Physical guided neural network-based p-GaN gate HEMT modeling method

Abstract

The invention relates to the field of gallium nitride high electron mobility transistor (GaN HEMT) compact modeling, and provides a p-GaN gate HEMT model establishment method based on a physical guided neural network (PGNN) in order to solve the problem of p-GaN gate HEMT compact model development, and the method comprises the following steps: constructing a p-GaN gate HEMT transistor model by using a TCAD simulation tool, generating a model sample, and extracting simulation data; constructing a DerivNet model and a loss function based on a physical guidance neural network; and finally, performing model training by using the simulation data to obtain the p-GaN gate HEMT model based on the physical guidance neural network.

Description

Technical Field

[0001] The present invention relates to a method for compact modeling of a gallium nitride high electron mobility transistor (GaN HEMT), and in particular to a method for establishing a p-GaN gate HEMT model based on a physics-guided neural network (PGNN). Background Art

[0002] GaN HEMT has the advantages of high breakdown voltage, large output current, and good thermal conductivity, and has become a hot spot in the research of power electronic devices. GaN HEMT power switching devices are widely used in fast charging and have entered high-end fields such as server power supplies and automotive electronics. The market needs enhancement mode (E-mode) GaN HEMT power switching devices. Currently, one of the methods to obtain enhancement mode (E-mode) transistors is to use a p-GaN gate, which increases the conduction band on the channel surface in equilibrium and realizes a normally-off device by depleting the two-dimensional electron gas (2DEG). The p-GaN gate HEMT device is the first commercialized E-mode device with a high and robust threshold voltage V th , high breakdown voltage V BV and low dynamic resistance R on . In recent years, the structure of p-GaN gate HEMT has become more complex, the power has increased, the frequency has increased, and new requirements for GaN integrated circuit design have been continuously proposed, which has brought new problems and challenges to the accurate development of device models and parameter extraction. The table-based model does not have the ability to extrapolate and it is difficult to expand the model prediction within the known data range, and it cannot be used for nonlinear circuit simulation and other problems. Empirical curve fitting models such as the Curtice model, the Statz model, and the Angelov-Chalmers model are not suitable for large signal modes. The DynaFET model is an innovative nonlinear model based on a nonlinear network analyzer (NVNA) and an artificial neural network (ANN), but it lacks predictive capabilities and is limited by the test system. The development of compact models for p-GaN gate HEMTs has become a recognized problem in industry and academia. It is necessary to propose new modeling techniques and develop new compact models to fully meet the needs of GaN HEMT device optimization and GaN integrated circuit design. Summary of the invention

[0003] The purpose of the present invention is to provide a modeling method based on a physical guided neural network (PGNN) to address the shortcomings of existing modeling technology for p-GaN gate HEMTs, aiming to solve the problem of difficult compact modeling of p-GaN gate HEMTs, thereby quickly and accurately realizing device simulation optimization and modeling.

[0004] The present invention proposes a physical guided neural network (PGNN) model, which uses data generated by a TCAD simulation tool to train the model and uses the first-order partial derivative of current with respect to voltage extracted from the PyTorch self-derivative process to To enhance the performance of artificial neural networks (ANN), the DerivNet model with physical guided neural networks (PGNN) can obtain accurate DC characteristics of the device. Based on the simple network, a more physically meaningful and accurate network model is constructed. The first-order partial derivatives of key physical parameters in the DerivNet model It is transmitted to the second ANN and learned simultaneously with the input parameters to complete the neural network model suitable for small samples, achieving good prediction accuracy in various requirements such as current and key indicators. Controlling the first-order partial derivative of a given data set during the learning process of the physics-guided neural network (PGNN) model can make full use of the hidden information of the sample set, thereby improving the learning efficiency of the model for the sample set. By considering the first-order partial derivatives in the network, the relationship between the input and output characteristics of the device is quickly captured. At the same time, due to its first-order partial derivative fitting function, it is particularly effective for analog circuit simulation. In addition, we introduce a physics-based loss function to guide the learning of the physics-guided neural network (PGNN) model, and take the physical consistency of the model output as the learning goal, which can achieve better generalization performance when the training data is small and not fully representative.

[0005] A p-GaN gate HEMT modeling method based on a physics-guided neural network comprises the following steps:

[0006] S1. Use TCAD simulation tools to build a p-GaN gate HEMT transistor model and generate several model samples. Each model sample includes the following device parameters: barrier layer AlGaN thickness t AlGaN , Al composition x of barrier layer AlGaN AlGaN 、pGaN gate layer thickness t pGaN 、Mg doping concentration N in pGaN layer pGaN , gate length L g ;

[0007] Perform IV characteristic simulation on each model sample and extract simulation data, the simulation data includes: input gate voltage V g , output drain voltage V d , drain current I d ;

[0008] S2, performing data preprocessing on the device parameters of each of the model samples in step S1;

[0009] S3, build the DerivNet model based on physical guided neural network;

[0010] Construct the device parameter network para-net and the total network composed of fully connected neural networks respectively;

[0011] The input of the device parameter network para-net is a device parameter combination; the device parameter network para-net also includes a short-circuit connection mechanism, which adds the input of each layer in the device parameter network para-net and the output of the previous layer and inputs them into the next layer of the device parameter network para-net;

[0012] Construct a bias network Bias-net composed of a fully connected neural network; there are four types of bias networks Bias-net, namely bias network 1, bias network 2, bias network 3, and bias network 4; the input of bias network 1 is the input gate voltage V g ; The input of bias network 2 is the input gate voltage V g and g m,rough ; g m,rough is the rough predicted value of drain current I d,rough_T The input gate voltage V g The first-order partial derivative of; the input of bias network 3 is the input gate voltage V g and the output drain voltage V d ; The input of bias network 4 is the input gate voltage V g , output drain voltage V d and conductivity g ds,rough ; g ds,rough is the rough predicted value of drain current I d,rough_O The output drain voltage V d The first partial derivative of

[0013] The DerivNet model includes: a network 1 for predicting device transfer characteristics and a network 2 for predicting device output characteristics;

[0014] Network 1 consists of 1 subnet 1 and 1 subnet 2;

[0015] Subnet 1 includes 1 device parameter network para-net, 1 bias network 1 and 1 total network; the output characteristics of the device parameter network para-net and the output characteristics of the bias network 1 are spliced ​​and input into the total network, and the total network outputs a rough drain current prediction value I of the transfer characteristic d,roughT ;

[0016] Using the drain current I output from subnet 1 d，rough_T V g Find the first-order partial derivative and get the transconductance g m,rough ;

[0017] Subnet 2 includes 1 device parameter network para-net, 1 bias network 2 and 1 total network; the output characteristics of the device parameter network para-net and the output characteristics of the bias network 2 are spliced ​​and input into the total network, and the total network outputs the precise value of the drain current of the transfer characteristic I d,precise_T ;

[0018] Using the drain current I d,precise_T V g Find the first-order partial derivative and get the transconductance g m,pred ;

[0019] Network 2 consists of 1 subnet 3 and 1 subnet 4;

[0020] Subnet 3 includes 1 device parameter network para-net, 1 bias network 3 and 1 total network;

[0021] The output characteristics of the device parameter network para-net and the output characteristics of the bias network 3 are spliced ​​and input into the total network, and the total network outputs a rough prediction value of the drain current of the output characteristic, the drain current I d，rough_O ;

[0022] Using the drain current I of the output of subnet three d，rough_O V d Find the first-order partial derivative and get the conductance g ds,rough ;

[0023] Subnet 4 includes 1 device parameter network para-net, 1 bias network 4 and 1 total network;

[0024] The output characteristics of the device parameter network para-net and the output characteristics of the bias network 4 are spliced ​​and input into the total network, and the total network outputs the precise value of the drain current of the output characteristic I d，precise_O ;

[0025] Use I d,precise_O V d Take the first-order partial derivative to get g ds,pred ;

[0026] S4. Construct the loss function Loss of network 1 DerivNet_T And the loss function Loss of network 2 DerivNet_O :

[0027] Loss DerivNet_T =wMSE(I d,TCAD_T , I d,pred_T )+(1-w)MSE(g m，TCAD , g m,pred )

[0028] Loss DerivNet_O=wMSE(I d，TCAD_O , I d,pred_O )+(1-w)MSE(g ds，TCAD , g ds,pred )

[0029] Among them, I d，TCAD_T is the TCAD simulation value of the drain current based on the reaction transfer characteristics obtained from the model sample in step S1; g m,TCAD is the I obtained based on the model sample in step S1 d,TCAD_T , obtained by taking the first-order partial derivative of the gate voltage of the model sample; I d，TCAD_O is the TCAD simulation value of the drain current of the reaction output characteristic obtained based on the model sample in step S1; g ds,TCAD is the I obtained based on the model sample in step S1 d，TCAD_T The drain voltage V d Obtained by taking the first-order partial derivative; w is the preset weight value;

[0030] S5 model training

[0031] Using the simulation data obtained in steps S1 and S2 and the standardized device parameters, construct the training set and the test set, and use the loss function Loss constructed in step S4 DerivNet_T and Loss DerivNet_O Train network 1 and network 2 separately.

[0032] Preferably, in step S1, the number of the model samples is 810 groups.

[0033] Preferably, in step S2, the data preprocessing specifically includes performing zero-mean normalization processing.

[0034] Preferably, in step S3, tanh is used as the activation function of the fully connected neural network; batch normalization is used in the device parameter network para-net; and group normalization is used in the bias network Bias-net and the total network.

[0035] Preferably, in step S4, w=0.99.

[0036] Beneficial effects of the present invention:

[0037] (1) To address the low time efficiency problem of TCAD simulation in semiconductor process development, the first-order partial derivative of voltage with respect to current is introduced into the network. This physical quantity can make full use of the hidden information of the sample set and improve the learning efficiency of the model for the sample set. The DerivNet model can use the TCAD simulation data of a small sample set for training and accurately predict the electrical characteristics of the device.

[0038] (2) The P-GaN gate HEMT modeling method based on physical guided neural network (PGNN) can be used to replace TCAD simulation tools to a certain extent for calculation, bringing great value to practical applications. This modeling technology only takes more than ten minutes to optimize the device, significantly shortening the time spent on process parameter optimization design, and more conveniently and quickly finding a combination of transistor structure parameters and process parameters that meets actual requirements;

[0039] (3) The equations of the p-GaN gate HEMT based on the physically guided neural network (PGNN) model can describe a large class of similar devices, satisfying the optimal design of GaN HEMT devices and power GaN integrated circuits. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Cross-sectional view of a p-GaN gate HEMT device, where AlGaN barrier represents the aluminum gallium nitride barrier layer; GaNchannel represents the gallium nitride channel layer; and AlGaN back barrier represents the aluminum gallium nitride back barrier layer.

[0041] Figure 2 TCAD simulation diagram of optimized design of p-GaN gate HEMT device with size and doping concentration variation (L gd =6um), wherein AlGaN represents aluminum gallium nitride; GaN represents gallium nitride; Si3N4 represents silicon nitride; Conductor represents a conductor.

[0042] Figure 3 Schematic diagram of Network 1 in the DerivNet model, which is used to predict device transfer characteristics.

[0043] Figure 4 Schematic diagram of Network 2 in the DerivNet model, which is used to predict device output characteristics.

[0044] Figure 5 Prediction curves of the DerivNet model for p-GaN-gate HEMT devices, where (a) is the transfer characteristic and (b) is the output characteristic.

[0045] Figure 6 Drain current I of the DerivNet model for p-GaN gate HEMT devices d The coefficient of determination R 2 Figure. Among them, (a) is the transfer characteristic (b) is the output characteristic, and points of different colors represent the data of different samples;

[0046] Figure 7 Threshold voltage V of the DerivNet model for p-GaN gate HEMT devicesth Trend curve. (a) is L g , (b) is t pGaN , (c) is N pGaN (d) is t AlGaN (e) is x AlGaN .

[0047] Figure 8 Voltage extrapolation curves of the DerivNet model for p-GaN gate HEMT devices. (a) is the transfer characteristic and (b) is the output characteristic.

[0048] Fig. 9 Parameters of the DerivNet model for p-GaN gate HEMT devices Gate length L g Extrapolate the curve.

[0049] Fig.10 Gummel test results of the DerivNet model for p-GaN-gate HEMT devices. DETAILED DESCRIPTION

[0050] The specific implementation is described below with reference to the accompanying drawings.

[0051] A neural network model suitable for small samples is designed. The specific implementation of establishing a compact DerivNet model of p-GaN gate HEMT devices based on physical information neural network is as follows:

[0052] S1. Use TCAD simulation tools to construct a p-GaN gate HEMT transistor such as Figure 1 , generating 810 model samples, each with 5 parameters: barrier layer AlGaN thickness t AlGaN , Al composition x of barrier layer AlGaN AlGaN 、pGaN gate layer thickness t pGaN 、Mg doping concentration N in pGaN layer pGaN , gate length L g Specific TCAD simulation structure Figure 2 And the simulation settings are shown in Table 1.

[0053] Table 1 TCAD parameter change setting table for p-GaN gate HEMT devices.

[0054]

[0055] In terms of network construction, since the amount of data given is small and the parameter dimension is not large, the design of the network structure should not be too complicated, but should be refined as much as possible to fully explore the sample information. The transfer characteristics and output characteristics of the device are predicted by constructing the DerivNet model based on the physical guided neural network (PGNN). For the physical guided neural network (PGNN) for predicting the transfer characteristics, the gate voltage V g and transistor parameters are shown in Table 1 as the input of the neural network, the drain current I d As the output of the neural network; for the physically guided neural network (PGNN) that predicts the output characteristics, the gate voltage V g , output drain voltage V d and transistor parameters as shown in Table 1 as network inputs, drain current I d as the output of the neural network.

[0056] S2. The data set given in step S1 includes various transistor parameter combinations and TCAD simulation data in.log files and out.log files corresponding to different bias condition combinations. Among transistor parameters, the order of magnitude of different parameters varies greatly. In order to facilitate the neural network to learn the influence of transistor parameters on transfer characteristics and output characteristics, the transistor parameters of each sample are normalized to zero mean, so that the data distribution of each type of transistor parameters is smoother, and at the same time, the order of magnitude of different types of parameters will not vary too much. The standardized calculation formula is as follows:

[0057]

[0058] X train ,X test Respectively represent the original values ​​of the parameters, Represent the standardized parameters, μ train , σ train are the mean and standard deviation of a certain type of parameter in the training set, respectively.

[0059] S3, based on the modeling idea of ​​physical guided neural network (PGNN), in simple neural network structures such as Figure 3 and 4 Based on the left and right networks, continuous and differentiable first-order partial derivatives are used: transconductance g m and conductivity g ds To represent the target variable drain current I d Bias voltage V d and V g The physical relationship between them is used to construct an accurate DerivNet model based on the physical guided neural network (PGNN). m and conductivity g ds Improved the accuracy of the model's current prediction.m Not only measures the input gate voltage V g The key indicator of current gain and threshold voltage V th It is calculated by transconductance. Conductance g ds It is a key parameter that describes the output characteristics of the device and affects the on-resistance R under forward bias conditions. on size.

[0060] S3-1, simple neural network structure such as Figure 3 Transfer characteristics and Figure 4 The output characteristics of the left network structure. In the simple neural network structure, num_in is the number of parameter types, and num_out is the number of output parameter types. The network uses a fully connected linear network. Figure 3 and Figure 4 The sample device parameters are different, but the bias voltage V g With V d The input form of is the same. The device parameters can be regarded as the intrinsic parameters of the sample, and the applied voltage can be regarded as the external parameters of the sample. The input of the device parameter network para-net is the device parameter combination, and the bias network Bias-net is used to input the bias combination. g With V d Feature extraction is performed through the neural network para-net and bias-net. The outputs of these two networks are spliced ​​together as the input of the total network, and finally the device drain current I is obtained by the simple neural network structure of the output and transfer characteristics. d The predicted value of . The activation function uses tanh, batch normalization is used in the device parameter network para-net, and group normalization is used in the network Bias-net and the total network. Compared with batch normalization, the network predicted volt-ampere characteristic curve is smoother after the network Bias-net and the total network use group normalization.

[0061] Drawing on the idea of ​​residual network, short-circuit connection is introduced to add the input (i.e. parameter data) and output of the network para-net, so that the neural network can well retain the parameter characteristics in the deep structure, and also reduce the occurrence of gradient disappearance. According to the experiment, adding short-circuit connection significantly improves the learning ability of the network.

[0062] S3-2. Physically guided neural network architecture for predicting device transfer characteristics Figure 3 As shown. It is composed of two simple neural network structures of the same structure, subnet 1 and subnet 2, which correspond to Figure 3The left and right networks in the figure, the input of the left network is the preprocessed gate voltage V g and device parameter data, the output is a rough prediction of the drain current I d，rough_T , reuse I d，rough_T The gate voltage V g Find the first-order partial derivative to get the transconductance Combine it with the gate voltage V g Spliced ​​as the input of bias network 2 in the right network, the exact value of drain current I of the transfer characteristic is predicted d,precise_T The output of the left network is I d,rough_T The purpose is to roughly calibrate the transconductance g m , without changing the transconductance g m It is transmitted between networks as a hidden variable.

[0063] S3-3. Predicting device output characteristics based on a physically guided neural network structure such as Figure 4 The output characteristics are based on the physical guidance of the neural network Figure 4 Physics-guided neural network architecture with similar transfer properties Figure 3 The difference is that the input of the output characteristic subnet three (the left network) is the pre-processed bias voltage V g 、V d and device parameter data and output as I d，rough_O , based on I d,rough_O Get the first-order partial derivative The output of subnet 4 (right network) of the output characteristic network is the exact value of the drain current I of the output characteristic. d,precise_O .

[0064] S4. According to step S1-2, the TCAD simulated DC characteristic curve of the transistor sample is obtained. It is observed that the variation ranges of the current and the first-order partial derivative of the current with respect to the voltage between the samples are quite different. Therefore, the following loss function calculation formula is used:

[0065]

[0066] Where i indicates the sample number, j indicates the current number within a sample, and y i,j and They are TCAD simulation values ​​I d,TCAD and the predicted value I d,precise ,max(y i ) is the maximum value of the data in the sample labeled i.

[0067] In the loss function, max(y i) is used as the denominator to normalize the predicted values ​​and TCAD simulation values, so that the predicted values ​​and TCAD simulation values ​​of different samples are at the same order of magnitude when calculating the loss function. The loss function designed in this way can better measure the prediction effect of the network and can keenly capture the possible abnormal values ​​in the data. In addition, considering the back propagation and parameter update process, the loss function after the normalization of the current value and the first-order partial derivative of the current to the voltage can also effectively increase the importance of transistor samples with smaller current and the first-order partial derivative of the current to the voltage in the network training process.

[0068] For the volt-ampere characteristic, considering that the network has multiple inputs and outputs, the loss function is not only the mean square error MSE (I d,TCAD_T , I d,pred_T ) and MSE(I d，TCAD_O , I d,pred_O ), add the first-order partial derivative g to predict the device transfer characteristics m MSE(g m,TCAD , g m,pred ), add the first-order partial derivative g to predict the device output characteristics ds MSE(g ds,TCAD , g ds,pred ), thereby constructing the physics-based loss function formulas (4) and (5) to solve the problem of physical consistency in learning of the DerivNet model.

[0069] In the DerivNet model, the loss function for predicting device transfer characteristics is

[0070] Loss DerivNet_T =0.99MSE(I d,TCAD_T , I d,pred_T )+0.01MSE(g m,TCAD , g m，pred ) (4)

[0071] In the DerivNet model, the loss function for predicting device output characteristics is

[0072] Loss DerivNet_O =0.99MSE(I d，TCAD_O , I d,，pred_O )+0.01MSE(g ds,TCAD , g ds,pred ) (5)

[0073] Since known physical laws are assumed to apply equally to any unknown data instance, the physical consistency of the first-order partial derivative of device current to voltage is used as a learning target in the DerivNet model, which can achieve better generalization performance even if the training data is small and not fully representative. In addition, the DerivNet model takes the first-order partial derivative of device current to voltage into account in the workflow of device design, which can achieve faster optimization of devices.

[0074] Step S5: Model training and screening:

[0075] S5-1. Divide the data into training set and test set, with a sample ratio of 8:2. Select Adam optimizer for network parameter optimization, set the learning rate to 0.001, and set the batch size to 128. After testing, when the number of training steps is 5000, the network with the best training effect can be obtained.

[0076] S5-2, DerivNet model verification, gives the fitting accuracy of the DerivNet model, the test of the physical characteristics of the device, the test of the extrapolation ability, the Gummel test results, and the device evaluation framework results.

[0077] S5-2-1. Table 2 shows the current prediction effect of the DerivNet model on device samples.

[0078] Table 2 Statistics of the prediction effect of the DerivNet model for p-GaN gate HEMT devices.

[0079]

[0080] As shown in Table 2 above, the mean square error of the DerivNet model for the data set is 0.24988, the maximum error is 0.8615, and the minimum error is 0.03957. The key indicators of the DerivNet model, especially the threshold voltage V th The prediction error is 0.062, and the DerivNet model achieves accurate prediction results.

[0081] In the test set, the prediction effect of the DerivNet model on the device is as follows: Figure 5 (a) and (b). The dots represent TCAD simulation values, and the solid lines represent DerivNet model prediction values. Due to the large range of drain current in the transfer characteristics, Figure 5 A logarithmic scale is used in (a), from Figure 5 It can be seen that the DerivNet model achieves accurate fitting effects in terms of both the transfer characteristics and the output characteristics of the device.

[0082] In further tests, we randomly shuffled the order of the sample sets and then redivided the training set and validation set (8:2). The prediction effects of the DerivNet model on current and various key indicators were very similar, which verified the robustness of the model to the sample space.

[0083] Figure 6 is the drain current I of the DerivNet model for the p-GaN gate HEMT device d The coefficient of determination R 2 Figure 1 shows the transfer characteristics (a) and the output characteristics (b). The TCAD simulation value and the predicted value are used as the x and y coordinates of each point, respectively. Most of the points are along the diagonal line, showing the drain current I of the device transfer characteristics and output characteristics predicted by the DerivNet model. d A good prediction effect was achieved.

[0084] S5-2-2. Further verify whether the DerivNet model satisfies the drain current I d directionality and has the correct trend curve.

[0085] Drain current I d The directionality of the drain current I d Must be greater than or equal to 0. Since the network predicts the logarithmic value of the current, the DerivNet model can satisfy the drain current I d directionality.

[0086] The correct trend curve requires that when other parameters are fixed, the threshold voltage V predicted by the DerivNet model th With t pGaN 、N pGaN , L g ,t AlGaN 、x AlGaN Monotonically changing. A test sample (t AlGaN :0.015umx AlGaN :0.23t pGaN :0.11umN pGaN :3e17cm -3 L g :1.4um) and Figure 7 The results are shown in (a) to (e). Figure 7 In, t pGaN and N pGaN The value of varies from the minimum value given in the sample to the maximum value. It can be seen that when other parameters are fixed, the threshold voltage V th With t pGaN 、N pGaN and L g Monotonically increasing with tAlGaN and x AlGaN Monotonically decreasing. Multiple device samples were randomly selected for testing, and the DerivNet model obtained the correct trend curve.

[0087] S5-2-3. During testing, the parameter range of the data may slightly overflow. By checking the DerivNet model for V g ,V d The extrapolation capability of the device parameters is used to verify the trend of the curve. First, the DerivNet model is tested for the bias voltage V g and V d extrapolation capability. Figure 8 In (a) and (b), the right side of the vertical line is the expanded voltage and predicted drain current I d It can be seen that the DerivNet model still presents a reasonable trend curve for the extrapolated bias.

[0088] Secondly, the extrapolation ability of the DerivNet model for device parameters is tested. g As an example, check the DerivNet model for the extrapolated L g The prediction effect is shown in Fig. 9 . Select a set of sample parameters in the test set and set L g The value is set to 1.9um, and the largest L g The value after overflowing 15% is 2.2um, which verifies the model for the extrapolated parameter L g Whether the trend curve is still maintained, the test results are as follows Fig. 9 . Fig. 9 In the figure, the dotted line represents the predicted value corresponding to the original parameters, and the solid line represents the predicted value after parameter extrapolation. g After extrapolation, the curve predicted by the model can still maintain a reasonable trend. g After increasing, the drain current I d The value becomes smaller, which is consistent with the physical characteristics of the device.

[0089] The drain current I predicted by the DerivNet model d The values ​​were subjected to a Gummel test and the first derivative of the curve was calculated. Fig.10 The prediction results of the DerivNet model are shown, with good fitting results.

[0090] S5-2-4. According to the evaluation framework, the test results of the DerivNet model are shown in Table 3.

[0091] Table 3 Results of the evaluation framework of the DerivNet model for p-GaN-gate HEMT devices.

[0092]

[0093] In Table 3, the threshold voltage V th The mean absolute error is calculated, and the on-resistance R on and the maximum drain current I dmax The average relative error is calculated. According to the evaluation results in Table 3, the DerivNet model has achieved good results overall. It should be noted that due to the resource usage of the server, the training time and prediction time listed in the table are not accurate and are not the shortest time.

[0094] To address the low timeliness of TCAD simulation in semiconductor process development, this paper constructs a physics-driven neural network DerivNet model from process parameters to device electrical characteristics. It can use limited TCAD simulation data for training, accurately predict smooth current-voltage curves, and accurately extract the key indicator threshold voltage V of the device. th , on-resistance R on and the maximum drain current I dmax By randomly dividing the test set and training set, appropriately spilling over the input voltage and device parameters, and testing different device types, the robustness of the model to the sample space and device type is verified, and the correctness of the DerivNet model based on the physics-guided neural network in characterizing the physical properties of the device is also verified.

Claims

1. A p-GaN gate HEMT modeling method based on a physics-guided neural network, characterized in that: The following steps are involved: S1 uses TCAD simulation tools to build a p-GaN gate HEMT transistor model and generate several model samples. Each model sample includes the following device parameters: barrier layer AlGaN thickness t AlGaN , Al composition x of barrier layer AlGaN AlGaN 、pGaN gate layer thickness t pGaN 、Mg doping concentration N in pGaN layer pGaN , gate length L g ; Perform IV characteristic simulation on each model sample and extract simulation data, the simulation data includes: input gate voltage V g , output drain voltage V d , drain current I d ; S2, performing data preprocessing on the device parameters of each of the model samples in step S1; S3, build the DerivNet model based on physical guided neural network; Construct the device parameter network para-net and the total network composed of fully connected neural networks respectively; The input of the device parameter network para-net is a device parameter combination; the device parameter network para-net also includes a short-circuit connection mechanism, which adds the input of each layer in the device parameter network para-net and the output of the previous layer and inputs them into the next layer of the device parameter network para-net; A bias network Bias-net composed of a fully connected neural network is constructed; there are four types of bias networks Bias-bet, namely bias network 1, bias network 2, bias network 3, and bias network 4; the input of bias network 1 is the input gate voltage V g ; The input of bias network 2 is the input gate voltage V g and g m,rough ; g m,rough is the rough predicted value of drain current I d,rough_T The input gate voltage V g The first-order partial derivative of; the input of bias network 3 is the input gate voltage V g and the output drain voltage V d ; The input of bias network 4 is the input gate voltage V g , output drain voltage V d and g ds,rough ; g ds,rough is the rough predicted value of drain current I d,rough_O The output drain voltage V d The first partial derivative of The DerivNet model includes: a network 1 for predicting device transfer characteristics and a network 2 for predicting device output characteristics; Network 1 consists of 1 subnet 1 and 1 subnet 2; Subnet 1 includes 1 device parameter network para-net, 1 bias network 1 and 1 total network; the output characteristics of the device parameter network para-net and the output characteristics of the bias network 1 are spliced ​​and input into the total network, and the total network outputs a rough drain current prediction value I of the transfer characteristic d,rough_T ; Using the I output of subnet 1 d,rough_T V g Take the first-order partial derivative to get g m,rough ; Subnet 2 includes 1 device parameter network para-net, 1 bias network 2 and 1 total network; the output characteristics of the device parameter network para-net and the output characteristics of the bias network 2 are spliced ​​and input into the total network, and the total network outputs the precise value of the drain current of the transfer characteristic I d,precise_T ; Use I d,precise_T V g Take the first-order partial derivative to get g m,pred ; Network 2 consists of 1 subnet 3 and 1 subnet 4; Subnet 3 includes 1 device parameter network para-net, 1 bias network 3 and 1 total network; The output characteristics of the device parameter network para-net and the output characteristics of the bias network 3 are spliced ​​and input into the total network, and the total network outputs a rough predicted value of the drain current of the output characteristic I d,rough_O ; Using the output of subnet 3 d,rough_O V d Take the first-order partial derivative to get g ds,rough ; Subnet 4 includes 1 device parameter network para-net, 1 bias network 4 and 1 total network; The output characteristics of the device parameter network para-net and the output characteristics of the bias network 4 are spliced ​​and input into the total network, and the total network outputs the precise value of the drain current of the output characteristic I d,precise_O ; Use I d,precise_O V d Take the first-order partial derivative to get g ds,pred ; S4 builds the loss function Loss of network 1 DerivNet_T And the loss function Loss of network 2 DerivNet_O : Loss DerivNet_T =wMSE(I d,TCAD_T ,I d,pred_T )+(1-w)MSE(g m,TCAD ,g m,pred ) Loss DerivNet_O =wMSE(I d,TCAD_O ,I d,pred_O )+(1-w)MSE(g ds,TCAD ,g ds,pred ) Among them, I d,TCAD_T is the TCAD simulation value of the drain current based on the reaction transfer characteristics obtained from the model sample in step S1; g m,TCAD is the I obtained based on the model sample in step S1 d,TCAD_T , obtained by taking the first-order partial derivative of the gate voltage of the model sample; I d,TCAD_O is the TCAD simulation value of the drain current based on the reaction output characteristics obtained from the model sample in step S1; g ds,TCAD is the I obtained based on the model sample in step S1 d,TCAD_T The drain voltage V d Obtained by taking the first-order partial derivative; w is the preset weight value; S5 model training Using the simulation data obtained in steps S1 and S2 and the standardized device parameters, construct the training set and the test set, and use the loss function Loss constructed in step S4 DerivNet_T and Loss DerivNet_O Train network 1 and network 2 separately.

2. A p-GaN gate HEMT modeling method based on a physical guided neural network as claimed in claim 1, characterized in that: In step S1, the number of the model samples is 810 groups.

3. A p-GaN gate HEMT modeling method based on a physical guided neural network as claimed in claim 1, characterized in that: In step S2, the data preprocessing specifically includes performing zero-mean normalization processing.

4. A p-GaN gate HEMT modeling method based on a physical guided neural network as claimed in claim 1, characterized in that: In step S3, The activation function of the fully connected neural network uses tanh; Using batch normalization in the device parameter network para-net; Group Normalization is used in the bias network Bias-net and the overall network.

5. A p-GaN gate HEMT modeling method based on a physical guided neural network as claimed in claim 1, characterized in that: In step S4, w=0.99.

Images