GaN CAVET modeling method based on physical guidance neural network

Abstract

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

Description

TECHNICAL FIELD

[0001] The application relates to the field of GaN CAVET modeling, in particular to a GaN CAVET modeling method based on a physically guided neural network. BACKGROUND

[0002] A vertical GaN current aperture transistor (CAVET) is an important direction of high-power electronic devices due to the advantages of combining lateral and vertical structures. As an important vertical structure in GaN devices, the CAVET is widely concerned due to its advantages such as vertical current conduction, compact structure, low on-resistance and the like, and a structural diagram is shown in Figure 1 Compared with a traditional lateral HEMT, the CAVET is more suitable for high-power density applications, and its vertical structure is more conducive to heat dissipation and chip integration, and at the same time, has higher breakdown voltage and current density. The CAVET utilizes the high conductivity brought by the two-dimensional electron gas (2DEG) in the AlGaN / GaN heterojunction, and at the same time realizes high breakdown voltage through a longitudinal current path, and has high integration and excellent electrical performance, and has great application prospect. However, the electrical performance of the CAVET device is influenced by factors such as material parameters, device structure and process, and the traditional modeling method depends on numerical simulation or physical modeling formula. This method can still obtain relatively accurate static characteristics, but has problems such as large calculation overhead, complex parameter debugging, weak generalization ability and the like, and even due to the grid division problem, the phenomenon of archer paradox is faced. The Angelov experience model is a mathematical formula (such as a polynomial, an exponential function) based on experimental data fitting, adjusts parameters to match the I-V characteristics of the CAVET, and cannot reflect the vertical electric field distribution, the current aperture effect and other key mechanisms of the CAVET. The TCAD numerical simulation cannot reflect the complex behaviors in the physical mechanism, such as the current disturbance caused by the capture of electrons by deep energy state traps, and in an extreme case, an error is generated, and the three-dimensional simulation of the TCAD needs a large number of grid divisions, and a single simulation takes a long time. The traditional drift-diffusion analytical model, the Shockley model and the Curtice model are oversimplified, and cannot simulate some dynamic behaviors. SUMMARY

[0003] The application aims at training a better model with a smaller data set under the consideration of the actual physical background in the initial design or structure parameter optimization process of the CAVET device, and the application provides a GaN CAVET modeling method based on a physically guided neural network.

[0004] The application provides a GaN CAVET modeling method based on a physically guided neural network, and the method comprises the following steps:

[0005] S1, a CAVET model is constructed by using a TCAD simulation tool, and a model sample is generated, each model sample including the following device parameters: aperture layer length L ap , gate overlap length L go , unintentional doping layer thickness t GaN1 , current blocking layer thickness t GaN2 , current blocking layer doping concentration N GaN2 , drift layer thickness t GaN3 , and drift and aperture layer doping concentration N GaN3 ;

[0006] Ionsimulation data is extracted, including: output drain voltage V ds , drain current I ds,TCAD ; based on the output drain voltage V ds , drain current I ds,TCAD , an output characteristic curve is obtained;

[0007] S2, the device parameters of each of the model samples in step S1 are standardized;

[0008] S3, a physical guide neural network (PGNN) model is constructed, including a shallow neural network and a hypernetwork (Hypernet);

[0009] The shallow neural network is constructed, and the shallow neural network is used to realize the operation process recorded in the following formula:

[0010] I ds,pred = a1tanh(a2V ds +a3)+a4tanh(a5V ds +a6)+a7tanh(a8V ds +a9)+a 10

[0011] Wherein, a1-a 10 represent 10 parameters of the shallow neural network;

[0012] Based on the multilayer perception machine MLP, the hypernetwork (Hypernet) is constructed, the network takes the device parameters as input, outputs a1-a 10 10 parameters, and then the shallow neural network takes the output drain voltage V ds as input and outputs the predicted drain current value I dx,pred .

[0013] Preferably, in S3, the Hypernet includes three Multi-Level Processing (MLPs). The device parameters first enter the first MLP. The output vector of the first MLP is divided into the first four dimensions and the remaining part. The remaining part is linearly transformed and then cross-multiplied with the first four dimensions before being input into the second MLP. The output vector of the second MLP is then divided into the first four dimensions of the second MLP and the remaining part of the second MLP. The remaining part of the second MLP is linearly transformed and then cross-divided with the first four dimensions of the second MLP before being input into the third MLP. The third MLP finally generates the output.

[0014] Preferably, step S3 further includes the following step:

[0015] The shallow neural network is trained using the output characteristic curve obtained from S1: the output drain voltage V corresponding to each point in the output characteristic curve is used. ds Drain current I ds A dataset was constructed to train a shallow neural network, yielding 10 parameters for the shallow neural network; for each set of device parameters, a set of shallow neural network parameters a1 to a1 were trained. 10 With 10 parameters, we obtain a dataset of 10 parameters for a shallow neural network;

[0016] Using a dataset of 10 parameters of a shallow neural network and its corresponding device parameters, a hypernet is trained to obtain a pre-trained network.

[0017] Preferably, the following steps are included after S3:

[0018] S4 predicts the leakage current value I based on the output of the Physically Guided Neural Network (PGNN) model. ds,pred and drain current I ds,TCAD The difference, and transconductance g m,TCAD and the transconductance g obtained based on the predicted leakage current value m,pred The difference is used to construct a loss function, and the Physically Guided Neural Network (PGNN) model is trained; whereby,

[0019] Preferably, in S4, the formula for the loss function is as follows:

[0020]

[0021] Where k is a preset weight value, and n is the number of points on the output characteristic curve under specific device parameters.

[0022] Preferably, the GaN CAVET device structure, from top to bottom, includes a gate metal, a lower gate n-type gate, and a lower gate n-type gate. +GaN contact layer, AlGaN barrier layer, intrinsic GaN channel layer, unintentionally doped (UID) GaN layer, current blocking layer (CBL), aperture layer, and drift layer with n + GaN drain contact layer.

[0023] Advantages of the present application:

[0024] (1) The present study proposes a physical guide neural network (PGNN) model that combines shallow neural networks, hypernetworks (Hypernet), and residual connections to achieve high-precision and robust device modeling for CAVET DC characteristic prediction. The introduction of transconductance in the loss function assists model training, not only improving the model's understanding of device physical characteristics, but also to some extent suppressing overfitting, improving the model's generalization ability and prediction accuracy. The physical guide neural network (PGNN) model can be trained using small sample sets of TCAD simulation data to accurately predict device electrical characteristics.

[0025] (2) The CAVET modeling method based on the physical guide neural network (PGNN) can to some extent replace the TCAD simulation tool for calculation, bringing great value to practical applications. This network combines the advantages of the physical guide neural network (PGNN) and the hypernetwork (Hypernet), achieving interpretability and scalability of the physical guide neural network (PGNN) without overfitting, while the combination with the hypernetwork (Hypernet) has strong fitting ability, good fitting effect, strong robustness, and is not prone to overfitting, ensuring the basic shape of the network output characteristics, effectively suppressing the abnormal fluctuations caused by excessive grid in TCAD simulation, and generating natural value range constraints.

[0026] (3) In the CAVET modeling method of the physical guide neural network (PGNN), the hypernetwork (Hypernet) learns a function to dynamically generate weight parameters, rather than directly learning the weight parameters themselves, which can significantly improve the model's generalization ability and the applicability of CAVET characteristic prediction. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a schematic diagram of a CAVET device structure, wherein the structure includes, from top to bottom, a gate metal, a gate-under n + GaN contact layer, AlGaN barrier layer, intrinsic GaN channel layer, unintentionally doped (UID) GaN layer, current blocking layer (CBL), aperture layer, and drift layer with n + GaN drain contact layer.

[0028] Figure 2 TCAD simulation plot for optimization design of CAVET device with varying dimensions and doping concentration, where AlGaN represents aluminum gallium nitride; GaN represents gallium nitride; Si3N4 represents silicon nitride; Conductor represents conductor; Electrodes represents electrodes.

[0029] Figure 3 Schematic diagram of a shallow neural network.

[0030] Figure 4 Schematic diagram of a hypernetwork.

[0031] Figure 5 Fluctuation caused by TCAD simulation error.

[0032] Figure 6 Error between physical-guided neural network (PGNN) model prediction and TCAD simulation when current is too small.

[0033] Figure 7 Correlation scatter plot.

[0034] Figure 8 Output characteristic curve of physical-guided neural network (PGNN) model prediction and TCAD simulation. DETAILED DESCRIPTION

[0035] The specific embodiments are described below with reference to the accompanying drawings.

[0036] The present application designs a neural network model suitable for small samples, and establishes a compact model of CAVET device based on physical-guided neural network:

[0037] A GaN CAVET modeling method based on physical-guided neural network (PGNN) comprises the following steps:

[0038] S1, using a TCAD simulation tool to construct a CAVET transistor such as Figure 2 , to generate model samples, each model sample comprising the following device parameters: aperture layer length L ap , gate overlap length L go , unintentional doping layer thickness t GaN1 , current blocking layer thickness t GaN2 , current blocking layer doping concentration N GaN2 , drift layer thickness t GaN3 , and drift and aperture layer doping concentration N GaN3 ;

[0039] Performing volt-ampere characteristic simulation on each model sample, extracting simulation data, the simulation data comprising: output drain voltage Vds , drain current I ds,TCAD ; based on output drain voltage V ds , drain current I ds,TCAD get output characteristic curve; simulation settings are shown in Table 1.

[0040] Table 1. TCAD parameter variation setting table of CAVET device.

[0041] Parameter Range Step 1 L ap [μm]]]> (1,4,10,15) No 2 L go [μm]]]> (1,3,5) 2 3 t GaN1 [μm]]]> (0.1,0.15,0.2) 0.05 4 t GaN2 [μm]]]> (0.1,0.3,0.5) 0.2 5 t GaN3 [μm]]]> (1,3,5) 2 6 N GaN2 [cm -3 ]]]> (8e16, 2e17, 8e17, 2e18) No 7 N GaN3 [cm-3]]]> (2e15, 2e16, 2e17) No 8 V ds [V]]]> (0,0.5,1,1.5,2,…,40) 0.5

[0042] In terms of network construction, due to the small amount of given data and the small dimension of parameters, the design of network structure should not be too complex, and the sample information should be fully mined. A physical guidance neural network (PGNN) model is constructed to predict the transfer characteristics and output characteristics of the device. For the physical guidance neural network (PGNN) based on the prediction of the volt-ampere characteristic, the gate voltage V gs , output drain voltage V ds and transistor parameters in Table 1 are used as network inputs, and the drain current I ds is used as the output of the neural network.

[0043] S2, the data set given in step S1 includes various combinations of transistor parameters and TCAD simulation data in.log file and out.log file corresponding to different bias voltage condition combinations. Among the transistor parameters, the orders of magnitude of different parameters differ greatly. In order to facilitate the neural network to learn the influence of transistor parameters on the transfer characteristics and output characteristics, the transistor parameters of each sample are processed by 0 mean value standardization, so that the data distribution of each type of transistor parameter is more smooth, and the order of magnitude difference of different types of parameters is not too large. The standardization calculation formula is as follows:

[0044]

[0045] X train ,X test represent the original value of the parameter, respectively represent the standardized parameters, μ train , σ train are the mean and standard deviation of a certain type of parameter in the training set.

[0046] S3, a PGNN model based on physical guidance neural network is constructed, including a shallow neural network and a hypernetwork (Hypernet);

[0047] A shallow neural network is constructed, and the Curtice model drain current empirical formula derived according to the carrier saturation theory is as follows:

[0048] I ds = I ds0 (1+λV dstanh(aV ds ) (3)

[0049] Considering that λ in (1 + λV ds ) is small, the change of tanh(aV ds ) is much larger than (1 + λV ds ) when V ds is small, and the change of (1 + λV ds ) is much larger than tanh(aV ds ) when V ds is large. In addition to the change of tanh(aV ds ) and some actual effects of the model, such as the transfer characteristic metastability and the like, the shallow neural network adopts a linear combination formula (4) of multiple tanh activation functions to realize the nonlinear mapping of the input voltage to the output current. The tanh function has good nonlinear fitting capability and can more effectively capture the smooth transition of the device between the on and off states to improve the expression capability of the model to the key physical characteristics of the device.

[0050] The shallow neural network formula (4) is as follows:

[0051] I ds,pred = a1tanh(a2V ds + a3) + a4tanh(a5V ds + a6) + a7tanh(a8V ds + a9) + a 10 (4)

[0052] Wherein, a1-a 10 represent 10 parameters of the shallow neural network, and will be learned and adjusted by training data using a hypernet.

[0053] The output characteristic curve obtained by using S1 is used to construct a data set with the output drain voltage V ds and the drain current I ds corresponding to each point in the output characteristic curve to train the shallow neural network and obtain 10 parameters of the shallow neural network; each set of device parameters is trained to obtain a set of 10 parameters of the shallow neural network, thereby obtaining a data set of 10 parameters of the shallow neural network;

[0054] The hypernet is used to dynamically generate the weights a1-a 10A deep neural network with a wide function expression capability is used to represent parameters, and each parameter has a similar trend under the same other parameters. Considering the proportional relationship between the voltage and the fitting parameters, a cross-multiplication structure is added to obtain an extremely effective structure. This network has strong adaptability to this task and has obtained relatively ideal results. And through the dropout layer, the robustness of the model is increased to prevent overfitting, such as Figure 4 .

[0055] The hypernetwork (Hypernet) includes three MLPs, the input of the network is the device parameters, and the output is a1-a 10 ; The specific structure is that the device parameters first enter the first MLP, the output vector of the first MLP is divided into the first 4 dimensions and the remaining part, the remaining part is linearly transformed, and then multiplied with the first 4 dimensions and input into the second MLP; The output vector of the second MLP is divided into the first 4 dimensions and the remaining part, the remaining part is linearly transformed, and then divided by the first 4 dimensions and input into the third MLP to finally generate the output.

[0056] The hypernetwork (Hypernet) is trained using a 10-parameter dataset of a shallow neural network and its corresponding device parameters to obtain a pre-trained network; After pre-training, the hypernetwork (Hypernet) is combined with the shallow neural network and compared with the data obtained from TCAD simulation to find that the effect reaches the expectation.

[0057] The key structure parameters of the semiconductor device include the aperture layer length L ap , the gate overlap length L go , the unintentional doping layer thickness t GaN1 , the current blocking layer thickness t GaN2 , the current blocking layer doping concentration N GaN2 , the drift layer thickness t GaN3 , and the drift / aperture layer doping concentration N GaN3 are input into the hypernetwork (Hypernet), and the weight parameters of the physically guided neural network (PGNN) are dynamically generated, and then the voltage-current physical constraint mapping relationship is constructed. In this architecture, the physically guided neural network (PGNN) is based on the parameterized weight output by the hypernetwork (Hypernet), combines the shape features of the real-time leakage current, and uses the powerful fitting capability of the neural network to realize the neural network prediction model of the device voltage-current characteristic.

[0058] S4 constructs a loss function and trains

[0059] The loss function Loss of the physically guided neural network (PGNN) model is constructed,

[0060] In order to reduce the numerical error caused by grid division in the TCAD simulation process, an error processing mechanism is adopted: the numerical value with error magnitude less than 10 -8 is regarded as zero, so as to effectively avoid the interference of the extremely small numerical error on the model training result, and a loss function

[0061] I ds,pred And g m,pred Is based on the output of the physical guided neural network model; I ds,TCAD And g m,TCAD Is the simulation data of TCAD, wherein gm is the transconductance, The physical guided neural network model is trained to obtain the trained physical guided PGNN model, k is a preset weight value, and n is the number of points on the output characteristic curve under a specific device parameter;

[0062] The shallow neural network is combined with the hypernetwork (Hypernet) for training, the device parameters of the device are input into the hypernetwork (Hypernet), the hypernetwork (Hypernet) outputs the network parameters of the shallow neural network, and then the shallow neural network takes the voltage parameter as input and outputs the predicted leakage current value.

[0063] In the hypernetwork (Hypernet) training stage, a simple and effective loss function is designed, that is, the sum of squares of a1-a 10 Of the PGNN network, the function itself is relatively complex, but the gradient calculation process is relatively simple. At the same time, only a single network needs to be optimized, so the gradient descent method can be directly used for solving, that is, an initial solution with good convergence can be obtained.

[0064] The above is only part of the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

[0065] Training and screening of S5 model:

[0066] S5-1, divide the data into training set and test set, and the sample ratio is 8:2. The parameter optimization of the network selects Adam optimizer, the learning rate is set to 0.001, and the batch size is set to 128. After testing, when the training step is 5000 steps, the network with the best training effect can be obtained.

[0067] S5-2, verification of the physical guided neural network (PGNN) model, and the anti-interference and fitting accuracy of the model are given.

[0068] Because of the simplicity of the physical guided neural network (PGNN) structure, there is no case of large fluctuations, and the unstable fluctuations in TCAD simulation can be automatically ignored. Further, it ensures that the model does not overfit, effectively preventing abnormal interference that may interfere with the model, such as Figure 5 :

[0069] For the case of no effective information for the basic break, the model will be basically in a constant state of 0, such as Figure 6 . This case mainly relies on the last link, relying on the expansibility of the model to fit. And because the basic break state has no practicality, a proper error will not have much impact.

[0070] In order to show the prediction performance of the model, we take the true value as the abscissa and the predicted value as the ordinate to draw the scatter plot of all data as Figure 7 . It can be seen that almost all data points are closely distributed near the ideal straight line x=y, indicating that the neural network has high prediction accuracy, and the R 2 >0.99.

[0071] The following selects samples with representative output characteristic curves. The curve connected by the true value and the corresponding predicted value scatter points together constitutes an image, which can more clearly show the fitting effect of the model under different characteristics. Among them, Figure 8 The parameters of the output curve are V gs =0V, L ap =15um, L go =3um, TGaN1=0.2um, TGaN2=0.3um, NGaN2=8e+017, TGaN3=5um, NGaN3=2e+016.

Claims

1. A GaN CAVET modeling method based on a physically guided neural network, characterized in that, Includes the following steps: S1. Construct a CAVET model using the TCAD simulation tool and generate model samples. Each model sample includes the following device parameters: aperture layer length L. ap Gate overlap length L go Unintentionally doped layer thickness t GaN1 , current blocking layer thickness t GaN2 Current blocking layer doping concentration N GaN2 , drift layer thickness t GaN3 And drift and aperture layer doping concentration n GaN3 ; For each model sample, perform volt-ampere characteristic simulation and extract simulation data. The simulation data includes: output drain voltage V. ds Drain current I ds,TCAD Based on the output drain voltage V ds Drain current I ds,TCAD Obtain the output characteristic curve; S2. Standardize the device parameters of each model sample in step S1; S3. Construct a Physically Guided Neural Network (PGNN) model, including shallow neural networks and hypernets; Construct a shallow neural network, which is used to implement the computation process described in the following formula: I ds,pred =a1tanh(a2V ds +a3)+a4tanh(a5V ds +a6)+a7tanh(a8V ds +a9)+a 10 Among them, a1~a 10 Ten parameters representing a shallow neural network; A hypernet is constructed based on a multilayer perceptron (MLP). This hypernet takes device parameters as input and outputs a1 to a2. 10 Ten parameters, then the shallow neural network outputs the drain voltage V. ds Input is the predicted leakage current value I. ds,pred .

2. The GaN CAVET modeling method based on a physically guided neural network as described in claim 1, characterized in that, In S3, the Hypernet includes three Multi-Level Processing (MLPs). The device parameters first enter the first MLP. The output vector of the first MLP is divided into the first four dimensions and the remaining part. The remaining part is linearly transformed and then cross-multiplied with the first four dimensions before being input into the second MLP. The output vector of the second MLP is then divided into the first four dimensions of the second MLP and the remaining part of the second MLP. The remaining part of the second MLP is linearly transformed and then cross-divided with the first four dimensions of the second MLP before being input into the third MLP. The third MLP finally generates the output.

3. The GaN CAVET modeling method based on a physically guided neural network as described in claim 2, characterized in that, S3 also includes the following steps: The shallow neural network is trained using the output characteristic curve obtained from S1: the output drain voltage V corresponding to each point in the output characteristic curve is used. ds Drain current I ds A dataset was constructed to train a shallow neural network, yielding 10 parameters for the shallow neural network; for each set of device parameters, a set of shallow neural network parameters a1 to a1 were trained. 10 With 10 parameters, we obtain a dataset of 10 parameters for a shallow neural network; Using a dataset of 10 parameters of a shallow neural network and its corresponding device parameters, a hypernet is trained to obtain a pre-trained hypernet.

4. The GaN CAVET modeling method based on a physically guided neural network as described in claim 3, characterized in that, S3 is followed by the following steps: S4 is based on the predicted leakage current value I ds,pred and drain current I ds,TCAD The difference, and transconductance g m,TCAD and the transconductance g obtained based on the predicted leakage current value m,pred The difference is used to construct a loss function, and the Physically Guided Neural Network (PGNN) model is trained; whereby, 5. The GaN CAVET modeling method based on a physically guided neural network as described in claim 4, characterized in that, In S4, the formula for the loss function is as follows: Where k is a preset weight value, and n is the number of points on the output characteristic curve under specific device parameters.

6. The GaN CAVET modeling method based on a physically guided neural network as described in claim 1, characterized in that, The device structure of the GaN CAVET, from top to bottom, includes a gate metal, a lower gate n-type gate, and a lower gate n-type gate. + GaN contact layer, AlGaN barrier layer, intrinsic GaN channel layer, undoped GaN layer, current blocking layer, current aperture layer, and drift layer with n + GaN drain contact layer.

Images