Method and system for analyzing defects in wafer manufacturing based on big data

Abstract

The present invention provides a method and system for analyzing defects in wafer manufacturing based on big data, which relates to the field of big data technology, including collecting process parameter information, defect scanning images, and manufacturing process data information, and obtaining fusion features after processing and fusion through a multi-scale feature pyramid network; analyzing the fusion features through a deep learning network, wherein the deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to obtain defect type recognition results and feature maps of defect spatial distribution; based on the feature map, when a process parameter abnormality is detected, a graph neural network containing a time series attention mechanism is used to analyze historical defect data; process parameter constraints are set in combination with an expert knowledge base, and a reinforcement learning method is used to adaptively adjust the process parameters. The present invention improves the accuracy of defect detection and classification.

Description

Technical Field

[0001] The present invention relates to big data technology, and in particular to a method and system for analyzing defects in wafer manufacturing based on big data. Background Art

[0002] In the wafer manufacturing process, various types of defects often occur due to fluctuations in process parameters, changes in equipment status, and environmental factors. These defects not only affect product yield, but also cause a lot of economic losses. The main defects and shortcomings of the existing technology are:

[0003] Existing wafer defect detection methods mainly rely on single image feature analysis, which makes it difficult to effectively utilize process parameters and manufacturing process data, resulting in low defect classification accuracy and inability to deeply understand the root causes of defect formation.

[0004] Traditional process parameter monitoring methods often use simple statistical analysis methods, which lack in-depth analysis of the causal relationship between process parameters and defect distribution, making it difficult to accurately predict and prevent the occurrence of defects.

[0005] Existing process parameter adjustment methods mostly rely on manual experience and lack adaptive learning capabilities. They are unable to automatically optimize process parameters according to the evolution of historical defect data, resulting in low process optimization efficiency and susceptibility to human factors. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for analyzing defects in wafer manufacturing based on big data, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Provides a method for analyzing defects in wafer manufacturing based on big data, including:

[0009] Collect process parameter information of wafer manufacturing, including photolithography process alignment parameters, photoresist coating thickness parameters, exposure energy parameters, etching gas flow parameters, RF power parameters, bias voltage parameters, thin film deposition rate parameters and chamber pressure parameters; collect defect scanning images of wafer surface through multiple distributed defect detection equipment; collect manufacturing process data information, including process yield information, wafer electrical test information and equipment status information;

[0010] Defect features are extracted from defect scanned images using a multi-scale feature pyramid network including a deformable convolution module, and the defect features, process parameter information, and manufacturing process data information are fused to obtain fused features; the fused features are analyzed using a deep learning network, which uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches, thereby obtaining defect type recognition results and feature maps of defect spatial distribution;

[0011] Based on the feature map, a causal graph of process parameter changes and defect distribution evolution is constructed using an abnormality monitoring network including a causal reasoning module. Process parameter anomalies are detected based on the causal graph. When process parameter anomalies are detected, historical defect data are analyzed using a graph neural network including a time series attention mechanism. The graph neural network captures the evolution law of defect distribution through a dynamic graph structure. Based on the causal graph and the evolution law, process parameter constraints are set in combination with an expert knowledge base, and reinforcement learning methods are used to adaptively adjust the process parameters.

[0012] In an optional embodiment,

[0013] The steps of extracting defect features from the defect scan image using a multi-scale feature pyramid network including a deformable convolution module, and fusing the defect features, process parameter information, and manufacturing process data information to obtain fused features include:

[0014] Construct a multi-scale feature pyramid network including a deformable convolution module; the deformable convolution module includes a dual-branch attention network, the dual-branch attention network includes a spatial branch for capturing local structural information and a channel branch for capturing global semantic information, an offset is calculated based on the spatial branch and the channel branch, and the offset is used to guide the feature extraction of the deformable convolution; the multi-scale feature pyramid network includes a plurality of feature extraction layers, each feature extraction layer extracts local features through the deformable convolution module to obtain a feature tensor, upsamples deep features, and fuses the upsampled features with the current layer feature tensor to obtain a feature pyramid output; a feature complementary module is set in the multi-scale feature pyramid network to connect adjacent level features, and the feature complementary module performs adaptive weighted fusion of high-level features and low-level features processed by convolution to obtain defect features;

[0015] Standardize the process parameter information to obtain process parameter features; extract features from the manufacturing process data information to obtain manufacturing data features;

[0016] A feature semantic mapping matrix is ​​constructed based on the knowledge of wafer manufacturing domain. The feature semantic mapping matrix is ​​obtained by calculating the distance between different features in a predefined semantic space, and the predefined semantic space includes a defect morphology feature space, a process control feature space, and a quality indicator feature space. The feature semantic mapping matrix is ​​used to characterize the semantic correlation between different modal features. According to the feature semantic mapping matrix, the defect features, process parameter features, and manufacturing data features are semantically projected to map different modal features to a unified semantic representation space. A bidirectional mapping function is used to align cross-modal features. The bidirectional mapping function includes a first mapping matrix and a second mapping matrix. The first mapping matrix is ​​used to align source modal features, and the second mapping matrix is ​​combined with the feature semantic mapping matrix to align target modal features. The features after bidirectional mapping alignment are residually connected with the original features and nonlinearly transformed to obtain fused features.

[0017] In an optional embodiment,

[0018] The fusion features are analyzed by a deep learning network. The deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches. The steps of obtaining defect type recognition results and feature maps of defect spatial distribution include:

[0019] Constructing a deep learning network to analyze the fusion features, wherein the deep learning network adopts a shared encoder-multi-branch decoder network structure to construct a multi-task learning framework, wherein the shared encoder extracts a deep representation of the fusion features, and the multi-branch decoder includes a defect classification branch, a defect segmentation branch, and a defect evolution prediction branch;

[0020] The defect classification branch converts the feature map into a vector representation through a global average pooling layer and then completes the classification with a fully connected layer. The defect classification branch includes a feature recalibration module based on an attention mechanism. The feature recalibration module is used to adjust the weights of different channel features and output a probability vector of the defect type. The defect segmentation branch adopts an encoder-decoder structure with a jump connection, and a channel attention module is set at the jump connection to multiply the binary defect mask obtained by segmentation with the original feature map to obtain the local features of the defect area. The defect evolution prediction branch receives the probability vector of the defect type and the local features of the defect area as input, and extracts the spatiotemporal features through a spatiotemporal convolutional network. The spatiotemporal convolutional network includes a temporal dependency modeling module based on an attention mechanism, and the temporal dependency modeling module is used to capture the long-term evolution law of the defect.

[0021] The graph structure is constructed based on the morphological features and spatial distribution of defects. The similarity of graph structures of different batches is calculated through a multi-view matching mechanism, and the batch consistency loss is obtained by combining uncertainty assessment and historical feature matching.

[0022] A task graph containing task relevance is constructed, and the task graph is modeled through a graph attention network to obtain inter-task dependencies. The task weights are dynamically adjusted according to changes in the losses of each task. The prediction results of the defect classification branch, the prediction results of the defect segmentation branch, the prediction results of the defect evolution prediction branch, and the batch consistency loss are jointly optimized to obtain a feature map containing defect type recognition results and defect spatial distribution.

[0023] In an optional embodiment,

[0024] The steps of constructing a graph structure based on the morphological features and spatial distribution of defects, calculating the similarity of graph structures of different batches through a multi-view matching mechanism, and combining uncertainty assessment and historical feature matching to obtain batch consistency loss include:

[0025] A hierarchical graph structure is constructed based on the morphological characteristics and spatial distribution of defects. The hierarchical graph structure includes a global layer, a sub-layer and a node layer. The nodes of the node layer represent defect instances, and the edges of the node layer represent the spatial relationships of defects. A density clustering algorithm is used to group spatially adjacent defect groups into a sub-graph structure. The sub-graph structure contains centrality features and intra-group sorting features. Local information is transferred from the bottom to the top between the global layer, the sub-layer and the node layer through a cross-hierarchical feature transfer mechanism, and global constraints are fed back from the top to the bottom. The connection strength of the edges of the node layer is dynamically adjusted based on the attention gating mechanism, and potential topological changes are predicted in combination with the defect evolution trend. The graph structure differences at adjacent moments are controlled within the target threshold range through topological structure smoothing constraints.

[0026] According to the constructed hierarchical graph structure, a multi-view matching mechanism including node-level matching, subgraph-level matching and global-level matching is constructed, and the matching results of different levels are fused through adaptive weighting; an adversarial generative network is used to generate matching samples and the feature encoding parameters of the multi-view matching mechanism are optimized through adversarial training, a temporal dependency graph between batches is established, and a gated recurrent unit is introduced to capture long-term evolution laws;

[0027] Based on the multi-view matching results, a hierarchical loss function including node loss, subgraph loss and global loss is constructed. The curriculum learning strategy is used to dynamically adjust the loss weights of each layer. The gradients of loss functions at different levels are collaboratively optimized through the cross-layer loss complementary mechanism. An uncertainty assessment module based on the Bayesian framework is established to adaptively adjust the loss weights of samples with different confidence levels. A historical batch feature library is constructed to store matching patterns and maintain and update them through the attention mechanism. The hierarchical loss function, uncertainty assessment results and historical feature matching information are fused and optimized to obtain the batch consistency loss.

[0028] In an optional embodiment,

[0029] The steps of constructing a task graph including task correlations, modeling the task graph through a graph attention network to obtain dependencies between tasks, dynamically adjusting task weights according to changes in the losses of each task, and jointly optimizing the prediction results of the defect classification branch, the defect segmentation branch, the defect evolution prediction branch, and the batch consistency loss to obtain a feature map including defect type recognition results and defect spatial distribution include:

[0030] A mutual information neural estimator is used to calculate the mutual information of features between tasks, positive and negative sample pairs are constructed by double-sample random sampling, and a separator network is used to estimate the lower bound of the mutual information. A task relationship graph including defect classification task nodes, defect segmentation task nodes, and defect evolution prediction task nodes is constructed based on the mutual information, and the mutual information is used as the edge weight of the task relationship graph;

[0031] Perform multi-layer graph attention modeling on the task relationship graph, use the dot product attention mechanism and combine it with position encoding to calculate the local dependencies between task nodes, use the graph convolution message passing mechanism and the gating unit to aggregate global context information, and learn the optimal combination of task features through a dynamic routing algorithm;

[0032] Based on the results of graph attention modeling, a meta-learning optimizer is built to calculate the gradient cosine similarity and Hessian matrix between tasks, evaluate the spatial consistency of task representation, and dynamically update task weights using a model-independent meta-learning algorithm to obtain the optimized weights of each task.

[0033] Construct a multi-level joint optimization framework, including: introducing cross entropy loss for the prediction results of the defect classification branch, introducing distance loss and boundary loss for the prediction results of the defect segmentation branch, and introducing temporal consistency loss for the prediction results of the defect evolution prediction branch; based on the optimization weights of each task, a weighted combination of the defect classification branch loss, the defect segmentation branch loss, the defect evolution prediction branch loss and the batch consistency loss is performed, and contrastive learning loss is introduced in the shared feature space to extract common features between tasks, and soft label information is transferred through knowledge distillation loss to obtain optimized multi-task features;

[0034] A hierarchical feature fusion network is constructed, and the optimized multi-task features are aligned using spatial pyramid pooling and channel attention mechanism. Dynamic convolution kernel and bilinear pooling are used to adaptively fuse multi-scale features to generate a feature map containing defect type recognition results and defect spatial distribution.

[0035] In an optional embodiment,

[0036] Based on the feature graph, a causal graph of process parameter changes and defect distribution evolution is constructed using an abnormality monitoring network including a causal reasoning module, and process parameter anomalies are detected according to the causal graph. When process parameter anomalies are detected, historical defect data is analyzed using a graph neural network including a time series attention mechanism. The steps of capturing the evolution law of defect distribution through a dynamic graph structure of the graph neural network include:

[0037] The feature map containing defect type and defect spatial distribution information is expanded in time dimension, and the feature map of each time step is processed by deep separable convolution to obtain local spatial context information. One-dimensional convolution is used to model the feature sequence along the time axis to obtain temporal dynamic features. The local spatial context information and the temporal dynamic features are fused through residual connection to obtain defect distribution features with the ability to characterize spatiotemporal evolution.

[0038] The process parameter sequence and the defect distribution characteristics are respectively used as nodes of a causal graph, and the process parameter sequence and the defect distribution characteristics are decoupled by using nonlinear independent component analysis to obtain potential causal variables. The causal strength between the potential causal variables is calculated based on the conditional mutual information criterion, and the causal graph structure is optimized by using a scoring search algorithm to obtain a causal graph including process parameter nodes and defect distribution characteristic nodes;

[0039] Causal reasoning is performed based on the causal graph, the causal effect of the process parameters on the defect distribution is calculated through intervention calculation, the degree of abnormality of the process parameters is evaluated by comparing the observed values ​​and the counterfactual predicted values, and the dynamic threshold method is used to perform abnormality detection in combination with the causal effect scores of multiple time steps;

[0040] When a process parameter abnormality is detected, the defect distribution features at different times in the historical defect data are used as graph nodes, and a time-varying adjacency matrix of the dynamic graph structure is constructed based on feature similarity. The long-term and short-term dependencies are captured through the bidirectional temporal attention layer of the time series attention mechanism, and the gated graph convolution is used to dynamically update the feature representation of the graph nodes to predict the evolution law of the defect distribution.

[0041] In an optional embodiment,

[0042] According to the causal graph and the evolution law, the process parameter constraints are set in combination with the expert knowledge base, and the steps of using the reinforcement learning method to adaptively adjust the process parameters include:

[0043] The sensitivity range of process parameter adjustment is determined based on the edge weights from process parameter nodes to defect nodes in the causal graph. The change rate limit of process parameters is set through a sliding time window based on the evolution law. Combined with expert experience, a multi-level constraint knowledge base including physical limits of process parameters, coupling relationships of process parameters and process rules is constructed.

[0044] Construct a reinforcement learning environment and design a state space, which includes the current process parameter value, process parameter change trend, process state of historical statistical characteristics, defect spatial distribution, defect type proportion, defect quantity statistics, defect feature state, production load, equipment state, and predicted state of evolution law; map the process parameter adjustment amount to the normalized interval and consider the coordinated adjustment of multiple process parameters to construct the action space, and design a reward function including defect improvement index, process parameter stability score and adjustment efficiency evaluation;

[0045] Design a dual network architecture, use a multi-head attention mechanism to process multimodal input features, perform knowledge fusion through soft constraint embedding and expert rule injection, generate a process parameter adjustment strategy distribution that satisfies the multi-level constraint knowledge base, use a temporal convolutional network to extract state time features and evaluate the long-term impact of adjustment actions, perform value correction in combination with the prediction of evolution laws, and calculate sample priorities based on time difference errors and historical sample occurrence frequencies to build an experience replay pool;

[0046] For the process parameter adjustment strategy distribution, a dynamic balance between exploration and utilization is achieved based on the process parameter sensitivity, and the sampled process parameter adjustment actions are subject to multiple constraint verifications of physical limits, expert rules and coupling relationships. The importance sampling method is used to optimize the sampling of the strategy distribution, and the proximal strategy optimization algorithm is used to gradient update the policy network and the value network. The process parameters are adaptively adjusted through online anomaly detection and adaptive learning rate.

[0047] According to a second aspect of the embodiments of the present invention,

[0048] Provides a system for analyzing defects in wafer manufacturing based on big data, including:

[0049] The first unit is used to collect process parameter information of wafer manufacturing, including photolithography process alignment parameters, photoresist coating thickness parameters, exposure energy parameters, etching gas flow parameters, RF power parameters, bias voltage parameters, thin film deposition rate parameters and chamber pressure parameters; collect defect scanning images of the wafer surface through multiple distributed defect detection devices; collect manufacturing process data information, including process yield information, wafer electrical test information and equipment status information;

[0050] The second unit is used to extract defect features from the defect scan image using a multi-scale feature pyramid network including a deformable convolution module, and fuse the defect features, process parameter information and manufacturing process data information to obtain fused features; the fused features are analyzed through a deep learning network, and the deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches, and obtain defect type recognition results and feature maps of defect spatial distribution;

[0051] The third unit is used to construct a causal graph of process parameter changes and defect distribution evolution based on the feature map using an abnormality monitoring network including a causal reasoning module, detect process parameter anomalies according to the causal graph, and when process parameter anomalies are detected, analyze historical defect data using a graph neural network including a time series attention mechanism, and capture the evolution law of defect distribution through a dynamic graph structure; and set process parameter constraints based on the causal graph and the evolution law in combination with an expert knowledge base, and use a reinforcement learning method to adaptively adjust the process parameters.

[0052] According to a third aspect of the embodiments of the present invention,

[0053] An electronic device is provided, comprising:

[0054] processor;

[0055] a memory for storing processor-executable instructions;

[0056] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0057] A fourth aspect of the embodiments of the present invention is:

[0058] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0059] The present invention extracts defect features by adopting a multi-scale feature pyramid network of a deformable convolution module, and fuses the defect features with process parameters and manufacturing process data through multi-source information fusion. By combining a multi-task learning framework and graph matching consistency loss, the present invention can accurately identify the defect type and obtain a defect spatial distribution feature map, thereby improving the accuracy of defect detection and classification.

[0060] The present invention constructs a causal graph of process parameters and defect distribution evolution based on an anomaly monitoring network, analyzes process parameter anomalies through a causal reasoning module, and uses a graph neural network with a time series attention mechanism to capture the evolution law of defect distribution, thereby achieving accurate identification and effective monitoring of abnormal process parameters in the wafer manufacturing process.

[0061] The present invention adopts a reinforcement learning method to adaptively adjust the process parameters, combines the constraints set by the expert knowledge base, optimizes the process parameters according to the cause-effect diagram and the defect evolution law, improves the intelligence level of the wafer manufacturing process, effectively reduces the defect incidence rate, and improves the product yield. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of a process of analyzing defects in wafer manufacturing based on big data according to an embodiment of the present invention;

[0063] Figure 2 It is a schematic diagram of the structure of a system for analyzing defects in wafer manufacturing based on big data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0066] Figure 1 FIG. 1 is a flow chart of a method for analyzing defects in wafer manufacturing based on big data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] S1. Collect process parameter information of wafer manufacturing, including photolithography process alignment parameters, photoresist coating thickness parameters, exposure energy parameters, etching gas flow parameters, RF power parameters, bias voltage parameters, film deposition rate parameters and chamber pressure parameters; collect defect scanning images of wafer surface through multiple distributed defect detection equipment; collect manufacturing process data information, including process yield information, wafer electrical test information and equipment status information;

[0068] S2. Extract defect features from the defect scanned image using a multi-scale feature pyramid network including a deformable convolution module, and fuse the defect features, process parameter information, and manufacturing process data information to obtain fused features; analyze the fused features through a deep learning network, which uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces a consistency loss based on graph matching to establish associations between defect samples from different batches, and obtain defect type recognition results and feature maps of defect spatial distribution;

[0069] S3. Based on the feature map, an abnormality monitoring network including a causal reasoning module is used to construct a causal graph of process parameter changes and defect distribution evolution, and process parameter anomalies are detected according to the causal graph. When process parameter anomalies are detected, historical defect data are analyzed using a graph neural network including a time series attention mechanism. The graph neural network captures the evolution law of defect distribution through a dynamic graph structure; and according to the causal graph and the evolution law, process parameter constraints are set in combination with an expert knowledge base, and reinforcement learning methods are used to adaptively adjust the process parameters.

[0070] In an optional embodiment,

[0071] The steps of extracting defect features from the defect scan image using a multi-scale feature pyramid network including a deformable convolution module, and fusing the defect features, process parameter information, and manufacturing process data information to obtain fused features include:

[0072] Construct a multi-scale feature pyramid network including a deformable convolution module; the deformable convolution module includes a dual-branch attention network, the dual-branch attention network includes a spatial branch for capturing local structural information and a channel branch for capturing global semantic information, an offset is calculated based on the spatial branch and the channel branch, and the offset is used to guide the feature extraction of the deformable convolution; the multi-scale feature pyramid network includes a plurality of feature extraction layers, each feature extraction layer extracts local features through the deformable convolution module to obtain a feature tensor, upsamples deep features, and fuses the upsampled features with the current layer feature tensor to obtain a feature pyramid output; a feature complementary module is set in the multi-scale feature pyramid network to connect adjacent level features, and the feature complementary module performs adaptive weighted fusion of high-level features and low-level features processed by convolution to obtain defect features;

[0073] Standardize the process parameter information to obtain process parameter features; extract features from the manufacturing process data information to obtain manufacturing data features;

[0074] A feature semantic mapping matrix is ​​constructed based on the knowledge of wafer manufacturing domain. The feature semantic mapping matrix is ​​obtained by calculating the distance between different features in a predefined semantic space, and the predefined semantic space includes a defect morphology feature space, a process control feature space, and a quality indicator feature space. The feature semantic mapping matrix is ​​used to characterize the semantic correlation between different modal features. According to the feature semantic mapping matrix, the defect features, process parameter features, and manufacturing data features are semantically projected to map different modal features to a unified semantic representation space. A bidirectional mapping function is used to align cross-modal features. The bidirectional mapping function includes a first mapping matrix and a second mapping matrix. The first mapping matrix is ​​used to align source modal features, and the second mapping matrix is ​​combined with the feature semantic mapping matrix to align target modal features. The features after bidirectional mapping alignment are residually connected with the original features and nonlinearly transformed to obtain fused features.

[0075] Exemplarily, a multi-scale feature pyramid network containing a deformable convolution module is first constructed. The network adopts a dual-branch attention structure, including a spatial branch and a channel branch. The spatial branch extracts local structural features through a three-layer convolutional network. The convolution kernel sizes are 3×3, 3×3 and 1×1, respectively, and the step size is 1. Each layer is followed by a batch normalization layer and a ReLU activation function. The channel branch uses a global average pooling layer to capture global information. After two layers of fully connected layers, the output dimension is the same as the number of channels of the input feature map. The output feature maps of the spatial branch and the channel branch are element-wise multiplied to obtain an attention weight map, which is used to calculate the sampling position offset of the deformable convolution.

[0076] In the feature extraction process, the network contains 5 feature extraction layers, which extract features of different scales. Each layer uses a deformable convolution module to extract local features, and the convolution kernel size is 3×3. For the features extracted from the i-th layer, the features of the i+1 layer are upsampled and fused with them. The feature complementation module uses an adaptive weighting method to fuse high-level semantic features and low-level detail features. Specifically, the low-level features are first reduced in dimension by 1×1 convolution, and then the similarity matrix is ​​calculated with the high-level features. The fusion weights are generated based on the similarity matrix, and finally the multi-scale defect features are obtained.

[0077] For process parameter information, including numerical parameters such as temperature, pressure, and gas flow, the maximum and minimum normalization method is used for standardization, and the values ​​are mapped to the range of 0-1. For discrete parameters, one-hot encoding is used to convert them into vector representation. Manufacturing process data includes time series data such as equipment status and maintenance records. Time series features are extracted through long short-term memory networks. The hidden layer dimension is set to 128, and the extracted feature dimension is 256.

[0078] When constructing the feature semantic mapping matrix based on the knowledge of wafer manufacturing domain, three semantic spaces including defect morphology, process control and quality indicators are predefined. Each feature vector calculates the cosine distance in these three semantic spaces to obtain a 3×3 semantic correlation matrix. This matrix is ​​used to guide the semantic alignment of different modal features, and bidirectional feature mapping is achieved through two mapping networks. Each mapping network contains two fully connected layers, and a dropout layer is added in the middle to prevent overfitting. Finally, the mapped features are residually connected with the original features, and the final fused features are obtained through the ReLU activation function.

[0079] The present invention uses deformable convolution and dual-branch attention mechanism to adaptively capture key features of defective areas, improve the accuracy and robustness of feature extraction, and realize the effective fusion of features of different scales by using multi-scale feature pyramid structure, thereby enhancing the detection capability of defects of different sizes. By using semantic mapping matrix based on domain knowledge, semantic associations between features of different modalities are established, and effective alignment of heterogeneous features is realized. The bidirectional mapping mechanism ensures the bidirectional flow of information in the feature fusion process, improves the expression capability of fused features, makes full use of multi-source information in the manufacturing process, deeply fuses defect image features with process parameters and manufacturing data, and provides more comprehensive defect representation. The whole processing flow has a high degree of automation, is suitable for actual application in industrial sites, and has strong practical value.

[0080] In an optional embodiment,

[0081] The fusion features are analyzed by a deep learning network. The deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches. The steps of obtaining defect type recognition results and feature maps of defect spatial distribution include:

[0082] Constructing a deep learning network to analyze the fusion features, wherein the deep learning network adopts a shared encoder-multi-branch decoder network structure to construct a multi-task learning framework, wherein the shared encoder extracts a deep representation of the fusion features, and the multi-branch decoder includes a defect classification branch, a defect segmentation branch, and a defect evolution prediction branch;

[0083] The defect classification branch converts the feature map into a vector representation through a global average pooling layer and then completes the classification with a fully connected layer. The defect classification branch includes a feature recalibration module based on an attention mechanism. The feature recalibration module is used to adjust the weights of different channel features and output a probability vector of the defect type. The defect segmentation branch adopts an encoder-decoder structure with a jump connection, and a channel attention module is set at the jump connection to multiply the binary defect mask obtained by segmentation with the original feature map to obtain the local features of the defect area. The defect evolution prediction branch receives the probability vector of the defect type and the local features of the defect area as input, and extracts the spatiotemporal features through a spatiotemporal convolutional network. The spatiotemporal convolutional network includes a temporal dependency modeling module based on an attention mechanism, and the temporal dependency modeling module is used to capture the long-term evolution law of the defect.

[0084] The graph structure is constructed based on the morphological features and spatial distribution of defects. The similarity of graph structures of different batches is calculated through a multi-view matching mechanism, and the batch consistency loss is obtained by combining uncertainty assessment and historical feature matching.

[0085] A task graph containing task relevance is constructed, and the task graph is modeled through a graph attention network to obtain inter-task dependencies. The task weights are dynamically adjusted according to changes in the losses of each task. The prediction results of the defect classification branch, the prediction results of the defect segmentation branch, the prediction results of the defect evolution prediction branch, and the batch consistency loss are jointly optimized to obtain a feature map containing defect type recognition results and defect spatial distribution.

[0086] Exemplarily, the infrastructure of the deep learning network is first constructed. The network adopts a shared encoder-multi-branch decoder structure. The shared encoder uses ResNet50 as the backbone network to extract multi-scale representations of input features. The encoder contains 5 stages, each of which consists of multiple residual blocks, which gradually reduce the spatial resolution of the feature map and increase the number of channels. Specifically, the input feature first passes through a 7x7 convolutional layer and a maximum pooling layer, and then passes through the conv2 to conv5 stages in sequence, and finally obtains a feature map of 1 / 32 size.

[0087] In the defect classification branch, the channel attention mechanism is introduced to recalibrate the features. In the specific implementation, the statistical information of each channel is first obtained through global average pooling, and then the correlation between channels is learned using a two-layer fully connected network to output a weight vector with the same dimension as the number of channels. The weight vector is multiplied by the original feature to achieve adaptive adjustment of the feature. Finally, the probability distribution of the defect type is obtained through global average pooling and a fully connected layer.

[0088] The defect segmentation branch adopts the U-Net structure. When the output features of each stage of the encoder are fused with the features of the corresponding level of the decoder, a channel attention module is added. This module calculates the importance weights of different channel features to highlight key information. The decoder gradually restores the resolution of the feature map through transposed convolution and finally outputs a binary segmentation mask of the same size as the input image. The mask is multiplied with the original feature map to extract the local representation of the defect area.

[0089] The defect evolution prediction branch designs a spatiotemporal convolution module, which first performs a one-dimensional convolution on the feature sequence along the time dimension to model the local associations of adjacent moments. Then, the self-attention mechanism is used to calculate the similarity matrix of features at different moments, and the long-range dependency is captured by weighted summation. Finally, the defect type probability vector and local features are combined to predict the evolution trend of the defect.

[0090] In terms of batch consistency loss, a graph structure is constructed based on the shape, size, position and other attributes of the defects. The nodes in the graph represent defect instances, and the edges represent the spatial relationship between defects. For graph structures of different batches, the node similarity and edge similarity are calculated from the three perspectives of appearance, position and topology, and combined with the matching results of the historical feature library, the consistency loss across batches is obtained.

[0091] Finally, a task graph is constructed to reflect the relevance of tasks, where nodes represent different tasks and edges represent dependencies between tasks. The weights of edges are learned using a graph attention network, and the task weights are dynamically adjusted according to the gradient direction and magnitude of the loss function of each task, thus achieving collaborative optimization of multiple tasks.

[0092] The present invention extracts universal feature representation through a shared encoder, and the multi-branch decoder completes different tasks in a targeted manner, thereby improving feature utilization efficiency and reducing computational overhead. The attention mechanism introduced in each branch can highlight important features and suppress the interference of irrelevant information. The batch consistency loss based on graph matching establishes the association between samples of different batches, which improves the stability and reliability of the model in practical applications. The multi-perspective matching mechanism comprehensively considers multiple attributes of defects to make the matching results more accurate. The introduction of task graphs realizes adaptive modeling of dependencies between tasks. The dynamic task weight adjustment mechanism ensures the balance of multi-task learning. It can simultaneously complete multiple tasks such as defect classification, segmentation and evolution prediction, providing comprehensive information support for defect analysis.

[0093] In an optional embodiment,

[0094] The steps of constructing a graph structure based on the morphological features and spatial distribution of defects, calculating the similarity of graph structures of different batches through a multi-view matching mechanism, and combining uncertainty assessment and historical feature matching to obtain batch consistency loss include:

[0095] A hierarchical graph structure is constructed based on the morphological characteristics and spatial distribution of defects. The hierarchical graph structure includes a global layer, a sub-layer and a node layer. The nodes of the node layer represent defect instances, and the edges of the node layer represent the spatial relationships of defects. A density clustering algorithm is used to group spatially adjacent defect groups into a sub-graph structure. The sub-graph structure contains centrality features and intra-group sorting features. Local information is transferred from the bottom to the top between the global layer, the sub-layer and the node layer through a cross-hierarchical feature transfer mechanism, and global constraints are fed back from the top to the bottom. The connection strength of the edges of the node layer is dynamically adjusted based on the attention gating mechanism, and potential topological changes are predicted in combination with the defect evolution trend. The graph structure differences at adjacent moments are controlled within the target threshold range through topological structure smoothing constraints.

[0096] According to the constructed hierarchical graph structure, a multi-view matching mechanism including node-level matching, subgraph-level matching and global-level matching is constructed, and the matching results of different levels are fused through adaptive weighting; an adversarial generative network is used to generate matching samples and the feature encoding parameters of the multi-view matching mechanism are optimized through adversarial training, a temporal dependency graph between batches is established, and a gated recurrent unit is introduced to capture long-term evolution laws;

[0097] Based on the multi-view matching results, a hierarchical loss function including node loss, subgraph loss and global loss is constructed. The curriculum learning strategy is used to dynamically adjust the loss weights of each layer. The gradients of loss functions at different levels are collaboratively optimized through the cross-layer loss complementary mechanism. An uncertainty assessment module based on the Bayesian framework is established to adaptively adjust the loss weights of samples with different confidence levels. A historical batch feature library is constructed to store matching patterns and maintain and update them through the attention mechanism. The hierarchical loss function, uncertainty assessment results and historical feature matching information are fused and optimized to obtain the batch consistency loss.

[0098] Exemplarily, first, a hierarchical graph structure is constructed based on the morphological characteristics and spatial distribution of defects. The hierarchical graph structure includes a global layer, a sublayer, and a node layer. In the node layer, each node represents a defect instance, and the edge between nodes represents the spatial relationship between defects. In specific implementation, the Euclidean distance or Manhattan distance can be used to measure the spatial relationship between defects, and a threshold is set to determine whether to establish an edge connection between nodes.

[0099] Next, a density clustering algorithm is used to group spatially adjacent defect groups into subgraph structures. The DBSCAN algorithm (Density-Based Spatial Clustering of Applications with Noise) can be used here to implement density clustering. The core idea of ​​the DBSCAN algorithm is to divide density-connected areas into a cluster under a given density threshold. For each subgraph structure, its centrality features and intra-group sorting features are calculated. Centrality features can include degree centrality, betweenness centrality, and closeness centrality, etc., which are used to describe the importance of each node in the subgraph. The intra-group sorting features can be sorted based on the attribute values ​​of the nodes (such as defect area, depth, etc.) to obtain the relative position of the nodes in the subgraph.

[0100] In order to realize cross-level feature transfer, a bottom-up and top-down bidirectional information transfer mechanism is designed. In the bottom-up transfer process, the features of the node layer are transferred to the sub-layer through an aggregation function (such as average pooling or maximum pooling), and then transferred from the sub-layer to the global layer. In the top-down feedback process, the features of the global layer are distributed to the sub-layer and node layer through the attention mechanism to realize the transfer of global constraints.

[0101] In terms of adjusting the connection strength of edges, an attention gating mechanism is introduced. Specifically, a multi-head attention mechanism can be used to dynamically adjust the weight of edges according to the similarity between nodes. At the same time, considering the evolution trend of defects, time series prediction models (such as ARIMA or Prophet) can be used to predict potential topological changes. In order to ensure the smoothness of the graph structure, a topological structure smoothing constraint is introduced, that is, the topological change is controlled by minimizing the difference in graph structure at adjacent moments. The graph edit distance or spectral distance can be used to measure the difference in graph structure, and a target threshold range can be set.

[0102] When building a multi-view matching mechanism, three levels of matching modules are designed: node-level matching, subgraph-level matching, and global-level matching. Node-level matching can use a twin neural network, input the feature vectors of two nodes, and output their similarity. Subgraph-level matching can use a graph neural network (such as a graph convolutional network or a graph attention network) to extract the structural features of the subgraph, and then use cosine similarity to calculate the similarity between subgraphs. Global-level matching can use a graph kernel method or a graph embedding method to calculate the similarity of the entire graph structure.

[0103] In order to adaptively fuse the matching results of different levels, an attention mechanism can be used. Specifically, an attention weight can be assigned to each matching level, and these weights are optimized through back propagation to achieve adaptive weighted fusion.

[0104] In order to enhance the robustness of the matching mechanism, a generative adversarial network (GAN) is introduced to generate matching samples. The generator is responsible for generating graph samples with similar structures but different details, while the discriminator is responsible for distinguishing between real samples and generated samples. Through adversarial training, the feature encoding parameters of the multi-view matching mechanism can be optimized and its generalization ability can be improved.

[0105] In terms of capturing long-term evolutionary laws, we first build a temporal dependency graph between batches. Each node represents a batch, and the edge represents the relationship between adjacent batches. Then, we use a gated recurrent unit (GRU) to process this temporal dependency graph. The hidden state of the GRU can capture long-term evolutionary laws, and its output can be used as a basis for predicting the features of the next batch.

[0106] When constructing the hierarchical loss function, node loss, subgraph loss and global loss are designed respectively. Node loss can use cross entropy loss to measure the accuracy of node-level matching. Subgraph loss can use structural similarity index (SSIM) to measure the accuracy of subgraph-level matching. Global loss can use Wasserstein distance to measure the accuracy of global-level matching.

[0107] In order to dynamically adjust the loss weights of each layer, a curriculum learning strategy is adopted. In the early stage of training, a larger weight is given to the node loss. As the training progresses, the weights of the subgraph loss and the global loss are gradually increased. This allows the model to learn simple local features first, and then gradually learn complex global features.

[0108] In terms of the cross-layer loss complementation mechanism, a gradient collaborative optimization module is designed. This module analyzes the gradient directions of loss functions at different levels. When it is found that the gradient directions of some levels are inconsistent, it is adjusted through methods such as gradient clipping or gradient normalization to achieve complementarity of losses at different levels.

[0109] In terms of uncertainty assessment, an evaluation module based on the Bayesian framework is established. Specifically, Monte Carlo dropout can be used to estimate the uncertainty of model predictions. For each sample, the model with dropout is run multiple times to obtain multiple prediction results. The variance of the prediction results can be used as a measure of uncertainty. According to the size of the uncertainty, the loss weight of the sample is adaptively adjusted, and a smaller weight is given to samples with high uncertainty.

[0110] In terms of the construction and maintenance of the historical batch feature library, a key-value storage structure can be used. The key is the identifier of the batch, and the value is the feature vector of the batch. Use the attention mechanism to update the feature library. Specifically, when a new batch arrives, calculate its similarity with the historical batch, assign attention weights based on the similarity, and then update the feature vector in the feature library in a weighted sum manner.

[0111] Finally, the hierarchical loss function, uncertainty assessment results, and historical feature matching information are fused and optimized to obtain the batch consistency loss. The fusion can be performed in a weighted sum manner, and the weights can be determined by grid search or Bayesian optimization. The fused result is used as the final consistency constraint to guide subsequent defect detection and classification tasks.

[0112] By constructing a hierarchical graph structure and a multi-view matching mechanism, the present invention can comprehensively capture the morphological characteristics and spatial distribution information of defects, thereby improving the matching accuracy of defect patterns between different batches; the introduction of uncertainty assessment and historical feature matching enhances the robustness and adaptability of the method, and can better handle complex and changeable industrial production environments; the adoption of a curriculum learning strategy and a cross-layer loss complementary mechanism optimizes the model training process, improves the model's convergence speed and generalization ability, and thereby achieves a more accurate batch consistency loss.

[0113] In an optional embodiment,

[0114] The steps of constructing a task graph including task correlations, modeling the task graph through a graph attention network to obtain dependencies between tasks, dynamically adjusting task weights according to changes in the losses of each task, and jointly optimizing the prediction results of the defect classification branch, the defect segmentation branch, the defect evolution prediction branch, and the batch consistency loss to obtain a feature map including defect type recognition results and defect spatial distribution include:

[0115] A mutual information neural estimator is used to calculate the mutual information of features between tasks, positive and negative sample pairs are constructed by double-sample random sampling, and a separator network is used to estimate the lower bound of the mutual information. A task relationship graph including defect classification task nodes, defect segmentation task nodes, and defect evolution prediction task nodes is constructed based on the mutual information, and the mutual information is used as the edge weight of the task relationship graph;

[0116] Perform multi-layer graph attention modeling on the task relationship graph, use the dot product attention mechanism and combine it with position encoding to calculate the local dependencies between task nodes, use the graph convolution message passing mechanism and the gating unit to aggregate global context information, and learn the optimal combination of task features through a dynamic routing algorithm;

[0117] Based on the results of graph attention modeling, a meta-learning optimizer is built to calculate the gradient cosine similarity and Hessian matrix between tasks, evaluate the spatial consistency of task representation, and dynamically update task weights using a model-independent meta-learning algorithm to obtain the optimized weights of each task.

[0118] Construct a multi-level joint optimization framework, including: introducing cross entropy loss for the prediction results of the defect classification branch, introducing distance loss and boundary loss for the prediction results of the defect segmentation branch, and introducing temporal consistency loss for the prediction results of the defect evolution prediction branch; based on the optimization weights of each task, a weighted combination of the defect classification branch loss, the defect segmentation branch loss, the defect evolution prediction branch loss and the batch consistency loss is performed, and contrastive learning loss is introduced in the shared feature space to extract common features between tasks, and soft label information is transferred through knowledge distillation loss to obtain optimized multi-task features;

[0119] A hierarchical feature fusion network is constructed, and the optimized multi-task features are aligned using spatial pyramid pooling and channel attention mechanism. Dynamic convolution kernel and bilinear pooling are used to adaptively fuse multi-scale features to generate a feature map containing defect type recognition results and defect spatial distribution.

[0120] Exemplarily, first, a task relationship graph is constructed through a mutual information neural estimator. For the three tasks of defect classification, segmentation, and evolution prediction, a double-sample random sampling method is used to extract sample pairs from the characteristics of each task. Specifically, two different batches of samples are randomly selected from the same task feature space to form a positive sample pair, and samples are randomly selected from different task feature spaces to form a negative sample pair. A multilayer perceptron is used as a separator network, the feature difference of the sample pair is input, and the mutual information estimate is output. A task relationship graph is constructed based on the calculated mutual information between tasks, where the nodes of the graph represent three different tasks, and the weights of the edges are the mutual information values ​​between the corresponding task pairs.

[0121] Next, the task relationship graph is modeled with graph attention. A multi-head attention mechanism is used, and each attention head independently calculates the attention score between task nodes. When calculating attention, the position information of the task node is encoded into a vector and concatenated with the node features, and the attention matrix is ​​obtained through dot product operation. A gated graph neural network is used for message passing, and each task node performs weighted aggregation of the information of neighbor nodes according to the attention score. The node representation is updated through dynamic routing iteration, so that more information can be transmitted between task nodes with strong correlation.

[0122] Then, a meta-learning optimizer is constructed based on the modeling results of graph attention. The cosine similarity between the gradient vectors of different tasks is calculated to evaluate the consistency of task learning objectives. At the same time, the second-order derivative (Hessian matrix) of the task loss function is estimated to analyze the difficulty of task learning. Combining gradient similarity and Hessian features, a meta-learning algorithm based on gradient descent is used to dynamically adjust the optimization weights of each task.

[0123] In the multi-level joint optimization stage, corresponding loss functions are designed for different task branches. The defect classification branch uses cross entropy loss to measure the difference between the predicted category and the true label; the defect segmentation branch combines distance transform field loss and boundary perception loss to improve segmentation accuracy; the defect evolution prediction branch introduces temporal consistency constraints. Construct contrastive learning loss in the feature space to bring the feature representations of similar samples closer and push the feature representations of samples of different categories apart. The soft label information of the teacher model is transferred to the student model through knowledge distillation.

[0124] Finally, a hierarchical feature fusion network is constructed. Spatial pyramid pooling is used to extract multi-scale features, and the channel attention mechanism is used to highlight important feature channels. Dynamic convolution kernels are designed to adaptively adjust convolution parameters according to input features. Bilinear pooling is used to fuse features of different scales to generate the final feature map, which contains the type information and spatial distribution information of the defects.

[0125] The present invention accurately captures the correlation and dependency between tasks through mutual information neural estimation and graph attention modeling, realizes efficient transmission and fusion of task features, and improves the overall performance of multi-task learning; based on the dynamic weight optimization strategy of meta-learning, it adaptively balances the learning difficulty of different tasks, avoids the problem of gradient vanishing or gradient exploding in difficult tasks during training, and ensures stable training of the model; the design of the multi-level joint optimization framework and the hierarchical feature fusion network realizes the collaborative optimization of defect detection, segmentation and evolution prediction tasks, makes full use of the complementary information between tasks, and significantly improves the prediction accuracy and generalization ability of various tasks.

[0126] In an optional embodiment,

[0127] Based on the feature graph, a causal graph of process parameter changes and defect distribution evolution is constructed using an abnormality monitoring network including a causal reasoning module, and process parameter anomalies are detected according to the causal graph. When process parameter anomalies are detected, historical defect data is analyzed using a graph neural network including a time series attention mechanism. The steps of capturing the evolution law of defect distribution through a dynamic graph structure of the graph neural network include:

[0128] The feature map containing defect type and defect spatial distribution information is expanded in time dimension, and the feature map of each time step is processed by deep separable convolution to obtain local spatial context information. One-dimensional convolution is used to model the feature sequence along the time axis to obtain temporal dynamic features. The local spatial context information and the temporal dynamic features are fused through residual connection to obtain defect distribution features with the ability to characterize spatiotemporal evolution.

[0129] The process parameter sequence and the defect distribution characteristics are respectively used as nodes of a causal graph, and the process parameter sequence and the defect distribution characteristics are decoupled by using nonlinear independent component analysis to obtain potential causal variables. The causal strength between the potential causal variables is calculated based on the conditional mutual information criterion, and the causal graph structure is optimized by using a scoring search algorithm to obtain a causal graph including process parameter nodes and defect distribution characteristic nodes;

[0130] Causal reasoning is performed based on the causal graph, the causal effect of the process parameters on the defect distribution is calculated through intervention calculation, the degree of abnormality of the process parameters is evaluated by comparing the observed values ​​and the counterfactual predicted values, and the dynamic threshold method is used to perform abnormality detection in combination with the causal effect scores of multiple time steps;

[0131] When a process parameter abnormality is detected, the defect distribution features at different times in the historical defect data are used as graph nodes, and a time-varying adjacency matrix of the dynamic graph structure is constructed based on feature similarity. The long-term and short-term dependencies are captured through the bidirectional temporal attention layer of the time series attention mechanism, and the gated graph convolution is used to dynamically update the feature representation of the graph nodes to predict the evolution law of the defect distribution.

[0132] Exemplarily, first, the feature map containing defect type and defect spatial distribution information is expanded in the time dimension. The feature map can be represented as a three-dimensional tensor, in which two dimensions represent spatial information and the third dimension represents the defect type. In order to capture the time evolution characteristics of the defect distribution, we expand the feature map in the time dimension to form a four-dimensional tensor, in which the fourth dimension represents the time step.

[0133] Next, the feature map of each time step is processed using depthwise separable convolution to obtain local spatial context information. Depthwise separable convolution includes two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs convolution operations on each input channel separately to capture spatial features; pointwise convolution performs 1x1 convolution on all channels to achieve information fusion between channels. This method can effectively reduce the number of parameters and improve computational efficiency.

[0134] Then, one-dimensional convolution is used to model the feature sequence along the time axis to obtain the temporal dynamic features. One-dimensional convolution can capture the correlation between different time steps, thereby extracting the dynamic features of the defect distribution changing over time.

[0135] The local spatial context information and temporal dynamic features are fused through residual connections to obtain defect distribution features with the ability to characterize spatiotemporal evolution. Residual connections can alleviate the gradient vanishing problem in deep networks while retaining the original feature information, which is beneficial to network training and performance improvement.

[0136] The process parameter sequence and defect distribution characteristics are used as nodes of the causal graph. The process parameter sequence can include time series data of key process parameters such as temperature, pressure, and flow rate. The defect distribution characteristics are the characteristics obtained in the previous step that have the ability to characterize spatiotemporal evolution.

[0137] Nonlinear independent component analysis is used to decouple the process parameter sequence and defect distribution characteristics to obtain potential causal variables. Nonlinear independent component analysis can decompose complex observation data into independent potential factors, which helps to discover potential causal relationships in the data.

[0138] The causal strength between potential causal variables is calculated based on the conditional mutual information criterion. Conditional mutual information can measure the degree of mutual dependence between two variables while considering the influence of other variables. By calculating the conditional mutual information between different potential causal variables, the causal strength between them can be quantified.

[0139] The scoring search algorithm is used to optimize the causal graph structure and obtain a causal graph containing process parameter nodes and defect distribution feature nodes. The scoring search algorithm continuously adjusts the graph structure through an iterative optimization process to maximize a certain scoring function (such as the Bayesian Information Criterion). In this way, a causal graph structure that can explain the observed data and has good generalization ability can be obtained.

[0140] Causal reasoning is performed based on the causal graph, and the causal effect of process parameters on defect distribution is calculated through intervention calculus. Intervention calculus simulates the situation of human intervention in a certain process parameter. By comparing the changes in defect distribution before and after the intervention, the causal effect of process parameters on defect distribution can be quantified.

[0141] Compare the observed values ​​and counterfactual predicted values ​​to evaluate the degree of abnormality of process parameters. Counterfactual predicted values ​​refer to the defect distribution predicted by the model under the assumption that the process parameters are not abnormal. By comparing the actually observed defect distribution with the counterfactual predicted values, the degree of abnormality of the process parameters can be evaluated.

[0142] Combined with the causal effect scores of multiple time steps, the dynamic threshold method is used for anomaly detection. The dynamic threshold method can adaptively adjust the threshold of anomaly detection according to the statistical characteristics of historical data, improving the accuracy and robustness of detection.

[0143] When a process parameter anomaly is detected, the defect distribution characteristics at different times in the historical defect data are used as graph nodes. Each node represents the defect distribution characteristics of a time step.

[0144] The time-varying adjacency matrix of the dynamic graph structure is constructed based on feature similarity. The similarity between defect distribution features at different time steps can be calculated using metrics such as cosine similarity or Euclidean distance, and the weight of the adjacency matrix can be set based on the similarity. The time-varying adjacency matrix can capture the dynamic relationship of defect distribution features over time.

[0145] The long-term and short-term dependencies are captured through the bidirectional temporal attention layer of the time series attention mechanism. The bidirectional temporal attention layer can consider both past and future information at the same time to better capture the long-term and short-term dependencies in the time series. The attention mechanism can adaptively assign weights to information at different time steps and highlight important time steps.

[0146] Gated graph convolution is used to dynamically update the feature representation of graph nodes. Gated graph convolution combines graph convolution and gating mechanism, and can dynamically update the feature representation of nodes based on the neighborhood information and historical information of the nodes. This method can effectively capture the propagation and evolution process of defect distribution features in the graph structure.

[0147] Finally, based on the updated node feature representation, the evolution law of the defect distribution is predicted. A multi-layer perceptron or other prediction model can be used to take the updated node features as input to predict the defect distribution characteristics of future time steps.

[0148] Suppose we have a data set of semiconductor wafer manufacturing process. The feature map has a size of 100x100, which represents the spatial distribution of the wafer surface. There are 5 types of defects, represented by integers from 0 to 4. The time dimension is expanded to 30 time steps, each time step corresponds to one hour of production data. The process parameters include temperature, pressure and gas flow rate.

[0149] In the depthwise separable convolution step, we use a 3x3 convolution kernel for depthwise convolution, followed by a 1x1 convolution kernel for pointwise convolution. The one-dimensional convolution uses a convolution kernel of length 5 with a stride of 1. Nonlinear independent component analysis decomposes the process parameters and defect distribution characteristics into 10 potential causal variables.

[0150] In the process of causal graph construction, we set the threshold of conditional mutual information to 0.05, and only causal relationships exceeding this threshold will establish edges in the graph. The scoring search algorithm uses the Bayesian information criterion as the scoring function and iterates 500 times to obtain the optimal causal graph structure.

[0151] In the anomaly detection phase, we use the causal effect scores of the past 24 hours to calculate a dynamic threshold. If the causal effect score of the current time step exceeds the mean of the scores of the past 24 hours plus two standard deviations, it is considered an anomaly.

[0152] In the graph neural network analysis phase, we construct a dynamic graph containing the past 72 hours (i.e., 72 nodes). The similarity between nodes is calculated using cosine similarity, and edges are established between nodes with similarity greater than 0.8. The bidirectional temporal attention layer uses 8 attention heads and the hidden layer dimension is 64. The gated graph convolution layer uses two layers, and the output dimension of each layer is 32.

[0153] Finally, we use a three-layer multi-layer perceptron with hidden layer dimensions of 64 and 32 and an output layer dimension of 500 (corresponding to the five defect types in the 100x100 feature map) to predict the defect distribution characteristics for the next hour.

[0154] The method proposed in the present invention realizes the precise detection of process parameter anomalies and the accurate prediction of defect distribution evolution by combining causal reasoning and graph neural network. It can capture the complex causal relationship between process parameters and defect distribution, and improve the accuracy and interpretability of anomaly detection. By using deep separable convolution and one-dimensional convolution, the spatiotemporal characteristics of defect distribution can be effectively extracted while maintaining low computational complexity. The application of residual connection further enhances the effect of feature extraction, so that the model can better capture the spatiotemporal evolution law of defect distribution. The dynamic graph structure and time series attention mechanism are adopted to adapt to the dynamic characteristics of defect distribution over time. This flexible modeling method improves the accuracy of defect distribution evolution prediction and provides strong support for production process optimization and quality control.

[0155] In an optional embodiment,

[0156] According to the causal graph and the evolution law, the process parameter constraints are set in combination with the expert knowledge base, and the steps of using the reinforcement learning method to adaptively adjust the process parameters include:

[0157] The sensitivity range of process parameter adjustment is determined based on the edge weights from process parameter nodes to defect nodes in the causal graph. The change rate limit of process parameters is set through a sliding time window based on the evolution law. Combined with expert experience, a multi-level constraint knowledge base including physical limits of process parameters, coupling relationships of process parameters and process rules is constructed.

[0158] Construct a reinforcement learning environment and design a state space, which includes the current process parameter value, process parameter change trend, process state of historical statistical characteristics, defect spatial distribution, defect type proportion, defect quantity statistics, defect feature state, production load, equipment state, and predicted state of evolution law; map the process parameter adjustment amount to the normalized interval and consider the coordinated adjustment of multiple process parameters to construct the action space, and design a reward function including defect improvement index, process parameter stability score and adjustment efficiency evaluation;

[0159] Design a dual network architecture, use a multi-head attention mechanism to process multimodal input features, perform knowledge fusion through soft constraint embedding and expert rule injection, generate a process parameter adjustment strategy distribution that satisfies the multi-level constraint knowledge base, use a temporal convolutional network to extract state time features and evaluate the long-term impact of adjustment actions, perform value correction in combination with the prediction of evolution laws, and calculate sample priorities based on time difference errors and historical sample occurrence frequencies to build an experience replay pool;

[0160] For the process parameter adjustment strategy distribution, a dynamic balance between exploration and utilization is achieved based on the process parameter sensitivity, and the sampled process parameter adjustment actions are subject to multiple constraint verifications of physical limits, expert rules and coupling relationships. The importance sampling method is used to optimize the sampling of the strategy distribution, and the proximal strategy optimization algorithm is used to gradient update the policy network and the value network. The process parameters are adaptively adjusted through online anomaly detection and adaptive learning rate.

[0161] Exemplarily, the sensitivity range of process parameter adjustment is first constructed based on the cause-effect graph. The influence of each process parameter on the defect is determined by analyzing the edge weights from the process parameter node to the defect node in the cause-effect graph. For example, the edge weight from the temperature parameter to the pore defect is 0.8, indicating that the temperature adjustment has a strong influence on the pore defect. The sensitivity range of parameter adjustment is set based on these weight values, such as the adjustment range of the temperature parameter is plus or minus 5% of the current value.

[0162] Set the parameter change rate limit in combination with the process evolution law. Use a 30-minute sliding time window to analyze the parameter change trend within the window and limit the parameter adjustment speed. For example, the maximum change rate of the temperature parameter within the window does not exceed 2 degrees per minute. At the same time, establish a multi-level constraint knowledge base, including physical limit constraints such as temperature range between 500-800 degrees, coupling relationship constraints such as the ratio of temperature to pressure change between 1.2-1.5, and process rule constraints such as the heating rate does not exceed the critical value.

[0163] When building a reinforcement learning environment, design a multidimensional state space. The process state includes the current values ​​of parameters such as temperature and pressure, the change trend in the past hour, and statistical characteristics such as the mean variance in the past 24 hours. The defect feature state includes information such as the distribution density map of defects on the workpiece surface, the proportion of each type of defect, and the time series change of the total number of defects. The environmental state includes operating parameters such as the production load percentage and the temperature of key equipment, as well as the process trend in the next 4 hours predicted based on the evolution law.

[0164] The action space design uses normalization processing to map the process parameter adjustment amount to the range from negative one to positive one. Considering the coordinated adjustment of multiple parameters, a joint action space is constructed. The reward function design includes three aspects: the defect improvement index accounts for 50%, which evaluates the degree of reduction in the number of defects after adjustment; the process parameter stability score accounts for 30%, which examines the fluctuation range of parameters; the adjustment efficiency evaluation accounts for 20%, which measures the number of adjustments required to achieve the goal.

[0165] In the network architecture design, a dual network structure is adopted. A multi-head attention mechanism is used to process different modal inputs, such as 8-head attention for process parameters and 4-head attention for defect features. Physical limit information is integrated through the soft constraint embedding layer, and process knowledge is integrated through the expert rule injection layer. A time series convolutional network is used to extract state sequence features with 3 convolutional layers to evaluate the long-term impact of adjustment actions. Value correction is performed based on the predicted value of the evolution law, and the priority is calculated based on the time difference error and sample frequency to build an experience replay pool with a capacity of 100,000 data.

[0166] Finally, the process parameters are adaptively adjusted. The exploration probability is set based on sensitivity, with an initial value of 0.3, which is gradually reduced as the training progresses. Multiple verifications are performed on the sampled adjustment actions to ensure that the physical limits, expert rules, and coupling relationship requirements are met. Importance sampling is used to optimize the strategy distribution, and the proximal strategy optimization algorithm is used to update the network parameters. Abnormal parameter fluctuations are identified through online anomaly detection, and the learning rate is adaptively adjusted to achieve stable training.

[0167] The present invention ensures that process parameter adjustment satisfies physical constraints and expert experience through a multi-level constraint knowledge base and a soft constraint embedding mechanism, improves the reliability and safety of the adjustment plan, and effectively avoids quality problems caused by improper parameter adjustment; based on the dual architecture of a multi-head attention mechanism and a temporal convolutional network, deep fusion of multimodal features and extraction of temporal features are achieved, which enhances the model's ability to understand complex process states and the accuracy of assessing long-term impacts, and significantly improves the timeliness of the adjustment strategy; a dynamic balance exploration mechanism and a multiple constraint verification method are adopted, combined with online anomaly detection and an adaptive learning rate, to achieve precise control of process parameters, improve product quality stability, reduce the defect rate, and at the same time ensure the smoothness and efficiency of the adjustment process.

[0168] Figure 2 FIG. 1 is a schematic diagram of a system for analyzing defects in wafer manufacturing based on big data according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0169] The first unit is used to collect process parameter information of wafer manufacturing, including photolithography process alignment parameters, photoresist coating thickness parameters, exposure energy parameters, etching gas flow parameters, RF power parameters, bias voltage parameters, thin film deposition rate parameters and chamber pressure parameters; collect defect scanning images of the wafer surface through multiple distributed defect detection devices; collect manufacturing process data information, including process yield information, wafer electrical test information and equipment status information;

[0170] The second unit is used to extract defect features from the defect scan image using a multi-scale feature pyramid network including a deformable convolution module, and fuse the defect features, process parameter information and manufacturing process data information to obtain fused features; the fused features are analyzed through a deep learning network, and the deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches, and obtain defect type recognition results and feature maps of defect spatial distribution;

[0171] The third unit is used to construct a causal graph of process parameter changes and defect distribution evolution based on the feature map using an abnormality monitoring network including a causal reasoning module, detect process parameter anomalies according to the causal graph, and when process parameter anomalies are detected, analyze historical defect data using a graph neural network including a time series attention mechanism, and capture the evolution law of defect distribution through a dynamic graph structure; and set process parameter constraints based on the causal graph and the evolution law in combination with an expert knowledge base, and use a reinforcement learning method to adaptively adjust the process parameters.

[0172] According to a third aspect of the embodiments of the present invention,

[0173] An electronic device is provided, comprising:

[0174] processor;

[0175] a memory for storing processor-executable instructions;

[0176] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0177] A fourth aspect of the embodiments of the present invention is:

[0178] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0179] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing defects in wafer manufacturing based on big data, characterized in that: include: Collect process parameter information of wafer manufacturing, including photolithography process alignment parameters, photoresist coating thickness parameters, exposure energy parameters, etching gas flow parameters, RF power parameters, bias voltage parameters, thin film deposition rate parameters and chamber pressure parameters; collect defect scanning images of wafer surface through multiple distributed defect detection equipment; collect manufacturing process data information, including process yield information, wafer electrical test information and equipment status information; Extracting defect features from the defect scanned image using a multi-scale feature pyramid network including a deformable convolution module, and fusing the defect features, process parameter information, and manufacturing process data information to obtain fused features; The fusion features are analyzed through a deep learning network. The deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches, and obtain defect type recognition results and feature maps of defect spatial distribution; Based on the feature map, a causal graph of process parameter changes and defect distribution evolution is constructed using an abnormality monitoring network including a causal reasoning module. Process parameter anomalies are detected based on the causal graph. When process parameter anomalies are detected, historical defect data are analyzed using a graph neural network including a time series attention mechanism. The graph neural network captures the evolution law of defect distribution through a dynamic graph structure. Based on the causal graph and the evolution law, process parameter constraints are set in combination with an expert knowledge base, and reinforcement learning methods are used to adaptively adjust the process parameters.

2. The method according to claim 1, characterized in that The steps of extracting defect features from the defect scan image using a multi-scale feature pyramid network including a deformable convolution module, and fusing the defect features, process parameter information, and manufacturing process data information to obtain fused features include: Construct a multi-scale feature pyramid network including a deformable convolution module; the deformable convolution module includes a dual-branch attention network, the dual-branch attention network includes a spatial branch for capturing local structural information and a channel branch for capturing global semantic information, an offset is calculated based on the spatial branch and the channel branch, and the offset is used to guide the feature extraction of the deformable convolution; the multi-scale feature pyramid network includes a plurality of feature extraction layers, each feature extraction layer extracts local features through the deformable convolution module to obtain a feature tensor, upsamples deep features, and fuses the upsampled features with the current layer feature tensor to obtain a feature pyramid output; a feature complementary module is set in the multi-scale feature pyramid network to connect adjacent level features, and the feature complementary module performs adaptive weighted fusion of high-level features and low-level features processed by convolution to obtain defect features; Standardize the process parameter information to obtain process parameter features; extract features from the manufacturing process data information to obtain manufacturing data features; A feature semantic mapping matrix is ​​constructed based on the knowledge of wafer manufacturing domain. The feature semantic mapping matrix is ​​obtained by calculating the distance between different features in a predefined semantic space, and the predefined semantic space includes a defect morphology feature space, a process control feature space, and a quality indicator feature space. The feature semantic mapping matrix is ​​used to characterize the semantic correlation between different modal features. According to the feature semantic mapping matrix, the defect features, process parameter features, and manufacturing data features are semantically projected to map different modal features to a unified semantic representation space. A bidirectional mapping function is used to align cross-modal features. The bidirectional mapping function includes a first mapping matrix and a second mapping matrix. The first mapping matrix is ​​used to align source modal features, and the second mapping matrix is ​​combined with the feature semantic mapping matrix to align target modal features. The features after bidirectional mapping alignment are residually connected with the original features and nonlinearly transformed to obtain fused features.

3. The method according to claim 1, characterized in that The fusion features are analyzed by a deep learning network. The deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches. The steps of obtaining defect type recognition results and feature maps of defect spatial distribution include: Constructing a deep learning network to analyze the fusion features, wherein the deep learning network adopts a shared encoder-multi-branch decoder network structure to construct a multi-task learning framework, wherein the shared encoder extracts a deep representation of the fusion features, and the multi-branch decoder includes a defect classification branch, a defect segmentation branch, and a defect evolution prediction branch; The defect classification branch converts the feature map into a vector representation through a global average pooling layer and then completes the classification with a fully connected layer. The defect classification branch includes a feature recalibration module based on an attention mechanism. The feature recalibration module is used to adjust the weights of different channel features and output a probability vector of the defect type. The defect segmentation branch adopts an encoder-decoder structure with a jump connection, and a channel attention module is set at the jump connection to multiply the binary defect mask obtained by segmentation with the original feature map to obtain the local features of the defect area. The defect evolution prediction branch receives the probability vector of the defect type and the local features of the defect area as input, and extracts the spatiotemporal features through a spatiotemporal convolutional network. The spatiotemporal convolutional network includes a temporal dependency modeling module based on an attention mechanism, and the temporal dependency modeling module is used to capture the long-term evolution law of the defect. The graph structure is constructed based on the morphological features and spatial distribution of defects. The similarity of graph structures of different batches is calculated through a multi-view matching mechanism, and the batch consistency loss is obtained by combining uncertainty assessment and historical feature matching. A task graph containing task relevance is constructed, and the task graph is modeled through a graph attention network to obtain inter-task dependencies. The task weights are dynamically adjusted according to changes in the losses of each task. The prediction results of the defect classification branch, the prediction results of the defect segmentation branch, the prediction results of the defect evolution prediction branch, and the batch consistency loss are jointly optimized to obtain a feature map containing defect type recognition results and defect spatial distribution.

4. The method according to claim 3, characterized in that: The steps of constructing a graph structure based on the morphological features and spatial distribution of defects, calculating the similarity of graph structures of different batches through a multi-view matching mechanism, and combining uncertainty assessment and historical feature matching to obtain batch consistency loss include: A hierarchical graph structure is constructed based on the morphological characteristics and spatial distribution of defects. The hierarchical graph structure includes a global layer, a sub-layer and a node layer. The nodes of the node layer represent defect instances, and the edges of the node layer represent the spatial relationships of defects. A density clustering algorithm is used to group spatially adjacent defect groups into a sub-graph structure. The sub-graph structure contains centrality features and intra-group sorting features. Local information is transferred from the bottom to the top between the global layer, the sub-layer and the node layer through a cross-hierarchical feature transfer mechanism, and global constraints are fed back from the top to the bottom. The connection strength of the edges of the node layer is dynamically adjusted based on the attention gating mechanism, and potential topological changes are predicted in combination with the defect evolution trend. The graph structure differences at adjacent moments are controlled within the target threshold range through topological structure smoothing constraints. According to the constructed hierarchical graph structure, a multi-view matching mechanism including node-level matching, subgraph-level matching and global-level matching is constructed, and the matching results of different levels are fused through adaptive weighting; an adversarial generative network is used to generate matching samples and the feature encoding parameters of the multi-view matching mechanism are optimized through adversarial training, a temporal dependency graph between batches is established, and a gated recurrent unit is introduced to capture long-term evolution laws; Based on the multi-view matching results, a hierarchical loss function including node loss, subgraph loss and global loss is constructed. The curriculum learning strategy is used to dynamically adjust the loss weights of each layer. The gradients of loss functions at different levels are collaboratively optimized through the cross-layer loss complementary mechanism. An uncertainty assessment module based on the Bayesian framework is established to adaptively adjust the loss weights of samples with different confidence levels. A historical batch feature library is constructed to store matching patterns and maintain and update them through the attention mechanism. The hierarchical loss function, uncertainty assessment results and historical feature matching information are fused and optimized to obtain the batch consistency loss.

5. The method according to claim 3, characterized in that: The steps of constructing a task graph including task correlations, modeling the task graph through a graph attention network to obtain dependencies between tasks, dynamically adjusting task weights according to changes in the losses of each task, and jointly optimizing the prediction results of the defect classification branch, the defect segmentation branch, the defect evolution prediction branch, and the batch consistency loss to obtain a feature map including defect type recognition results and defect spatial distribution include: A mutual information neural estimator is used to calculate the mutual information of features between tasks, positive and negative sample pairs are constructed by double-sample random sampling, and a separator network is used to estimate the lower bound of the mutual information. A task relationship graph including defect classification task nodes, defect segmentation task nodes, and defect evolution prediction task nodes is constructed based on the mutual information, and the mutual information is used as the edge weight of the task relationship graph; Perform multi-layer graph attention modeling on the task relationship graph, use the dot product attention mechanism and combine it with position encoding to calculate the local dependencies between task nodes, use the graph convolution message passing mechanism and the gating unit to aggregate global context information, and learn the optimal combination of task features through a dynamic routing algorithm; Based on the results of graph attention modeling, a meta-learning optimizer is built to calculate the gradient cosine similarity and Hessian matrix between tasks, evaluate the spatial consistency of task representation, and dynamically update task weights using a model-independent meta-learning algorithm to obtain the optimized weights of each task. Construct a multi-level joint optimization framework, including: introducing cross entropy loss for the prediction results of the defect classification branch, introducing distance loss and boundary loss for the prediction results of the defect segmentation branch, and introducing temporal consistency loss for the prediction results of the defect evolution prediction branch; based on the optimization weights of each task, a weighted combination of the defect classification branch loss, the defect segmentation branch loss, the defect evolution prediction branch loss and the batch consistency loss is performed, and contrastive learning loss is introduced in the shared feature space to extract common features between tasks, and soft label information is transferred through knowledge distillation loss to obtain optimized multi-task features; A hierarchical feature fusion network is constructed, and the optimized multi-task features are aligned using spatial pyramid pooling and channel attention mechanism. Dynamic convolution kernel and bilinear pooling are used to adaptively fuse multi-scale features to generate a feature map containing defect type recognition results and defect spatial distribution.

6. The method according to claim 1, characterized in that Based on the feature graph, a causal graph of process parameter changes and defect distribution evolution is constructed using an abnormality monitoring network including a causal reasoning module, and process parameter anomalies are detected according to the causal graph. When process parameter anomalies are detected, historical defect data is analyzed using a graph neural network including a time series attention mechanism. The steps of capturing the evolution law of defect distribution through a dynamic graph structure of the graph neural network include: The feature map containing defect type and defect spatial distribution information is expanded in time dimension, and the feature map of each time step is processed by deep separable convolution to obtain local spatial context information. One-dimensional convolution is used to model the feature sequence along the time axis to obtain temporal dynamic features. The local spatial context information and the temporal dynamic features are fused through residual connection to obtain defect distribution features with the ability to characterize spatiotemporal evolution. The process parameter sequence and the defect distribution characteristics are respectively used as nodes of a causal graph, and the process parameter sequence and the defect distribution characteristics are decoupled by using nonlinear independent component analysis to obtain potential causal variables. The causal strength between the potential causal variables is calculated based on the conditional mutual information criterion, and the causal graph structure is optimized by using a scoring search algorithm to obtain a causal graph including process parameter nodes and defect distribution characteristic nodes; Causal reasoning is performed based on the causal graph, the causal effect of the process parameters on the defect distribution is calculated through intervention calculation, the degree of abnormality of the process parameters is evaluated by comparing the observed values ​​and the counterfactual predicted values, and the dynamic threshold method is used to perform abnormality detection in combination with the causal effect scores of multiple time steps; When a process parameter abnormality is detected, the defect distribution features at different times in the historical defect data are used as graph nodes, and a time-varying adjacency matrix of the dynamic graph structure is constructed based on feature similarity. The long-term and short-term dependencies are captured through the bidirectional temporal attention layer of the time series attention mechanism, and the gated graph convolution is used to dynamically update the feature representation of the graph nodes to predict the evolution law of the defect distribution.

7. The method according to claim 1, characterized in that According to the causal graph and the evolution law, the process parameter constraints are set in combination with the expert knowledge base, and the steps of using the reinforcement learning method to adaptively adjust the process parameters include: The sensitivity range of process parameter adjustment is determined based on the edge weights from process parameter nodes to defect nodes in the causal graph. The change rate limit of process parameters is set through a sliding time window based on the evolution law. Combined with expert experience, a multi-level constraint knowledge base including physical limits of process parameters, coupling relationships of process parameters and process rules is constructed. Construct a reinforcement learning environment and design a state space, which includes the current process parameter value, process parameter change trend, process state of historical statistical characteristics, defect spatial distribution, defect type proportion, defect quantity statistics, defect feature state, production load, equipment state, and predicted state of evolution law; map the process parameter adjustment amount to the normalized interval and consider the coordinated adjustment of multiple process parameters to construct the action space, and design a reward function including defect improvement index, process parameter stability score and adjustment efficiency evaluation; Design a dual network architecture, use a multi-head attention mechanism to process multimodal input features, perform knowledge fusion through soft constraint embedding and expert rule injection, generate a process parameter adjustment strategy distribution that satisfies the multi-level constraint knowledge base, use a temporal convolutional network to extract state time features and evaluate the long-term impact of adjustment actions, perform value correction in combination with the prediction of evolution laws, and calculate sample priorities based on time difference errors and historical sample occurrence frequencies to build an experience replay pool; For the process parameter adjustment strategy distribution, a dynamic balance between exploration and utilization is achieved based on the process parameter sensitivity, and the sampled process parameter adjustment actions are subject to multiple constraint verifications of physical limits, expert rules and coupling relationships. The importance sampling method is used to optimize the sampling of the strategy distribution, and the proximal strategy optimization algorithm is used to gradient update the policy network and the value network. The process parameters are adaptively adjusted through online anomaly detection and adaptive learning rate.

8. A system for analyzing defects in wafer manufacturing based on big data, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect process parameter information of wafer manufacturing, including photolithography process alignment parameters, photoresist coating thickness parameters, exposure energy parameters, etching gas flow parameters, RF power parameters, bias voltage parameters, film deposition rate parameters and chamber pressure parameters; and collect defect scanning images of the wafer surface through multiple distributed defect detection devices; Collect manufacturing process data information, including process yield information, wafer electrical test information and equipment status information; The second unit is used to extract defect features from the defect scanning image using a multi-scale feature pyramid network including a deformable convolution module, and fuse the defect features, process parameter information and manufacturing process data information to obtain fused features; The fusion features are analyzed through a deep learning network. The deep learning network uses a multi-task learning framework to simultaneously complete defect classification tasks, defect segmentation tasks, and defect evolution prediction tasks, and introduces consistency loss based on graph matching to establish associations between defect samples from different batches, and obtain defect type recognition results and feature maps of defect spatial distribution; The third unit is used to construct a causal graph of process parameter changes and defect distribution evolution based on the feature map using an abnormality monitoring network including a causal reasoning module, detect process parameter anomalies according to the causal graph, and when process parameter anomalies are detected, analyze historical defect data using a graph neural network including a time series attention mechanism, and capture the evolution law of defect distribution through a dynamic graph structure; and set process parameter constraints based on the causal graph and the evolution law in combination with an expert knowledge base, and use a reinforcement learning method to adaptively adjust the process parameters.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images