles *ex vivo*. A-D) UMAP plots showing total read counts, features (genes), mitochondrial and ribosomal gene percentages per cell. E) UMAP plot of integrated dataset of 3 different virus treatment from all timepoints showing a total 86,893 cells. Twenty-three cell clusters were identified. F) UMAP plot of integrating cells showing 9 inferred cell type, and marker genes were used to annotate the sensory and non-sensory cell types. G) Heatmap showing differentially expressed genes among the 23 cell clusters. The top differentially-expressed genes of each cluster are shown on the right. A full list of these genes is found in Table S2. The heatmap is colored by relative expression from -2 (magenta) to 2 (yellow). H-I) Expression of established markers of hair cells (*Lhx3*) and supporting cell (*Agr3*) within the cell clusters. J-K) UMAP plots showing ATOH1-EGFP-transgene and endogenous *ATOH1* expression.
+
+- [SuppFigure5scRNAinvitrorelatedtofigure3V1001.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/c165cc1ca4fc7685c903da22.png)  
+  Supplementary figure 5. Single-cell transcriptomes of regenerating hair cells *ex vivo*. A) Violin plot of ATOH1-EGFP-transgene expression level after CMV, GFAP or RLBP1-ATOH1 treatment. B) The fraction of cells from different timepoints with GFAP and RLBP1-ATOH1 treatment assigned to each cell cluster. C) Violin plot of endogenous *ATOH1* expression level in 6 different cell clusters. D) UMAP plot of endogenous *ATOH1* expression level. E) Ridge plots of notch activity score of 6 different cell clusters. F) Dot plots showing top 13 highly enriched notch activity genes from each cell cluster. G) Dot plots showing top 6 highly enriched bundle genes from each cell cluster. J) Dot plots showing top 4 highly enriched synapse genes from each cell cluster.
+
+- [Suppfigure6IDPNdamagemodelV401.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/526779ff286a2ad7d82cdf8a.png)  
+  Supplementary figure 6. IDPN damage model *in vivo*. A) Utricles from 4-10-week-old mice were injected with IDPN (4 mg/g), organs were examined 7 days after IDPN injection. B-C’’) Fewer hair cells were noted in both the extrastriolar and striolar regions 7 days after IDPN treatment compared to saline control. D) Quantification showing significantly fewer hair cells in IDPN treated utricles. E) Few type I hair cells survived after IDPN treatment. F-G) Fewer hair cells were observed in cristae after IDPN treatment. Data shown as mean±S.D., compared using unpaired Student’s t-tests. ***p<0.001. Scale bars: B-C’’, F-G 100mm, B’-C’’ 20mm.
+
+- [Suppfigure7serotypeofAAVV301.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/7178ab939c0c7fad3520bc2b.png)  
+  Supplementary figure 7. Tropism of different AAV serotypes in vestibular organs *in vivo*. A) 4-10-week-old mice were injected with IDPN first, then treated with different AAV serotypes carrying CAG-H2B-EGFP. Organs were analyzed 2 weeks after treatment. B-E’) AAV8 and AAV9 has higher transduction efficiency compared to AAV1 in undamaged utricles. No GFP signal was observed in contralateral ear. F-H’) AAV8 and AAV9 have high transduction efficiency after IDPN induced hair cell loss. No GFP signal was observed in contralateral ear. I-J’) Representative high magnification images showed GFP⁺ hair cells and supporting cells in both saline and IDPN treated utricles. K-L) Relative to AAV1, AAV8 and AAV9 transduced significantly more supporting cells and hair cells in undamaged utricles. With IDPN treatment, AAV8 and AAV9 transduced most of hair cell and supporting cells, significantly higher than saline controls. L-O’) AAV8 and AAV9 has higher transduction efficiency in undamaged cristae compared to AAV1. No GFP signal was observed in contralateral ear. P-R’) AAV8 and AAV9 have high transduction efficiency in cristae after IDPN treatment. No GFP signal was observed in contralateral ear. Data shown as mean±S.D., compared using one-way ANOVA. *p<0.05, **p<0.01, ***p<0.001. Scale bars: B-H’, L-R’ 100mm, I-J’ 20mm.
+
+- [Suppfigure8AAV8CMVGFPhighandlowdoseV201.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/9af57539651cdd5b112e9f98.png)  
+  Supplementary figure 8. Characterizing the effects of dose on the transduction efficiency of AAV8-GFAP-H2B-EGFP. A) High dose (3.0x10¹⁰ vg/ml) and low dose (1.0x10¹⁰ vg/ml) of AAV8-GFAP-H2B-EGFP were given into the inner ear of 4-10-week-old mice. Utricles were analyzed 2 weeks later. B) Immunohistochemistry showed GFAP promoter restricts transgene expression to supporting cells. Both hair cells and supporting cells were transduced with the CMV promoter. C-D’) GFP signal was observed in both high dose and low dose AAV8-GFAP-H2B-EGFP treatment. E-F) Few hair cells and many supporting cells were GFP positive. There was no significant difference between the two doses. F) GFP signal intensity in supporting cells has no significant difference in high and low doses treatment groups. Data shown as mean±S.D., compared using unpaired Student’s t-tests. Scale bars: C-D’, 100mm.
+
+- [Suppfigure9AAV8GFAPAtoh1highandlowdoseV201.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/bf7f34ccb8743e645a982454.png)  
+  Supplementary figure 9. AAV8-GFAP-ATOH1-H2B-EGFP induced regeneration of hair cells in the adult mouse utricle *in vivo*. A) Low dose (5.0x10⁹ vg/ml) and high dose (1.0x10¹⁰ vg/ml) of AAV8-GFAP-ATOH1-H2B-EGFP were administered to 4-10-week-old mice. Utricles were analyzed 2 weeks later. B-E’) More hair cells were observed in injected ear compared to contralateral ear in both high and low doses GFAP-ATOH1 group. Insets showing GFP signals in B-E. F) Quantification showing significantly more hair cells in injected ears compared to contralateral ears. G) Higher dose GFAP-ATOH1 induced significantly more hair cells than low dose. H) GFP intensity in supporting cells was significantly higher in high dose treated group. Data shown as mean±S.D., compared using unpaired Student’s t-tests. *p<0.05, ***p<0.001. Scale bars: B-E’, 100mm.
+
+- [SuppFigure10AAVinvivohighmagandcristaeV401.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/08a543194018b0d1d7adb8a4.png)  
+  Supplementary figure 10. AAV8-GFAP-ATOH1-H2B-EGFP induced hair cell regeneration in adult mouse utricle and cristae *in vivo*. A) 4-10-week-old mice were injected with AAV-GFAP-ATOH1-H2B-EGFP, organs were examined at 2 weeks and 1 months after surgery. B) Robust regeneration of hair cells (Myosin7a, red) in injected ears compared to contralateral ear at 2 weeks after AAV8-GFAP-ATOH1-H2B-EGFP treatment. Striola shown. Some dim GFP+ hair cells were observed (arrowhead). C) More hair cells were observed in both the extrastriola and striola of utricles from injected ear compared to contralateral ear at 1 month. Some dim GFP+ hair cells were observed (arrowhead). D) Regenerated hair cells increased from 2 weeks to 1 month. E) Diagram illustrating central zone (CZ) and peripheral zone (PZ) of cristae. F, H) At 2 weeks and 1 month after GFAP-ATOH1 treatment, injected ears have more hair cells in both peripheral and central zones compared to contralateral ear. G, I) Hair cell number were higher in injected ears in both the peripheral and central zone at 2 weeks and 1 month. Data shown as mean±S.D., compared using paired Student’s t-tests. *p<0.05, **p<0.01. Scale bars: B-C, F, H, 20mm.
+
+- [SuppFigure11scRNAinvivorelatedV401.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/176f04d3c520eb3cf2c04598.png)  
+  Supplementary figure 11. Single-cell transcriptome of regenerating mature mouse utricle *in vivo*. A-D) UMAP plots showing total read counts, features (genes), mitochondrial and ribosomal gene percentages per cell. E) UMAP plot of integrated dataset of 3 different treatment groups (control, IDPN, IDPN and GFAP-ATOH1) from both timepoints showing 38,387 single cells following all quality control. Twenty-six cell clusters were identified. F) UMAP plot of integrating cells showing 19 inferred cell type, and marker genes were used to annotate the sensory and non-sensory cell types. G) Heatmap showing differentially expressed genes among the 26 cell clusters. The top differentially expressed genes of each cluster are shown on the right side of the heatmap. A full list of these genes is found in Table S5. The heatmap is colored by relative expression from -2 (magenta) to 2 (yellow). H-I) Expression of established markers of hair cells (*Lhx3*) and supporting cell (*Agr3*) within the cell clusters. J-K) UMAP plots showing ATOH1-EGFP-transgene and endogenous ATOH1 expression.
+
+- [SuppFigure12scRNAinvivrelatedV601.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/840dbd349063374d2a9fdd09.png)  
+  Supplementary figure 12. Single-cell RNA transcriptome of regenerating hair cells in the mouse utricle *in vivo*. A) Violin plot of endogenous *ATOH1* expression in 6 different cell clusters *in vivo*. B) UMAP plot of endogenous ATOH1 expression level *in vivo*. C) Ridge plots of notch activity score of 6 different cell clusters. D) Dot plots showing top 13 highly enriched notch activity genes from each cell type. E) Dot plots showing top 6 highly enriched bundle genes from each cell cluster. F) Dot plots showing top 4 highly enriched synapse genes from each cell cluster.
+
+- [Suppfigure13scRNAinvivocristaV301.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/65b32da28b5d576abc93fa01.png)  
+  Supplementary figure 13. Sensory cell types in cristae after IDPN damage and GFAP-ATOH1 treatment *in vivo*. A) 4-10-week-old mice were treated with IDPN (5mg/g) at day 1, and AAV8-GFAP-ATOH1-H2B-EGFP (1.71x10¹³ vg/mL) was injected at day 15. Single cell RNA sequencing was performed at 2 weeks and 1 month after injection. B) UMAP plot of integrated dataset of 3 different groups (control, IDPN, IDPN and GFAP-ATOH1) from both timepoints showing 7,855 single cells including only hair cells and supporting cells. Six cell clusters including supporting cells, stage 1, stage 2, stage 3 regenerated hair cells, type II and type I hair cells were identified, and marker genes were used to annotate hair cells and supporting cells. C) UMAP plot colored by control, IDPN or IDPN and GFAP-ATOH1 treatment. D) Heatmap showing differentially expressed genes among the 6 cell clusters. The top differentially expressed genes of each cluster are shown on the right side of the heatmap. A full list of these genes is found in Table S6. The heatmap is colored by relative expression from -2 (magenta) to 2 (yellow). E) UMAP plot showing the Monocle 3 pseudotime value of each cell, progressing from gray to purple. F) Shown the fraction of cells from different treatment groups assigned to each cell type. G) Shown the fraction of cells from different timepoints assigned to each cell type. H) The fraction of cells from different timepoints with GFAP-ATOH1 treatment assigned to each cell type.
+
+- [SuppFigure14scRNAinvivorelatedcristaV201.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/488a57ecc49af013d73d176b.png)  
+  Supplementary figure 14. Single-cell transcriptomes of regenerating cristae *in vivo*. A-D) UMAP plots showing total read counts, features (genes), mitochondrial and ribosomal gene percentages per cell. E) UMAP plot of integrated dataset of 3 different treatment groups (control, IDPN, IDPN and GFAP-ATOH1) from both timepoints showing 78,101 single cells following quality control steps. Thirty-six cell clusters were identified. Clusters designated as doublets were removed. F) UMAP plot of integrating cells showing 22 inferred cell types, and marker genes were used to annotate the sensory and non-sensory cell types. G) Heatmap showing differentially expressed genes among the 36 cell clusters. The top differentially expressed genes of each cluster are shown on the left side of the heatmap. A full list of these genes is found in Table S7. G) The heatmap is colored by relative expression from -2 (magenta) to 2 (yellow). H-I) Expression of established markers of hair cells (*Lhx3*) and supporting cell (*Agr3*) within the cell clusters. J-K) UMAP plots showing *ATOH1-EGFP*-transgene and endogenous *ATOH1* expression.
+
+- [Suppfigure15scRNAinvivocrista01.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/8221b1d65fe160d2d7538d83.png)  
+  Supplementary figure 15. Supporting-cell-specific promoters drive regenerated hair cell maturation in the cristae *in vivo*. A) Violin plot of ATOH1-eGFP-transgene expression level in 6 cell clusters. B) UMAP plot of *ATOH1-eGFP*-transgene expression level. C-F) Ridge plots of hair cell maturity, immature hair cells, type II and type I hair cell scores of 6 different cell clusters. G) Dot plots showing top 18 highly enriched mature hair cell genes from each cell clusters. H) Dot plots showing top 10 highly enriched immature hair cell genes from each cell clusters.
+
+- [SuppFigure16scRNAinvivocristaV201.png](https://assets-eu.researchsquare.com/files/rs-3190105/v1/6ad6bda0f24cce9fce3abcc2.png)  
+  Supplementary figure 16. Notch activity score in cristae *in vivo*. A) Violin plot of endogenous *ATOH1* expression in 6 different cell clusters *in vivo*. B) UMAP plot of endogenous ATOH1 expression level *in vivo*. C) Ridge plots of notch activity score of 6 different cell clusters. D) Dot plots showing top 13 highly enriched notch activity genes from each cell cluster clusters. E) Dot plots showing top 6 highly enriched bundle genes from each cell cluster. F) Dot plots showing top 4 highly enriched synapse genes from each cell cluster.
+
+- [SupplementalTable1Fig2DV1.xlsx](https://assets-eu.researchsquare.com/files/rs-3190105/v1/184cf6533a1812f12bc9ab1c.xlsx)  
+  Supplementary Table 1. Differential gene expression of all utricular sensory cells *ex vivo* (related to Figure 2D)
+
+- [SupplementalTable2FigureS4G.xlsx](https://assets-eu.researchsquare.com/files/rs-3190105/v1/eb8c5186f8292572c0f422d7.xlsx)  
+  Supplementary Table 2. Differential gene expression of all utricular cells *ex vivo* (related to Figure S4G)
+
+- [SupplementalTable3signatures.xlsx](https://assets-eu.researchsquare.com/files/rs-3190105/v1/67f500cd29513c3ba058f68a.xlsx)  
+  Supplementary Table 3. Signature scores (related to Figure 3 and 6, Figure S5, 12, 15 and 16)
+
+- [SupplementalTable4Fig5DV1.xlsx](https://assets-eu.researchsquare.com/files/rs-3190105/v1/6db73ed0ff31fdb55d4f07fe.xlsx)  
+  Supplementary Table 4. Differential gene expression of all utricular sensory cells *in vivo* (related to Figure 5D)
+
+- [SupplementalTable5FigureS11G.xlsx](https://assets-eu.researchsquare.com/files/rs-3190105/v1/995d7449f1d9c9d6bdd38d7c.xlsx)  
+  Supplementary Table 5. Differential gene expression of all utricular cells *in vivo* (related to Figure S11G)
+
+- [SupplementalTable6FigureS13DV1.xlsx](https://assets-eu.researchsquare.com/files/rs-3190105/v1/f82ea3bd76ab57cfa3263164.xlsx)  
+  Supplementary Table 6. Differential gene expression of all cristae sensory cells *in vivo* (related to Figure S13D)
+
+- [SupplementalTable7FigureS14G.xlsx](https://assets-eu.researchsquare.com/files/rs-3190105/v1/718c09e506c925bc2a16572b.xlsx)  
+  Supplementary Table 7. Differential gene expression of all cristae cells *in vivo* (related to Figure S14G)

0dfca8da50fdbbca7658136bc95928fa3678a29b12b79ef1499816d98b6a8374/preprint/images_list.json

@@ -0,0 +1,42 @@
+[
+    {
+        "type": "image",
+        "img_path": "images/Figure_1.png",
+        "caption": "Structural and morphological changes in lead-tin perovskite films during aging. a) Diagram of degradation phase formation observed in FA0.83Cs0.17Pb0.5Sn0.5I3 upon aging in inert and ambient atmospheres. b) OM and SEM images of perovskite films deposited on glass, before and after encapsulated aging under 65 \u00b0C and simulated full spectrum sun light (76 mW cm-2) irradiance for 2590\u00a0hours. OM images were taken with diascopic illumination and SEM images were acquired with a 2\u00a0kV accelerating voltage. c) EDX-SEM analysis. i) SEM image of the area prior to measurement. New crystallites are marked with a yellow border. ii) Ratio of mean EDX signal intensity measured within new crystallites (yellow lines) to mean signal intensity in background (rest of image). d) SEM image of a cross-section of the lead-tin perovskite absorber layer inside a device after encapsulated aging under 65 \u00b0C and simulated full spectrum sun light (76 mW cm-2) irradiance for 627\u00a0hours, acquired with a 5\u00a0kV accelerating voltage.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_2.png",
+        "caption": "Opto-electronic changes in lead-tin perovskite films during encapsulated aging under 65 \u00b0C and simulated full spectrum sun light (76 mW cm-2) irradiance. a) Sum of mobilities and b) background charge carrier density from TPC measurements of lead-tin perovskite films deposited on glass. The median is plotted with error bars representing Q1/Q3 (n = 8 for controls, n = 4 for EDAI2 passivated). c) Intensity-dependent PLQE of lead-tin perovskite films deposited on glass, glass/PEDOT:PSS or glass/PTAA/. d) Dark conductance measured laterally across 300 and 500 \u03bcm channels on lead-tin perovskite films deposited on glass or thin layers of PEDOT:PSS or PTAA. The median is plotted with error bars representing Q1/Q3 (n = 8).",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_3.png",
+        "caption": "Performance degradation of lead-tin perovskite solar cells during encapsulated aging under 65 \u00b0C and simulated full spectrum sun light (76 mW cm-2) irradiance. a) Performance metrics for encapsulated ITO/HTL/FA0.83Cs0.17Pb0.5Sn0.5I3/PCBM/ BCP/Cr/Au solar cells measured at room temperature under AM1.5\u00a0100\u00a0mW\u00a0cm-2 simulated sun light at various time intervals during aging using PEDOT:PSS or PTAA HTLs, showing i) maximum power point power conversion efficiency, ii) steady-state open circuit voltage, iii) steady-state short circuit current, and iv) steady-state fill factor of devices during aging. Transparent solid lines represent individual devices, whilst the dark solid line represents the median value with error bars denoting Q1/Q3 (n = 8). b) Open circuit voltage of lead-tin perovskite solar cells using PEDOT:PSS or PTAA hole transport layers compared to QFLS (from PLQE) of perovskite films on glass, measured after various periods of aging. The median is plotted with error bars representing Q1/Q3 (n = 4 for QFLS measurements). c) Current density under AM1.5\u00a0100\u00a0mW\u00a0cm-2 simulated sun light over time at short circuit of a champion ITO/HTL/FA0.83Cs0.17Pb0.5Sn0.5I3/ PCBM/BCP/Cr/Au device using a PEDOT:PSS or PTAA/Al2O3 HTL, measured after various periods of aging.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_4.png",
+        "caption": "Impact of mobile ions on lead-tin perovskite device performance during encapsulated aging under 65 \u00b0C and simulated full spectrum sun light (76 mW cm-2) irradiance. a) Fast 752\u00a0V\u00a0s-1 and slow 0.18\u00a0V\u00a0s-1 J-V scans of a ITO/HTL/ FA0.83Cs0.17Pb0.5Sn0.5I3/EDAI2/PCBM/BCP/Cr/Au device under AM1.5\u00a0100\u00a0mW\u00a0cm-2 simulated sun light after 0 and 288\u00a0hours of aging, using either a i) PEDOT:PSS or a ii) PTAA HTL. b) Plot of the PCE loss extracted from the forward J-V scans of the champion devices shown in a), after various periods of aging. Losses are divided into \u2018mobile ion\u2019 losses, derived from the difference between fast (752\u00a0V\u00a0s-1) and slow (0.18\u00a0V\u00a0s-1) scans at each aging time, and \u2018other\u2019 losses, derived from the progressive losses observed in the fast (752\u00a0V\u00a0s-1) scans over time. c) Simulated J-V curves produced using SCAPS-1D for the devices described in a), varying i) background carrier density, ii) perovskite mid-gap/deep trap density, and iii) perovskite mobility in devices using a PEDOT:PSS HTL, and iv), v), vi) the same for devices using a PTAA HTL. Real device data from a) is overlaid, showing fast (752\u00a0V\u00a0s-1) J-V scans of a fresh device (blue dotted line) and 288 hour aged device (red dotted line).",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    },
+    {
+        "type": "image",
+        "img_path": "images/Figure_5.png",
+        "caption": "Unnumbered image in the Methods section.",
+        "footnote": [],
+        "bbox": [],
+        "page_idx": -1
+    }
+]

0dfca8da50fdbbca7658136bc95928fa3678a29b12b79ef1499816d98b6a8374/preprint/preprint.md

@@ -0,0 +1,283 @@
+# Abstract
+
+Narrow bandgap mixed lead-tin perovskites are critical for efficient all-perovskite multi-junction solar cells, but their poor stability under operating conditions represents a major barrier to implementation. In this work, we explore the causes of this instability under combined heat and light stress (ISOS L-2 conditions). The phase, absorbance, morphology, and background hole density in lead-tin perovskite films are observed to be stable beyond the usual timescales associated with device degradation. We measure a moderate increase in non-radiative recombination during stressing, but device simulations demonstrate that this can only account for a small portion of the observed steady-state performance loss. Variable rate current-voltage scanning of devices instead reveals an increasing impact of mobile ions to be the major cause of early-time performance degradation. This impact is found to be significantly mitigated by selecting an alternative hole transport layer. Over longer aging times, we also identify the growth of impurity phases as well as hole transport material-dependent changes in the electronic properties of the perovskite. By quantifying the impact of these changes on device performance, we identify the most dominant degradation pathway at each aging time for different device architectures, defining a clear direction for future stability improvements.
+
+Physical sciences/Energy science and technology/Renewable energy/Solar energy/Photovoltaics/Solar cells  
+Physical sciences/Materials science/Materials for energy and catalysis/Solar cells  
+Perovskites  
+Solar Cells  
+Stability  
+Lead-Tin  
+Ion Migration
+
+# 1. Introduction
+
+The ability to rapidly and affordably expand solar power generation capacities will be essential for a successful global transition to renewable energy sources.<sup>[1]</sup> Hybrid organic-inorganic perovskites are a promising class of emergent photovoltaic materials for this task.<sup>[2, 3]</sup> as they benefit from a high absorption coefficient,<sup>[4]</sup> spontaneous exciton dissociation,<sup>[5]</sup> and long charge carrier diffusion.<sup>[6]</sup> Additionally, the bandgap of perovskites is highly tunable by varying the material composition,<sup>[7]</sup> enabling the development of perovskites with well-suited bandgaps for multi-junction architectures. So called ‘all-perovskite tandem’ solar cells are a promising new technology due to their potential for high efficiencies (feasibly 33.8% for double junctions or 36.6% for triple junctions),<sup>[8]</sup> ease of processing, and promise of lower embodied energy in manufacturing as compared to silicon-based single-junction or tandem cells.<sup>[9]</sup>
+
+The lowest bandgap (∼1.25 eV) perovskites can be achieved by alloying Pb and Sn,<sup>[10, 11]</sup> making lead-tin perovskites an ideal material for the low energy absorber in such tandems. However, whilst efficiencies of lead-tin perovskite-based devices now exceed 23% in single junction devices,<sup>[12]</sup> 28.5% in all-perovskite double junctions,<sup>[13]</sup> and 24.3% in triple junctions,<sup>[14]</sup> reports of high stability under operating conditions remain rare. The development of more stable lead-tin perovskite absorber layers and device stacks is hence essential for the realization of field-deployable all-perovskite tandems.
+
+The use of Sn<sup>2+</sup> in perovskites can have a two-fold detrimental effect on stability. The facile oxidation of Sn<sup>2+</sup> to Sn<sup>4+</sup> at the perovskite surface upon exposure to air, along with other chemical processes, can break down the original perovskite phase and form recombination centers.<sup>[15, 16]</sup> Second, tin-containing perovskites are expected to have a different defect chemistry compared to neat Pb perovskites resulting in increased p-doping,<sup>[16]</sup> which can affect recombination dynamics as well as energetic alignment at interfaces within photovoltaic devices. Encouragingly, Sn<sup>2+</sup> oxidation has been suggested to proceed slower in the mixed-metal perovskites.<sup>[17]</sup> However, as Sn<sup>2+</sup> oxidation is likely initiated at surfaces, surface degradation could significantly affect device performance even if the bulk remains intact.<sup>[15, 18]</sup>
+
+The instability of lead-tin perovskite under air exposure will likely become less relevant as advances in device packaging are made.<sup>[19]</sup> Indeed, multiple studies have reported an improvement in the stability and performance of lead-tin perovskite-based devices through the use of protective capping layers that block atmospheric oxygen and moisture such as sputtered indium-doped tin oxide (ITO),<sup>[20]</sup> indium-doped zinc oxide<sup>[21]</sup> or atomic-layer-deposition tin oxide.<sup>[22]</sup> However, lead-tin PSCs have been found to degrade rapidly under operating conditions even when air exposure is prevented. Interventions found to improve the stability of lead-tin perovskite based solar cells under these conditions have been the inclusion of a discontinuous Al<sub>2</sub>O<sub>3</sub>-nanoparticle ‘buffer’ layer between the perovskite and electron transport layer,<sup>[23]</sup> the increase of polycrystalline perovskite grain size<sup>[21]</sup> and the removal of the poly(3,4-ethylenedioxythiophene) polystyrene sulfonate (PEDOT:PSS) hole transport layer (HTL).<sup>[21, 24]</sup> Notable examples of unusually stable lead-tin perovskite devices are the achievement of 1000 hours of stable maximum power point (MPP) operation in N<sub>2</sub> of a HTL-free single junction lead-tin device,<sup>[21]</sup> as well as the good operational stability over hundreds of hours of some all-perovskite tandems employing lead-tin perovskites under similar aging conditions (ISOS-L1<sup>[25]</sup>).<sup>[22, 26]</sup> However, the consensus in the field remains that lead-tin perovskite solar cells tend to have inferior stability compared to their neat-lead counterparts, with significant performance losses during the first tens to hundreds of hours of aging under light exposure at elevated temperatures.<sup>[27]</sup>
+
+It remains unclear which factors ultimately limit device stability, and whether these are inherent to the lead-tin perovskite material itself or stem from another aspect of the device. Herein, we aim to address these questions by disentangling the degradation occurring in various components in the device stack as well as in fully fabricated solar cells. We investigate the evolving properties of lead-tin perovskite films and devices during encapsulated aging under combined light and thermal stressing. Our findings show that in the case of the isolated lead-tin perovskite absorber layer, changes in morphology and phase purity only become significant after thousands of hours of aging under simulated sun light at 65°C. While we find the absorber layer to be structurally and opto-electronically stable over hundreds of hours of aging, the corresponding photocurrent generated from solar cells degrades significantly during this time. We attribute this rapid device degradation to largely be the result of increased mobile ion-induced losses, highlighting an important future focus for research aiming to enable long-term stable lead-tin perovskite solar cells. We additionally reveal that the magnitude of this mobile ion-induced loss can be greatly reduced by selecting an alternative HTL. However, contact of the perovskite with different HTLs is found to induce additional changes to the optoelectronic properties of the perovskite during aging. We quantify the impact of each of these factors on device performance over time, revealing the most relevant degradation pathways for each architecture. Based on this, we make recommendations for future stability optimization.
+
+# 2. Results
+
+## 2.1 Structural and morphological changes during aging
+
+We begin by investigating structural changes that occur in the bulk perovskite absorber layers during aging when deposited either on glass or in ‘half-stacks’ of ITO/HTL/perovskite. In this study we use a ‘methylammonium (MA)-free’ perovskite composition that instead mainly uses formamidinium (FA), FA₀.₈₃Cs₀.₁₇Pb₀.₅Sn₀.₅I₃, with a photovoltaic bandgap of 1.25 eV determined from external quantum efficiency (EQE) spectra (Figure S1). Although MA-containing lead-tin perovskites are responsible for most record device efficiencies, they have been shown to be significantly less thermally stable than MA-free devices. SnF₂ salt and Sn⁰ powder additives are used to reduce Sn⁴⁺ species in the precursor solution and Sn²⁺ vacancies in the fresh films. Films and devices were encapsulated with UV-cured epoxy and glass cover slips and aged at elevated temperatures (65°C) and illumination (0.76 suns equivalent from a AM1.5 xenon lamp aging box) in an ambient environment (ISOS L-2 conditions). Throughout this paper, we refer to these stressing conditions simply as ‘aging’ for simplicity.
+
+To characterize crystallographic changes in the films during aging, we performed X-ray diffraction (XRD) on lead-tin perovskite films on glass or HTLs (Figure S2). All XRD patterns can be fit to a pseudo-cubic perovskite structure, both before and after 2000 hours of aging. New XRD peaks emerge in the aged films and can be assigned to a single degradation product, which we identify as the orthorhombic yellow δ-CsSnI₃ or δ-CsPbI₃ phase (Fig. 1a). This phase appears in perovskite aged on glass (unencapsulated, aged in N₂), on ITO/PEDOT:PSS and on ITO/poly(bis(4-phenyl)(2,4,6-trimethylphenyl)amine) (PTAA) half-stacks (encapsulated, aged in ambient conditions) (Figure S2).
+
+When the perovskite films are not encapsulated and rather exposed directly to ambient conditions, we instead detect Cs₂SnI₆ after only a few hours of air exposure, even without exposure to light and heat (Fig. 1a). Cs₂SnI₆ has been previously observed by others in lead-tin perovskites exposed to air, and can be spontaneously formed from δ-CsSnI₃ under ambient conditions at room temperature. Since we do not detect any Cs₂SnI₆ in aged, encapsulated samples, we infer that our encapsulation successfully inhibits the formation of large amounts of Sn⁴⁺ within these films during prolonged aging.
+
+We then performed XRD measurements on complete devices with an ITO/PEDOT:PSS/FA₀.₈₃Cs₀.₁₇Pb₀.₅Sn₀.₅I₃/ethylenediammonium diiodide (EDAI₂)/phenyl-C61-butyric acid methyl ester (PCBM)/bathocuproine (BCP)/Cr/Au architecture which had been aged for 627 hours, to check whether the perovskite absorber layers in devices degrade in the same way as the films on glass and half-stacks on HTLs (Figure S2). New peaks appearing at 2θ values of 12.8° and 53.5° can be ascribed to PbI₂, and other new peaks at 13.2° and 26.5° can be ascribed to either δ-CsSnI₃/δ-CsPbI₃ or Cs₂SnI₆. Despite the appearance of these degradation products, the relative peak intensity and width of the original perovskite phase does not change significantly, indicating that the perovskite bulk is left largely intact during 600 + hours of device aging. We do however consistently observe a slight expansion in the cubic lattice volume of the main perovskite phase during aging, for films on glass, HTLs and complete devices (Table S1). This would be consistent with the formation of a slightly more FA- and Pb-rich primary phase as δ-CsSnI₃ or Cs₂SnI₆ evolve from the initial perovskite composition.
+
+Next, we examine the morphological qualities of films of lead-tin perovskite deposited on glass and on HTLs during aging. Fresh films appear smooth and homogenous in both optical microscope (OM) and scanning electron microscope (SEM) images. However, after long periods of aging (2000 + hours) we observe by OM bright yellow regions a few micrometers in size (Fig. 1b). SEM images reveal that the polycrystalline form of the perovskite has changed in these regions, being composed of smaller crystalline domains than the remaining perovskite bulk. We find that these new crystallites start to become visible on the surface of perovskite films after ∼1000 hours of aging (Figure S3a), although it is possible that small regions begin to form away from the surface of the film before this time. We also observe the emergence of these features in OM images of aged lead-tin perovskite films deposited on ITO/PEDOT:PSS and ITO/PTAA after ∼1000 hours of aging, where the regions appear larger when PEDOT:PSS is used as the underlying layer (Figure S3b).
+
+To confirm the chemical identity of these regions we perform spatially resolved energy-disperse X-ray scanning electron microscopy (EDX-SEM) imaging (full maps are shown in Figure S4). EDX indicates that a higher density of Cs and Sn is present within the new crystallites as compared to the rest of the film, whilst the Pb density is similar between the crystallites and the background (Fig. 1c). Identifying these regions as δ-CsSnI₃ is also consistent with the observations of increased transmittance of light through these regions made by OM, as δ-CsSnI₃ has a very wide optical bandgap of 2.55 eV, whilst δ-CsPbI₃ has a bandgap of 1.7 eV. Hence, there is significant evidence that the observed new crystallites are the δCsSnI₃ phase detected by XRD measurements.
+
+The new crystallites also contain a slightly increased density of iodine. It is possible that some other degradation products such as I₂, SnI₂ or SnI₄ also form in these regions, but any crystalline domains of these other possible impurities are too small to be detected by XRD. Additionally, since δCsSnI₃ has a significantly higher density (4.82 g cm⁻³) than the lead-tin perovskite (3.87 g cm⁻³ for MAPb₀.₅₄Sn₀.₄₆I₃), its formation within the film is expected to be accompanied by significant strain leading to void formation. SEM imaging of cross-sections of aged devices confirms this – after 627 hours of aging, we observe voids filled with smaller crystallites in the active layer of aged devices (Fig. 1d), which we don’t observe in fresh devices.
+
+Despite the slow evolution of these degradation products, the original lead-tin perovskite phase remains stable over long periods of light stressing at elevated temperature, even in half-stacks with HTLs or in complete devices. In agreement with this, the absorption onset of films on glass changes only very slightly over hundreds to thousands of hours of aging (Figure S5) and EQE of aged devices showed no significant change in photovoltaic bandgap during the first ∼300 hours of aging (Figure S1). However, as smaller-scale degradation processes could be occurring on faster timescales, we next investigate how the optoelectronic properties of the perovskite change during aging.
+
+## 2.2 Opto-electronic changes during aging
+
+Although the perovskite’s bulk structure and optical absorption properties remain stable during the first few hundred hours of aging, smaller-scale changes not captured by the previous methods may affect the electronic properties of the film during this time. This especially includes the evolution of point defects which may affect background carrier density, mobility, and rates of defect-mediated recombination. Lead-tin perovskites are slightly p-type with a background hole density p₀ around 10¹⁴ cm⁻³ (previously determined by others using Hall effect measurements and 2-point probe conductivity measurements). The magnitude of p₀ is important to device functionality, as it is expected to affect the energetic alignment within devices when it is larger than ∼10¹⁵ cm⁻³, and to impact recombination kinetics when it is larger than the photoexcited carrier density. Although p₀ has been shown to lie below these critical densities in fresh devices, it remains unclear to what extent it may increase during aging.
+
+To measure background carrier density and long-range mobility, we probe the lateral conductivity across 300–500 µm channels of perovskite thin films, in the dark and upon photoexcitation. The TPC method is described in detail elsewhere, but briefly, we excite a perovskite film with a pulse of light to induce a temporary increase in free-carrier density and in turn in conductivity across the sample (conductivity traces are shown in Figure S6, and extracted dark- and maximum photo-conductance are plotted in Figure S7). Since conductivity is the product of carrier density (n, p) and mobility (μ) (Eq. 1), measuring both the dark conductivity (Eq. 2) and the photoconductivity after a known increase in free charge carrier density (Eq. 3) allows us to determine the sum of charge carrier mobilities (Eq. 4) and estimate an upper limit for the background hole density, p₀ (Eq. 5). As lead-tin perovskites are expected to be p-type, the background electron density n₀ will be many orders of magnitude smaller than the photoexcited carrier densities Δn (10¹⁶–10¹⁹ cm⁻³) and we can approximate n = Δn + n₀ ≈ Δn.
+
+$$
+\sigma =e({\mu }_{n}n+{\mu }_{p}p)
+$$
+(1)
+
+$$
+\frac{{\sigma }_{dark}}{e}={\mu }_{p}\left({p}_{0}\right)
+$$
+(2)
+
+$$
+\frac{{\sigma }_{photo}}{e}={\mu }_{n}\left(\varDelta n\right)+{\\mu }_{p}({p}_{0}+\\varDelta n)
+$$
+(3)
+
+$$
+\frac{{\\sigma }_{photo}-{\\sigma }_{dark}}{e\\varDelta n}={\\varvec{\\mu }}_{\\varvec{n}}+{\\varvec{\\mu }}_{\\varvec{p}}
+$$
+(4)
+
+$$
+{\\frac{{\\sigma }_{dark}}{e{\\mu }_{p}}=\\varvec{p}}_{0}
+$$
+(5)
+
+Our estimation of the sum of mobilities of the perovskite films on glass versus aging time is presented in Fig. 2a. The initial sum of mobilities measured by us is roughly 30x smaller than measured by others using THz conductivity, which is likely due to TPC probing charge transport on a longer length scale which includes transport through grain boundaries. We observe that the sum of mobilities initially slightly increases, but overall does not change significantly during 150 hours of aging.
+
+To calculate p₀ from dark conductivity values, we assume here that the electron and hole mobilities are equal (as TPC is only sensitive to the sum of the mobilities). We present our estimated p₀ as a function of aging time in Fig. 2b. Critically, p₀ also does not vary significantly and remains below 10¹⁵ cm⁻³ during the 150-hour aging period. While the absolute value of p₀ may be slightly higher or lower than estimated here based on the actual electron to hole mobility ratio, the observation that the perovskite doping density does not significantly change remains valid. This further confirms that the alloying of Sn with Pb successfully avoids high levels of self-doping, not just immediately after fabrication, but also during at least 150 hours of aging. Additionally, we do not observe any significant difference in the trends of the mobility or background carrier density for samples with an EDAI₂ surface passivation (commonly used for lead-tin perovskites to improve device performance).
+
+To assess whether charge carrier recombination dynamics change during aging, we measure the PLQE of lead-tin perovskite films on glass at excitation fluences ranging from 0.4 to 4 suns equivalent. We find that the PLQE continuously decreases during aging, without much change in the shape of the variation in PLQE across excitation fluences (Fig. 2c). This is consistent with a moderate increase in non-radiative recombination during aging. Interestingly, this is not accompanied by a significant decrease in long-range mobility or change in background carrier density, which would indicate that any defects formed do not act as effective charge scattering centers nor dope the perovskite. When probing changes in the properties of lead-tin perovskites upon exposure to air using THz conductivity measurements, others also observed this increase in nonradiative recombination without a significant accompanying change in mobility.
+
+To understand possible effects related to interaction of the perovskite with the HTL during aging, we also investigate perovskite films deposited on glass/PEDOT:PSS and glass/PTAA (Fig. 2c). The PLQE of the fresh lead-tin perovskite on PEDOT:PSS is significantly lower than the PLQE of fresh films on glass or PTAA, which could be related to the larger hole density in PEDOT:PSS as a strongly p-doped material causing increased interfacial recombination. The PLQE of PEDOT:PSS/perovskite films increases 10-fold during the first 19 hours of aging, then decreases again, but is never reduced to a value lower than the fresh sample. In comparison to the behavior of the isolated perovskite films, this suggests significant chemical changes at the PEDOT:PSS/perovskite interface during aging, which may include PEDOT:PSS forming complexes with I⁻ and A-cations, or reacting with Sn to form SnS.
+
+On PTAA, however, the PLQE is similar in magnitude and follows a similar trend to that of the perovskite on glass. This indicates that in these samples, bulk recombination dominates over interfacial recombination at the PTAA/perovskite interface under open-circuit conditions. Interestingly, we also observe a significant reduction in the gradient of the PLQE against excitation intensity of PTAA/perovskite stacks (Fig. 2c), which could be due to a significant increase in the background carrier density of the perovskite. When the background carrier density becomes larger than the steady-state photoexcited carrier density, the rate of radiative recombination tends towards monomolecular rather than bimolecular, making the PLQE independent of excitation intensity for the measured fluence range.
+
+To further investigate this potential increase in the p₀ of the perovskite when aging on PTAA, we measure the lateral conductance of perovskite films deposited on a thin layers of HTLs on glass (Fig. 2d). Due to challenges in deconvolving the contribution to the total conductance from the perovskite layer and the HTL individually, we could not determine an absolute value for the mobility and p₀. Hence, we consider only relative changes in the absolute dark conductance of the HTL/perovskite stacks in our analysis (presented in Fig. 2d, with dark conductance of all stacks as well as isolated HTLs shown in Figure S7). We observe a 5-fold increase in dark conductance through the glass/PEDOT:PSS/perovskite stack during the first 20 hours of aging, but no significant changes after this. The dark conductance through glass/PTAA/perovskite stacks, however, shows a much larger increase from ∼1 x 10⁻⁸ to ∼5 x 10⁻⁷ S between 100 and 300 hours of aging. The conductance of the isolated perovskite and the isolated HTL are found to be orders of magnitude lower than this, and do not vary during aging. Given this, we may infer that either the perovskite or the PTAA becomes significantly more conductive by being in contact during aging. For this 4 x 10⁻⁷ S conductance to be mainly due to conduction through the PTAA layer (estimated thickness < 10 nm), a conductivity of at least 0.4 S cm⁻¹ would be required. This is unlikely to be the case since the conductivity of PTAA is typically not reported to exceed 10⁻⁴ S cm⁻¹ even when extrinsically doped. We hence propose that the observed ∼50-fold increase in conductance is dominated by changes in the perovskite, namely due to an increased p₀. We observe this increase on the same timescale as the aforementioned reduction in the gradient of the intensity-dependent PLQE of PTAA/perovskite stacks. These observations together strongly indicate that the presence of PTAA during aging induces increased p-doping in the lead-tin perovskite. This is surprising, as undoped PTAA has been used to fabricate relatively stable neat-Pb PSCs.
+
+Overall, we find that the p₀ and mobility of isolated lead-tin perovskite films on glass are stable over 300 hours of aging. Non-radiative recombination rates are moderately increased during aging, but do not become worse at the PEDOT:PSS/perovskite interface. Finally, aging in contact with PTAA seems to significantly increase p₀ in the perovskite after 100 + hours, which is not observed with PEDOT:PSS. We next consider the effect which these changes could have on device performance during aging.
+
+## 2.3 Device performance and simulations
+
+Next, we investigate lead-tin perovskite devices. We fabricate and encapsulate devices with an architecture of ITO/PEDOT:PSS or PTAA/FA₀.₈₃Cs₀.₁₇Pb₀.₅Sn₀.₅I₃/EDAI₂/PCBM/BCP/Cr/Au and age these cells under combined thermal and light stressing (65°C, 0.76 suns equivalent) at open-circuit conditions. In devices using PEDOT:PSS, the MPP tracked power conversion efficiency (PCE) decreases by more than 50% during the first 108 hours of aging (Fig. 3a i). The open-circuit voltage (Voc) (Fig. 3a ii) is relatively stable during this time, while both the short-circuit current (Jsc) (Fig. 3a iii) and fill factor (FF) (Fig. 3a iv) rapidly decay. In addition, a notable increase in hysteresis (Figure S8a) can be observed over time when comparing the reverse (0.9 to 0.2 V) and forward (0.2 to 0.9 V) current density-voltage (J-V) curve sweeps. The dominance of FF and Jsc decay during the degradation of lead-tin perovskite devices with PEDOT:PSS HTLs has been previously observed, even with different lead-tin perovskite compositions. We also fabricate devices using PTAA and find that their initial performance is comparable to that of devices using PEDOT:PSS (Fig. 3a), with significantly less hysteresis (Figure S8a). Devices using PTAA are more stable during the first 100 hours of aging, mainly due to a much lower Jsc loss during this time (Fig. 3a iii), but continuously degrade and perform worse than devices using PEDOT:PSS after 288 hours of aging. EQE and dark J-V traces are shown in Figure S8, and time-resolved traces of maximum power point, steady-state Jsc and Voc tracking are shown in Figure S9.
+
+Both devices using PEDOT:PSS and PTAA show a relatively stable Voc during the first 100 hours of aging, despite a decrease in quasi-Fermi level splitting (QFLS) of the bulk perovskite (derived from the PLQE of films on glass at 1 sun equivalent excitation fluence, Fig. 3c). The difference between the bulk QFLS and the device Voc can be ascribed to interfacial losses, which our observations indicate do not change during aging, pinning the Voc. However, increased bulk recombination rates may still affect device performance by decreasing charge collection efficiency through shorter charge carrier diffusion lengths. For devices using PTAA, the Voc decreases significantly after 100 hours of aging. This is likely related to the increase in the perovskite p₀ which we observed in glass/PTAA/perovskite half-stacks in the previous section, and we discuss this further in section 2.5.
+
+We also find that for both HTLs used, the EQE of devices decreases relatively uniformly over all wavelengths during aging, and the current expected by integrating the EQE spectrum with the AM 1.5 spectrum corresponds well to the measured current densities in devices (Figure S8b). This indicates that the degradation of device performance during aging is dominated by a decrease in charge collection efficiency rather than bulk material degradation. Previous investigations by others have also suggested the poor stability of lead-tin PSCs to be due to the formation of a charge extraction barrier, possibly as a result of the formation of degradation products, a reaction with PEDOT:PSS at elevated temperatures, and/or a conductivity drop at the perovskite grain boundaries. We instead postulate that the formation of this charge extraction barrier arises from the effects of mobile ions in the perovskite absorber. In devices using PEDOT:PSS, we observe a small decay in steady-state Jsc during the first few seconds at short-circuit, which grows significantly to ∼11 mA cm⁻² after 288 hours of aging. This decay in Jsc at short-circuit has previously been ascribed to the aggregation of mobile ions at interfaces during device operation, hindering charge extraction. Our measurements show that this loss not only gets much worse during aging, but also that the extent to which this loss grows varies significantly with HTL used, since devices with a PTAA show a much smaller current decay of only ∼3 mA cm⁻² after 288 hours of aging. We next attempt to quantify these losses due to mobile ions in comparison to other causes of device degradation.
+
+## 2.4 Impact of mobile ions on device performance during aging
+
+It has been understood for several years now that the diffusion of I⁻ vacancies, and likely also the slower diffusion of A-site cation vacancies, is possible in PSCs with a variety of perovskite compositions. The diffusion of these mobile ions has been shown to cause performance losses which get worse during aging under illumination. This has been rationalized to be caused by the aggregation of mobile ions at interfaces during device operation, resulting in screening of the bulk electric field across the absorber. In addition to this, various mobile ion candidates such as I⁻ vacancies may also be able to trap charges, locally shift the Fermi level of the perovskite, and provide sites for the initiation of degradation reactions where their density is high. In lead-tin PSCs, mobile ions have already been shown to cause significant performance losses in fresh devices by Thiesbrummel et al., but their impact during aging has not yet been explored.
+
+To quantify the evolving influence of mobile ions during aging, we perform J-V scans at a wide range of scan rates on fresh and aged devices. Devices are pre-biased above Voc (1.1 V) for 5 seconds before a reverse and forward J-V sweep are carried out. At this prebiasing condition, potential drops at interfaces are expected to be small and negative, such that positively charged mobile ions are pulled slightly towards the electron transport layer. During the scan towards −0.2 V, the drop in potential between the electrodes is reversed. This drives mobile ions in the opposite direction, causing positive ions accumulate at the HTL interface.
+
+When a scan is performed at a faster rate than the characteristic time of ion motion and accumulation, ions do not have sufficient time to aggregate at interfaces under the influence of the electric field. A sufficiently fast J-V scan will hence reveal the characteristics of a device with ions ‘frozen’ in their pre-biasing position, whilst a slow J-V scan will reveal the characteristics of a device with ions that have redistributed in response to the applied voltage at every voltage point. We note that the processes that mobile ions undergo in PSCs, which may include adsorption or reaction with charge transport layers (CTLs) in addition to diffusion across the absorber layer, are not yet fully understood. To be certain that the slowest scan at 0.18 V s⁻¹ (7 s⁻¹) allows sufficient time for all the processes relevant to steady-state device performance to occur, we compare parameters extracted from this scan to the steady-state device performance measured previously (Figure S10). This shows that these are mostly very similar, deviating slightly only for the Voc of devices using PTAA HTLs after 100 hours of aging. We compare device performance during these ‘slow’ scans to ‘fast’ scans at 752 V s⁻¹ (2 ms⁻¹), in which we did not observe any hysteresis, indicating that device performance is no longer significantly affected by ion redistribution above this scan speed. Comparing these scan rate enables us to identify the extent to which observed losses during aging can be ascribed to the redistribution of mobile ions in comparison to other effects.
+
+We first focus on devices using a PEDOT:PSS HTL. Figure 4a shows J-V scans obtained at the ‘slow’ (0.18 V s⁻¹) and ‘fast’ (752 V s⁻¹) scan speeds described above, with the J-V parameters extracted during each aging step summarized in Figure S10. Surprisingly, at fast scan speeds, the large Jsc loss observed in devices using PEDOT:PSS after aging for 288 hours is almost entirely eliminated. Although there remains a small progressive loss in FF under fast scans, the performance degradation is minor compared to that under slow scans. This is strong evidence that the rapid degradation observed in lead-tin perovskite devices using PEDOT:PSS is dominated by effects from aggregation of mobile ions at interfaces, as opposed to optoelectronic degradation of the perovskite absorber layer.
+
+Recent work by others has suggested the possibility of a reduced impact of ion migration in lead-tin perovskites compared to neat lead perovskites, inferred from the observation of lower hysteresis in field effect transistors (FETs) and the computational prediction of a higher energy barrier to I⁻ migration with an increased density of Sn²⁺ vacancies. We demonstrate here that this is not the case in lead-tin PSCs with PEDOT:PSS HTLs, where in fact effects from mobile ions dominate early-time performance degradation. This discrepancy could be due to a variety of factors including device architecture being different from FETs, the possibility of a relatively lower density of Sn²⁺ vacancies in our lead-tin perovskite resulting in a lesser energetic barrier to I⁻ migration, as well as variations in mobile ion density.
+
+In devices with PTAA HTLs, we still observe a difference between the slow and fast scans after aging for 288 hours, indicating that mobile ions also contribute to device performance degradation when a PTAA HTL is used (Fig. 4a). However, the fast scan performance after aging, which is not affected by the aggregation of mobile ions, is much worse than for devices using PEDOT:PSS. In Fig. 4b we plot the absolute difference in PCE derived from the fast and slow J-V scans over time, which we refer to as “ionic losses”, and the difference between the PCE derived from the fast J-V scan before aging and the fast JV scans over time, which we refer to as “non-ionic losses”. The performance loss due to mobile ions in devices using PTAA is initially smaller than in devices using PEDOT:PSS, and increases much less during the first 20 hours. After this time, the performance degradation caused by mobile ions increases with a similar rate for devices using PTAA and PEDOT:PSS but remains approximately twice as large for devices with PEDOT:PSS.
+
+To explain this difference in the impact of mobile ions during aging between devices using different HTLs, we consider how the properties of the CTL/perovskite interface may affect the extent and impact of ion accumulation during steady-state device operation. Simulations by Courtier et al. suggest that the product of doping density and permittivity of CTLs modulates mobile ion accumulation at interfaces during device operation, with highly doped CTLs causing a larger density of ions to accumulate in the perovskite at the interface. The strongly p-doped nature of PEDOT:PSS as compared to undoped PTAA could hence contribute to an initial difference in the extent of ion accumulation during operation. Additionally, the variation in PLQE during aging measured earlier shows that non-radiative recombination at the perovskite/HTL interface at open-circuit conditions is much slower on PTAA as compared to PEDOT:PSS. It also suggests that the PEDOT:PSS/perovskite interface undergoes significant changes during aging, which may affect mobile ion density or energetic alignment at the interface during aging. Finally, the ability of mobile ions to transfer into the HTL during aging could also affect their impact on device performance over time – PEDOT:PSS is a polyelectrolyte and an ionic conductor whilst PTAA is not, which may influence how rapidly and to what extent ‘foreign’ ions from the perovskite absorber layer can accumulate and even migrate within the HTL layer. Further research is clearly needed to understand which of these factors is most relevant to mitigating the impact of mobile ions.
+
+It is highly encouraging that the detrimental effect of mobile ions on device performance during aging can be largely mitigated by using PTAA instead of PEDOT:PSS as the HTL in lead-tin PSCs. However, devices using PTAA experience much larger non-mobile-ion-related FF and Jsc losses, which result in overall worse steady-state PCE after 288 hours of aging. Hence, we next focus on identifying the causes of the non-mobile-ion-related performance degradation observed.
+
+## 2.5 Impact of electronic changes on device performance during aging
+
+Although mobile ions strongly contribute to the performance degradation of lead-tin perovskite devices, some losses remain in the fast scans which must be due to other factors. For devices using PTAA, these non-ionic losses match the magnitude of ionic losses after 288 hours of aging. To visualize how changes in material properties could affect device performance in the absence of mobile ions, we perform drift-diffusion simulations using the SCAPS-1D simulation package (parameters used are detailed in table S2). We find that the p₀, mobility and deep trap density in the perovskite absorber layer have the most significant effect on simulated device J-V curves, which we present in Fig. 4c as a function of parameter values. Changes in shallow trap density in the perovskite, trap density at the HTL/perovskite and perovskite/ETL interfaces, and changes in HTL doping density and mobility have a smaller effect upon the J-V characteristics, and are presented in Figure S11. The ‘immobile ion’ fast J-V scans of champion devices after 0 and 288 hours of aging are overlaid on the simulation results in blue and red in Fig. 4c. We do not neccesarily expect simulations to perfectly replicate the measured J-V curves, as many device parameters likely undergo small changes during aging.
+
+For devices using PEDOT:PSS, simulations show that the small FF loss observed in the measured fast J-V scans can be accounted for by a moderate change in bulk deep trap density (Fig. 4cii), which we had experimentally determined earlier via PL measurements, and/or a slight increase in p₀ during aging (Fig. 4ci). In devices using PTAA, however, the non-ionic losses continuously increase during aging and are much larger than those in devices using PEDOT:PSS after 288 hours of aging. We previously inferred a ∼50-fold increase in the p₀ of the perovskite absorber after 288 hours of aging in contact with PTAA. Indeed, when comparing the effect of changes in p₀, trap density and mobility on simulated J-V curves, the shape of the measured fast J-V curves after 288 hours of aging can be most closely reproduced by increasing the p₀ of the lead-tin perovskite from ∼10¹⁴ to ∼10¹⁶ cm⁻³ (Fig. 4c iv). Some increase in deep trap density may also contribute to the observed non-ionic losses. Notably, the performance loss due to ion migration in devices using PTAA is not accelerated during the time in which the doping density was observed to significantly increase (150–300 hours), indicating that these processes are not connected.
+
+Overall, we find that for devices using PEDOT:PSS, the small non-ionic FF losses during aging can be explained by the moderate increase in bulk non-radiative recombination previously observed in isolated perovskite films. For devices using PTAA, however, the much larger non-ionic FF and current density losses can be largely explained by the increase in the p₀ of the perovskite, previously observed in PTAA/perovskite half-stacks. This highlights the need for a better understanding of HTL-perovskite interactions during aging, as well as screening of novel hole transport materials to prevent electronic degradation of the perovskite as a result of HTL interactions during aging.
+
+# 3. Discussion
+
+In conclusion, we reveal that the rapid degradation in device performance of lead-tin PSCs is dominated by an increasing impact of mobile ions on the device under combined heat and light stressing. The impact of mobile ions is significantly mitigated when using the hole-conductor PTAA instead of PEDOT:PSS, indicating that it is not inherent to lead-tin perovskites but rather an effect related to other material choices. Determining more precisely which charge transport material properties most affect the impact of mobile ions during aging will likely be essential to improving stability of lead-tin PSCs.
+
+Although lead-tin PSCs employing PTAA are less affected by mobile-ion related losses during aging, the electronic losses resulting from an unexpected acceleration in p-doping during aging lead to comparable overall degradation as devices using PEDOT:PSS over a few hundred hours. This indicates there are complicated interactions occurring at the HTL/perovskite interface, which require understanding and controlling to deliver competitively stable devices. Charge transport materials should hence be screened both on their effect upon the impact of mobile ions during aging, and on any potential interactions with the perovskite absorber leading to opto-electronic degradation.
+
+Of secondary importance is the need to suppress longer term changes to the defect density in the perovskite absorber layer itself, causing increased defect-mediated recombination. Here, a better understanding of the type of recombination-active defects that develop in lead-tin perovskites during aging and the design of targeted passivation strategies could help minimize these changes. Understanding the potential impact of the δ-CsSnI₃ degradation phase we observed and avoiding its formation may also contribute to suppressing longer-term degradation over timescales of weeks to months.
+
+The emerging picture is that lead-tin perovskites are not fundamentally unstable, and do not experience significant Sn²⁺ oxidation during combined heat and light stressing if they are well-protected from air exposure. Unexpectedly, the increased impact from mobile ions appears to be the key driver for rapid performance degradation, but this can be significantly reduced through adjustments to device architecture. If the impact of mobile ions as well as unfavorable interactions with CTLs during stressing can be further mitigated, we expect that lead-tin PSCs will be no less stable than standard lead-based PSCs. These results are highly encouraging for the realization of efficient and stable multi-junction thin-film PSCs and define a clear direction for future efforts to improve the stability of lead-tin PSCs.
+
+# 4. Methods
+
+## Preparation of perovskite precursor
+
+To make a 1.8M FA₀.₈₃Cs₀.₁₇Pb₀.₅Sn₀.₅I₃ perovskite precursor, FAI (1.49 mmol, Greatcellsolar), CsI (0.31 mmol, Alfa Aesar, 99.9%), PbI₂ (0.90 mmol, Thermo Fisher Scientific, ultra dry 99.999%), SnI₂ (0.90 mmol, Thermo Fisher Scientific, ultra dry 99.999%), SnF₂ (0.09 mmol, Aldrich, 99%), and metallic Sn powder (10 mg, Aldrich, 99.5%) were stirred in DMF (0.800 ml, Sigma Aldrich, anhydrous) and DMSO (0.200 ml, Sigma Aldrich, anhydrous) for 4 days at room temperature in a glovebox. Solutions were filtered with a 0.45 µm PTFE filter shortly before spincoating.
+
+## Fabrication of films and devices
+
+Glass or ITO substrates (Biotain, 10–15 Ω cm⁻²) were cleaned by scrubbing with dishwashing soap, then sonicated for 10 min in 1 vol% Decon90 in DI water. Substrates were then rinsed with DI water and sonicated in DI water for 10 minutes, then sonicated in acetone and isopropyl alcohol (IPA) for 5 minutes each. Substrates were dried with N₂ and exposed to UV ozone for 10 minutes immediately before further processing.
+
+For films deposited on glass or PTAA, a layer of Al₂O₃ nanoparticles was added to the substrate to improve wetting. Al₂O₃ nanoparticle suspension (90 µl, Sigma Aldrich, < 50 nm, 20 wt% in IPA, diluted 1:150 in IPA) was deposited on the substrate by dynamic dropping during spinning at 5000 rpm for 30 s and subsequently dried for 2 minutes at 100°C.
+
+PEDOT:PSS (Heraeus, AL3083) was mixed with IPA in a 1:2 ratio and filtered with a 0.40 µm PVDF filter shortly before spin-coating. PEDOT:PSS solution (250 µl) was statically dropped and spread onto the ITO substrate, then spin-coated at 4000 rpm and 1333 acc for 30 s. Films were then annealed at 150°C for 15 minutes in air, and a further 15 minutes in a nitrogen-filled glovebox. PTAA (Xi’an Polymer Light Corp, Mw 10,000–100,000 g mol⁻¹) in toluene (1.5 mg ml⁻¹) was stirred overnight to dissolve, then filtered with a 0.45 µm PTFE filter. PTAA was spin-coated onto ITO substrates in a nitrogen glovebox by dynamically dripping 150 µl solution onto the substrate at the start of a 6000 rpm, 1000 acc, 30 s spin-coating cycle. Films were then annealed at 100°C for 10 min. To improve wetting on the PTAA surface, once the substrates had cooled Al₂O₃ nanoparticles were deposited as described above.
+
+To deposit perovskite films, perovskite precursor (80 µl) was statically dropped and spread onto the substrate, then spun at 5000 rpm and 1000 acc for 60 s. Anisole (200 ul, anhydrous, Sigma-Aldrich) antisolvent was dropped onto the substrate after 30 s. The substrates were then annealed at 100°C for 15 minutes.
+
+For full device fabrication, the perovskite film was passivated with EDAI₂. EDAI₂ (0.5 mg ml⁻¹, Merck) was stirred in 1:1 toluene:IPA for 4 days and subsequently filtered with a 0.45 µm PTFE filter. 80 µl of this solution was dropped onto the perovskite surface, and once the solution had spread to the edges of the film it was spun at 5000 rpm and 5000 acc for 20 seconds and annealed at 100°C for 1 minute.
+
+PCBM (20 mg ml⁻¹, Ossila) was dissolved in 1:3 chlorobenzene:dichlorobenzene by stirring overnight, and filtered with a 0.45 µm PTFE filter. PCBM solution (80 µl) was dynamically deposited on the perovskite whilst spinning at 2000 rpm for 20 s and subsequently annealed at 100°C for 3 minutes. BCP (0.5 mg ml⁻¹, Xi’an Polymer Light Technology Corp) was dissolved in IPA by stirring for 4 days and filtered with a 0.45 µm PTFE filter. BCP solution (80 µl) was dynamically deposited on the perovskite whilst spinning at 5000 rpm for 30 s and subsequently annealed at 100°C for 1 minute.
+
+Finally, 3.5 nm Cr and 100 nm Au were sequentially thermally deposited onto the device (initial rate 0.02 nm s⁻¹, ∼10⁻⁷ Pa vacuum). Films and devices were not exposed to air at any point during the fabrication process.
+
+## Aging
+
+Samples were encapsulated with a glass slide attached to the substrate with UV-activated epoxy which was cured for 3 minutes (Everlight Eversolar AB-341). For films, a recessed cavity glass with epoxy only deposited at the encapsulation edge was used to allow for optical measurements, and for devices the active area was fully covered by the epoxy. Before encapsulation, perovskite material was removed at the epoxy edge for optimal adhesion. Samples were aged in an ambient, illuminated aging chamber (Atlas Suntest XLS+) at 65°C and simulated full spectrum sun light (76 mW cm⁻²) irradiance under open circuit conditions. Before any measurement was performed, samples were removed from the aging environment and allowed to come to room temperate without direct illumination for 20 minutes. For XRD and SEM measurements of aged samples, the glass edges of the sample were cut to remove the encapsulation glass and epoxy.
+
+## J-V measurements
+
+The J-V characteristics of devices were measured under AM1.5G illumination (WaveLabs Sinus-220 solar simulator) with 100 mW cm⁻² equivalent irradiance (certified by KG3-filtered Si reference photodiode). Voltage was swept from 0.9 V to −0.2 V and back at a rate of 0.61 V s⁻¹. The device areas were defined by shadow masks, and each substrate contained 3 devices with an area of 0.25 cm² and one device with an area of 1.00 cm². For steady-state Vₒc and Jₛc measurements, devices were held at 0 V or Vₒc for 30 s. For steady-state PCE measurements, an MPP tracker based on a gradient descent algorithm was employed to measure the maximum power for 60 s. A quasi-steady-state FF was calculated by dividing the product of V and J at the max power measured by the steady-state PCE measurement by the product of steady-state Vₒc and Jₛc.
+
+## UV-vis absorption measurements
+
+The total transmittance and total reflectance of encapsulated samples were measured in a Cary 5000 spectrophotometer using an internal diffuse reflectance accessory. Absorbance was calculated according to  
+$A=-\text{l}\text{n}(1-T-R)$.
+
+## Optical microscope imaging
+
+Samples were imaged in an optical microscope (Nikon Eclipse LV100ND) using a 20x objective in bright-field mode. Samples were illuminated from the bottom.
+
+## Scanning electron microscope and energy dispersive X-ray spectroscopy measurements
+
+SEM images were obtained with a FEI Quanta 600 FEG SEM using an accelerating voltage of 2, 5 or 20 kV as specified in the main text. Images were always collected on fresh areas to avoid degradation under to the electron beam. EDX-SEM maps were acquired with the same FEI Quanta 600 FEG SEM, operating at an accelerating voltage of 20 kV at a working distance of 10 mm. The total acquisition time was 7 minutes to minimise beam damage.
+
+Note on contrast: We observed the strongest contrast between the bulk and the new crystallites observed at a low imaging voltage of 2 keV. Significant Z-contrast between FA₀.₈₃Cs₀.₁₇Pb₀.₅Sn₀.₅I₃ and δ-CsSnI₃ is not expected at normal accelerating voltages of 5–20 keV due to similar electron backscattering coefficients (η = 0.433 and 0.423, respectively, calculated as in Goldstein et al. [60]). The stronger contrast between the bulk and the new crystallites observed at the lower voltage of 2 keV may hence be due to topography or local compositional differences on the surface. [60] The darker appearance of the perovskite around these grains in SEM images under low accelerating voltages could suggest a difference in surface composition or roughness.
+
+## X-ray diffraction measurements
+
+XRD measurements were performed on a Panalytical X’pert Pro XRD diffractometer using a Cu-k(alpha) radiation source with a wavelength of 1.54 eV and a generator voltage of 40 V and current of 40 mA. To track any effects related to oxidation during measurements, we perform several successive scans on each sample. During the 2.5-hour measurement duration we did not observe formation of additional peaks, but did observe a very slight (< 0.1°) peak shift of the perovskite phase to lower angles, which can be explained by the formation of a slightly more FA- and Pb-rich primary phase as small amounts of Cs₂SnI₆ or δ-CsSnI₃ evolve from the initial perovskite composition.
+
+## Time-resolved photoconductivity measurements
+
+Films were removed from aging conditions 20 minutes before each measurement and allowed to cool to room temperature in dark conditions. We used the method detailed in recent work by Lim et al. [39]. A Nd:YAG laser (Ekspla NT342A) excitation source tuned to a wavelength of 600 nm and pumped at 10 Hz with 3.74 ns pulses (full-width-half-maximum, FWHM) was used, attenuated by an optical density filter to achieve a range of fluences. This pulsed light illuminated the entire sample area of 0.25 cm² to uniformly excite the film from the substrate side. A photodetector (Thorlabs, FDS015) was used to detect the pulse source for the optical trigger for transient measurements using a digital oscilloscope. A small DC bias (< 5 mV µm⁻¹, which is three orders of magnitude smaller than that for solar cell characterization, ~ 3 V µm⁻¹) was applied across the in-plane (lateral) electrodes, while the current was monitored by an oscilloscope. As the contact resistance between the perovskite film and Au electrode is fairly small compared to the sample resistance, we employed a two-wire conductivity measurement. A fixed resistor was put in series with the sample in the circuit to be < 1% of the sample resistance. We monitored the voltage drop across the variable series resistor through a parallel oscilloscope (1 MΩ input impedance) to determine the potential drop across the two in-plane Au electrodes on the sample. The measured photo-conductivity value was obtained after 1 min illumination to minimize the time-dependent photo-doping (total measurement time for one sample is < 5 minutes). Transient conductivity (σ) was calculated by the equation  
+$\\sigma =\\frac{{V}_{R}}{{R}_{R}({V}_{app}-{V}_{R})}\\bullet \\frac{l}{wt}$, where Vᵣ is the monitored voltage drop through the fixed resistor, Rᵣ, Vₐₚₚ is the applied bias voltage, l is the channel-to-channel length, w is the channel width, and t is the film thickness. The dark conductivity was determined by taking the mean conductivity measured during the 320 ns before the light pulse.
+
+Significant recombination of photogenerated charges is expected to already take place during the length of the light pulse (3.74 ns). We partially corrected for this by fitting the photoconductivity decay to a mono-exponential decay and extrapolating this back to the beginning of the pulse (t₀) to determine the maximum photoconductivity reached. However, higher-order recombination as well as diffusion- and trap filling-related processes are also expected to take place during the pulse duration, which we do not correct for here. Hence the photoconductivity and hence the sum of mobilities that we determine are likely underestimates and should be understood as lower bounds. We measured the photoconductivity at multiple fluences and selected the fluence which results in the highest sum of mobilities calculated.
+
+## Photoluminescence quantum efficiency measurements
+
+Films were removed from aging conditions 20 minutes before each measurement and allowed to cool to room temperature in dark conditions. PLQE of samples was determined according to the method of de Mello et al. [61]. Samples were placed inside an integrating sphere and excited from the substrate side with a 657 nm continuous wave laser excitation source (Thorlabs) with a large spot size of 0.15 cm². The resulting PL signal was collected via a fibre bundle (Ocean Optics QR600 7 SR125BX) coupled with a spectrometer (QE Pro, Ocean Optics), and a stray light correction was applied to recorded spectra after measuring. Two different spots were measured on each substrate. QFLS was calculated with the following equation:  
+$QFLS= {QFLS}_{rad}+{k}_{B}T\\text{ln}\\left(PLQE\\right)$
+
+## External quantum efficiency measurements
+
+The external quantum efficiency of our devices was determined using Fourier transform photocurrent spectroscopy. Our custom-built set up is based on a Bruker Vertex 80 v Fourier transform interferometer. The solar cells were masked with a metal aperture such that the whole active area was illuminated by a tungsten halogen lamp. To determine the EQE, the photocurrent spectrum of the device under test was divided by that of a calibrated Si reference cell (Newport) of a known EQE. The acquisition time for each photocurrent spectrum was ~ 60 s.
+
+To determine the equivalent short-circuit current density under 1 sun irradiance from the EQE measurements, the overlap integral of the AM1.5 photon flux (φ_AM1.5) spectrum with the EQE was calculated. Explicitly, this is given by  
+$${J}_{sc}=q{\\int }_{0}^{\\infty }d\\lambda EQE\\left(\\lambda \\right){\\phi }_{AM1.5}\\left(\\lambda \\right)$$  
+where q is the elementary charge and λ is the wavelength.
+
+Bandgaps were determined from EQE by plotting (E*EQE)² against E, then finding the point at which a linear fit to the EQE onset goes to 0. [63]
+
+## Variable rate J-V scanning/fast hysteresis measurements
+
+To investigate the transient impact of mobile ions our devices, an in-house-built fast JV setup was used.
+
+The device was connected to a function generator (RSDG 1032X) coupled to a generic operational amplifier, with a 10 ohm resistor placed in series. The forward and reverse voltage sweep was induced by sending a triangular pulse at a set frequency to the device, with the amplifier providing the power to drive the cell. The current was measured by measuring the potential drop across the resistor in series, which was converted to a current by Ohms law. This was done using a digitizer (Picoscope 204A), resulting in a current-to-time trace. A trigger pulse from the function generator was coupled into the digitizer, allowing us to link the trace time to the applied voltage to infer the JV curve. JV curves were measured over a large range of frequencies in logarithmically spaced intervals. Devices were pre-biased at 1.1 V for 5 s, before scanning to – 0.2 V and back to 1.1 V at the specified rate. To reduce noise, 10 repeats for each applied frequency were taken and averaged, and a spline fit to the J-V curves was performed to extract parameters (Vₒc, Jₛc, FF, PCE).
+
+## SCAPS simulations
+
+We carried out J-V curve simulations using the SCAPS-1D (v 3310) drift diffusion package. [59] This package does not take any effects from ion migration into account. The full list of the parameters used (when not varied) is giving in SI table 2. The generation profile used was calculated using a transfer matrix optical model, adjusted to match the short circuit current measured.
+
+# References
+
+1. International Energy Agency, *World Energy Outlook 2022*, **2022**.
+2. A. Kojima, K. Teshima, Y. Shirai, T. Miyasaka, *J. Am. Chem. Soc.* **2009**, *131*, 6050.
+3. M. M. Lee, J. Teuscher, T. Miyasaka, T. N. Murakami, H. J. Snaith, *Science* **2012**, *338*, 643.
+4. Q. Lin, A. Armin, R. C. R. Nagiri, P. L. Burn, P. Meredith, *Nature Photon* **2015**, *9*, 106.
+5. C. S. Ponseca, T. J. Savenije, M. Abdellah, K. Zheng, A. Yartsev, T. Pascher, T. Harlang, P. Chabera, T. Pullerits, A. Stepanov, J.-P. Wolf, V. Sundström, *J. Am. Chem. Soc.* **2014**, *136*, 5189.
+6. S. D. Stranks, G. E. Eperon, G. Grancini, C. Menelaou, M. J. P. Alcocer, T. Leijtens, L. M. Herz, A. Petrozza, H. J. Snaith, *Science* **2013**, *342*, 341.
+7. M. Saliba, J.-P. Correa-Baena, M. Grätzel, A. Hagfeldt, A. Abate, *Angew. Chem. Int. Ed.* **2018**, *57*, 2554.
+8. M. T. Hörantner, T. Leijtens, M. E. Ziffer, G. E. Eperon, M. G. Christoforo, M. D. McGehee, H. J. Snaith, *ACS Energy Lett.* **2017**, *2*, 2506.
+9. X. Tian, S. D. Stranks, F. You, *Sci. Adv.* **2020**, *6*, eabb0055.
+10. M. T. Klug, R. L. Milot, J. B. Patel, T. Green, H. C. Sansom, M. D. Farrar, A. J. Ramadan, S. Martani, Z. Wang, B. Wenger, J. M. Ball, L. Langshaw, A. Petrozza, M. B. Johnston, L. M. Herz, H. J. Snaith, *Energy Environ. Sci.* **2020**, *13*, 1776.
+11. K. J. Savill, A. M. Ulatowski, L. M. Herz, *ACS Energy Lett.* **2021**, *6*, 2413.
+12. S. Hu, K. Otsuka, R. Murdey, T. Nakamura, M. A. Truong, T. Yamada, T. Handa, K. Matsuda, K. Nakano, A. Sato, K. Marumoto, K. Tajima, Y. Kanemitsu, A. Wakamiya, *Optimized Carrier Extraction at Interfaces for 23.6% Efficient Tin–Lead Perovskite Solar Cells*, In Review, **2021**.
+13. R. Lin, Y. Wang, Q. Lu, B. Tang, J. Li, H. Gao, Y. Gao, H. Li, C. Ding, J. Wen, P. Wu, C. Liu, S. Zhao, K. Xiao, Z. Liu, C. Ma, Y. Deng, L. Li, F. Fan, H. Tan, *Nature* **2023**, DOI 10.1038/s41586-023-06278-z.
+14. Z. Wang, L. Zeng, T. Zhu, H. Chen, B. Chen, D. J. Kubicki, A. Balvanz, C. Li, A. Maxwell, E. Ugur, R. Dos Reis, M. Cheng, G. Yang, B. Subedi, D. Luo, J. Hu, J. Wang, S. Teale, S. Mahesh, S. Wang, S. Hu, E. D. Jung, M. Wei, S. M. Park, L. Grater, E. Aydin, Z. Song, N. J. Podraza, Z.-H. Lu, J. Huang, V. P. Dravid, S. De Wolf, Y. Yan, M. Grätzel, M. G. Kanatzidis, E. H. Sargent, *Nature* **2023**, *618*, 74.
+15. L. E. Mundt, J. Tong, A. F. Palmstrom, S. P. Dunfield, K. Zhu, J. J. Berry, L. T. Schelhas, E. L. Ratcliff, *ACS Energy Lett.* **2020**, *5*, 3344.
+16. D. Meggiolaro, D. Ricciarelli, A. A. Alasmari, F. A. S. Alasmary, F. De Angelis, *J. Phys. Chem. Lett.* **2020**, *11*, 3546.
+17. T. Leijtens, R. Prasanna, A. Gold-Parker, M. F. Toney, M. D. McGehee, *ACS Energy Lett.* **2017**, *2*, 2159.
+18. S. Hu, J. A. Smith, H. J. Snaith, A. Wakamiya, *Precision Chemistry* **2023**, *1*, 69.
+19. K. Domanski, E. A. Alharbi, A. Hagfeldt, M. Grätzel, W. Tress, *Nat Energy* **2018**, *3*, 61.
+20. T. Leijtens, R. Prasanna, K. A. Bush, G. E. Eperon, J. A. Raiford, A. Gold-Parker, E. J. Wolf, S. A. Swifter, C. C. Boyd, H.-P. Wang, M. F. Toney, S. F. Bent, M. D. McGehee, *Sustainable Energy Fuels* **2018**, *2*, 2450.
+21. R. Prasanna, T. Leijtens, S. P. Dunfield, J. A. Raiford, E. J. Wolf, S. A. Swifter, J. Werner, G. E. Eperon, C. de Paula, A. F. Palmstrom, C. C. Boyd, M. F. A. M. van Hest, S. F. Bent, G. Teeter, J. J. Berry, M. D. McGehee, *Nat Energy* **2019**, *4*, 939.
+22. K. Xiao, R. Lin, Q. Han, Y. Hou, Z. Qin, H. T. Nguyen, J. Wen, M. Wei, V. Yeddu, M. I. Saidaminov, Y. Gao, X. Luo, Y. Wang, H. Gao, C. Zhang, J. Xu, J. Zhu, E. H. Sargent, H. Tan, *Nat Energy* **2020**, *5*, 870.
+23. H. Jin, M. D. Farrar, J. M. Ball, A. Dasgupta, P. Caprioglio, S. Narayanan, R. D. J. Oliver, F. M. Rombach, B. W. J. Putland, M. B. Johnston, H. J. Snaith, *Adv Funct Materials* **2023**, 2303012.
+24. P. Wu, J. Wen, Y. Wang, Z. Liu, R. Lin, H. Li, H. Luo, H. Tan, *Advanced Energy Materials* **2022**, *12*, 2202948.
+25. M. V. Khenkin, E. A. Katz, A. Abate, G. Bardizza, J. J. Berry, C. Brabec, F. Brunetti, V. Bulović, Q. Burlingame, A. Di Carlo, R. Cheacharoen, Y.-B. Cheng, A. Colsmann, S. Cros, K. Domanski, M. Dusza, C. J. Fell, S. R. Forrest, Y. Galagan, D. Di Girolamo, M. Grätzel, A. Hagfeldt, E. von Hauff, H. Hoppe, J. Kettle, H. Köbler, M. S. Leite, S. Liu, Y.-L. Loo, J. M. Luther, C.-Q. Ma, M. Madsen, M. Manceau, M. Matheron, M. McGehee, R. Meitzner, M. K. Nazeeruddin, A. F. Nogueira, Ç. Odabaşı, A. Osherov, N.-G. Park, M. O. Reese, F. De Rossi, M. Saliba, U. S. Schubert, H. J. Snaith, S. D. Stranks, W. Tress, P. A. Troshin, V. Turkovic, S. Veenstra, I. Visoly-Fisher, A. Walsh, T. Watson, H. Xie, R. Yıldırım, S. M. Zakeeruddin, K. Zhu, M. Lira-Cantu, *Nat Energy* **2020**, *5*, 35.
+26. R. Lin, J. Xu, M. Wei, Y. Wang, Z. Qin, Z. Liu, J. Wu, K. Xiao, B. Chen, S. M. Park, G. Chen, H. R. Atapattu, K. R. Graham, J. Xu, J. Zhu, L. Li, C. Zhang, E. H. Sargent, H. Tan, *Nature* **2022**, *603*, 73.
+27. S. Lv, W. Gao, Y. Liu, H. Dong, N. Sun, T. Niu, Y. Xia, Z. Wu, L. Song, C. Ran, L. Fu, Y. Chen, *Journal of Energy Chemistry* **2022**, *65*, 371.
+28. P. K. Nayak, S. Mahesh, H. J. Snaith, D. Cahen, *Nat Rev Mater* **2019**, *4*, 269.
+29. J. Pascual, M. Flatken, R. Félix, G. Li, S. Turren‐Cruz, M. H. Aldamasy, C. Hartmann, M. Li, D. Di Girolamo, G. Nasti, E. Hüsam, R. G. Wilks, A. Dallmann, M. Bär, A. Hoell, A. Abate, *Angew. Chem. Int. Ed.* **2021**, anie.202107599.
+30. R. Lin, K. Xiao, Z. Qin, Q. Han, C. Zhang, M. Wei, M. I. Saidaminov, Y. Gao, J. Xu, M. Xiao, A. Li, J. Zhu, E. H. Sargent, H. Tan, *Nat Energy* **2019**, *4*, 864.
+31. P. Mauersberger, F. Huber, *Acta Crystallogr B Struct Sci* **1980**, *36*, 683.
+32. I. Chung, J.-H. Song, J. Im, J. Androulakis, C. D. Malliakas, H. Li, A. J. Freeman, J. T. Kenney, M. G. Kanatzidis, *J. Am. Chem. Soc.* **2012**, *134*, 8579.
+33. V. J. ‐Y. Lim, A. M. Ulatowski, C. Kamaraki, M. T. Klug, L. Miranda Perez, M. B. Johnston, L. M. Herz, *Advanced Energy Materials* **2022**, 2200847.
+34. X. Qiu, B. Cao, S. Yuan, X. Chen, Z. Qiu, Y. Jiang, Q. Ye, H. Wang, H. Zeng, J. Liu, M. G. Kanatzidis, *Solar Energy Materials and Solar Cells* **2017**, *159*, 227.
+35. S. Xu, A. Libanori, G. Luo, J. Chen, *iScience* **2021**, *24*, 102235.
+36. C. C. Stoumpos, C. D. Malliakas, M. G. Kanatzidis, *Inorg. Chem.* **2013**, *52*, 9019.
+37. J. Thiesbrummel, V. M. Le Corre, F. Peña‐Camargo, L. Perdigón‐Toro, F. Lang, F. Yang, M. Grischek, E. Gutierrez‐Partida, J. Warby, M. D. Farrar, S. Mahesh, P. Caprioglio, S. Albrecht, D. Neher, H. J. Snaith, M. Stolterfoht, *Adv. Energy Mater.* **2021**, 2101447.
+38. F. Pena-Camargo, J. Thiesbrummel, H. Hempel, A. Musiienko, V. M. L. Corre, J. Diekmann, J. Warby, T. Unold, F. Lang, D. Neher, M. Stolterfoht, *Applied Physics Reviews* **2022**.
+39. J. Lim, M. Kober-Czerny, Y.-H. Lin, J. M. Ball, N. Sakai, E. A. Duijnstee, M. J. Hong, J. G. Labram, B. Wenger, H. J. Snaith, *Nat Commun* **2022**, *13*, 4201.
+40. C. L. Davies, M. R. Filip, J. B. Patel, T. W. Crothers, C. Verdi, A. D. Wright, R. L. Milot, F. Giustino, M. B. Johnston, L. M. Herz, *Nat Commun* **2018**, *9*, 293.
+41. C. Kamaraki, M. T. Klug, V. J. ‐Y. Lim, N. Zibouche, L. M. Herz, M. S. Islam, C. Case, L. Miranda Perez, *Advanced Energy Materials* **2024**, 2302916.
+42. S. Lee, M. Y. Woo, C. Kim, K. W. Kim, H. Lee, S. B. Kang, J. M. Im, M. J. Jeong, Y. Hong, J. W. Yoon, S. Y. Kim, K. Heo, K. Zhu, J.-S. Park, J. H. Noh, D. H. Kim, *Chemical Engineering Journal* **2024**, *479*, 147587.
+43. J. Zillner, H. Boyen, P. Schulz, J. Hanisch, N. Gauquelin, J. Verbeeck, J. Küffner, D. Desta, L. Eisele, E. Ahlswede, M. Powalla, *Adv Funct Materials* **2022**, *32*, 2109649.
+44. R. L. Milot, G. E. Eperon, T. Green, H. J. Snaith, M. B. Johnston, L. M. Herz, *J. Phys. Chem. Lett.* **2016**, *7*, 4178.
+45. M. Stolterfoht, C. M. Wolff, Y. Amir, A. Paulke, L. Perdigon, P. Caprioglio, D. Neher, *Energy & Environmental Science* **2017**, *10*, 1530.
+46. F. M. Rombach, S. A. Haque, T. J. Macdonald, *Energy Environ. Sci.* **2021**, *14*, 5161.
+47. M. Saliba, T. Matsui, K. Domanski, J.-Y. Seo, A. Ummadisingu, S. M. Zakeeruddin, J.-P. Correa-Baena, W. R. Tress, A. Abate, A. Hagfeldt, M. Gratzel, *Science* **2016**, *354*, 206.
+48. C. Eames, J. M. Frost, P. R. F. Barnes, B. C. O’Regan, A. Walsh, M. S. Islam, *Nat Commun* **2015**, *6*, 7497.
+49. S. R. Pering, P. J. Cameron, *Mater. Adv.* **2022**, *3*, 7918.
+50. W. Tress, N. Marinova, T. Moehl, S. M. Zakeeruddin, M. K. Nazeeruddin, M. Grätzel, *Energy Environ. Sci.* **2015**, *8*, 995.
+51. J. Thiesbrummel, S. Shah, E. Gutierrez-Partida, F. Zu, F. Camargo, S. Zeiske, J. Diekmann, F. Ye, K. Peters, K. Brinkmann, J. Warby, Q. Jeangros, F. Lang, Y. Wu, S. Albrecht, T. Riedl, A. Armin, D. Neher, N. Koch, V. Corre, H. Snaith, M. Stolterfoht, *Ion Induced Field Screening Governs the Early Performance Degradation of Perovskite Solar Cells*, In Review, **2023**.
+52. C. Li, A. Guerrero, S. Huettner, J. Bisquert, *Nat Commun* **2018**, *9*, 5113.
+53. N. Aristidou, C. Eames, I. Sanchez-Molina, X. Bu, J. Kosco, M. S. Islam, S. A. Haque, *Nat Commun* **2017**, *8*, 15218.
+54. V. M. Le Corre, J. Diekmann, F. Peña-Camargo, J. Thiesbrummel, N. Tokmoldin, E. Gutierrez-Partida, K. P. Peters, L. Perdigón-Toro, M. H. Futscher, F. Lang, J. Warby, H. J. Snaith, D. Neher, M. Stolterfoht, *Solar RRL* **2022**, *6*, 2100772.
+55. H. Wang, A. Guerrero, A. Bou, A. M. Al-Mayouf, J. Bisquert, *Energy Environ. Sci.* **2019**, *12*, 2054.
+56. S. P. Senanayak, K. Dey, R. Shivanna, W. Li, D. Ghosh, Y. Zhang, B. Roose, S. J. Zelewski, Z. Andaji-Garmaroudi, W. Wood, N. Tiwale, J. L. MacManus-Driscoll, R. H. Friend, S. D. Stranks, H. Sirringhaus, *Nat. Mater.* **2023**, *22*, 216.
+57. K. Dey, D. Ghosh, M. Pilot, S. R. Pering, P. Deswal, S. P. Senanayak, M. S. Islam, S. D. Stranks, **n.d.**
+58. N. E. Courtier, J. M. Cave, J. M. Foster, A. B. Walker, G. Richardson, *Energy Environ. Sci.* **2019**, *12*, 396.
+59. M. Burgelman, P. Nollet, S. Degrave, *Thin Solid Films* **2000**, *361–362*, 527.
+60. J. I. Goldstein, D. E. Newbury, J. R. Michael, N. W. M. Ritchie, J. H. J. Scott, D. C. Joy, *Scanning Electron Microscopy and X-Ray Microanalysis*, Springer New York, New York, NY, **2018**.
+61. J. C. de Mello, H. F. Wittmann, R. H. Friend, *Adv. Mater.* **1997**, *9*, 230.
+62. P. Caprioglio, M. Stolterfoht, C. M. Wolff, T. Unold, B. Rech, S. Albrecht, D. Neher, *Adv. Energy Mater.* **2019**, *9*, 1901631.
+63. R. Carron, C. Andres, E. Avancini, T. Feurer, S. Nishiwaki, S. Pisoni, F. Fu, M. Lingg, Y. E. Romanyuk, S. Buecheler, A. N. Tiwari, *Thin Solid Films* **2019**, *669*, 482.
+
+# Supplementary Files
+
+- [SupplementaryInformation.docx](https://assets-eu.researchsquare.com/files/rs-4502930/v1/b528a994b0ec81131975fc9c.docx)