A method for predicting output current of a hydrogenated amorphous silicon thin film transistor based on a physical information neural network
By combining a physical information neural network with a hybrid loss function of an optimized analytical model, the problem of high-precision prediction of the output current of hydrogenated amorphous silicon thin-film transistors was solved, achieving efficient, accurate, and physically constrained current prediction under limited data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-12
Smart Images
Figure CN122197772A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of semiconductor device simulation and modeling technology, and relates to a method for predicting the output current of hydrogenated amorphous silicon thin film transistors based on physical information neural networks. Background Technology
[0002] In the manufacture of display panels and large sensor arrays, hydrogenated amorphous silicon thin-film transistors (TFTs) are widely used as pixel switches or driving elements due to their mature fabrication process and relatively low cost. The electrical characteristics of the device, especially the output current (IV characteristic), directly determine the function and performance of the circuit. Therefore, the ability to quickly and accurately predict the output current of the thin-film transistor is crucial during the device design, process optimization, and circuit simulation stages.
[0003] Currently, the main methods for obtaining device characteristics include experimental measurement, process and device simulation (TCAD), and data-driven modeling. While experimental measurement methods are accurate, they are time-consuming, labor-intensive, and costly, making them unsuitable for large-scale design space exploration. TCAD simulation predicts device behavior by solving semiconductor physics equations (such as the Poisson equation and the carrier continuity equation), but achieving ideal simulation results often requires extensive manual debugging experience and significant time costs, presenting a high technical barrier.
[0004] To reduce reliance on TCAD simulations, researchers have proposed various deep learning models based on TCAD simulation data to predict parameters such as the IV curve and threshold voltage of thin-film transistors (TFTs). These purely data-driven methods bypass the empirical requirements of TCAD simulations to some extent; however, their performance heavily depends on the scale and density of the training data, requiring the collection of large amounts of evenly distributed data. This constitutes a cost and time barrier to data acquisition in practical R&D. Furthermore, purely data-driven models may have insufficient generalization ability in sparse data regions, and their predictions may violate fundamental physical laws.
[0005] Physics-Informed Neural Networks (PINNs) offer a novel approach to device modeling in small-data scenarios. This method embeds known physical laws (often in the form of partial differential equations or analytical models) as soft constraints into the loss function of the neural network, guiding the model to follow physical laws during training, thus achieving good generalization performance even on small datasets. For example, some studies have applied PINN to current prediction in field-effect transistors (FETs), constructing a loss function containing complex coupled partial differential equations such as the Poisson equation and the Boltzmann transport equation through automatic differentiation; however, the solution process is complex and computationally intensive. Other studies have used PINN to predict the IV curve of metal-oxide-semiconductor field-effect transistors (MOSFETs). The core of this approach is to use a neural network to predict a correction factor to correct a complex compact model (such as BSIM-CMG); however, this compact model itself is structurally complex and specific to MOSFETs, not directly applicable to hydrogenated amorphous silicon thin-film transistors.
[0006] Therefore, there is a clear need in the existing technology: to develop a modeling method specifically for hydrogenated amorphous silicon thin-film transistors that can achieve high-precision current prediction with limited experimental or simulation data support, and has clear physical constraints and high computational efficiency. Summary of the Invention
[0007] In view of this, the purpose of the present invention is to provide a method for predicting the output current of hydrogenated amorphous silicon thin film transistors based on a physical information neural network.
[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network includes the following steps: S1: Obtain the original dataset containing the structural parameters and corresponding output currents of hydrogenated amorphous silicon thin-film transistors, and preprocess the original dataset to obtain the training set and validation set. The structural parameters include the thickness of the source and drain amorphous silicon (AS) active layers. Thickness of the amorphous silicon remain (AS remain) active layer in the channel Gate Insulator (GI) thickness Gate-source voltage The output current is the drain current. ; S2: Constructing a data-driven neural network prediction model; S3: Obtain the optimized analytical model of the amorphous silicon thin-film transistor. The optimized analytical model of the amorphous silicon thin-film transistor is obtained by introducing structural parameter adjustment factors into the RPI model or other commonly used SPICE models. and The model, wherein the structural parameter adjustment factor and The calculation formulas are as follows: , and These represent the thickness of the active layer in the channel and the thickness of the active layers at the source and drain, respectively. and It is a constant; S4: Construct a total loss function that includes data item loss and physical item loss. The data item loss is the mean square error (MSE) between the predicted current value output by the data-driven neural network prediction model and the actual current value in the training set. The physical item loss is the mean square error between the predicted current value output by the data-driven neural network prediction model and the calculated current value obtained by the optimized physical analytical model based on the same structural parameters. S5: Train the data-driven neural network prediction model using the training set, and minimize the total loss function through an optimization algorithm to update the network parameters of the data-driven neural network prediction model; S6: The trained data-driven neural network prediction model is tested using the test set to test the final model used to predict the output current of hydrogenated amorphous silicon thin-film transistors.
[0009] Furthermore, in S2, the data-driven neural network prediction model is a convolutional neural network, a backpropagation neural network, or a deep neural network.
[0010] Furthermore, when the data-driven neural network prediction model is a convolutional neural network, it sequentially includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, and a fully connected layer. The number of filters in the first convolutional layer, the second convolutional layer, and the third convolutional layer are 64, 128, and 256, respectively, and the convolutional kernel size is (2,1). The first pooling layer, the second pooling layer, and the third pooling layer are all max pooling layers with a size of 2.
[0011] Furthermore, the convolutional neural network prediction model also includes a Dropout layer with a dropout probability of 60%.
[0012] Furthermore, in the training hyperparameters of the convolutional neural network prediction model, the initial learning rate is set to 0.03, the learning rate decay factor is set to 0.9, and the batch size is set to 70.
[0013] Furthermore, in S3, the optimized physical analysis model is an optimized RPI model, which is represented by different formulas according to different working regions. The working regions include the cutoff region, the subthreshold region, the linear region, and the saturation region.
[0014] Furthermore, the optimized RPI model is expressed as: Calculate the current value in the cutoff region. for: ,in, To minimize leakage current, To describe the effect of current on drain-source voltage Parameters of dependency, To describe the effect of current on gate-source voltage Dependency parameters, V ds This represents the small-signal increment component of the source-drain voltage. Determined by the resolution of the measuring equipment; In the subthreshold region, the current value is calculated. for: ,in, V dse From the formula definition, K Boltzmann's constant, T Absolute temperature For unit charge, For carrier mobility, For the thickness of the gate insulation layer , For the thickness of the active layer in the channel , The width of the transistor channel. The length of the transistor channel. The dielectric constant of the gate insulating layer material is . The dielectric constant of hydrogenated amorphous silicon is... Threshold voltage, This represents the dark state carrier concentration. It is a flat band voltage. Characteristic voltage, V e From the formula definition; Calculate the current value in the linear and saturation regions. for: ,in, For field-effect mobility, It is a gate insulator capacitor. Where is the threshold voltage; V dse From the formula definition,V sate Defined as , The saturation proportionality constant is set to 0.6. This is used to prevent a sudden change in current when transitioning from the linear region to the saturation region. In the linear region, i.e., V DS < V sate hour, V dse = V DS In the saturation region, i.e. V DS > V sate hour, V dse = V sate .
[0015] Furthermore, in S3, the optimized physical analysis model is replaced by an Amorphous Model Level 35 model, a Hack-Shaw model, a Nathan low-leakage model, or a charge-controlled model.
[0016] Furthermore, in step S4, the physical term loss accounts for 41% of the weight in the total loss function; in step S1, the preprocessing includes taking the logarithm of the structural parameters and output current in the original data, and normalizing the logarithm-processed data, and the training set accounts for 70% of the total dataset after preprocessing, and the validation set accounts for 30%; the method also includes step S7: selecting transistor data with channel active layer thicknesses of 80nm and 20nm as test sets, and inputting them into the final model for prediction testing.
[0017] Furthermore, when the data-driven neural network prediction model is a BP neural network or a deep neural network, the total loss function constructed in S4, which includes data item loss and physical item loss, is also coupled to the loss function of the BP neural network or deep neural network model.
[0018] The beneficial effects of this invention are as follows: (1) This invention couples an optimized physical analytical model (such as an RPI model with structural parameter adjustment factors) as a physical constraint term into the training process of a data-driven neural network in the form of a loss function. This physical information neural network framework enables the model to not only learn the statistical regularities in limited experimental or simulation data, but also to be forced to a solution space that conforms to the basic physical principles of the device. Therefore, even in regions where training data is scarce (especially for extreme size parameters, such as ultrathin channel layers), the model can maintain excellent prediction accuracy and generalization ability, effectively overcoming the problem of pure data-driven models' dependence on large amounts of data and their sharp performance drop in sparse data regions.
[0019] (2) The physical term loss function directly measures the deviation between the neural network prediction and the physical model calculation, ensuring that the model output will not seriously violate known physical laws under any input conditions. This makes the prediction results not only numerically accurate, but also clearly physically interpretable, avoiding the absurd predictions that may be generated by pure black box models that are physically unreasonable, thereby greatly improving the reliability of the prediction results and their practical value in device design and simulation.
[0020] (3) Compared to directly solving complex coupled physical partial differential equations (such as TCAD simulation) or using extremely complex compact models for fitting, the optimized physical analytical model adopted in this invention is relatively simple, and the forward prediction process of the neural network is extremely fast after training. This method integrates physical knowledge in the model training stage and enjoys the efficiency of data-driven models in the application stage, providing an efficient tool for rapid, large-scale device performance prediction and design space exploration.
[0021] (4) This invention specifically addresses the impact of device structural parameters (source-drain active layer thickness and channel active layer thickness) on device characteristics of hydrogenated amorphous silicon thin film transistors. In the physical model, a structural parameter adjustment factor directly related to the source-drain active layer thickness and channel active layer thickness is introduced. and This optimization allows physical constraints to more precisely reflect the impact of critical dimensional changes on current transport characteristics in actual processes, thereby making physical guidance more accurate and significantly improving the prediction accuracy of the final model, especially its sensitivity to changes in thickness parameters.
[0022] (5) On the one hand, this invention provides a series of verified preferred implementation schemes, from network architecture (such as the specific number of convolutional neural network layers and filters), regularization methods (Dropout rate), training hyperparameters to physical loss weights (41%), which have clear operability. On the other hand, the claims also clarify the flexibility of the framework: the data-driven network can be CNN, BPNN, or DNN; the physical model can be a variety of choices such as optimized RPI or Amorphous Level 35 model. This feature of "fixed core framework and replaceable components" enables the method of this invention to adapt to different data conditions, accuracy requirements, and computing resources, and has good universality and scalability.
[0023] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 The basic architecture of the prediction network model; Figure 2 This is a model structure diagram of an a-Si TFT; Figure 3 shows the prediction results of PINN and CNN without physical information constraints; Figure 3(a) shows the prediction results with a channel layer thickness of 80 nm; Figure 3(b) shows the prediction results with a channel layer thickness of 20 nm. Detailed Implementation
[0025] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0026] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0027] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0028] Example 1 This embodiment provides a method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network. For example... Figure 1 and Figure 2 As shown, in Figure 1 and Figure 2 China , , Substitute , , To optimize image representation. This method specifically includes the following steps: S101. Dataset Acquisition and Preprocessing.
[0029] Experimental measurement data of hydrogenated amorphous silicon thin-film transistors fabricated using a standard 5-mask process were obtained to form the original dataset, in which data were set to... The value is 10V. The original dataset includes multiple sets of input features and corresponding output labels. The input features are key dimensional parameters and electrical bias parameters of the device, specifically including: the thickness of the active layer below the source and drain electrodes. , thickness of active layer in the trench Thickness of the gate insulation layer Gate-source voltage The output label corresponds to the drain current. .
[0030] First, the original dataset is cleaned to remove obviously anomalous and outdated data. Next, all input features ( , , , ) and output labels The data is processed by taking a base-10 logarithm to standardize the order of magnitude of all parameters. Then, a random permutation function (such as `randperm`) is used to shuffle the data order to improve the generalization ability of the subsequently trained model. The shuffled dataset is then divided into training and validation sets at a ratio of 70% and 30%, respectively. A reshaping function (such as `reshape`) is used to convert the input features of each sample into a two-dimensional array of size (4,1) to suit the input requirements of the convolutional neural network. Finally, the data is normalized and scaled to the [0,1] interval.
[0031] S102. Construct a convolutional neural network prediction model.
[0032] Construct a convolutional neural network for regression prediction. The specific architecture of this network is as follows: Figure 1 As shown, it includes, in sequence: The input layer receives input data of size (4,1).
[0033] The first convolutional layer uses 64 filters of size (2×1) and employs "same" padding to preserve the feature map size.
[0034] The first pooling layer is the max pooling layer with a pooling size of 2.
[0035] The second convolutional layer uses 128 filters of size (2×1).
[0036] The second pooling layer is the max pooling layer with a pooling size of 2.
[0037] The third convolutional layer uses 256 filters of size (2×1).
[0038] The third pooling layer is the maximum pooling layer with a pooling size of 2.
[0039] The flattening layer flattens multidimensional features into a one-dimensional vector.
[0040] The fully connected layer ultimately outputs a scalar value, namely the predicted logarithm of the current.
[0041] A dropout layer with a dropout probability of 60% is introduced into the network to mitigate the risk of overfitting. Simultaneously, L2 regularization is applied to the network weights. The hyperparameters for model training are set as follows: initial learning rate of 0.03, learning rate decay factor of 0.9, batch size of 70, mean squared error as the base loss function, and the Adam optimizer.
[0042] S103. Obtain and optimize the physical constraint model.
[0043] The RPI model was chosen as the basis for physical constraints. This is to more accurately reflect the structural characteristics of hydrogenated amorphous silicon thin-film transistors (such as...). Figure 2 As shown), in particular, the influence of the thickness of the source / drain active layer and the channel active layer on the current is addressed by introducing two structural parameter adjustment factors into the basic RPI model. and The model was optimized. The optimized model formula is as follows: The adjustment factor is defined as:
[0044] in, and These are the normal values obtained by fitting experimental data.
[0045] The optimized physical model is described according to the device's working region as follows: 1) Cut-off area:
[0046] To minimize leakage current, To describe the effect of current on drain-source voltage Dependency parameters, To describe the effect of current on gate-source voltage Dependency parameters, It is determined by the resolution of the measuring equipment.
[0047] 2) Subthreshold region:
[0048] in, For unit charge, For carrier mobility, For the thickness of the gate insulation layer , For the thickness of the active layer in the channel , The width of the transistor channel. The length of the transistor channel. The dielectric constant of the gate insulating layer material is . The dielectric constant of hydrogenated amorphous silicon is... Threshold voltage, This represents the dark state carrier concentration. It is a flat band voltage. This is the characteristic voltage.
[0049] 3) Linear and saturated regions:
[0050] in, It is the field-effect mobility. It is the capacitance per unit area of the gate insulator.
[0051] S104. Construct a hybrid loss function for a physical information neural network.
[0052] Design the total loss function of a physical information neural network. It is caused by the loss of data items. and physical item loss Weighted composition:
[0053] in, The weighting coefficient for the physical term loss is set to 0.7 in this embodiment. The data term loss is the neural network prediction value. Mean square error between the measured value y and the actual measured value y: The physical term loss is the neural network prediction value. Comparison with the calculated values of the optimized AIM-SPICE physical model Mean square error between: .here represent , represent , represent .
[0054] S105, Model Training.
[0055] The training and validation sets obtained in step S101 are input into the convolutional neural network constructed in step S102. The predicted current is calculated using forward propagation, and the physical term loss is calculated using the physical model from step S103. Then, using the total loss function defined in step S104, all weights and bias parameters of the network are automatically optimized and updated via the error backpropagation algorithm. Validation is performed at each epoch to prevent overfitting. The training process continues until the loss function converges.
[0056] S106, Model Testing.
[0057] To test the model's generalization ability on small datasets, especially its prediction ability for extreme sizes not present in the training set, the thickness of the active layer in the channel was selected. Transistor data at 80nm and 20nm were used as independent test sets and input into the finally trained model for prediction. The prediction results of the Physical Information Neural Network (PINN) in this embodiment were compared with the prediction results of a regular Convolutional Neural Network (CNN) without physical constraints. As shown in Figures 3(a) and 3(b), when the channel layer thickness of the a-Si TFT is less than 30nm, the device performance prediction shows a significant deviation. Two sets of test datasets were used to test the accuracy and generalization ability of the model, namely the structural parameter input data for channel thicknesses of 20nm and 80nm. Figures 3(a) and 3(b) compare the transfer characteristic curves predicted by the PINN (red dashed line) and CNN (green dashed line) models for channel layer values of 20nm and 80nm, and compare them with the actual measured transfer characteristic current curves of the device (black scatter dots). Regardless of the channel layer thickness, PINN always achieves higher fitting accuracy. In contrast, although the CNN model performs accurately at 80nm, it shows a significant prediction error at 20nm. Figure 3(a) shows the prediction data for a channel layer thickness of 80 nm. Both the CNN and PINN models perform well, with PINN showing better fitting in the cutoff and subthreshold regions. However, overall, there is a small difference in accuracy between the CNN and PINN models. This may be due to the abundance of data in the 30–150 nm range, coupled with the relatively small impact of channel thickness variations on device performance within this range. Therefore, physical constraints have a weaker influence on model predictions, resulting in only a slight difference in accuracy between PINN and CNN. Figure 3(b) shows the prediction data for a channel layer thickness of 20 nm. PINN's accuracy is significantly higher than CNN's. This is likely because data is scarce in the range less than 30 nm, and the physical constraints embedded in PINN play a more significant role here, making its predictions more consistent with experimental results.
[0058] Example 2 The main difference between this embodiment and Embodiment 1 is that the convolutional neural network prediction model in step S102 is replaced with a backpropagation neural network model.
[0059] Specifically, after data preprocessing, a BP neural network model is created using a feedforward neural network function (such as the `feedforwardnet` function in MATLAB), specifying the number of hidden layer neurons (e.g., 10). The hybrid loss function containing physical constraints, constructed in step S104, is coupled into the training process of this BP neural network. The network's training parameters are then configured, including the training algorithm, maximum number of iterations, and target error. Finally, the BP neural network model with physical constraints is trained and tested using the same training, validation, and test sets, thereby enabling the prediction of the output current of hydrogenated amorphous silicon thin-film transistors.
[0060] Example 3 The main difference between this embodiment and Embodiment 1 is that the convolutional neural network prediction model in step S102 is replaced with a deep neural network model.
[0061] Specifically, a deep neural network architecture is constructed by stacking multiple fully connected layers. The hybrid loss function containing physical constraints, constructed in step S104, is coupled into the loss calculation of this DNN model. Training parameters such as optimizer, learning rate, and number of training epochs are configured. Subsequently, the physical information-constrained deep neural network model is trained and tested using the same training set, validation set, and test set, thereby enabling the prediction of the output current of hydrogenated amorphous silicon thin-film transistors.
[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
[0063] Example 4 The main difference between this embodiment and Embodiment 1 is that the RPI physics model in step S103 is replaced with the RPI AmorphousModle (Level 35) model.
[0064] Specifically, the leakage current model of this model is as follows:
[0065] in
[0066]
[0067]
[0068] RThis is the sum of the effective resistances of the source and drain. V ds Source-drain voltage, M It's the sharpness of the knee area. N s The concentration of charge carriers varies, and the physical mechanisms affecting charge carrier density differ in different regions. N s different.
[0069] Example 5 The main difference between this embodiment and Embodiment 1 is that the RPI physical model in step S103 is replaced with the Hack-Shaw model, the Nathan low-leakage model, or the charge-controlled model.
[0070] The specific Hack-Shaw model establishes an analytical model of drift current based on the bandgap state and trap charge of amorphous silicon, distinguishing between the linear region and the saturation region; the Nathan low-leakage model focuses on subthreshold and reverse leakage, equating the drain terminal to a parasitic pin diode, and focuses on describing low-current mechanisms such as heat generation and trap-assisted tunneling; the charge-controlled model uses channel charge as the core variable, uniformly describing DC and dynamic characteristics, naturally including capacitance and transient behavior, and is suitable for high-precision circuit simulation.
[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network, characterized in that: Includes the following steps: S1: Obtain the original dataset containing the structural parameters and corresponding output currents of hydrogenated amorphous silicon thin-film transistors, and preprocess the original dataset to obtain a training set and a validation set. The structural parameters include the thickness of the source and drain amorphous silicon active layers (AS). The thickness of the amorphous silicon active layer AS in the channel GI gate insulation layer thickness Gate-source voltage The output current is the drain current. ; S2: Constructing a data-driven neural network prediction model; S3: Obtain the optimized analytical model of the amorphous silicon thin-film transistor. The optimized analytical model of the amorphous silicon thin-film transistor is obtained by introducing structural parameter adjustment factors into the RPI model or other commonly used SPICE models. and The model, wherein the structural parameter adjustment factor and The calculation formulas are as follows: , and These represent the thickness of the active layer in the channel and the thickness of the active layers at the source and drain, respectively. and It is a constant; S4: Construct a total loss function that includes data item loss and physical item loss. The data item loss is the mean square error (MSE) between the predicted current value output by the data-driven neural network prediction model and the actual current value in the training set. The physical item loss is the mean square error between the predicted current value output by the data-driven neural network prediction model and the calculated current value obtained by the optimized physical analytical model based on the same structural parameters. S5: Train the data-driven neural network prediction model using the training set, and minimize the total loss function through an optimization algorithm to update the network parameters of the data-driven neural network prediction model; S6: The trained data-driven neural network prediction model is tested using the test set to test the final model used to predict the output current of hydrogenated amorphous silicon thin-film transistors.
2. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 1, characterized in that: In S2, the data-driven neural network prediction model is a convolutional neural network, a backpropagation neural network, or a deep neural network.
3. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 2, characterized in that: When the data-driven neural network prediction model is a convolutional neural network, it sequentially includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, and a fully connected layer. The number of filters in the first convolutional layer, the second convolutional layer, and the third convolutional layer are 64, 128, and 256, respectively, and the convolutional kernel size is (2,1). The first pooling layer, the second pooling layer, and the third pooling layer are all max pooling layers with a size of 2.
4. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 3, characterized in that: The convolutional neural network prediction model also includes a Dropout layer with a dropout probability of 60%.
5. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 3 or 4, characterized in that: In the training hyperparameters of the convolutional neural network prediction model, the initial learning rate is set to 0.03, the learning rate decay factor is set to 0.9, and the batch size is set to 70.
6. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 1, characterized in that: In S3, the optimized physical analysis model is the optimized RPI model, which is represented by different formulas according to different working regions. The working regions include the cutoff region, the subthreshold region, the linear region, and the saturation region.
7. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 6, characterized in that: The optimized RPI model is represented as follows: Calculate the current value in the cutoff region. for: ,in, To minimize leakage current, To describe the effect of current on drain-source voltage Dependency parameters, To describe the effect of current on gate-source voltage Dependency parameters, V ds This represents the small-signal increment component of the source-drain voltage. Determined by the resolution of the measuring equipment; In the subthreshold region, the current value is calculated. for: ,in, V dse From the formula definition, K Boltzmann's constant, T Absolute temperature For unit charge, For carrier mobility, Thickness of the gate insulation layer , For the thickness of the active layer in the channel , The width of the transistor channel. The length of the transistor channel. The dielectric constant of the gate insulating layer material is . The dielectric constant of hydrogenated amorphous silicon is... Threshold voltage, The dark state carrier concentration, It is a flat band voltage. Characteristic voltage, V e From the formula definition; Calculate the current value in the linear and saturation regions. for: ,in, For field-effect mobility, It is a gate insulator capacitor. Where is the threshold voltage; V dse From the formula definition, V sate Defined as , The saturation proportionality constant is set to 0.
6. This is used to prevent a sudden change in current when transitioning from the linear region to the saturation region. In the linear region, i.e., V DS < V sate hour, V dse = V DS In the saturation region, i.e. V DS > V sate hour, V dse = V sate .
8. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 1, characterized in that: In step S3, the optimized physical analysis model is replaced by an Amorphous Model Level 35 model, a Hack-Shaw model, a Nathan low-leakage model, or a charge-controlled model.
9. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 1, characterized in that: In step S4, the physical term loss accounts for 41% of the weight in the total loss function; in step S1, the preprocessing includes taking the logarithm of the structural parameters and output current in the original data, and normalizing the logarithm-processed data, wherein the training set accounts for 70% of the total dataset after preprocessing, and the validation set accounts for 30%; the method further includes step S7: selecting transistor data with channel active layer thicknesses of 80nm and 20nm as test sets, and inputting them into the final model for testing.
10. The method for predicting the output current of hydrogenated amorphous silicon thin-film transistors based on a physical information neural network according to claim 2, characterized in that: When the data-driven neural network prediction model is a BP neural network or a deep neural network, the total loss function constructed in S4, which includes data item loss and physical item loss, is also coupled to the loss function of the BP neural network or deep neural network model.