Single-Particle Effect Data Generation Method Based on Generative Adversarial Networks
By combining generative adversarial networks with TCAD simulation, the problem of insufficient single-event effect data was solved, generating high-quality single-event effect data, reducing simulation costs and improving the representativeness and accuracy of the data.
Patent Information
- Application Number
- CN202510218116.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-26
- Publication Date
- 2025-12-02
- Estimated Expiration
- 2045-02-26
AI Technical Summary
In existing technologies, single-event effect simulation is costly and time-consuming to acquire data, which limits the accuracy and generalization ability of machine learning modeling and makes it difficult to meet the modeling requirements of large-scale integrated circuits.
A generative adversarial network (CGAN) combined with TCAD simulation was used to generate single-event effect data through conditional generative adversarial network training. A large amount of high-quality single-event transient current data was generated using a small amount of real data, including the construction of an NMOS device structure model, segmented sampling, and conditional generative adversarial network training.
It achieves efficient generation of single-event effect data consistent with the distribution of real data, reduces simulation time costs, and improves the data sampling characterization and the quality of generated data.
Smart Images
Figure CN120046447B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of machine learning technology, specifically relating to a method for generating single-particle effect data based on generative adversarial networks. Background Technology
[0002] Currently, the main methods for analyzing and evaluating single-event effects (SEE) in aerospace integrated circuits include ground-based simulation experiments and computer simulations. Since ground-based radiation simulation experiments struggle to capture the internal response of circuits, SEE simulation technology is more suitable for analyzing and understanding the impact of SEE on aerospace integrated circuits and the effectiveness of radiation-hardened designs. Among these methods, the SEE model, as a core factor affecting the accuracy of SEE simulations, has attracted significant attention and research from numerous researchers.
[0003] In existing technologies, traditional single-event effect simulation modeling methods typically employ device simulation techniques based on physical mechanisms. While these methods can deeply analyze the transient response inside the circuit after a particle impact, their high computational complexity and slow simulation speed make them unsuitable for modeling large-scale integrated circuits. Therefore, numerical fitting methods are widely used, extracting patterns from experimental data to achieve rapid modeling at a lower computational cost. However, these methods still have significant limitations in terms of modeling complexity and accuracy, failing to fully reflect the true response of the circuit.
[0004] In recent years, machine learning-based modeling techniques have gradually become the mainstream research direction for single-event effects due to their powerful pattern recognition capabilities and good predictive performance. Machine learning techniques obtain more accurate single-event effect models through learning from large amounts of data.
[0005] However, machine learning technology initially relied heavily on high-quality and large-scale single-event effect datasets. The high cost and long time required to acquire single-event effect data resulted in insufficient data, which became the main bottleneck limiting modeling accuracy and generalization ability, and also hindered its widespread 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 address the aforementioned problems in existing technologies, this invention provides a method for generating single-event effect data based on generative adversarial networks, thus solving the current problem of insufficient data in single-event effect modeling based on machine learning.
[0008] The objective of this invention can be achieved through the following technical solution: a method for generating single-particle effect data based on generative adversarial networks, comprising the following steps:
[0009] S1: Construct a training dataset based on TCAD simulation:
[0010] Establish and calibrate the NMOS device structure model.
[0011] A single-event effect model was constructed using the TCAD heavy ion model, with the LET value of heavy ions set to 4-100 MeV·cm. 2 / mg, incident angle of 0-90°, bombardment location is NMOS drain, and the recombination model parameters are calibrated;
[0012] The single-particle transient current is segmented based on the rate of change of data. Sampling points are allocated according to the characteristics of the single-particle transient current in a preset ratio. Training dataset is generated by combining the gradient rate of change and interpolation.
[0013] S2: Data generation based on conditional generative adversarial networks:
[0014] A CGAN model is constructed, with LET value, incident position and angle as conditional information and latent space noise as input, to generate single-particle transient current data.
[0015] The generator and discriminator were trained using 16 sets of transient current data of particles at different heavy ion LET values, incident angles and incident positions obtained by TCAD simulation as the training set.
[0016] Preferably, after S2, S3 is also included: verifying the accuracy and efficiency of the generated data using random comparison, wherein the random comparison verification includes evaluating the consistency between the generated data and the real data distribution through KL divergence and JS divergence, and statistically analyzing the generation efficiency.
[0017] Preferably, in S1, segmented sampling of the single-particle transient current includes:
[0018] According to [T0, T max ], [T max T 0.1max ], [T 0.1max T 0.01max The transient interruption was divided into three intervals, with sampling points allocated at 40%, 40%, and 20% respectively.
[0019] Where T0 is the ion bombardment time, T max When the transient current reaches its maximum value, T 0.1max T is the moment when the transient current decays to 10% of its maximum value. 0.01max The moment when the transient current is less than 1% of the maximum current value;
[0020] Divide each interval into N / 2 sub-intervals evenly, and dynamically allocate sampling points according to the gradient change rate to ensure that each sub-interval contains at least 1 sampling point;
[0021] A consistent dataset is generated by correcting the number of sampling points through interpolation.
[0022] Preferably, in S1, the parameters of the NMOS device structure model include: channel length 130nm, width 420nm, gate oxide thickness 2.58nm, and source / drain doping concentration 1×10⁻⁶. 20 cm -3 Channel doping density 1.57 × 10 17 cm -3 The gate voltage is fixed at 0V, and the drain voltage ranges from 0 to 1.8V.
[0023] Preferably, in S2, the generator and discriminator of the CGAN are both composed of 6-layer fully connected networks, the latent space noise is a Gaussian distributed random variable, and the condition information of the generator and discriminator includes LET value, incident position coordinates and incident angle.
[0024] Preferably, in S2, the model optimizer is the Adam optimizer, the learning rate is set to 0.00002, and the network training Epoch parameter is set to 50000.
[0025] Preferably, the generator uses ReLU as its activation function, the discriminator uses Sigmoid as its activation function in the last layer, and LeakyReLU is used in the other layers.
[0026] Preferably, the single-particle effect model adopts a composite model, which includes the Shockley–Read–Hall (SRH) model, the density of states model, the Fermi-Dirac statistics, the collisional ionization model, and the Lucent model.
[0027] The beneficial effects of this invention are as follows:
[0028] 1. Better data sampling representation: Compared with the equidistant sampling method, the data obtained by segmented sampling based on the rate of change of data is sampled more finely in the interval of drastic data change, and the sample data obtained can better represent the characteristics of single-particle transient current.
[0029] 2. High-quality data can be efficiently obtained with a small number of training samples: This invention uses a conditional generative adversarial network (GAN) for training. With a small amount of real simulation sample data as input, it determines whether the current data deviates based on given single-particle transient conditions. The resulting data's mean distribution closely approximates the real data distribution. This allows for the generation of a large amount of high-quality data from a small sample, thus saving significant simulation time and costs. Attached Figure Description
[0030] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.
[0031] Figure 1 Flowchart of a single-particle effect data generation method based on generative adversarial networks:
[0032] Figure 2 A schematic diagram of a 3D model of an NMOS device;
[0033] Figure 3 This is a diagram showing the electron density distribution after heavy ion bombardment of the drain region.
[0034] Figure 4 This is a schematic diagram of a piecewise sampling algorithm based on the numerical rate of change.
[0035] Figure 5 This is a schematic diagram of a data generation model based on Conditional Generative Adversarial Network (CGAN).
[0036] Figure 6 This diagram illustrates the changes in the model loss function during the training of a conditional generative adversarial network (CGAN).
[0037] Figure 7 This diagram illustrates a comparison of the time costs of existing TCAD-based modeling and simulation methods with the method proposed in this application.
[0038] Figure 8 A comparison and data distribution diagram of single-particle transient current data generated by TCAD simulation under existing technology and the method proposed in this application are presented.
[0039] Figure 9 This is a probability density histogram of the mean of the real and generated data for ten sets of single-particle transient current data. Detailed Implementation
[0040] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.
[0041] Please see Figures 1-9 This embodiment provides a method for generating single-event effect data based on generative adversarial networks (GANs). Combining TCAD simulation and the CGAN model, it generates large-scale, high-quality single-event transient current data using a small amount of real data. The method includes the following steps:
[0042] S1: Construct a training dataset based on TCAD simulation:
[0043] S11: Establish and calibrate the NMOS device structure model.
[0044] An NMOS 3D model was constructed based on SMIC's 130nm process, as shown in the attached image. Figure 2 As shown. The device has a channel length of 130 nm, a width of 420 nm, a gate oxide thickness of 2.58 nm, and source / drain doping of 10%. 20 cm -3 The channel doping density is 1.57 × 10⁻⁶. 17 cm -3 The device's gate voltage is fixed at 0V, the source and substrate are grounded, and the drain voltage is set to 0-1.8V.
[0045] Then, the SMIC 130nm PDK process design kit is used to calibrate the static characteristics of the device (such as threshold voltage and drain current) to ensure consistency between the model and the actual process.
[0046] S12: Construction and Simulation of Single-Event Effect Model: The heavy-ion bombardment simulation uses the heavy-ion model from TCAD, and employs a combination of Shockley–Read–Hall (SRH), Auger, and direct (orradiative) to construct the single-event effect model.
[0047] Among them, the SRH carrier lifetime exhibits doping-dependent and temperature-dependent characteristics. Other applicable models include the effective density of states model, Fermi-Dirac statistics, collisional ionization, and the Lucent model. The LET for heavy ions is set to 4-100 MeV·cm⁻¹. 2 / mg, the bombardment location is selected as the drain region of the NMOS device. The heavy ion incident trajectory length is set to 10μm, and the heavy ion incident angle is 0-90° (0° is perpendicular to the device surface).
[0048] The simulation experimental group was set up using a three-factor, four-level method, changing the LET value of heavy ions, the incident angle, and the incident position, based on L... 16 (4 3 An orthogonal array can cover all combinations of three factors and four levels in just 16 experiments, thus effectively reducing the number of experiments and ensuring the balance of experimental results.
[0049] Simulation output, extracting the transient current waveform, and the electron density distribution of the device 25 ps after heavy ion bombardment perpendicularly to the drain center of the NOMS are shown in the attached figure. Figure 3 As shown;
[0050] S13: Acquisition of segmented sampling data of single-event transient current based on data change rate:
[0051] To ensure consistency of single-event transient current sampling points used in the training dataset, a piecewise sampling algorithm based on the rate of change of data is proposed, taking into account the characteristics of single-event transient current sources. The flowchart of this algorithm is attached. Figure 4 As shown.
[0052] a: Input the raw data and set the specific requirements for sampling;
[0053] b: Divide the data into several intervals and preprocess the data, such as performing normalization; determine the number of sampling points to be allocated to each interval based on the gradient rate of change of each interval, and calculate the gradient rate of change of each interval.
[0054] c: Sort the intervals according to the gradient rate of change and assign corresponding sampling points;
[0055] d: Check whether the sampling results meet the set sampling requirements. If they do, use the difference method to generate sampling point data and output the sampling point data that meets the requirements.
[0056] If it does not meet the requirements, the sampling point data will be readjusted.
[0057] One method for dividing the data into several intervals is based on the characteristics of single-particle transient current, according to [T0, T...]. max ], [T max T 0.1max ], [T 0.1max T 0.01max The interval is divided into three intervals, where T0 is the ion bombardment time, T... max When the transient current reaches its maximum value, T 0.1max T is the moment when the transient current decays to 10% of its maximum value. 0.01max The moment when the transient current is less than 1% of the maximum current value;
[0058] S2: Data generation based on conditional generative adversarial networks:
[0059] S21: Construct the CGAN model, which includes a generator and a discriminator. The generator's input is the latent space (randomly generated noise) + condition vector (LET value, incident position, incident angle). The network structure is a 6-layer fully connected network with a resolution of 256×512×512×512×256×100. The activation function is ReLU. The output is single-particle transient current data.
[0060] Discriminator: The input consists of real data and single-particle transient current data generated by the generator. The structure is a 6-layer fully connected network with a size of 256×512×512×512×256×1. The activation function of the last layer is Sigmoid, and the activation functions of the other layers are LeakyReLU. The output is the probability of the data being real.
[0061] Using LET value, incident position and angle as conditional information, and latent space noise as input, single-particle transient current data are generated.
[0062] S22: Training generator and discriminator. The model training data uses 16 sets of particle transient current data obtained from TCAD simulation at different heavy ion LET values, incident angles and incident positions as the training set.
[0063] S3: Verify the authenticity of the data:
[0064] The accuracy and efficiency of the generated data were verified by random comparison. The random comparison verification included evaluating the consistency between the generated data and the real data distribution through KL divergence, JS divergence and Wasserstein distance, and statistically analyzing the generation efficiency.
[0065] A schematic diagram of the CGAN data generation model is shown below. Figure 5 As shown.
[0066] This invention achieves the generation of high-quality data closely resembling the real data distribution using a single-particle data generation model composed of a generator and a discriminator. This allows for the acquisition of a large amount of near-realistic data with only a small number of training samples. The invention employs a conditional generative adversarial network (GAN) for training, using a small amount of real simulation sample data as input. Based on given single-particle transient conditions, it determines whether the current data deviates from the real data distribution, ultimately obtaining a mean distribution that closely approximates the real data distribution. This allows for the generation of a large amount of high-quality data from a small sample size, thus saving significant simulation time and costs.
[0067] Because the single-event transient current changes drastically in the first two intervals, in this embodiment, the number of sampling points is allocated in a ratio of 40%, 40%, and 20%, that is:
[0068] Interval 1: [T0, T max ], sampling points accounted for 40%;
[0069] Interval 2: [T max T 0.1max The sampling points accounted for 40%;
[0070] Interval 3: [T 0.1max T 0.01max The sampling points accounted for 20%;
[0071] Based on the dynamic characteristics of each stage, high-density sampling is performed in key areas to ensure the capture of key features such as transient spikes and abrupt slope changes. This allows for differentiated allocation of sampling resources, avoiding oversampling or undersampling caused by uniform sampling. In practice, this ratio can be adjusted according to the data conditions.
[0072] To prevent some sub-intervals from being completely ignored due to excessive gradient differences, in one embodiment, the number of sampling points is first allocated based on the magnitude of the gradient change rate of each sub-interval; then, based on the number of sampling points N in each interval, the sampling interval is evenly divided into N / 2 sub-intervals and the gradient change rate is calculated. Moreover, to avoid a large number of sub-intervals having no sampling points due to excessive differences in gradient change rates, each sub-interval is required to have at least one sampling point.
[0073] This approach results in better data sampling characterization: data obtained through segmented sampling based on the rate of change of data is sampled more finely in the range of drastic data changes, and the resulting sample data can better characterize the features of single-particle transient currents.
[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 is built using PyTorch, and the stability of model training and the accuracy of generated data are ensured by monitoring the loss functions (G_loss and D_loss) of the generator and discriminator. Please refer to the appendix. Figure 6 , Figure 6 This is a schematic diagram of the loss function under the method of this application. As can be seen from the figure, when the Epoch is greater than 15000, the loss functions of the generator and the discriminator remain basically stable, and the accuracy of the proposed single-particle transient current data generation model based on GAN no longer improves.
[0075] Please see Figure 7 , Figure 7 The table compares the time costs of traditional TCAD-based modeling and simulation with the method proposed in this application, taking the single-event transient current modeling for three influencing factors—LET value, incident angle, and incident position—as an example. If 10 different values are selected for each factor as conditions for sample generation, the sample size is 103. If all samples are obtained using TCAD modeling and simulation, it would take 1000-2000 hours to obtain the required sample data. However, using the GAN-based single-event effect data generation technique proposed in this study, only 1%-10% of the samples are needed as training samples for the GAN model to complete the sample data acquisition. Furthermore, the advantages of this method become more significant as the sample size increases.
[0076] Regarding the quality of the generated data, please refer to Figure 8 , Figure 8 To compare the single-particle transient current data obtained from TCAD simulation with the data generated by the GAN-based data generation model proposed in this study, under the condition of randomly selecting four different sets of data, Figure 8In this context, Agl represents the incident angle, with 0° perpendicular to the device surface; LET stands for Linear Energy Transfer, measured in MeV·cm²·mg⁻¹; Loc is the offset of the incident position along the direction from the drain center to the source center, with 0 representing the drain center position of the NMOS device, measured in μm. Figure 8 As can be seen from this, the proposed GAN-based data generation model is consistent with the real data obtained from TCAD simulation.
[0077] To further illustrate the distribution of data generated by the data generation model proposed in this study compared to real data, please refer to [link / reference]. Figure 9 , Figure 9 The probability density histograms of the mean values of the real and generated data were plotted for ten randomly selected sets of single-particle transient current data. Figure 9 As shown, Density is the probability density, which is equal to the number of times the sample appears / (total number of samples × binwidth). binwidth = the data range of the mean / the number of interval divisions. In this paper, the number of interval divisions is set to 30.
[0078] Therefore, by Figure 8 and Figure 9 It is evident that the method proposed in this application generates data that is very close to real data, and the model quality is excellent.
[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. A method for generating single-event effect data based on generative adversarial networks, characterized in that: Includes the following steps: S1: Construct a training dataset based on TCAD simulation, and establish and calibrate the NMOS device structure model. A single-event effect model was constructed by heavy ion bombardment simulation using TCAD. The heavy ion LET value was set to 4-100 MeV·cm² / mg, the incident angle was 0-90°, the bombardment position was the NMOS drain, and the composite model parameters were calibrated. The single-particle effect model is a composite model including the Shockley-Read-Hall model, Auger and direct models. Other applied models include the effective density of states model, Fermi-Dirac statistics, collisional ionization and Lucent model. The single-particle transient current is segmented based on the rate of change of data. Sampling points are allocated according to the characteristics of the single-particle transient current in a preset ratio. Training dataset is generated by combining the gradient rate of change and interpolation. S2: Data generation based on conditional generative adversarial networks: A CGAN model is constructed, with LET value, incident position and angle as conditional information and latent space noise as input, to generate single-particle transient current data. The generator and discriminator were trained. The model training data used 16 sets of particle transient current data at different heavy ion LET values, incident angles and incident positions obtained by TCAD simulation as the training set to generate and judge high-quality data that is close to the distribution of real data. 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 current is divided into three intervals, with sampling points allocated at ratios of 40%, 40%, and 20%. Where T0 is the ion bombardment time, T max When the transient current reaches its maximum value, T 0.1max T is the moment when the transient current decays to 10% of its maximum current value. 0.01max The moment when the transient current is less than 1% of the maximum current value; Divide each interval into N / 2 sub-intervals evenly, and dynamically allocate sampling points according to the gradient change rate to ensure that each sub-interval contains at least 1 sampling point; A consistent dataset is generated by correcting the number of sampling points through interpolation.
2. The method for generating single-event effect data based on generative adversarial networks according to claim 1, characterized in that: S2 is followed by S3: using random comparison to verify the accuracy and efficiency of the generated data. The random comparison verification includes evaluating the consistency between the generated data and the real data distribution through KL divergence and JS divergence, and statistically analyzing the generation efficiency.
3. The method for generating single-event effect data based on generative adversarial networks according to claim 1, characterized in that: In S1, the parameters of the NMOS device structure model include: channel length 130 nm, width 420 nm, gate oxide thickness 2.58 nm, and source / drain doping concentration 1 × 10⁻⁶. 20 cm -3 Channel doping density 1.57 × 10 17 cm -3 The gate voltage is fixed at 0 V, and the drain voltage ranges from 0 to 1.8 V.
4. The method for generating single-event effect data based on generative adversarial networks according to claim 1, characterized in that: In S2, the generator and discriminator of the CGAN are both composed of 6 fully connected networks, the latent space noise is a Gaussian distributed random variable, and the condition information of the generator and discriminator includes LET value, incident position coordinates and incident angle.
5. The method for generating single-event effect data based on generative adversarial networks according to claim 4, 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.
6. The method for generating single-event effect data based on generative adversarial networks according to claim 4, characterized in that: The generator uses ReLU as its activation function, the discriminator uses Sigmoid as its activation function in the last layer, and LeakyReLU in the other layers.