Power VDMOSFET single event effect prediction method based on artificial neural network

Abstract

The invention discloses a power VDMOSFET single event effect prediction method based on an artificial neural network. The method comprises the following steps: step 1, establishing a power Si-based VDMOSFET device model based on TCAD, and collecting single event effect simulation data of the Si-based VDMOSFET; 2, establishing a data set based on TCAD simulation data, and dividing the data set into a training set, a cross validation set and a test set; step 3, establishing a deep learning prediction model based on the convolutional neural network; 4, inputting the training set into a deep learning prediction model, measuring the difference between a predicted value and a real value through a loss function, and optimizing the deep learning prediction model; and step 5, test set data are input into the optimized deep learning prediction model, the prediction accuracy and generalization ability of the model are evaluated, and the model meeting requirements can be used for single event effect prediction of the Si-based VDMOSFET device under different parameters and different working conditions. The method has the characteristics of high prediction precision and short prediction time.

Description

Technical Field

[0001] The invention belongs to the technical field of power VDMOSFET devices, and in particular relates to a power VDMOSFET single particle effect prediction method based on an artificial neural network. Background Art

[0002] Due to the limitations of power supply and energy storage, the pursuit of high efficiency of power systems in some fields is increasing. Low-loss power VDMOSFET has the potential to replace traditional power devices due to its excellent performance and efficient power handling capabilities. However, in special environments, such as harsh space and nuclear radiation conditions, the radiation resistance of the device is severely challenged. If its radiation effects are not studied, the device will not be able to work stably in these environments, and may even fail, causing the entire system to fail to operate normally, resulting in serious consequences. Therefore, the single-particle effect (SEE) of power VDMOSFET devices has become an important issue affecting the reliability of modern electronic systems in space and ground applications. Single-particle effects include single-particle upset (SEU), single-particle transient (SET), single-particle latch-up (SEL), single-particle burnout (SEB) and single-particle gate penetration (SEGR). The basic mechanisms of these effects are similar. High-energy particles affect the charged position of sensitive areas. When the charges are released, they are transferred to the device, causing the sensitive areas of the device to collect these charges. At present, researchers often use TCAD numerical simulation tools such as Sentaurus and Silvaco to study the single-particle effects of VDMOSFET. However, the traditional VDMOSFET radiation simulation process is complex and has problems such as non-convergence. It takes a long time for researchers to master TCAD software and conduct radiation characteristics research on VDMOSFET.

[0003] In outer space, the radiation effects on power VDMOSFET are mainly divided into total dose radiation effect and single particle effect. The total dose radiation effect will cause the device electrical performance to drift, while the single particle effect may directly cause the device to fail, greatly challenging its reliability. At present, the research on the single particle effect of Si-based VDMOSFET devices usually relies on TCAD simulation software. In addition, researchers have also begun preliminary research on the single particle effect of SiC VDMOSFET and have achieved corresponding research results. However, in a short period of time, Si-based VDMOSFET will still occupy the main market of power VDMOSFET for a long time and has a very broad development prospect and application space. However, the traditional research method of Si-based VDMOSFET based on Sentaurus TCAD software has problems such as slow simulation speed, non-convergence of curves and high threshold.

[0004] Establishing a power Si-based VDMOSFET device model based on TCAD can realize the simulation and research of single particle effects, which is the main research method for current power R&D personnel. However, this process is not only complicated and time-consuming, but also requires researchers to have a solid knowledge of semiconductor physics and power device irradiation. In addition, when using Sentaurus TCAD software to simulate VDMOSFET devices, due to the large amount of data, it is more likely to have simulation non-convergence problems compared to FinFET and GAAFET, which often leads to simulation difficulties and low research efficiency. Therefore, the existing Si-based VDMOSFET single particle effect simulation method has the problems of long time consumption, low efficiency, and being greatly affected by human factors. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a power VDMOSFET single particle effect prediction method based on artificial neural network, which has the characteristics of high prediction accuracy and fast prediction time. With the help of a carefully trained prediction model, researchers can bypass those complex and time-consuming TCAD simulation processes. By simply inputting the device structure parameters at the input end of the network model, the source current pulse curve of the power VDMOFET device and the dynamic characteristics of the single particle effect can be efficiently predicted.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A power VDMOSFET single particle effect prediction method based on artificial neural network includes the following steps:

[0008] Step 1: Establish a power Si-based VDMOSFET device model based on TCAD and collect the single-particle effect simulation data of Si-based VDMOSFET;

[0009] Step 2, establishing a data set based on the simulation data in step 1, and dividing the data set into a training set, a cross-validation set, and a test set;

[0010] Step 3, establishing a deep learning prediction model based on convolutional neural network;

[0011] Step 4: Input the training set into the deep learning prediction model, measure the gap between the predicted value and the true value through the loss function, and optimize the deep learning prediction model;

[0012] Step 5: Input the test set data into the optimized deep learning prediction model to evaluate its prediction accuracy and generalization ability. The model that meets the requirements can be used to predict the single particle effect of Si-based VDMOSFE devices under different parameters and different working conditions.

[0013] The step 1 simulates the phenomenon of charge collection generated when high-energy particles pass through the device, and then studies the influence of single particle effect on power Si-based VDMOSFET;

[0014] The basic structure of the power Si-based VDMOSFET device includes an N-channel region in the upper part and a P-base region in the lower part. In TCAD modeling, the selected device material is Silicon. The N-channel region is a conductive region, and the gate voltage is used to control the opening and closing of the channel to control the flow of current; the P-base region is a reverse withstand voltage region, which is used to bear high voltage. The breakdown voltage of the simulated VDMOSFET is 280V. In step 2, based on the device model built in step 1, four groups of incident parameters are pulled to change the design conditions, linear energy transfer LET, particle incident position Location (x0), particle incident angle θ and drain bias voltage (Vd) to obtain 3600 sets of data sets, and the data sets are divided into training sets, cross-validation sets and test sets in a ratio of 8:1:1 to obtain data sets by changing the design conditions.

[0015] In step 3, the deep learning prediction model is divided into three modules, namely, a dimension expansion module, a feature extraction module and a prediction output module;

[0016] The dimension expansion module is used to expand the dimension of the input data to provide more information for the subsequent feature extraction module. The dimension expansion increases the dimension of the data by using a fully connected layer;

[0017] The feature extraction module is used for convolution operation to extract features between data. The convolution layer learns the local features of the data, and the pooling layer reduces the number of parameters and extracts important features.

[0018] The prediction output module is used to reduce the data dimension and convert the extracted features into the final prediction results. The prediction output module includes several fully connected layers, and the last layer uses an activation function to generate the prediction output.

[0019] The dimension expansion module includes two fully connected layers and one transposed convolution layer. The input data passes through the fully connected layer and the transposed convolution layer in sequence. The transposed convolution layer further expands the dimension of the input data on the basis of the full connection, which is convenient for subsequent convolution.

[0020] The feature extraction module includes two identical submodules, each of which is composed of a convolution layer with a convolution kernel size of 1×3 and a double-branch structure unit connected in sequence, and the input data area observed by each convolution layer or pooling layer;

[0021] The size of the convolution kernel is 1×3, and the size of the pooling layer is 1×2;

[0022] The dual-branch structure includes two channels, one channel contains two cascaded one-dimensional convolutional layers with a convolution kernel size of 1×3, and the other channel contains a one-dimensional convolutional layer with a convolution kernel size of 1×3. The output features of the two channels are then concatenated, and the concatenated features are used as the output of the dual-branch convolutional layer;

[0023] In addition, in the feature extraction module, a batch normalization layer is cascaded after each convolution layer to prevent overfitting.

[0024] In step 4, 360 groups of cross-validation sets divided into 8:1:1 are input into the deep learning prediction model. After the training is completed, the accuracy of the prediction model is improved by comparing its loss function curve with the loss function curve of the training set, while the adverse effects caused by overfitting of the training are weakened. The loss calculation function during the training process is as follows:

[0025]

[0026] Where n is the total number of samples, y i and They represent the predicted value and true value of the i-th sample respectively.

[0027] The loss function curve is determined by setting the code in Pycharm, observing the loss function decline curves of the training set and the validation set, and judging the training effect.

[0028] In the step 5, the different parameters and the different working conditions are the linear energy transfer LET, the incident position Location (x0) of the particle, the incident angle θ of the particle and the drain bias voltage (Vd).

[0029] Beneficial effects of the present invention:

[0030] 1. Compared with the prior art, the present invention establishes a data set through TCAD simulation software as the input of the neural network prediction model, and utilizes the ability of convolutional neural network to analyze data and extract features, so as to accurately and quickly predict the single particle effect characteristics of VDMOSFET devices. It solves the problems of long time consumption, low efficiency and great influence of artificial factors in the traditional method of studying the characteristics of Si-based VDMOSFET devices based on simulation software, and expands the possibility of device prediction implementation.

[0031] 2. The network structure adopted in the present invention integrates the convolution layer and the dual-branch structure into a new module. By cross-utilizing the data extraction feature method of the convolution layer and the dual-branch convolution unit, the prediction accuracy and prediction time of the radiation characteristics prediction of the Si-based VDMOSFET device are significantly improved, and the final prediction accuracy is 98.91%, and the training time is only 17.9182s. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is the structure diagram of the VDMOSFET device.

[0033] Figure 2 .Structure diagram of VDMOSFET device after particle injection.

[0034] Figure 3 Code for the HeavyIon module

[0035] Figure 4 This is the structure diagram of the dual-branch convolution unit

[0036] Figure 5 This is a diagram of the neural network structure.

[0037] Figure 6 This is a comparison chart of the prediction results of a group of randomly selected single-particle effect current transient pulse curves.

[0038] Figure 7 The figure is a comparison chart of the prediction results of two groups of randomly selected single particle effect current transient pulse curves.

[0039] Figure 8 The figure is a comparison chart of the prediction results of three groups of randomly selected single particle effect current transient pulse curves.

[0040] Fig. 9 The figure is a comparison chart of the prediction results of 4 groups of randomly selected single particle effect current transient pulse curves.

[0041] Fig.10 This is a comparison chart of the prediction results of the total collected charge Q0 for single particle effects.

[0042] Fig.11 This is a comparison chart of the single particle effect peak current I0 prediction results. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0044] The present invention aims to predict the single particle effect of Si-based VDMOSFET devices using deep learning. The prediction includes two aspects: one is to predict the source current pulse of the device, and the other is to predict the dynamic characteristic parameters of the single particle effect.

[0045] First, the single particle effect of Si-based VDMOSFET devices is simulated using device simulation software, and the number of input and output neurons of the neural network is determined based on this. Next, the network model will be established and optimized to ensure the prediction speed and accuracy, and the dynamic characteristic parameters of the source current pulse and single particle effect of Si-based VDMOSFET devices will be predicted.

[0046] Step 1: Establish a power Si-based VDMOSFET device model based on TCAD. Use the Sentaurus SDE tool in the Sentaurus TCAD tool suite to complete the structural modeling of the Si-based VDMOSFET device, simulate the process characteristics, complete the doping and meshing of each part of the model, and the device structure is as follows: Figure 1 As shown: the current flows upward from the drain substrate region through the drift region (Epi) and the JFET region, and then flows horizontally through the channel region to the source (Source). Compared with the lateral structure of the traditional MOSFET, the voltage and current resistance capabilities are greatly improved. For the enhanced NMOS, when the device is in the off state (i.e. VGS = 0V, VDS> 0V), the PN junction formed by the P-type body region (Pbdoy) and the N-type drift region (Nepi) provides a large depletion layer to withstand high voltage through the lightly doped drift region. The source and base regions are short-circuited by metal connections to suppress the parasitic triode formed by the N-type source region, P-type body region and N-type drift region in the MOS structure. The channel region is generally formed by self-aligned injection and push-back junction, and the channel length refers to the distance between the diffusion boundary of the Body region and the Source region after diffusion. Since the current flow direction is vertical and the channel is formed by double diffusion, this process is called vertical double diffusion power MOSFET. The device structure parameters selected by the present invention are: cell width Wcell=5um, gate oxide layer thickness is 0.2um, drift region doping concentration is 7.5×10 14 cm -3 , the drift region thickness is 12um, and the substrate doping concentration is 1×10 19 cm -3 , the doping of the pbase region and the source and drain regions is set to Gaussian distribution, and the peak concentration is 1.7×10 17 cm -3 and 2×10 18 cm -3 .

[0047] After the device structure model is established, the single particle effect of the device can be simulated by writing code in the HeavyIon module of the SDevice tool, and then the source current pulse can be extracted, and key single particle effect parameters can be extracted from it, such as the source peak current I0 and the total collected charge Q0, the four sets of incident parameters of the pull-bias, the linear energy transfer LET, the particle's incident position Location (x0), the particle's incident angle θ and the drain bias voltage (Vd) to obtain 3600 sets of data sets for subsequent neural network training.

[0048] like Figure 3As shown in Figure 1, the code of this module allows the adjustment of parameters such as the impact position, direction, angle and radius of the particles. Table 1 lists in detail the specific meaning of each parameter in the code and the parameter values ​​used in this article. The Picocoulomb option in the code is used for unit conversion. When this option is enabled, the units of length and incident radius will change from centimeters to micrometers, and the unit of LET_f will be converted from cm-3 to pc / μm. When high-energy particles penetrate integrated circuit materials, they collide with atoms, causing atomic ionization and excitation, releasing electrons and holes, and forming electron-hole pairs, which will affect the working state of the device.

[0049] Formulas (1.1) and (1.2) can be used to calculate the peak current I0 and the total collected charge Q0. The charge collection mechanism of the single particle effect mainly includes drift collection and diffusion collection. When the excess carriers deposited by the charged particles are located in the applied electric field in the device (for example, the electric field of the pn junction depletion layer), the carriers will drift rapidly to the collection node under the action of the electric field; and when the excess carriers deposited by the charged particles are located in the area of ​​the device without electric field or with low electric field strength (such as the substrate), the carriers will migrate to the collection node by diffusion driven by the concentration difference. These two mechanisms jointly affect the response characteristics of the device under single particle irradiation, and play an important role in the formation of transient current pulses and the extraction of characteristic parameters.

[0050]

[0051]

[0052] The single event incident model of the device is as follows Figure 2 As shown:

[0053] By changing the design parameters through TCAD simulation software, two sets of data sets are obtained for deep learning model prediction, which are the source current curve and dynamic characteristic parameters of VDMOSFET devices. According to previous researchers' research on single particle effects, linear energy transfer LET, particle incident position Location (x0), particle incident angle θ and drain bias voltage (Vd) are selected as the influencing factors of single particle effects, and used as inputs for neural network prediction, which are simulation inputs here. By changing these irradiation conditions to simulate the single particle effect of the device, the specific values ​​are shown in Table 1. A total of 3600 sets of data are collected, which are drain transient current pulse curves, current peaks and total collected charges.

[0054] Both datasets 1 and 2 use the structural parameters of the VDMOSFET device as the input of the network prediction model. The output corresponding to dataset 1 is 201 data extracted from the average of the single particle effect transient pulse curve of the device. The output corresponding to dataset 2 is the dynamic characteristic parameters of the device, peak transient current I0, total collected charge Q0

[0055] Table 1 SEE characteristic variable values

[0056]

[0057]

[0058] Step 2: Establish a data set based on the TCAD simulation results. Change the design conditions based on the device model established in step 1 to obtain a data set. Divide the data set into 2880 training sets, 360 cross-validation sets, and 360 test sets in a ratio of 8:1:1.

[0059] Step 3: Establish a deep learning prediction model based on convolutional neural network. The neural network model structure is as follows: Figure 3 As shown in the figure, the network structure for predicting the source transient curve is basically the same as the structure for predicting the characteristic parameters. The difference is that the output of predicting the source transient curve is 201 curve values, while the output of predicting the characteristic parameters is 2 characteristic parameters. The prediction purpose can be achieved by adjusting the number of fully connected neurons.

[0060] The entire neural network prediction model structure is divided into three modules, namely, dimension expansion module, feature extraction module and prediction output module;

[0061] like Figure 5 As shown in the figure, the dimension expansion module contains two fully connected layers and one transposed convolution layer. The transposed convolution layer can further expand the dimension of the input data on the basis of full connection, which is convenient for subsequent convolution.

[0062] The fully connected layer can be used to increase the dimension of the data to expand the dimension of the input vector. General convolution operations will reduce the dimension of the input vector. Depending on the size of the convolution kernel and the convolution step, the output size varies greatly. The transposed convolution layer acts in the opposite way to the convolution layer and can expand the data dimension. For example, in image semantic segmentation, the feature size of the final output is often required to be the same as the original input size, but the size of the feature map gradually decreases during the network convolution and pooling process. In this case, transposed convolution is needed. Therefore, a one-dimensional transposed convolution layer is added after the fully connected layer to further combine and expand the input information, thereby increasing the operating space for the subsequent convolution operation.

[0063] The feature extraction module contains two identical submodules, which are composed of a convolution layer with a convolution kernel size of 1×3 and a dual-branch structural unit. Figure 4This is a diagram of the structure of a dual-branch convolution unit. Generally speaking, the choice of convolution kernel size directly affects the receptive field of the network, that is, the area of ​​input data that can be observed by each convolution layer or pooling layer. For convolution layers, when processing time series data or one-dimensional signals, a smaller kernel size, such as 3 or 5, is usually selected. Such a choice helps the network effectively capture local features and perform well on one-dimensional data. For pooling layers, the size of the convolution kernel usually depends on the data dimension you want to reduce and the feature information you need to retain. Larger pooling kernels can reduce data dimensions and retain more global features, while smaller pooling kernels are more suitable for retaining local details and reducing computational load. Through training comparison, it was finally decided that the size of the convolution kernel is 1×3 and the size of the pooling layer is 1×2, which can maximize the training speed while ensuring training accuracy. In addition, the dual-branch structure includes two channels, one channel includes two one-dimensional convolution layers with a convolution kernel size of 1×3 cascaded in sequence, and the other channel includes a one-dimensional convolution layer with a convolution kernel size of 1×3. The output features of the two channels are then spliced, and the spliced ​​features are used as the output of the dual-branch convolution layer; in addition, a batch normalization layer is added between each convolution layer to prevent overfitting. The purpose of designing the network structure in this way is to fully extract data features, because the internal characteristic mechanism of the single particle effect of Si-based VDMOSFET is complex, and the amount of data for the prediction curve and irradiation characteristic parameters is large. The transposed convolution layer is used to further expand the input dimension while the dual-branch structure is used to efficiently extract data features, which can significantly improve the prediction accuracy. At the same time, the dual-branch structure, the number of convolution layers and fully connected layers is not easy to be too many, which is easy to cause overfitting. The present invention adjusts the network structure multiple times and finally determines that the network structure uses two submodules, which can ensure the highest prediction accuracy of 98.91%, and the training time is only 17.9182s.

[0064] Step 4: Input the training set into the prediction model, measure the gap between the predicted value and the true value through the loss function, optimize the prediction model, and input the cross-validation set into the prediction model. By comparing its loss function curve with the training set loss function curve, the accuracy of the prediction model can be improved while weakening the adverse effects caused by overfitting. The loss calculation function in the training process is as follows:

[0065]

[0066] Where n is the total number of samples, y i and They represent the predicted value and true value of the i-th sample respectively.

[0067] Step 5: Input the test set data into the prediction model to evaluate its prediction accuracy and generalization ability. The model that meets the requirements can be used to predict the single particle effect of Si-based VDMOSFE devices under different parameters and different working conditions.

[0068] Example:

[0069] In one embodiment of the present invention, the VDMOSFET device structure established by the present invention is as follows Figure 1 As shown. Among them, the cell width Wcell = 5um, the gate oxide layer thickness is 0.2um, and the drift region doping concentration is 7.5×10 14 cm -3 , the drift region thickness is 12um, and the substrate doping concentration is 1×10 19 cm -3 , the doping of the pbase region and the source and drain regions is set to Gaussian distribution, and the peak concentration is 1.7×10 17 cm -3 and 2×10 18 cm -3 .

[0070] Figure 6 to Figure 9 This is a comparison chart of the single-particle effect source transient current curves of VDMOSFET randomly selected from 360 test sets and the corresponding simulation data. As shown in the figure, the curve has a high degree of fit and the prediction model has a very good prediction effect.

[0071] Fig.10 Fig.11 This is a comparison chart of the electrical parameters of the power VDMOSFET in the 360 ​​test sets in ascending order with the corresponding simulation data. As shown in the figure, the results of the real value and the predicted value have a high degree of fit, and the prediction effect of the prediction model is good.

[0072] The present invention obtains 3600 sets of single particle effect characteristics of Si-based VDMOSFE devices by means of TCAD device model simulation, including source current curves and two sets of single particle effect dynamic characteristic parameters I0 and Q0. The input characteristic parameters of the present invention select source current curves and single particle effect dynamic characteristic parameters under different device design parameters to form a data set of a neural network prediction model. This method meets the application needs of current researchers and has certain research and promotion value.

[0073] The present invention uses TCAD numerical simulation software to model Si-based VDMOSFET devices, which is a common practice. This method can help researchers deeply understand the behavior of devices under different conditions. Through TCAD software, the process of interaction between high-energy particles and devices can be simulated, thereby collecting single-particle effect simulation data of Si-based VDMOSFET. These data are crucial for building deep learning models because they can be used as training data to help neural networks learn how to predict single-particle effects.

[0074] In the training process of the deep learning model, the present invention uses these simulation data to train the neural network prediction model. During the training process, the model continuously adjusts the network parameters through forward propagation, loss function calculation, back propagation and optimization algorithm to minimize the prediction error.

[0075] Once the model is trained, researchers can use this deep learning prediction model to predict single-particle effects of new VDMOSFET devices without the need for time-consuming TCAD irradiation modeling simulations. This approach can significantly save computing resources and improve research efficiency. In this way, researchers can devote more time and energy to the design and optimization of new devices instead of spending it on repetitive simulation calculations. In addition, the generalization ability of the deep learning model also means that it can be predicted under different devices and conditions, which further enhances the practicality of the model. With the improvement of computing power and the continuous optimization of algorithms, the performance and efficiency of deep learning models in predicting single-particle effects will be further improved.

[0076] The single particle effect modeling research method used in the present invention is relatively common and is the irradiation research process method currently generally followed by researchers. However, this method has the problems of difficulty in modeling and long simulation time. However, there is no invention that introduces deep learning into the irradiation modeling of Si-based VDMOSFET. In addition, the network input parameters of the present invention are different from traditional research methods. Four groups of electrical parameters that have a greater impact on the single-particle effect of Si-based VDMOSFET are selected, and the drain bias voltage (Vd) is selected as 9V, 10V, 11V, 12V, and 13V. The LET value is selected as 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1. The incident angle is selected as 0, 15, 30, 45, 60, and 75. The incident position varies according to the device structure, which are 0, 0.4, 0.8, 1.1, 1.4, 1.7, 2.5, 3, 3.3, 4.1, 4.3, and 4.7, respectively, involving the channel region, JFET region, Nplus injection region, Pplus injection region, and Pbase region.

[0077] The deep learning prediction model adopted by the present invention is based on a convolutional neural network structure. Unlike the traditional method of relying on TCAD software for simulation research, the end-to-end powerful learning ability of the neural network can skip the cumbersome and complex software operation links and directly learn the simulation principles of the underlying software. This method greatly improves the simulation efficiency of single-particle effects in the field of Si-based VDMOSFE, and provides a more efficient and convenient tool for related research. At the same time, the present invention creatively combines the convolution layer and the dual-branch convolution unit into a new module. This cross-use of the convolution layer and the dual-branch convolution unit training structure greatly improves the training efficiency and training accuracy compared to the traditional method of relying only on the convolution layer to train data layer by layer.

Claims

1. A power VDMOSFET single-particle effect prediction method based on artificial neural network, characterized in that: The steps include: Step 1: Establish a power Si-based VDMOSFET device model based on TCAD and collect the single-particle effect simulation data of Si-based VDMOSFET; Step 2, establishing a data set based on the simulation data in step 1, and dividing the data set into a training set, a cross-validation set, and a test set; Step 3, establishing a deep learning prediction model based on convolutional neural network; Step 4: Input the training set into the deep learning prediction model, measure the gap between the predicted value and the true value through the loss function, and optimize the deep learning prediction model; Step 5: Input the test set data into the optimized deep learning prediction model to evaluate its prediction accuracy and generalization ability. The model that meets the requirements can be used to predict the single particle effect of Si-based VDMOSFE devices under different parameters and different working conditions.

2. The method for predicting single-particle effects of power VDMOSFET based on artificial neural network according to claim 1, characterized in that: The basic structure of the power Si-based VDMOSFET device in step 1 includes an N-channel region in the upper part and a P-base region in the lower part. In TCAD modeling, the material of the device is selected as Silicon. The N-channel region is a conductive region, and the opening and closing of the channel is controlled by the gate voltage to control the flow of control current; the P-base region is a reverse voltage withstand region, which is used to withstand high voltage.

3. The power VDMOSFET single-particle effect prediction method based on artificial neural network according to claim 1, characterized in that: In step 2, based on the device model constructed in step 1, the design conditions are changed by pulling four groups of incident parameters, linear energy transfer LET, particle incident position Location (x0), particle incident angle θ and drain bias voltage (Vd) to obtain a data set, and the data set is divided into a training set, a cross-validation set and a test set.

4. The method for predicting single-particle effects of power VDMOSFET based on artificial neural network according to claim 1, characterized in that: In step 3, the deep learning prediction model is divided into three modules, namely, a dimension expansion module, a feature extraction module and a prediction output module; The dimension expansion module is used to expand the dimension of the input data to provide more information for the subsequent feature extraction module. The dimension expansion increases the dimension of the data by using a fully connected layer; The feature extraction module is used for convolution operation to extract features between data. The convolution layer learns the local features of the data, and the pooling layer reduces the number of parameters and extracts important features. The prediction output module is used to reduce the data dimension and convert the extracted features into the final prediction results. The prediction output module includes several fully connected layers, and the last layer uses an activation function to generate the prediction output.

5. The method for predicting single-particle effects of power VDMOSFET based on artificial neural network according to claim 4, characterized in that: The dimension expansion module includes two fully connected layers and one transposed convolutional layer. The input data passes through the fully connected layer and the transposed convolutional layer in sequence. The transposed convolutional layer further expands the dimension of the input data on the basis of full connection to facilitate subsequent convolution.

6. The method for predicting single particle effects of power VDMOSFET based on artificial neural network according to claim 4, characterized in that: The feature extraction module includes two identical submodules, each of which is composed of a convolution layer with a convolution kernel size of 1×3 and a double-branch structure unit connected in sequence, and the input data area observed by each convolution layer or pooling layer; The size of the convolution kernel is 1×3, and the size of the pooling layer is 1×2; The dual-branch structure includes two channels, one channel contains two cascaded one-dimensional convolutional layers with a convolution kernel size of 1×3, and the other channel contains a one-dimensional convolutional layer with a convolution kernel size of 1×3. The output features of the two channels are then concatenated, and the concatenated features are used as the output of the dual-branch convolutional layer; In addition, in the feature extraction module, a batch normalization layer is cascaded after each convolution layer to prevent overfitting.

7. The method for predicting single-particle effects of power VDMOSFET based on artificial neural network according to claim 1, characterized in that: In step 4, the divided cross-validation set is input into the deep learning prediction model. After the training is completed, the accuracy of the prediction model is improved by comparing its loss function curve with the loss function curve of the training set, while the adverse effects caused by overfitting of the training are weakened. The loss calculation function during the training process is as follows: Where n is the total number of samples, y i and They represent the predicted value and true value of the i-th sample respectively.

8. The method for predicting single-particle effects of power VDMOSFET based on artificial neural network according to claim 7, characterized in that: The loss function curve is determined by setting the code in Pycharm, observing the loss function decline curves of the training set and the validation set, and judging the training effect.

9. The method for predicting single-particle effects of power VDMOSFET based on artificial neural network according to claim 1, characterized in that: In the step 5, the different parameters and the different working conditions are the linear energy transfer LET, the incident position Location (x0) of the particle, the incident angle θ of the particle and the drain bias voltage (Vd).

Images