Hotspot discovery method based on lithography process window

Abstract

The application provides a hotspot discovery method based on a photolithography process window, relates to the field of photolithography technology, and comprises a processing system, wherein the processing system comprises a control module, a Metro module, a model simulation module, a Retarget module, a PatternMatch module and a PWA module. The control module is a control main body and performs the following control process: initializing the system, loading chip layout data; sending a measurement point determination instruction to the Metro module; receiving measurement result data of the Metro module and transmitting the measurement result data to the model simulation module; triggering the model simulation module to generate simulation measurement results; transmitting the measurement result data and the simulation measurement results to the PWA module; receiving common process window data and hotspot point data of the PWA module; and if the common process window meets a preset threshold value, terminating the process, wherein the Retarget module generates corrected graphic data by using CycleGAN, and the PatternMatch module searches for approximate graphics by using a convolutional neural network, so that the common process window is effectively expanded.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of photolithography, and particularly to a hotspot discovery method based on a photolithography process window. BACKGROUND

[0002] With the advancement of semiconductor manufacturing technology to advanced processes, photolithography technology, as the core of chip production, directly determines the performance and manufacturing yield of chips. Influenced by factors such as optical resolution, lens difference, and photoresist light acid diffusion, the chip circuit pattern manufactured by photolithography often deviates from the design pattern, and the deviation is too large to cause chip failure. The process window is a key factor affecting the yield of chip manufacturing, which refers to the range of illumination dose, focusing position and other parameters in the manufacturing process that can produce qualified patterns within a certain range, and exceeding the range will result in unqualified patterns and chip failure. Different shapes of circuits have different process windows, and photolithography of a chip needs to ensure that all patterns in the entire layout are within the common process window. Therefore, accurately identifying and correcting hotspot points that cause the common process window to be too small, and optimizing the process window to improve the yield, have become an urgent need in the field of photolithography manufacturing.

[0003] Although the traditional hotspot detection method plays a certain role in identifying failed patterns, it still has significant shortcomings:

[0004] Ignoring the interactive influence of the common process window: the traditional method only analyzes the hotspot pattern on the failed chip, without considering that the pattern is caused to fail due to the limitation of the process window of other patterns. This results in too many hotspot points that need to be corrected, increasing the process complexity and cost, and failing to effectively expand the common process window.

[0005] Low detection efficiency and lack of systematic analysis: the traditional detection relies on manual inspection or simple simulation, lacks systematic analysis of the process window of the entire layout, and is difficult to quickly identify special patterns that have the greatest impact on the common process window, the detection and correction process is time-consuming, and cannot meet the efficient needs of advanced processes.

[0006] These shortcomings limit the application effect of the traditional method in advanced photolithography manufacturing, so a hotspot discovery method based on a photolithography process window is needed to solve the above problems. SUMMARY

[0007] The technical problem solved is:

[0008] In view of the shortcomings of the prior art, the present application provides a hotspot discovery method based on a photolithography process window, which solves the problems in the above background art.

[0009] Technical scheme:

[0010] To achieve the above object, the present application is implemented by the following technical solutions: a hotspot discovery method based on a photolithography process window, comprising a processing system containing a control module, a Metro module, a model simulation module, a Retarget module, a PatternMatch module and a PWA module;

[0011] The control module controls the following control flow: initializing the system, loading chip layout data; sending a measurement point determination instruction to the Metro module; receiving measurement result data from the Metro module and transmitting it to the model simulation module; triggering the model simulation module to generate simulated measurement results; transmitting the measurement result data and the simulated measurement results to the PWA module; receiving common process window data and hotspot point data from the PWA module; if the common process window meets a preset threshold, terminating the flow; otherwise, triggering the Retarget module and the PatternMatch module, updating the layout data, and iteratively analyzing until the threshold is met.

[0012] The Metro module generates measurement result data and transmits it to the control module; the model simulation module generates simulated measurement results and transmits them to the control module; the PWA module contains a PW submodule to generate process window data, a CPW submodule to generate common process window data, and a PWHotspot submodule to generate hotspot point data, which are transmitted to the control module; the Retarget module generates corrected pattern data, and the PatternMatch module generates approximate pattern data, which are transmitted to the control module.

[0013] The discovery method of the processing system comprises the following steps:

[0014] S1: determining a plurality of measurement points on the chip layout, each measurement point corresponding to a key width of a typical pattern;

[0015] S2: determining a process parameter combination, including illumination dose and focus position;

[0016] S3: generating measurement result data and simulated measurement results for each measurement point under each set of process parameters, and fusing the measurement result data and the simulated measurement results;

[0017] S4: generating a process window for each measurement point based on the fused data;

[0018] S5: generating a common process window based on the process windows of all measurement points;

[0019] S6: determining whether the common process window meets a preset threshold; if it does, terminating the flow; if it does not, performing step S7;

[0020] S7: Analyze the measurement points that have the greatest impact on the common process window, generate hotspot point data;

[0021] S8: Generate correction pattern data based on the hotspot point data, search for an approximate pattern based on the correction pattern data, generate approximate pattern data, and update the chip layout data;

[0022] S9: Recalculate the process window based on the updated layout data, repeat steps S5 to S8 until the common process window meets the preset threshold.

[0023] Preferably, the Metro module extracts pattern features of the measurement points through a capsule network to generate a feature vector; processes time series data of process parameters such as illumination dose and focus position through a time series convolution network to generate time series features; adjusts fusion weights according to the distribution density of the measurement points through an adaptive fusion network, fuses the feature vector and the time series features, and generates measurement result data; the measurement result data is transmitted to the control module for use by the model simulation module and the PWA module.

[0024] Preferably, the model simulation module constructs a lithography simulation model through a variational autoencoder to generate simulation measurement results; processes high-frequency and low-frequency noise through a double-scale filter, determines a segmentation threshold according to pixel distribution through adaptive threshold segmentation, and generates denoised simulation measurement results; the simulation measurement results are transmitted to the control module for use by the PWA module.

[0025] Preferably, the PW sub-module of the PWA module processes the measurement result data and the simulation measurement result based on a self-attention mechanism through a long short-term memory network with a Transformer architecture to generate process window data; the CPW sub-module generates common process window data based on lithography process rules and process window data through a Markov logic network; the PWHotspot sub-module generates hotspot point data based on neural network approximation of Q function by analyzing process window data and common process window data through a deep reinforcement learning network; the data is transmitted to the control module.

[0026] Preferably, the Retarget module generates correction pattern data based on hotspot point data through a variant CycleGAN of a generative adversarial network to update the geometric dimensions of the pattern corresponding to the hotspot points; the PatternMatch module extracts a pattern feature vector of the hotspot points through a convolutional neural network, searches for an approximate pattern based on the feature vector, and generates approximate pattern data; the correction pattern data and the approximate pattern data are transmitted to the control module.

[0027] Preferably, the PWA module comprises an adaptive optimization submodule, which prunes redundant parameters of a deep reinforcement learning network through a genetic algorithm to generate a simplified hotspot point analysis model, and migrates knowledge of a complex analysis model to the simplified model through transfer learning to generate a hotspot point analysis model adapted to different process nodes, and hotspot point data of the model is transmitted to the control module.

[0028] Preferably, the control module performs the following steps in the system training stage: pre-training of a Metro module, a model simulation module, a PWA module, a Retarget module and a PatternMatch module to generate initial parameters, joint fine-tuning of parameters of each module through a back propagation algorithm to optimize system performance based on an overall loss function, and generation of trained system parameters through an adaptive learning rate adjustment mechanism according to gradient changes, and the parameters are stored in the control module.

[0029] Preferably, the control module performs the following steps in the system running stage: receiving hotspot point data and common process window data of the PWA module to calculate a manufacturing yield, comparing the yield with a preset target to generate deviation data, adjusting feature extraction parameters of the Metro module, simulation parameters of the model simulation module and analysis parameters of the PWA module based on the deviation data, and generating updated system parameters through incremental learning to fuse new layout data and process parameter data, and the parameters are stored in the control module.

[0030] Preferably, the Metro module generates an adversarial sample containing process parameter fluctuations and layout noise through a generative adversarial network, trains a capsule network and a temporal convolution network to generate a feature extraction model after adversarial training, the model simulation module processes simulation measurement results through random disturbance and data augmentation to train a variational autoencoder to generate a data-enhanced simulation model, and the models are stored in the control module.

[0031] Beneficial effects:

[0032] The application provides a hotspot discovery method based on a lithography process window.

[0033] 1、The technical solution uses a deep reinforcement learning network of a PWHotspot submodule of a PWA module to analyze hotspot points that have the greatest impact on a common process window, and only a small number of key patterns need to be corrected.

[0034] 2、PWA module's PW submodule generates process window data through the Transformer network, and the CPW submodule generates common process window data by using the Markov logic network to accurately identify the limiting effect of special patterns. The system ensures that the window meets the preset threshold through iterative analysis. Overcomes the shortcomings of traditional methods that cannot effectively expand the common process window.

[0035] 3、The scheme adopts an automatic process, the Metro module quickly generates measurement result data through the capsule network, the time sequence convolution network and the adaptive fusion network, and the model simulation module generates efficient simulation data by using the variational autoencoder and the double-scale filter. Compared with traditional manual inspection or simple simulation, the efficiency is greatly improved to meet the efficient needs of advanced processes. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a system framework diagram of the present application;

[0037] Figure 2 is a process flowchart of the present application. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. Specific embodiment one:

[0040] As shown in Figures 1-2 A hotspot discovery method based on a lithography process window includes a processing system, which includes a control module, a Metro module, a model simulation module, a Retarget module, a PatternMatch module and a PWA module.

[0041] The control module is a control subject and performs the following control process: initializing the system, loading chip layout data; sending a measurement point determination instruction to the Metro module; receiving measurement result data from the Metro module and transmitting it to the model simulation module; triggering the model simulation module to generate simulated measurement results; transmitting the measurement result data and the simulated measurement results to the PWA module; receiving common process window data and hotspot point data from the PWA module; if the common process window meets the preset threshold, terminating the process; otherwise, triggering the Retarget module and the PatternMatch module, updating the layout data, and iterating until the threshold is met.

[0042] The Metro module generates measurement result data and transmits it to the control module; the model simulation module generates simulation measurement results and transmits them to the control module; the PWA module includes the PW submodule to generate process window data, the CPW submodule to generate common process window data, and the PWHotspot submodule to generate hotspot point data, all of which are transmitted to the control module; the Retarget module generates corrected graphic data, and the PatternMatch module generates approximate graphic data, all of which are transmitted to the control module.

[0043] The discovery method for processing systems includes the following steps:

[0044] S1: Determine multiple measurement points on the chip layout, with each measurement point corresponding to the critical width of a typical pattern;

[0045] S2: Determine the combination of process parameters, including light dose and focusing position;

[0046] S3: Generate measurement result data and simulation measurement results for each measurement point under each set of process parameters, and merge the measurement result data and simulation measurement results;

[0047] S4: Generate the process window for each measurement point based on the fused data;

[0048] S5: Generate a common process window based on the process windows of all measurement points;

[0049] S6: Determine whether the common process window meets the preset threshold; if it does, terminate the process; if it does not, proceed to step S7.

[0050] S7: Analyze the measurement points that have the greatest impact on the common process window and generate hotspot data;

[0051] S8: Generate corrected graphic data based on hotspot point data, search for approximate graphics based on corrected graphic data, generate approximate graphic data, and update chip layout data;

[0052] S9: Recalculate the process window based on the updated layout data, and repeat steps S5 to S8 until the common process window meets the preset threshold.

[0053] The Metro module extracts graphical features of measurement points through a capsule network to generate feature vectors; it processes time-series data of process parameters such as light dose and focus position through a temporal convolutional network to generate time-series features; and it adjusts the fusion weights according to the distribution density of measurement points through an adaptive fusion network to fuse feature vectors and time-series features to generate measurement result data. The measurement result data is then transmitted to the control module for use by the model simulation module and the PWA module.

[0054] The model simulation module constructs a lithography simulation model through a variational autoencoder and generates simulation measurement results; it processes high-frequency and low-frequency noise through dual-scale filtering and determines the segmentation threshold based on pixel distribution through adaptive threshold segmentation to generate noise-reduced simulation measurement results; the simulation measurement results are transmitted to the control module for use by the PWA module.

[0055] The PW submodule of the PWA module uses a Transformer architecture with a long short-term dependency network and a self-attention mechanism to process measurement result data and simulation measurement results to generate process window data; the CPW submodule uses a Markov logic network to generate common process window data based on lithography process rules and process window data; the PWHotspot submodule uses a deep reinforcement learning network to analyze process window data and common process window data based on neural network approximation of the Q function and generate hotspot point data; the data is transmitted to the control module.

[0056] The Retarget module generates corrected graphic data based on hotspot point data using CycleGAN, a variant of generative adversarial network, and updates the geometric dimensions of the graphic corresponding to the hotspot points. The PatternMatch module extracts feature vectors of the hotspot point graphics using a convolutional neural network, searches for approximate graphics based on the feature vectors, and generates approximate graphic data. The corrected graphic data and approximate graphic data are then transmitted to the control module.

[0057] The PWA module includes an adaptive optimization submodule, which uses a genetic algorithm to prune redundant parameters of the deep reinforcement learning network to generate a simplified hotspot analysis model; it then uses transfer learning to transfer knowledge from the complex analysis model to the simplified model, generating a hotspot analysis model adapted to different process nodes; the hotspot data of the model is then transmitted to the control module.

[0058] The control module performs the following steps during the system training phase: pre-training the Metro module, model simulation module, PWA module, Retarget module, and PatternMatch module to generate initial parameters; jointly fine-tuning the parameters of each module through the backpropagation algorithm to optimize system performance based on the overall loss function; adjusting the learning rate according to gradient changes through an adaptive learning rate adjustment mechanism to generate trained system parameters; and storing the parameters in the control module.

[0059] During system operation, the control module performs the following steps: receiving hotspot data and common process window data from the PWA module, and calculating manufacturing yield; comparing the yield with the preset target to generate deviation data; adjusting the feature extraction parameters of the Metro module, the simulation parameters of the model simulation module, and the analysis parameters of the PWA module based on the deviation data; fusing new layout data and process parameter data through incremental learning to generate updated system parameters; and storing the parameters in the control module.

[0060] The Metro module generates adversarial samples containing process parameter fluctuations and layout noise through a generative adversarial network, trains capsule networks and temporal convolutional networks, and generates a feature extraction model after adversarial training; the model simulation module processes simulation measurement results through random perturbation and data augmentation, trains variational autoencoders, and generates a data-augmented simulation model; the model is stored in the control module.

[0061] When determining measurement points, the Metro module extracts features from the chip layout using a convolutional neural network to generate a layout complexity distribution map. Based on a clustering algorithm, it divides the layout into multiple regions and automatically selects representative measurement points according to the shape complexity and critical width distribution within each region. This generates measurement point data containing location coordinates and critical width values, which is then transmitted to the control module. The control module uses a regression model trained on historical lithography data to predict the initial combination of process parameters for illumination dose and focus position. The Metro module analyzes time-series data of process parameters using a temporal convolutional network to generate parameter fluctuation trends. The control module dynamically adjusts the process parameter combinations based on these trends and layout complexity, generating optimized parameter combination data, which is then transmitted to the model simulation module and the PWA module. Finally, the control module uses a support vector machine algorithm to calculate fusion weights based on the confidence levels of the measurement results and simulation results. It dynamically adjusts these weights according to the lithography equipment's operating status and the density of the layout region's graphics. A weighted average algorithm is then used to fuse the measurement results and simulation results, generating fused data containing critical width deviations and process parameter responses, which is then transmitted to the PWA module. When the PW submodule calculates the process window using the Transformer network, it sets boundary conditions based on lithography process rules, including the range of illumination dose, focus position, and critical width deviation threshold. If the measurement point data exceeds the boundary conditions, an anomaly marker is generated and transmitted to the control module. The CPW submodule then uses a Markov logic network to remove the anomaly markers from the measurement point data, generating common process window data. The control module records the anomaly markers, generates an anomaly report, and transmits it to external storage. The PWHotspot submodule uses a deep reinforcement learning network to calculate the constraint weights of each measurement point on the common process window. It then uses a sorting algorithm to generate a priority list by descending constraint weights. Based on preset corrected resource constraints, it selects the measurement point with the highest priority as the hotspot point, generating hotspot point data containing the measurement point location and constraint weights, which is then transmitted to the control module. After generating corrected graphic data, the Retarget module uses an optical nearest neighbor effect simulation algorithm to verify the manufacturability of the corrected graphic, generating verification results. If the preset manufacturing threshold is met, the results are transmitted to the control module; otherwise, the corrected graphic data is regenerated using CycleGAN. After generating approximate graphic data, the PatternMatch module uses feature vector distance calculation to verify the matching degree, generating matching degree data, which is then transmitted to the control module. During iterative analysis, the control module records the common process window data and hotspot data for each iteration. It calculates the convergence trend of the common process window using the gradient descent algorithm. If the convergence trend is lower than the preset convergence threshold or the number of iterations reaches the preset maximum value, the iteration is terminated. An optimization log containing the number of iterations, common process window data, and hotspot data is generated and transmitted to the external storage unit.The control module verifies the integrity and consistency of data from each module using a verification algorithm. If a data error is detected, an error code is generated, triggering the corresponding module to regenerate the data. The module's operating status, data transmission time, and an operation log containing timestamps, module identifiers, and data summaries of the error codes are recorded and transmitted to external storage. During system initialization, the control module loads a process node configuration file containing the lithography equipment type, process node parameters, and layout complexity thresholds. Based on the configuration file, it adjusts the feature extraction parameters of the Metro module, the simulation model parameters of the model simulation module, and the analysis parameters of the PWA module. The adjusted parameters are stored in the control module. A configuration file update mechanism periodically loads the new process node configuration file to generate updated system parameters. The processing system includes a data storage module that receives chip layout data, measurement result data, simulation measurement results, process window data, hotspot data, corrected graphic data, approximate graphic data, optimization logs, and operation logs transmitted from the control module. It organizes the data through a database management system, generating a data index table for external systems to access via a standard interface. Specific Implementation Example 2:

[0063] like Figures 1-2 As shown, the key algorithm mentioned in Example 1 will be analyzed in detail below, including its core mathematical formulas and explanations:

[0064] Metro module: Capsule network (feature extraction):

[0065] The formula is as follows:

[0066]

[0067]

[0068]

[0069]

[0070] in, For low-level capsule output, For the transformation matrix, For the prediction vector, For dynamic routing coupling coefficients, As an initial logarithmic prior, For high-level capsule infusion, For high-level capsule output.

[0071] The Metro module uses capsule networks to extract graphical features from measurement points. It takes pixel data from the input chip layout as input and generates low-level capsule representations. Through a dynamic routing mechanism, the feature vectors of these low-level capsules are... The vector is passed to a higher-level capsule to generate a feature vector. The feature vector contains key width and shape information of the measurement points and is transmitted to the adaptive fusion network.

[0072] The formula for temporal convolutional networks (time series processing) is as follows:

[0073]

[0074] in, To input the time series of process parameters, For convolution kernel, For the size of the kernel, Void factor, For weights, For bias, To output time series features.

[0075] The Metro module processes time-series data of light dose and focus position. Input process parameter sequence. Time series features are generated through causal convolution and dilated convolution. Capture long-term dependencies of parameters; features It is transmitted to an adaptive fusion network and fused with graphic features.

[0076] Metro Module: Adaptive Fusion Network (Feature Fusion):

[0077] The formula is as follows:

[0078]

[0079] in, The graphical feature vector of the capsule network, For time-series features of temporal convolutional networks, The weight matrix, For bias, The sigmoid activation function is used. To integrate weights, This is a feature of fusion.

[0080] Metro module integrates graphical features and time series features Adjust the weights based on the distribution density of the measurement points. Generate fusion features The measurement results data are transmitted to the control module for use by the model simulation module and the PWA module.

[0081] Model simulation module: Variational autoencoder (lithography simulation):

[0082] The formula is as follows:

[0083] in, To input measurement point and process parameter data, For latent variables, For encoder distribution, For decoder distribution, For the mean and variance, Let KL divergence be the KL divergence. This is the loss function.

[0084] The model simulation module constructs a lithography simulation model. Inputting measurement data, the encoder generates the latent distribution. The decoder generates simulation measurement results. By minimizing the loss The model is optimized, and the noise-reduced simulation measurement results are generated and transmitted to the control module.

[0085] PWA module (PW submodule):

[0086] Transformer Long Short-Term Dependency Network (Process Window Calculation):

[0087] The formula is as follows:

[0088]

[0089] in, To input measurement results and simulation measurement results, For querying key-value matrices, This is the weight matrix. For the key dimension, Attention is the attention output.

[0090] PW submodule calculation process window. Input fused data. It generates attention output through a self-attention mechanism to capture the dependence of measurement points under different process parameters; it outputs process window data, including the acceptable range of light dose and focus position, and transmits it to the CPW submodule.

[0091] PWA module (CPW submodule):

[0092] Markov logic network (common process window computation):

[0093] The formula is as follows:

[0094]

[0095] in, For process window data, For logical rule characteristic functions, As weight, The normalization constant is It is a joint probability distribution.

[0096] The CPW submodule calculates the common process window. It inputs the process window data and defines characteristic functions based on lithography process rules (such as dose range constraints). By maximizing the joint probability Generate common process window data and transmit it to the control module and the PWHotspot submodule.

[0097] PWA module (Adaptive Optimization Submodule):

[0098] Transfer learning (model adaptation):

[0099] The formula is as follows:

[0100]

[0101] in, Output as the target. For model prediction, These are the current model weights. For pre-trained model weights, The regularization coefficient is . This is the loss function.

[0102] The adaptive optimization submodule adapts to different process nodes. It inputs weights from a complex analysis model. By minimizing Optimize and simplify model weights Generate an adapted model and output hotspot data. Specific Implementation Example 3:

[0104] like Figures 1-2 As shown, the following describes the specific application logic steps of each module and algorithm in the hotspot discovery method based on the photolithography process window:

[0105] Control module:

[0106] The control module coordinates system operation through a scheduling algorithm, loads chip layout data and process parameter configurations, generates initialization parameters, and transmits them to the Metro module; it sends measurement point determination instructions to the Metro module, receives measurement result data generated by the Metro module, and transmits it to the model simulation module; it triggers the model simulation module to generate simulation measurement results and receives the simulation measurement results; it calculates the fusion weights of the measurement result data and simulation measurement results using a support vector machine algorithm, adjusts the weights based on the lithography equipment operating status and layout pattern density, fuses the data using a weighted average algorithm, generates fused data, and transmits it to the PWA module; it receives common process window data and hotspot point data from the PWA module, determines whether the common process window meets a preset threshold; if it does, it generates a termination signal and stores the final layout data in the data storage module; if not, it terminates the process window if the common process window does not meet a preset threshold. The system initiates several steps: First, it triggers the Retarget module to generate corrected graphical data and the PatternMatch module to generate approximate graphical data, updating the chip layout data. Second, it calculates the convergence trend of the common process window using a gradient descent algorithm. If the trend falls below a preset convergence threshold or the number of iterations reaches its maximum, the iteration terminates, generating an optimization log containing the iteration count, common process window data, and hotspot point data, which is then transmitted to the data storage module. Third, it verifies the integrity of data from each module using a validation algorithm. If an error is detected, an error code is generated, triggering the corresponding module to regenerate data, recording the running status, data transmission time, and error code, generating a running log, and transmitting it to the data storage module. Fourth, it loads the process node configuration file during initialization, adjusts the parameters of each module, periodically updates the configuration file, generates updated system parameters, and stores them in the control module.

[0107] Metro module:

[0108] The Metro module processes chip layout image data using a convolutional neural network to extract layout complexity features and generate a complexity distribution map. It then divides the layout into multiple regions using a clustering algorithm, selecting representative measurement points based on the shape complexity and critical width distribution within each region to generate measurement point data containing location coordinates and critical width values. Next, it processes the measurement point image data using a capsule network to extract low-level edge features and uses a dynamic routing mechanism to pass these features to higher-level capsules, generating feature vectors representing critical width and shape. Finally, it processes time-series data of process parameters such as illumination dose and focus position using a temporal convolutional network, employing causal convolution to ensure reliance on historical data. The system uses dilated convolutions to capture long-term fluctuation trends and generate time-series features. An adaptive fusion network calculates fusion weights based on the distribution density of measurement points, fusing feature vectors and time-series features to generate measurement result data containing key width deviations and process parameter responses. A regression model predicts initial process parameter combinations based on historical lithography data, generating parameter combination data. A generative adversarial network generates adversarial examples containing process parameter fluctuations and layout noise, training capsule networks and temporal convolutional networks to generate an adversarially trained feature extraction model. The measurement result data, measurement point data, and parameter combination data are then transmitted to the control module for use by the model simulation module and the PWA module.

[0109] Model simulation module:

[0110] The model simulation module processes measurement result data and process parameter combination data through a variational autoencoder to construct a lithography simulation model. The encoder learns the potential distribution of optical nearest neighbor effect and photoacid diffusion, while the decoder generates a simulation image, optimizes reconstruction error and KL divergence, and generates simulation measurement results containing critical width deviations. The simulation image is then processed through dual-scale filtering, using a high-frequency Gaussian kernel to preserve image edges and a low-frequency Gaussian kernel to smooth noise. The high-frequency and low-frequency weights are dynamically adjusted according to pixel distribution to generate a denoised simulation image. The denoised image is then processed through adaptive threshold segmentation, calculating the segmentation threshold based on the maximum inter-class variance to separate the image region from the background, generating segmentation data containing critical width deviations. The simulation measurement results are processed through random perturbation and data augmentation to train the variational autoencoder and generate a data-augmented simulation model. The simulation measurement results and segmentation data are then transmitted to the control module for use by the PWA module.

[0111] PWA Module - PW Submodule:

[0112] The PW submodule processes measurement result data and simulation measurement results through a long short-term dependency network based on the Transformer architecture. It calculates the dependency relationship of measurement points under different illumination doses and focus positions based on a self-attention mechanism, and uses multi-head attention to fuse multi-dimensional features to generate process window data that includes process window boundaries. It sets boundary conditions for illumination dose range, focus position range, and critical width deviation threshold based on lithography process rules. If the measurement point data exceeds the boundary conditions, an anomaly marker is generated and transmitted to the control module. Finally, the process window data and anomaly markers are transmitted to the CPW submodule.

[0113] PWA module - CPW submodule:

[0114] The CPW submodule processes the process window data through a Markov logic network, defines logical constraints based on lithography process rules, calculates the process window intersection probability of all measurement points, removes measurement point data marked as abnormal, and generates common process window data; the common process window data is then transmitted to the PWHotspot submodule and the control module.

[0115] PWA Module - PWHotspot Submodule:

[0116] The PWHotspot submodule processes process window data and common process window data using a deep reinforcement learning network. It defines the state as the constraint weight of the measurement point, the action as selecting a measurement point, and the reward as the increase in the area of ​​the common process window. It approximates the Q-function through a neural network to calculate the constraint weight of each measurement point to the common process window. A quicksort algorithm is used to sort the measurement points in descending order of constraint weight, generating a priority list. Based on preset modified resource constraints, the highest priority measurement point is selected, generating hotspot data containing the measurement point's location and constraint weight. A genetic algorithm is used to prune redundant parameters of the deep reinforcement learning network, generating a simplified hotspot analysis model. Transfer learning is used to transfer knowledge from the complex analysis model to the simplified model, generating a hotspot analysis model adapted to different process nodes. Finally, the hotspot data is transmitted to the control module.

[0117] Retarget module:

[0118] The Retarget module processes hotspot data using CycleGAN, a variant of generative adversarial networks. The generator maps the hotspot pattern to a manufacturable pattern domain, the discriminator verifies the authenticity of the corrected pattern, optimizes the adversarial loss and cycle consistency loss, and generates corrected pattern data with updated pattern geometry. The light intensity distribution of the corrected pattern is calculated using an optical nearest neighbor effect simulation algorithm, and the deviation from the manufacturability threshold is compared to generate a verification result. If the threshold is met, the corrected pattern data is transmitted to the control module; otherwise, the corrected pattern data is regenerated. The corrected pattern data is then transmitted to both the control module and the PatternMatch module.

[0119] PatternMatch module:

[0120] The PatternMatch module processes hotspot point graphics and chip layout data through a convolutional neural network, extracts edge and shape features of the graphics, and generates feature vectors for the hotspot point graphics. Based on the feature vectors, it calculates the distance between the graphics in the layout, searches for approximate graphics, and generates approximate graphic data containing the positions of the approximate graphics. It verifies the matching degree of the approximate graphics by calculating the distance between the feature vectors and generates matching degree data. Finally, it transmits the approximate graphic data and matching degree data to the control module.

[0121] Data storage module:

[0122] The data storage module receives chip layout data, measurement result data, simulation measurement results, process window data, hotspot data, correction graphic data, approximate graphic data, optimization logs, operation logs, and anomaly reports transmitted by the control module. It organizes the data through the database management system, generates a data index table, and stores the data to local or cloud storage units for external systems to access through standard interfaces. Specific Implementation Example 4:

[0124] like Figures 1-2 As shown, the following is a detailed description of the hardware composition and hardware specifications of each module in Embodiment 1:

[0125] Control module:

[0126] The hardware components include a central processing unit (CPU), random access memory (RAM), read-only memory (ROM), solid-state memory (SSD), input / output interfaces, a clock module, and a network interface module. The CPU employs a multi-core architecture, executing system scheduling algorithms and handling initialization, instruction sending, data receiving, fusion, judgment, and iteration logic. It runs a support vector machine algorithm to calculate fusion weights, a gradient descent algorithm to determine iteration termination, and a verification algorithm to verify data integrity. RAM caches chip layout data, measurement results, simulation measurement results, common process window data, hotspot data, corrected graphic data, and approximate graphic data, supporting high-speed data processing. ROM stores the system startup program and process node configuration files, including lithography equipment type, process node parameters, and layout complexity thresholds. SSD stores operation logs, optimization logs, anomaly reports, and system parameters for system maintenance and updates. Input / output interfaces connect to the Metro module, model simulation module, PWA module, Retarget module, PatternMatch module, and data storage module, transmitting instructions and data via a high-speed bus. The clock module provides a high-precision clock signal to synchronize data transmission and algorithm execution between modules. The network interface module supports the Ethernet protocol, connecting external storage units and lithography equipment to transmit configuration files and log data. The control module receives measurement result data from the Metro module, transmits it to the model simulation module, fuses the simulation measurement results, and then transmits them to the PWA module. It receives common process window data and hotspot point data from the PWA module, triggers the Retarget module and PatternMatch module to update the layout data, and stores the final data to the data storage module.

[0127] Metro module:

[0128] The hardware components include a graphics processing unit (GPU), a field-programmable gate array (FPGA), random access memory (RAM), flash memory, and a data acquisition interface. The GPU runs a convolutional neural network to extract layout complexity features and generate a complexity distribution map; it runs a capsule network to extract graphical features of measurement points and generate feature vectors; it runs a sequential convolutional network to process time-series data of process parameters and generate time-series features; it runs an adaptive fusion network to fuse feature vectors and time-series features to generate measurement result data; it runs a regression model to predict initial process parameter combinations; and it runs a generative adversarial network to generate adversarial examples. The FPGA accelerates clustering algorithms, dividing the layout into regions, selecting representative measurement points, and generating measurement point data. The RAM caches layout image data, process parameter sequences, feature vectors, time-series features, and measurement result data, supporting high-throughput computation. Flash memory stores pre-trained neural network model parameters and adversarial training model parameters, ensuring fast loading. The data acquisition interface connects to the lithography equipment sensors to collect light dose and focus position data, which is then transmitted to the sequential convolutional network. The Metro module transmits measurement point data, measurement result data, and parameter combination data to the model simulation module and PWA module via the high-speed bus of the control module, supporting measurement point determination and data generation in steps S1-S3.

[0129] Model simulation module:

[0130] The hardware components include a graphics processing unit (GPU), an application-specific integrated circuit (ASIC), random access memory (RAM), solid-state memory (SSD), and an image processing interface. The GPU runs a variational autoencoder (VAE) to build a lithography simulation model, generates simulation measurement results, performs dual-scale filtering to process high-frequency and low-frequency noise, performs adaptive threshold segmentation to separate the graphic region from the background, and trains the simulation model using random perturbation and data augmentation. The ASIC optimizes the encoding and decoding calculations of the VAE, accelerating the lithography process simulation and simulating optical proximity effects and photo-acid diffusion. The RAM caches measurement result data, process parameter combination data, simulation images, and segmentation data, supporting real-time image processing. The SSD stores simulation model parameters and data augmentation model parameters, ensuring rapid model deployment. The image processing interface receives measurement result data from the Metro module and outputs denoised simulation measurement results. The model simulation module transmits the simulation measurement results and segmentation data to the PWA module via the control module, supporting the generation of simulation data in step S3.

[0131] PWA Module - PW Submodule:

[0132] The hardware components include a graphics processing unit (GPU), random access memory (RAM), and a cache. The GPU runs a long short-term dependency network based on the Transformer architecture, processing measurement and simulation results based on a self-attention mechanism to generate process window data, perform boundary condition checks, and generate anomaly markers. The RAM caches input data and intermediate features, storing boundary conditions for lithography process rules, including illumination dose range, focus position range, and critical width deviation thresholds. The cache accelerates self-attention calculations and improves the efficiency of multi-head attention fusion. The PW submodule transfers the process window data and anomaly markers to the CPW submodule, supporting process window generation in step S4.

[0133] PWA module - CPW submodule:

[0134] The hardware components include a central processing unit (CPU), random access memory (RAM), and non-volatile memory. The CPU runs a Markov logic network, calculates the intersection probability of common process windows based on lithography process rules and process window data, removes measurement point data marked as abnormal, and generates common process window data. The RAM caches the process window data and logic rule weights, and stores intermediate results of the intersection calculation. The non-volatile memory stores predefined lithography process rules, ensuring fast rule loading. The CPW submodule transmits the common process window data to the PWHotspot submodule and the control module, supporting the common process window generation in step S5.

[0135] PWA Module - PWHotspot Submodule:

[0136] The hardware components include a graphics processing unit (GPU), a field-programmable gate array (FPGA), random access memory (RAM), and solid-state memory (SSD). The GPU runs a deep reinforcement learning network, calculates the constraint weights of measurement points on a common process window, generates hotspot data, runs a genetic algorithm to prune network parameters, and runs transfer learning to optimize model adaptation. The FPGA accelerates a fast sorting algorithm, sorting measurement points in descending order of constraint weights, generating a priority list, and selecting the highest-priority hotspot. The RAM caches process window data, common process window data, and the priority list, storing constraint weights and the reward function. The SSD stores a simplified hotspot analysis model and model parameters adapted to different process nodes. The PWHotspot submodule transmits hotspot data to the Retarget module via the control module, supporting hotspot analysis in step S7.

[0137] Retarget module:

[0138] The hardware components include a graphics processing unit (GPU), random access memory (RAM), solid-state memory (SSD), and a simulation interface. The GPU runs a variant of the generative adversarial network (GAN), CycleGAN, to generate corrected image data, optimize adversarial loss and cycle consistency loss, and runs an optical nearest neighbor effect simulation algorithm to verify the manufacturability of the corrected image, generating verification results. The RAM caches hotspot data, corrected image data, and light intensity distribution data, supporting high-frequency generation and verification. The SSD stores the generator and discriminator model parameters, ensuring fast loading. The simulation interface connects to lithography simulation software, inputting a mask pattern and outputting the light intensity distribution. The Retarget module transmits the verified corrected image data to the control module and the PatternMatch module, supporting image correction in step S8.

[0139] PatternMatch module:

[0140] The hardware components include a graphics processor (GPU), random access memory (RAM), and a high-speed search interface. The GPU runs a convolutional neural network to extract feature vectors from hotspot points, calculates the distance between feature vectors of the graphics in the layout, generates approximate graphic data, verifies the matching degree through distance calculation, and generates matching degree data. The RAM caches chip layout data, feature vectors, and approximate graphic positions, supporting fast searching. The high-speed search interface accelerates feature vector matching, connects to the layout database, and accepts input layout data. The PatternMatch module transmits the approximate graphic data and matching degree data to the control module, supporting the approximate graphic search in step S8.

[0141] Data storage module:

[0142] The hardware components include solid-state storage (SSD), a database processor, a network interface, and a redundant storage array. The SSD stores chip layout data, measurement results, simulation measurement results, process window data, hotspot data, corrected graphical data, approximate graphical data, optimization logs, operation logs, and anomaly reports, providing high-capacity storage. The database processor runs the database management system, generates data index tables, organizes data structures, and supports fast queries. The network interface connects to external systems via Ethernet, transmits data index tables, and responds to standard interface access requests. The redundant storage array backs up all data, employing RAID configuration to ensure data reliability. The data storage module receives data transmitted from the control module and stores it locally or in the cloud for system maintenance, debugging, and external integration. Specific Implementation Example 5:

[0144] like Figures 1-2 As shown, the following are specific use cases of the entire technical solution:

[0145] Scenario 1: 7nm SRAM layout (EUV process):

[0146] Background: A semiconductor factory is using an ASML Twinscan NXE:3400B EUV lithography machine (NA=0.33, wavelength 13.5nm) to produce 7nm FinFET process SRAM chips with a layout size of 100μm × 100μm and a target yield of ≥95%.

[0147] Operation: The control module initializes, loading the SRAM layout (GDSII format) and the 7nm process configuration file (7nm linewidth, dose 20-30 mJ / cm²). The Metro module selects 1000 measurement points (critical width 7nm) using a convolutional neural network and K-means clustering to generate measurement point data. The control module predicts 25 sets of process parameters (5 doses × 5 focuses), and the Metro module analyzes parameter fluctuations using a temporal convolutional network to generate optimized parameter combinations. The Metro module extracts graphic features (128-dimensional) using a capsule network, generates time-series features (64-dimensional) using a temporal convolutional network, and fuses the data using an adaptive fusion network (0.7 weight for dense regions) to generate measurement result data. The model simulation module simulates lithography using a variational autoencoder to generate simulated measurement results, performs dual-scale filtering (3×3, 5×5 kernels) for noise reduction, and adaptive threshold segmentation to generate the final simulation data. The control module fuses measurement and simulation data using a support vector machine (weights 0.6:0.4) to generate fused data. The PWA module's PW submodule uses a Transformer to calculate 1000 process windows (boundary: deviation ±1nm), removing 10 outliers. The CPW submodule uses a Markov logic network to generate a common process window (initial area 480 mJ·nm). The control module determines that the threshold (500 mJ·nm) is not met, triggering the PWHotspot submodule, which uses deep reinforcement learning to select 30 hotspot points (weight > 0.8). The Retarget module uses CycleGAN to adjust the linewidth (7nm to 7.5nm) and verify manufacturability (deviation < 0.5nm). The PatternMatch module uses a convolutional neural network to search for 50 approximate patterns (matching degree 0.95). The control module updates the layout, iterating twice, until the common process window reaches 576 mJ·nm, meeting the threshold. The log records the iteration count and hotspot points, and the runtime log records errors (5 data retries), storing them in the data storage module.

[0148] Results: Yield increased from 90% to 96%, runtime was 30 minutes, and the data index table supports MES system integration.

[0149] Scenario 2: 5nm logic cell (EUV process):

[0150] Background: The factory uses an ASML NXE:3600D EUV lithography machine (NA=0.55) to produce logic chips using the 5nm FinFET process, with a layout size of 50μm × 50μm and a target yield of ≥97%.

[0151] Operation: The control module loads the 5nm process configuration file (5nm linewidth, dose 15-25 mJ / cm²) and initializes the logic layout. The Metro module selects 800 measurement points (critical width 5nm) and generates measurement point data. The control module predicts 16 sets of process parameters (4 doses × 4 focuses), and the Metro module optimizes the parameters. The Metro module extracts graphical features (256 dimensions) and time-series features (128 dimensions), and adaptively fuses them to generate measurement result data (sparse region weight 0.4). The model simulation module generates simulation data using a variational autoencoder, performs dual-scale filtering for noise reduction, and adaptive threshold segmentation. The control module fuses the data (weight 0.5:0.5). The PW submodule calculates 800 process windows (boundary: deviation ±0.8nm) and removes 8 outliers. The CPW submodule generates a common process window (initial area 320 mJ·nm). The control module determines that the threshold (350 mJ·nm) is not met. The PWHotspot submodule selects 20 hotspot points (weight >0.85). The Retarget module adjusted the linewidth (from 5nm to 5.3nm) and added auxiliary patterns, which passed verification. The PatternMatch module searched for 40 approximate patterns (matching degree 0.96). The control module updated the layout, iterated once, and the common process window reached 360 mJ·nm, meeting the threshold. Optimization logs and runtime logs (with 3 error retries) are stored in the data storage module, which is accessed by MES via API.

[0152] Results: Yield increased from 92% to 97.5%, runtime was 20 minutes, and it was adapted to high-density patterns in 5nm process.

[0153] The following are the specific experimental data:

[0154] Scenario 7nm SRAM (EUV) 5nm Logic (EUV) Layout size (pm x pm) 100 × 100 50 × 50 Critical width (nm) 7 5 Measurement point number 1000 800 Process parameter combination 25 (5 dose x 5 focus) 16 (4 dose x 4 focus) Dose range (mJ / cm2) 20-30 15-25 Focus range (nm) ±50 ±40 Abnormal point number 10 8 Initial common process window area (mJ.nm) 480 320 Threshold process window area (mJ.nm) 500 350 Hotspot point number 30 20 Approximate pattern number 50 40 Iteration number 2 1 Final common process window area (mJ.nm) 576 360 Initial yield (%) 90 92 Final yield (%) 96 97.5 Yield improvement (%) 6 5.5 Run time (minutes) 30 20 Error retry number 5 3 Data storage amount (MB) 150 100

[0155] The table above demonstrates that the hotspot discovery method based on the photolithography process window is effective and practical. Key conclusion:

[0156] Process window optimization: The area of ​​the common process window is increased by 12.5%-20%, resolving hotspot point limitations.

[0157] High efficiency and adaptability: The system is compatible with 7nm to 5nm processes, with a runtime of 20-30 minutes and 1-2 iterations, resulting in high efficiency.

[0158] Precise correction: accurate correction of hotspot points (20-30) and approximate figures (40-50), with a matching degree of 0.95-0.98.

[0159] Reliable and stable: Few outliers (1%-1.25%), 3-5 retries for errors, and efficient data management.

[0160] This technical solution significantly improves yield through accurate hotspot identification and correction, is compatible with advanced processes, operates efficiently and reliably, and is suitable for industrial applications.

[0161] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0162] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hotspot discovery method based on a photolithography process window, implemented by a processing system, characterized in that: The processing system includes a control module, a Metro module, a model simulation module, a Retarget module, a PatternMatch module, and a PWA module; The control module executes the following control flow for the control body: initializes the system and loads chip layout data; sends a measurement point determination command to the Metro module; receives measurement result data from the Metro module and transmits it to the model simulation module; triggers the model simulation module to generate simulation measurement results; and transmits the measurement result data and simulation measurement results to the PWA module. Receive common process window data and hotspot point data from the PWA module; if the common process window meets the preset threshold, terminate the process; otherwise, trigger the Retarget module and PatternMatch module to update the layout data and iteratively analyze until the threshold is met. The Metro module generates measurement result data and transmits it to the control module; the model simulation module generates simulation measurement results and transmits them to the control module; the PWA module includes a PW submodule that generates process window data, a CPW submodule that generates common process window data, and a PWHotspot submodule that generates hotspot point data, all of which are transmitted to the control module; the Retarget module generates corrected graphic data, and the PatternMatch module generates approximate graphic data, all of which are transmitted to the control module. The discovery method of the processing system includes the following steps: S1: Determine multiple measurement points on the chip layout, with each measurement point corresponding to the critical width of a typical pattern; S2: Determine the combination of process parameters, including light dose and focusing position; S3: Generate measurement result data and simulation measurement results for each measurement point under each set of process parameters, and merge the measurement result data and simulation measurement results; S4: Generate the process window for each measurement point based on the fused data; S5: Generate a common process window based on the process windows of all measurement points; S6: Determine whether the common process window meets the preset threshold; if it does, terminate the process; if it does not, proceed to step S7. S7: Analyze the measurement points that have the greatest impact on the common process window and generate hotspot data; S8: Generate corrected graphic data based on hotspot point data, search for approximate graphics based on corrected graphic data, generate approximate graphic data, and update chip layout data; S9: Steps S7 and S8 involve the PWHotspot submodule of the PWA module using a deep reinforcement learning network to systematically analyze the hotspot points that have the greatest impact on the common process window; the Retarget module using CycleGAN to generate corrected graphic data; the PatternMatch module using a convolutional neural network to search for approximate graphics; the PW submodule of the PWA module using a Transformer network to generate process window data; and the CPW submodule using a Markov logic network to generate common process window data. Based on the updated layout data, the process window is recalculated, and steps S5 to S8 are repeated until the common process window meets the preset threshold.

2. The hotspot discovery method based on photolithography process window according to claim 1, characterized in that: The Metro module extracts the graphic features of the measurement points through a capsule network to generate feature vectors; it processes the time-series data of process parameters such as light dose and focus position through a temporal convolutional network to generate time-series features; and it adjusts the fusion weights according to the distribution density of the measurement points through an adaptive fusion network to fuse the feature vectors and time-series features to generate measurement result data. The measurement result data is transmitted to the control module for use by the model simulation module and the PWA module.

3. The hotspot discovery method based on photolithography process window according to claim 2, characterized in that: The model simulation module constructs a lithography simulation model through a variational autoencoder and generates simulation measurement results; it processes high-frequency and low-frequency noise through dual-scale filtering and determines the segmentation threshold based on pixel distribution through adaptive threshold segmentation to generate noise-reduced simulation measurement results; the simulation measurement results are transmitted to the control module for use by the PWA module.

4. The hotspot discovery method based on photolithography process window according to claim 3, characterized in that: The PW submodule of the PWA module uses a long short-term dependency network based on the Transformer architecture to process measurement result data and simulation measurement results based on a self-attention mechanism, and generates process window data. The CPW submodule generates common process window data based on lithography process rules and process window data using a Markov logic network; the PWHotspot submodule generates hotspot data by analyzing process window data and common process window data using a deep reinforcement learning network and approximating the Q function through a neural network; the data is then transmitted to the control module.

5. The hotspot discovery method based on photolithography process window according to claim 4, characterized in that: The Retarget module generates corrected graphic data based on hotspot point data by using CycleGAN, a variant of the generative adversarial network, and updates the geometric dimensions of the graphic corresponding to the hotspot points. The PatternMatch module extracts the graphic feature vectors of the hotspot points through a convolutional neural network, searches for approximate graphics based on the feature vectors, and generates approximate graphic data. The corrected graphic data and approximate graphic data are transmitted to the control module.

6. The hotspot discovery method based on photolithography process window according to claim 5, characterized in that: The PWA module includes an adaptive optimization submodule, which uses a genetic algorithm to prune redundant parameters of the deep reinforcement learning network to generate a simplified hotspot analysis model; it then uses transfer learning to transfer knowledge from the complex analysis model to the simplified model, generating a hotspot analysis model adapted to different process nodes; the hotspot data of the model is then transmitted to the control module.

7. The hotspot discovery method based on photolithography process window according to claim 6, characterized in that: The control module performs the following steps during the system training phase: pre-training the Metro module, model simulation module, PWA module, Retarget module, and PatternMatch module to generate initial parameters; jointly fine-tuning the parameters of each module through the backpropagation algorithm to optimize system performance based on the overall loss function; and adjusting the learning rate according to gradient changes through an adaptive learning rate adjustment mechanism to generate trained system parameters; the parameters are stored in the control module.

8. The hotspot discovery method based on photolithography process window according to claim 7, characterized in that: The control module performs the following steps during system operation: receiving hotspot data and common process window data from the PWA module, and calculating manufacturing yield; comparing the yield with a preset target to generate deviation data; adjusting the feature extraction parameters of the Metro module, the simulation parameters of the model simulation module, and the analysis parameters of the PWA module based on the deviation data; and generating updated system parameters by merging new layout data and process parameter data through incremental learning; the parameters are stored in the control module.

9. The hotspot discovery method based on photolithography process window according to claim 8, characterized in that: The Metro module generates adversarial samples containing process parameter fluctuations and layout noise through a generative adversarial network, trains a capsule network and a temporal convolutional network, and generates a feature extraction model after adversarial training. The model simulation module processes the simulation measurement results through random perturbation and data augmentation, trains the variational autoencoder, and generates a data-augmented simulation model; the model is stored in the control module.

Images