Single event effect data generation method based on generative adversarial network

Abstract

The invention belongs to the technical field of machine learning, and particularly relates to a single event effect data generation method based on a generative adversarial network (GAN), which can generate high-quality simulation data consistent with actual data distribution on the basis of a small amount of real experimental data through the efficient data generation capability of the GAN, thereby reducing the high dependence on the experimental data. According to the method, diversified and high-precision training data can be provided for the machine learning model, the data acquisition time and economic cost can be remarkably reduced, and the defects of a traditional modeling method and an existing machine learning technology are effectively overcome. Through the application of the method, the precision and efficiency of single event effect modeling can be remarkably improved, important technical support is provided for the design of an anti-radiation hardening circuit, and the actual requirements of the design of a large-scale aerospace integrated circuit are met.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and particularly relates to a method for generating single-event effect data based on a generative adversarial network. Background Art

[0002] At present, the main means for analyzing and evaluating the single-event effect of aerospace integrated circuits are ground simulation tests and computer simulation. Since it is difficult to obtain the internal response of the circuit in ground radiation simulation tests, the single-event effect simulation technology is more suitable for analyzing and understanding the process of the single-event effect on aerospace integrated circuits and the effectiveness analysis of radiation hardening design. Among them, the single-event effect model, as the core factor affecting the accuracy of single-event effect simulation, has been concerned and studied by a large number of researchers.

[0003] In the prior art, traditional single-event effect simulation modeling methods usually adopt device simulation means based on physical mechanisms. Although these methods can deeply analyze the transient response inside the circuit after particle impact, due to high computational complexity and slow simulation speed, it is difficult to meet the modeling requirements of large-scale integrated circuits. For this reason, methods based on numerical fitting are widely used, and by extracting the laws of experimental data, rapid modeling is achieved at a low computational cost. However, this method still has great limitations in terms of modeling complexity and accuracy, and it is difficult to fully reflect the true response of the circuit.

[0004] In recent years, modeling techniques based on machine learning have gradually become the mainstream direction of single-event effect research due to their powerful pattern recognition ability and good prediction performance. Machine learning technology obtains a more accurate single-event effect model through learning a large amount of data.

[0005] However, machine learning technology heavily relies on high-quality and large-scale single-event effect data sets in the early stage. And the acquisition cost of single-event effect data is high and the cycle is long, resulting in insufficient data becoming the main bottleneck restricting the modeling accuracy and generalization ability, and also hindering its popularization and application in radiation hardening design.

[0006] Therefore, there is an urgent need for an efficient, low-cost and widely applicable single-event effect data generation technology to further promote the development of related fields. Summary of the Invention

[0007] To solve the above problems existing in the prior art, the present invention provides a method for generating single-event effect data based on a generative adversarial network, which solves the problem of insufficient data faced by current single-event effect modeling based on machine learning.

[0008] The object of the present invention can be achieved by the following technical solutions: A method for generating single-event effect data based on a generative adversarial network, comprising the following steps:

[0009] S1: Construct a training dataset based on TCAD simulation:

[0010] Establish and calibrate the NMOS device structure model,

[0011] Use the TCAD heavy ion model to construct a single event effect model, set the heavy ion LET value to 4 - 100 MeV·cm 2 / mg, the incident angle is 0 - 90°, the bombardment position is the NMOS drain, and calibrate the composite model parameters;

[0012] Perform segmented sampling on the single event transient current based on the data change rate, allocate sampling points according to the preset ratio according to the characteristics of the single event transient current, and generate a training dataset by combining the gradient change rate and interpolation;

[0013] S2: Data generation based on conditional generative adversarial network:

[0014] Construct a CGAN model, use the LET value, incident position and angle as conditional information, and the latent space noise as input to generate single event transient current data;

[0015] Train the generator and discriminator, and use 16 groups of particle transient current data under different heavy ion LET values, incident angles and incident positions obtained from TCAD simulation as the training set for model training.

[0016] Preferably, after S2, it further includes S3: Use random comparison to verify the accuracy and efficiency of the generated data. The random comparison verification includes evaluating the consistency of the generated data and the real data distribution through KL divergence and JS divergence, and statistically calculating the generation efficiency.

[0017] Preferably, in S1, the segmented sampling of the single event transient current includes:

[0018] Divide the transient current interruption into three intervals according to [T 0 , T max , [T max , T 0.1max , [T 0.1max , T 0.01max , and the sampling point allocation ratio is 40%, 40%, 20%;

[0019] Where T 0 is the ion bombardment moment, T max is the moment when the transient current reaches the maximum value, T 0.1max is the moment when the transient current decays to 10% of the maximum current value, T 0.01max is the moment when the transient current is less than 1% of the maximum current value;

[0020] Evenly divide N / 2 subintervals within each interval, and dynamically allocate sampling points according to the gradient change rate to ensure that each subinterval contains at least 1 sampling point;

[0021] Generate a consistent dataset by interpolating and correcting the number of sampling points.

[0022] Preferably, in S1, the parameters of the NMOS device structure model include: channel length 130 nm, width 420 nm, gate oxide thickness 2.58 nm, source-drain doping concentration 1×10 20 cm -3 、channel doping density 1.57×10 17 cm -3 , the gate voltage is fixed at 0 V, and the drain voltage range is 0 - 1.8 V.

[0023] Preferably, in S2, both the generator and discriminator of the CGAN are composed of 6-layer fully connected networks. The latent space noise is a Gaussian distribution random variable. The conditional information of the generator and discriminator both includes the LET value, incident position coordinates, and incident angle.

[0024] Preferably, in S2, the model optimizer uses the Adam optimizer, the learning rate is set to 0.00002, and the network training Epoch parameter is set to 50000.

[0025] Preferably, the activation function of the generator uses ReLU, the activation function of the last layer of the discriminator uses Sigmoid, and other layers use LeakyReLU.

[0026] Preferably, the single event effect model uses a composite model, and the composite model includes the Shockley–Read–Hall (SRH) model, effective density of states model, Fermi-Dirac statistics, impact ionization model, and Lucent model.

[0027] The beneficial effects of the present invention are as follows:

[0028] 1. The data sampling is more representative: Compared with the equidistant sampling method, the data obtained by segmented sampling based on the data change rate is sampled more densely in the intervals where the data changes violently. The sample data obtained thereby can better characterize the characteristics of single event transient current.

[0029] 2. A large amount of high-quality data can be efficiently obtained with a small number of training samples: The present invention is trained by a conditional generative adversarial network. With a small number of real simulation sample data as input, it discriminates whether the current data deviates according to the given single event transient conditional information. Finally, the mean distribution of the obtained data is very close to the real data distribution. A large amount of well-quality generated data can be obtained through a small number of sample data, thus saving a large amount of simulation time cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] For the convenience of those skilled in the art to understand, the present invention will be further described below with reference to the accompanying drawings.

[0031] Figure 1 It is a flowchart of a single event effect data generation method based on a generative adversarial network:

[0032] Figure 2 It is a schematic diagram of the 3D model of the NMOS device;

[0033] Figure 3 It is a distribution diagram of electron density after heavy ion bombardment of the drain region;

[0034] Figure 4 It is a schematic diagram of a piecewise sampling algorithm based on the numerical change rate;

[0035] Figure 5 It is a schematic diagram of a data generation model based on a conditional generative adversarial network (CGAN);

[0036] Figure 6 It is a schematic diagram of the change of the model loss function during the training process of the conditional generative adversarial network (CGAN);

[0037] Figure 7 It is a schematic diagram of the time cost comparison between the TCAD-based modeling and simulation of the prior art and the method proposed in this application;

[0038] Figure 8 It is a comparison and data distribution diagram of the single event transient current data generated by the TCAD simulation of the prior art and the method proposed in this application;

[0039] Figure 9 It is a probability density histogram of the means of the real data and the generated data of ten groups of single event transient current data. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will describe in detail the specific embodiments, structures, features and their effects of the present invention with reference to the accompanying drawings and preferred embodiments.

[0041] Please refer to Figures 1 - 9 , this embodiment provides a single event effect data generation method based on a generative adversarial network, which combines TCAD simulation and a CGAN model to generate a large amount of high-quality single event transient current data by using a small amount of real data. It includes the following steps:

[0042] S1: Construct a training data set based on TCAD simulation:

[0043] S11: Establish and calibrate the NMOS device structure model,

[0044] Build an NMOS 3D model based on the SMIC 130nm process, as shown in the appendix Figure 2 The channel length of the device is 130nm, the width is 420nm, the gate oxide thickness is 2.58nm, and the source / drain doping is 10 20 cm -3 , and the channel doping density is 1.57×10 17 cm -3 . The gate voltage of the device is fixed at 0V, the source and substrate are grounded, and the drain is set to 0 - 1.8V.

[0045] Then, use the SMIC 130nm PDK process design kit to calibrate the static characteristics of the device (such as threshold voltage, drain current) to ensure the consistency between the model and the actual process.

[0046] S12: Build a single - event effect model and simulation: For the heavy - ion bombardment simulation, use the heavy - ion model in TCAD, and build a single - event effect model using Shockley–Read–Hall (SRH), Auger, and direct (or radiative) recombination

[0047] Among them, the SRH carrier lifetime has doping - dependence and temperature - dependence. Other models applied include the effective density of states model, Fermi - Dirac statistics, impact ionization, and Lucent model. The LET of the heavy ion is set to 4 - 100 MeV·cm 2 / mg, and the bombardment position is selected as the drain region of the NMOS device. The incident trajectory length of the heavy ion is set to 10μm, and the incident angle of the heavy ion is 0 - 90° (0° is perpendicular to the device surface).

[0048] Set up the simulation experimental groups according to the method of three factors and four levels, change the LET value, incident angle, and incident position of the heavy ion. According to the L 16 (4 3 ) orthogonal table, all combinations of the three factors and four levels can be covered in only 16 experiments, thus effectively reducing the number of experiments and ensuring the balance of experimental results.

[0049] For the simulation output, extract the transient current waveform. The electron density distribution of the device 25ps after the heavy - ion bombardment vertically bombards the center of the NOMS drain is shown in the appendix Figure 3 as follows;

[0050] S13: Obtain segmented sampling data of single - event transient current based on the data change rate:

[0051] To meet the requirement that the sampling points of single-particle transient current for the training data set be consistent, a piecewise sampling algorithm based on the data change rate is proposed here according to the characteristics of the single-particle transient current source. The flowchart of this algorithm is as shown in the appendix Figure 4 as follows.

[0052] a: Input the original data and set the specific sampling requirements;

[0053] b: Divide the data into several intervals, and preprocess the data, such as performing a normalization operation; determine the number of sampling points to be allocated for each interval according to the gradient change rate of each interval, and calculate the gradient change rate of each interval;

[0054] c: Sort the intervals according to the gradient change rate and allocate the corresponding sampling points;

[0055] d: Check whether the sampling result meets the set sampling requirements. If it meets, generate the sampling point data using the difference method and output the sampling point data that meets the requirements;

[0056] If it does not meet, readjust the sampling point data.

[0057] Among them, the method of dividing the data into several intervals: According to the characteristics of the single-particle transient current, divide it into three intervals according to [T 0 , T max , [T max , T 0.1max , [T 0.1max , T 0.01max . Among them, T 0 is the ion bombardment moment, T max is the moment when the transient current reaches the maximum value, T 0.1max is the moment when the transient current decays to 10% of the maximum current value, and T 0.01max is the moment when the transient current is less than 1% of the maximum current value;

[0058] S2: Data generation based on conditional generative adversarial network:

[0059] S21: Construct a CGAN model. The model includes a generator and a discriminator. Generator: The input is the latent space (randomly generated noise) + conditional vector (LET value, incident position, incident angle). The network structure is a 6-layer fully connected network of 256×512×512×512×256×100. The activation function is ReLU, and the output is single-particle transient current data.

[0060] Discriminator: The input is real data + single-particle transient current data generated by the generator. It has a 6-layer fully connected network structure of 256×512×512×512×256×1. The activation function of the last layer is Sigmoid, and the activation functions of other layers are LeakyReLU, outputting the probability of data authenticity.

[0061] Taking the LET value, incident position, and angle as conditional information and latent space noise as the input, generate single-particle transient current data;

[0062] S22: Train the generator and discriminator. The model training data uses 16 groups of particle transient current data under different heavy ion LET values, incident angles, and incident positions obtained by TCAD simulation as the training set.

[0063] S3: Verify the authenticity of the data:

[0064] Use random comparison to verify the accuracy and efficiency of the generated data. Random comparison verification includes evaluating the distribution consistency between the generated data and the real data through KL divergence, JS divergence, and Wasserstein distance, and counting the generation efficiency.

[0065] The schematic diagram of the CGAN data generation model is as Figure 5 shown.

[0066] Implement a single-particle data generation model composed of a generator and a discriminator to generate high-quality data with a distribution close to the real data, enabling the acquisition of a large amount of data close to the real data with a small number of training samples: Through conditional generative adversarial network training, with a small amount of real simulation sample data as the input, this invention discriminates whether the current data deviates according to the given single-particle transient conditional information, and finally obtains a mean distribution of the data that is very close to the real data distribution. A large amount of well-quality generated data can be obtained through a small amount of sample data, thus saving a large amount of simulation time cost.

[0067] Since the change of single-particle transient current is relatively drastic in the first two intervals, in this embodiment, the sampling point quantity adopts an allocation ratio of 40%, 40%, 20%, that is:

[0068] Interval 1: [T 0 , T max , the sampling point proportion is 40%;

[0069] Interval 2: [T max , T 0.1max , the sampling point proportion is 40%;

[0070] Interval 3: [T 0.1max , T 0.01max , the sampling point proportion is 20%;

[0071] According to the dynamic characteristics of each stage, high-density sampling is carried out in key areas to ensure the capture of key features such as transient spikes and slope mutations, realizing differential allocation of sampling resources and avoiding "oversampling" or "undersampling" caused by uniform sampling. Specifically, this ratio can be adjusted according to the data situation.

[0072] In order to prevent some sub-intervals from being completely ignored due to excessive gradient differences, in one embodiment, first, the number of sampling points is allocated according to the magnitude of the gradient change rate of each sub-interval; then, according to the number of sampling points N in each interval, N / 2 sub-intervals are evenly divided in the sampling interval and the gradient change rate is calculated. Moreover, in order to avoid having no sampling points in a large number of sub-intervals due to excessive gradient change rate differences, it is required that there is at least 1 sampling point in the sub-interval.

[0073] This makes the data sampling more representative: the data obtained through segmented sampling based on the data change rate is sampled more densely in the intervals where the data changes violently, and the sample data obtained thereby can better represent the characteristics of the single-particle transient current.

[0074] In one embodiment, in S22, the model optimizer uses the Adam optimizer, the learning rate is set to 0.00002, the network training Epoch is set to 50000, the model construction is implemented using Pytorch, and by monitoring the loss functions (G_loss and D_loss) of the generator and discriminator, the stability of model training and the accuracy of the generated data are ensured. Please refer to the appendix Figure 6 , Figure 6 is a schematic diagram of the loss function under the method of this application. It can be seen from the figure that when Epoch is greater than 15000, the loss functions of the generator and discriminator are basically stable, and the accuracy of the proposed GAN-based single-particle transient current data generation model no longer improves.

[0075] Please refer to Figure 7 , Figure 7 The table shows the comparison of the time costs between traditional TCAD-based modeling and simulation and the method proposed in this application, taking the single-particle transient current modeling of three influencing factors, namely LET value, incident angle, and incident position, in this invention as an example. If 10 different values are selected for each factor as the conditions for sample generation, the number of samples is 103. If all are obtained through TCAD modeling and simulation, it takes 1000 - 2000h to obtain the required sample data. If the GAN-based single-particle effect data generation technology proposed in this study is used, only 1% - 10% of the samples are required as the training samples of the GAN model to train the model, and the sample data can be obtained, and the advantage of this method becomes more significant as the sample size increases.

[0076] In terms of the generation quality of the data, please refer to Figure 8 ,Figure 8 In the case of randomly selecting four groups of different data, the single-particle transient current data obtained by TCAD simulation is compared with the data generated by the GAN-based data generation model proposed in this study. Figure 8 In it, Agl represents the incident angle, 0° is perpendicular to the device surface; LET is the linear energy transfer, with the unit of MeV·cm2·mg-1, Loc is the offset of the incident position along the direction from the drain center to the source center, 0 represents the drain center position of the NMOS device, and the unit is μm. From Figure 8 It can be seen that the GAN-based data generation model proposed by this method is consistent with the real data obtained by TCAD simulation.

[0077] To further show the distribution of the data generated by the data generation model proposed in this study and the real data, please refer to Figure 9 , Figure 9 which are the probability density histograms of the means of the real data and the generated data respectively plotted for ten randomly selected single-particle transient current data. Figure 9 As shown in it, Density is the probability density, equal to the number of sample occurrences / (total number of samples × binwidth), binwidth = data range of the mean / number of interval divisions, and the number of interval divisions is set to 30 in this article.

[0078] Therefore, from Figure 8 and Figure 9 it can be clearly concluded that the method proposed in this application is very close in the quality of the generated data, and the generated data and the real data are very close, and the model quality is good.

[0079] The above is only a preferred embodiment of the present invention, and it does not impose any form of limitation on the present invention. Although the present invention has been disclosed above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to be equivalent embodiments within the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A single-particle effect data generation method based on a generative adversarial network, characterized by: The following steps are involved: S1: Constructing a training dataset based on TCAD simulation: Build and calibrate the NMOS device structure model, The single-event effect model was constructed using the TCAD heavy ion model, and the heavy ion LET value was set to 4-100 MeV·cm 2 / mg, the incident angle is 0-90°, the bombardment position is the NMOS drain, and the composite model parameters are calibrated; The single-particle transient current is sampled in segments based on the data change rate, the sampling points are allocated in a preset ratio according to the characteristics of the single-particle transient current, and a training data set is generated by combining the gradient change rate and interpolation; S2: Data generation based on conditional generative adversarial networks: A CGAN model is constructed, which uses LET value, incident position and angle as conditional information and latent spatial noise as input to generate single-particle transient current data; The generator and discriminator are trained. The model training data uses 16 sets of particle transient current data under different heavy ion LET values, incident angles and incident positions obtained by TCAD simulation as the training set.

2. The method for generating single-event effect data based on a generative adversarial network according to claim 1, characterized in that: S2 also includes S3: using random comparison to verify the accuracy and efficiency of the generated data, the random comparison verification includes evaluating the distribution consistency between the generated data and the real data through KL divergence and JS divergence, and statistically analyzing the generation efficiency.

3. The method for generating single-event effect data based on a generative adversarial network according to claim 2, characterized in that: In S1, segmented sampling of the single particle transient current includes: According to [T0, T max ],[T max , T 0.1max ],[T 0.1max , T 0.01max ] The transient interruption is divided into three intervals, and the sampling points are allocated in the proportions of 40%, 40%, and 20%; Where T0 is the ion bombardment time, T max is the moment when the transient current reaches its maximum value, T 0.1max The moment when the transient current decays to 10% of the maximum current value, T 0.01max It is the moment when the transient current is less than 1% of the maximum current value; Each interval is evenly divided into N / 2 subintervals, and sampling points are dynamically allocated according to the gradient change rate to ensure that each subinterval contains at least one sampling point; The number of sampling points is corrected by interpolation to generate a consistent data set.

4. The method for generating single event effect data based on a generative adversarial network according to claim 3, characterized in that: In S1, the parameters of the NMOS device structure model include: channel length 130nm, width 420nm, gate oxide thickness 2.58nm, source and drain doping concentration 1×10 20 cm -3 , channel doping density 1.57×10 17 cm -3 , the gate voltage is fixed at 0V, and the drain voltage range is 0-1.8V.

5. The method for generating single event effect data based on a generative adversarial network according to claim 4, characterized in that: In S2, the generator and discriminator of the CGAN are both composed of a 6-layer fully connected network, the latent space noise is a Gaussian distributed random variable, and the conditional information of the generator and discriminator includes LET value, incident position coordinates and incident angle.

6. The method for generating single event effect data based on a generative adversarial network according to claim 5, characterized in that: In S2, the model optimizer uses the Adam optimizer, the learning rate is set to 0.00002, and the network training Epoch parameter is set to 50000.

7. The method for generating single event effect data based on a generative adversarial network according to claim 3, characterized in that: The activation function of the generator uses ReLU, the activation function of the last layer of the discriminator uses Sigmoid, and the other layers use LeakyReLU.

8. The method for generating single event effect data based on a generative adversarial network according to claim 4, characterized in that: The single particle effect model adopts a composite model, which includes a Shockley-Read-Hall model, an effective state density model, a Fermi-Dirac statistics, an impact ionization model and a Lucent model.

Images