Machine learning based semiconductor yield collaborative prediction method, system, device, and medium

Abstract

This application relates to a machine learning-based collaborative prediction method, system, device, and medium for semiconductor yield. The method includes: first, acquiring time-series data of machine sensor data and wafer defect image data for each process in semiconductor manufacturing; then, extracting the time feature vectors and image feature vectors of the two types of data; combining wafer identifiers and processing timestamps to complete cross-process feature alignment, resulting in an enhanced cross-process feature set; based on a constructed spatial network topology, mapping the enhanced cross-process features to the topology to generate spatiotemporal graph data; extracting the time-series features of the spatiotemporal graph to obtain a dynamic evolution sequence of process parameters; using a federated averaging algorithm to encrypt and aggregate multi-faceted gradients to train a yield prediction model; finally, inputting the enhanced cross-process features and the current spatial network state into the model to locate high-risk process machines and time intervals, and outputting defect probability prediction results. This method can improve wafer yield and production efficiency while reducing costs.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing, and in particular relates to a method, system, device and medium for collaborative prediction of semiconductor yield based on machine learning. Background Technology

[0002] With the continuous development of advanced semiconductor manufacturing technologies, wafer fabrication processes are becoming increasingly precise and complex. As a result, production line sensing and monitoring equipment, as well as visual inspection equipment, are widely used. Data-driven wafer yield inspection technology is gradually becoming a core management tool in the industry. This type of technology relies on machine process parameters and wafer surface defect images to achieve automated yield determination. It features high efficiency and can replace manual sampling inspection. Currently, the industry generally adopts the traditional yield management model of single-machine offline monitoring and centralized post-defect inspection.

[0003] In traditional semiconductor yield control solutions, the production end typically monitors the process parameters of a single manufacturing machine independently at thresholds. After the wafer completes the entire process, defect data is collected uniformly through visual inspection equipment and the yield is calculated. Model training often adopts a local training method that centrally aggregates data in the factory area, and only performs post-event statistical analysis on the defects that have already occurred, without incorporating the wafer transfer process between different machines into the yield influencing factors system.

[0004] This traditional approach of independent control of individual machines and centralized testing afterward has significant limitations: on the one hand, defect discovery is significantly delayed, making it impossible to trace the key processes and time points where defects occur; on the other hand, data from each machine is fragmented, forming data silos, and centralized data processing also poses risks to data security and privacy compliance. Furthermore, the analysis model based on a single data dimension is difficult to accurately identify potential yield risks and cannot achieve early prediction and location of high-risk processes and abnormal periods. Summary of the Invention

[0005] Therefore, it is necessary to provide a machine learning-based semiconductor yield collaborative prediction method, system, equipment, and medium that can accurately capture the process coupling relationship between equipment and the impact of wafer transfer on yield, effectively solving the shortcomings of traditional defect detection that is lagging and unable to trace the source of defects.

[0006] Firstly, this application provides a machine learning-based collaborative prediction method for semiconductor yield, including:

[0007] The system acquires time-series data from machine sensors and wafer defect image data for each process in semiconductor manufacturing, forming a multi-source heterogeneous dataset.

[0008] Temporal feature vectors and image feature vectors are extracted from multi-source heterogeneous datasets respectively. Cross-process feature alignment is performed on the temporal feature vectors and image feature vectors based on wafer identifiers and processing timestamps to obtain an enhanced cross-process feature set.

[0009] Based on the flow path of wafers in the production line, a spatial network topology is constructed with manufacturing equipment as nodes and wafer flow relationships as edges. The enhanced cross-process feature set is mapped to the spatial network topology according to the equipment nodes through which the wafers flow, generating spatiotemporal graph data with node features and edge features.

[0010] Temporal features are extracted from the spatiotemporal graph data to obtain the dynamic evolution sequence of process parameters. Based on the dynamic evolution sequence, a collaborative model is trained and the gradients of multiple models are encrypted and aggregated using a federated averaging algorithm to obtain the trained yield prediction model.

[0011] By enhancing the cross-process feature set and inputting the current spatial network state into the trained yield prediction model, the high-risk process machines and time intervals where potential defects occur are located, and the defect probability prediction results are obtained.

[0012] In one embodiment, time feature vectors and image feature vectors are extracted from the multi-source heterogeneous dataset, respectively. Cross-process feature alignment is performed on the time feature vectors and image feature vectors based on the wafer identifier and processing timestamp to obtain an enhanced cross-process feature set, including:

[0013] The process parameter time series are extracted from the multi-source heterogeneous dataset, and a fixed-length process parameter time series is truncated using the sliding window method to obtain the time feature vector.

[0014] Semiconductor wafer defect images are collected from multi-source heterogeneous datasets. The wafer defect images are input into a preset convolutional neural network. The deep features of the wafer defect morphology are extracted through the convolutional neural network to obtain image feature vectors.

[0015] Using the wafer identifier as the unique matching identifier and the processing timestamp as the time sequence matching basis, the time feature vector and image feature vector are aligned across processes to obtain the aligned cross-process feature set.

[0016] The intersection of time features and image features is extracted from the aligned cross-process feature set as a fusion feature subset. The weights of each feature component of the fusion feature subset are assigned based on the attention mechanism, and the weighted feature vector is obtained after weighted calculation.

[0017] The weighted feature vector is input into the preset defect classification model to output the defect classification label. If the defect probability value corresponding to the defect classification label exceeds the preset defect judgment threshold, an alarm signal is generated.

[0018] The associated semiconductor manufacturing process parameters are extracted from the alarm signals and used as associated process parameters. The gradient descent algorithm is then used to iteratively optimize the associated process parameters to obtain the optimized set of process parameters.

[0019] Adjust the processing timestamps of the corresponding processes based on the optimized set of process parameters, and substitute the adjusted processing timestamps into the truncation process of the sliding window method to obtain the updated time feature vector.

[0020] Subsequence features are extracted from the updated temporal feature vector. The extracted subsequence features and image feature vectors are then re-aligned according to the cross-process alignment method to obtain an enhanced cross-process feature set.

[0021] In one embodiment, the enhanced cross-process feature set is mapped to a spatial network topology based on the machine nodes through which the wafer flows, generating spatiotemporal graph data with node and edge features, including:

[0022] Acquire the wafer transfer path data in the production line; the transfer path data includes manufacturing machine nodes and wafer transfer edge information between machines.

[0023] A spatial network topology is constructed based on the flow path data; the spatial network topology uses manufacturing machines as nodes and wafer flow relationships as edges.

[0024] The enhanced cross-process feature set is mapped into the spatial network topology according to the manufacturing machine nodes through which the wafer flows.

[0025] In the spatial network topology, each manufacturing machine node is assigned a corresponding node feature, and each wafer transfer edge is assigned a corresponding edge feature, generating spatiotemporal graph data with node and edge features.

[0026] In one embodiment, after generating spatiotemporal graph data with node and edge features, the method further includes:

[0027] A graph neural network is used to process the spatiotemporal graph data to obtain the node state vectors corresponding to each manufacturing machine node.

[0028] Based on the node state vector and the edge characteristics of the corresponding wafer transfer edge, calculate the state transition probability of the wafer on the transfer path.

[0029] The critical flow path of the wafer between each manufacturing machine node is determined based on the state transition probability.

[0030] In one embodiment, the state transition probability of the wafer along the transfer path is calculated using the following formula:

[0031]

[0032] in, Indicates machine node To machine node The state transition probability, Represents a node The node state vector, Represents the adaptive projection matrix of edge features. Indicates machine node to The edge features corresponding to the wafer transfer edges between them This indicates the attention weight of process coupling between machines. Represents a small constant for numerical smoothing. Indicates historical wafers from arrive The actual circulation frequency, Indicates the wafer is in , Average turnover time between Indicates the time-series flow weighting coefficient. This represents the Sigmoid activation function.

[0033] In one embodiment, the trained yield prediction model is obtained through the following process:

[0034] Obtain the interaction and association information between each manufacturing machine node in the spatiotemporal graph data, and construct a node dependency graph based on the interaction and association information.

[0035] Interactive association information includes the strength of process coupling between nodes, the frequency of data interaction, and the flow association relationship.

[0036] A temporal graph neural network is used to extract temporal features from the node dependency graph and key flow paths, capturing the dynamic evolution sequence of the process parameters corresponding to each manufacturing machine node over time.

[0037] The dynamic evolution sequence of process parameters includes the drift trend and time series correlation characteristics of process parameters.

[0038] Collaborative model training is performed based on the dynamic evolution sequence of process parameters, and encrypted gradient updates are obtained by encrypting the multi-party local model gradient data generated during the training process.

[0039] The encrypted gradient updates are globally aggregated using a federated averaging algorithm to obtain the global gradient update result.

[0040] If the global gradient update result reaches the preset gradient convergence threshold, the trained yield prediction model is obtained.

[0041] In one embodiment, the yield prediction model trained with enhanced cross-process feature sets and the current spatial network state is used to locate high-risk process machines and time intervals where potential defects may occur, and to obtain defect probability prediction results, including:

[0042] Obtain the dynamic process parameters corresponding to the current spatial network status.

[0043] The random forest algorithm is used to fuse and analyze dynamic process parameters and enhanced cross-process feature sets to obtain preliminary defect risk indicators.

[0044] By inputting the preliminary defect risk index into the trained yield prediction model, the defect probability prediction results for future batches of wafers are obtained.

[0045] Based on the defect probability prediction results, correlation analysis is performed on the process data of each critical flow path covering the process to determine whether there are abnormal fluctuations in process parameters of each high-risk process machine.

[0046] If it is determined that there are abnormal fluctuations in the process parameters of a machine tool in a high-risk process, then the time interval in which the potential defects occurred corresponding to the abnormal fluctuations is located.

[0047] Extract historical process log data of the corresponding high-risk process equipment within the time interval of potential defect occurrence, and generate process optimization and adjustment instructions based on the historical process log data.

[0048] The current spatial network status is dynamically updated using process optimization adjustment commands to obtain optimized defect probability prediction results.

[0049] Secondly, this application also provides a machine learning-based semiconductor yield collaborative prediction system, the system comprising:

[0050] The data acquisition module is used to acquire time-series data of machine sensor data and wafer defect image data of each process in semiconductor manufacturing, forming a multi-source heterogeneous dataset.

[0051] The feature extraction module is used to extract time feature vectors and image feature vectors from multi-source heterogeneous datasets respectively. Based on the wafer identifier and processing timestamp, the time feature vectors and image feature vectors are aligned across processes to obtain an enhanced cross-process feature set.

[0052] The spatiotemporal graph generation module is used to construct a spatial network topology structure with manufacturing equipment as nodes and wafer flow relationships as edges based on the wafer flow path in the production line. It maps the enhanced cross-process feature set to the spatial network topology structure according to the equipment nodes through which the wafer flows, and generates spatiotemporal graph data with node features and edge features.

[0053] The model training module is used to extract time-series features from spatiotemporal graph data to obtain the dynamic evolution sequence of process parameters. Based on the dynamic evolution sequence, collaborative model training is performed, and the gradients of multiple models are encrypted and aggregated through the federated averaging algorithm to obtain the trained yield prediction model.

[0054] The defect localization module is used to input the enhanced cross-process feature set and the current spatial network status into the trained yield prediction model to locate the high-risk process machines and time intervals where potential defects may occur, and obtain the defect probability prediction results.

[0055] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0056] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method.

[0057] The aforementioned machine learning-based semiconductor yield collaborative prediction method, system, computer equipment, and storage medium first acquire time-series data of machine sensor data and wafer defect image data from each semiconductor manufacturing process, integrating these two types of data to form a multi-source heterogeneous dataset. Then, time feature vectors and image feature vectors are extracted from this multi-source heterogeneous dataset. Using wafer identifiers as the unique matching criterion and processing timestamps as the time-series matching standard, cross-process feature alignment processing is performed on the extracted time feature vectors and image feature vectors to eliminate feature misalignment between different processes, ultimately obtaining an enhanced cross-process feature set. Based on the actual flow path of the wafer in the production line, a spatial network topology is constructed with manufacturing machines as nodes and wafer flow relationships as edges. The obtained enhanced cross-process feature set is mapped one by one to this spatial network topology according to the machine nodes through which the wafer flows, assigning each machine node an attribute. Assigning corresponding node features and edge features to each wafer flow edge generates spatiotemporal graph data with node and edge features. Temporal features are extracted from the generated spatiotemporal graph data to capture the changing patterns of process parameters of each manufacturing machine over time, obtaining a dynamic evolution sequence of process parameters. Based on this dynamic evolution sequence, collaborative model training is performed. During training, the local model gradient data generated by multiple parties is encrypted, and the encrypted gradient updates are globally aggregated using a federated averaging algorithm to obtain the global gradient update result. Once the global gradient update result reaches a preset convergence threshold, the trained yield prediction model is obtained. Finally, the enhanced cross-process feature set and the current spatial network state are input into the trained yield prediction model. Through model computation, high-risk process machines and corresponding time intervals where potential defects occur are located, ultimately outputting the defect probability prediction result for future batches of wafers. This method integrates and aligns the features of multi-source heterogeneous data, effectively fusing time-series data from machine sensors with wafer defect image data. This solves the problems of fragmented multi-source data and low feature utilization in traditional methods. By constructing a spatial network topology and generating spatiotemporal graph data, it organically combines wafer transfer paths with process parameter features, accurately capturing the process coupling relationship between machines and the impact of wafer transfer on yield. The federated averaging algorithm is used to encrypt and aggregate the gradients of multiple models, ensuring data privacy, security, and compliance across multiple plants while enabling collaborative training of models and improving their generalization ability and prediction accuracy. Through the yield prediction model, it enables early location of high-risk process machines and defect time intervals, effectively solving the shortcomings of traditional defect detection, such as lag and inability to trace the source of defects. This improves wafer production yield and efficiency while reducing production costs. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart of a machine learning-based semiconductor yield collaborative prediction method provided in an embodiment of the present invention;

[0060] Figure 2 The diagram below shows the structure of a machine learning-based semiconductor yield collaborative prediction system provided in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] In one embodiment, such as Figure 1 As shown, this application provides a machine learning-based semiconductor yield co-prediction method, which may include the following steps:

[0063] Step S101: Obtain time series data of machine sensor data and wafer defect image data of each process in semiconductor manufacturing to form a multi-source heterogeneous dataset.

[0064] Specifically, sensor devices deployed inside the manufacturing equipment and around the production line collect real-time time-series data directly related to process execution. This includes equipment operating parameters (such as rotation speed, temperature, pressure, current, and voltage), process environment parameters (such as chamber humidity, gas flow rate, and concentration), and processing status parameters (such as wafer clamping position and processing time). This data is stored in continuous time series format, containing clear timestamps, parameter values, and collection frequency information, completely recording the entire process changes of a single batch of wafers from entering the process to leaving the process. Simultaneously, after each process is completed, online visual inspection equipment or offline sampling inspection equipment collects surface image data of the corresponding wafers. This image data covers information such as wafer surface microstructure, defect distribution, defect types (such as scratches, particles, pits, and cracks), defect location coordinates, and defect size. The aforementioned time-series data from machine sensors and wafer defect image data differ significantly in data format, data dimension, acquisition logic, and sampling frequency: time-series data consists of numerical one-dimensional or multi-dimensional sequences, while image data is matrix-type multi-dimensional visual information; time-series data is continuously acquired during equipment operation, while image data is acquired in a triggered manner based on wafer processing batches. This step associates and aggregates the time-series data and defect image data corresponding to the same wafer batch, converts them into a standardized data format, and classifies and stores them according to wafer identifier, processing batch, and process node, ultimately forming a multi-source heterogeneous dataset covering the temporal characteristics of the process and the visual characteristics of defect results.

[0065] Step S102: Extract time feature vectors and image feature vectors from the multi-source heterogeneous datasets respectively. Align the time feature vectors and image feature vectors across processes based on the wafer identifier and processing timestamp to obtain an enhanced cross-process feature set.

[0066] Furthermore, feature extraction is performed on the sensor time-series data in the multi-source heterogeneous dataset to obtain time feature vectors representing the time-series variation of process parameters. Feature extraction is also performed on the wafer defect images in the dataset to obtain image feature vectors representing the morphology and distribution of defects. Using the unique wafer identifier as the basis for batch association and the processing timestamp as the time-series benchmark, the time features and image features corresponding to the same wafer under different processes are matched and aligned to eliminate time-series deviations and feature misalignments between cross-process data. Through feature integration and regularization, a more complete cross-process feature set with enhanced information dimensions is formed.

[0067] Step S103: Based on the flow path of the wafer in the production line, construct a spatial network topology with manufacturing equipment as nodes and wafer flow relationship as edges. Map the enhanced cross-process feature set to the spatial network topology according to the equipment nodes through which the wafer flows, and generate spatiotemporal graph data with node features and edge features.

[0068] Schematic, based on the actual processing flow sequence of wafers on the production line, the manufacturing equipment corresponding to each process is abstracted as network nodes, and the flow relationship between wafers and equipment is abstracted as network edges, thus constructing a spatial network topology. The process features and defect features related to each equipment in the enhanced cross-process feature set are mapped to the corresponding network nodes, and the flow association features between wafers and equipment are assigned to the network edges, so that the feature data is combined with the spatial structure of the production line, and finally, spatiotemporal graph data that integrates spatial topology relationships, process features and time sequence information is generated.

[0069] Step S104: Extract time-series features from the spatiotemporal graph data to obtain the dynamic evolution sequence of process parameters. Based on the dynamic evolution sequence, perform collaborative model training and encrypt and aggregate the gradients of multiple models using the federated averaging algorithm to obtain the trained yield prediction model.

[0070] The temporal characteristics of the changes in process parameters of each machine tool with the processing progress are extracted from the spatiotemporal graph data to form a dynamic evolution sequence of process parameters that can reflect the continuous change law of process status. Based on this sequence, collaborative model training is carried out in a distributed data scenario. After each participant completes the forward calculation and gradient calculation of the model locally, the local gradient data is encrypted. Then, the encrypted gradients of multiple parties are globally aggregated and the model parameters are updated through the federated averaging algorithm. When the global gradient update result meets the preset convergence condition, the training is terminated and the trained yield prediction model is obtained.

[0071] Step S105 involves inputting the enhanced cross-process feature set and the current spatial network status into the trained yield prediction model to locate the high-risk process machines and time intervals where potential defects may occur, thereby obtaining the defect probability prediction result.

[0072] First, the enhanced cross-process feature set obtained from the processing is used as the basic input feature set. This feature set includes time feature vectors and image feature vectors after cross-process alignment, comprehensively representing the temporal variation of process parameters, defect morphology correlation features, and inter-process feature matching relationships of the same wafer in each process. Simultaneously, the spatial network state representing the actual operating status of the current production line is collected as supplementary input. This spatial network state includes real-time operating parameters of each manufacturing machine (such as current process progress, equipment load status, and process parameter fluctuation values), real-time wafer flow correlation between machines, node activity, and real-time strength of process coupling between processes, fully reflecting the dynamic changes in the production line's spatial topology. These two types of input features are synchronously transmitted to the trained yield prediction model. The model internally performs deep feature fusion of the input enhanced cross-process features and the current spatial network state, and combines the "machine process parameters-wafer defects-process coupling relationship" correlation rules learned during the training phase to perform defect induction probability inference calculations for each manufacturing machine. Based on the inference results, the model identifies manufacturing machines whose defect induction probability exceeds a preset risk assessment threshold, classifying them as high-risk process machines with potential defect generation. Simultaneously, by combining the wafer processing timestamps on the production line, the processing time periods and sequence flow of each high-risk process machine, the model accurately locates the potential defect generation time interval corresponding to each high-risk process machine, clarifying the start and end processing nodes and the specific processing stage of the corresponding process within that time interval. Finally, the model outputs quantitative defect probability prediction results, including the probability of defect occurrence in future batches of wafers within the corresponding time intervals of each high-risk process machine, the risk level score of each high-risk process machine, and the anomaly confidence level of the time interval.

[0073] The aforementioned machine learning-based semiconductor yield collaborative prediction method first acquires time-series data of machine sensor data and wafer defect image data from each semiconductor manufacturing process, integrating them to form a multi-source heterogeneous dataset. Then, it extracts the temporal and image feature vectors from both types of data, combining wafer identifiers and processing timestamps to align cross-process features, resulting in an enhanced cross-process feature set. Based on the wafer production line flow path, it constructs a spatial network topology consisting of machine nodes and flow edges, mapping the enhanced cross-process features to the topology to generate a spatiotemporal graph with node and edge features. It extracts the temporal features of the spatiotemporal graph to obtain a dynamic evolution sequence of process parameters, collaboratively training the model based on this sequence. The model is then trained by encrypting and aggregating multi-faceted gradients using a federated averaging algorithm. Finally, it inputs the enhanced cross-process features and the current spatial network state into the model to locate high-risk process machines and time intervals, outputting the defect probability prediction result. This method achieves effective fusion of multi-source data, solving the problems of traditional data fragmentation and low feature utilization; it accurately captures the coupling relationship of equipment by combining wafer transfer paths and process characteristics; it protects data privacy through federated averaging algorithms, improves model generalization ability and prediction accuracy; it locates defect risks in advance, solving the shortcomings of traditional detection lag and inability to trace the source, thereby improving wafer yield and production efficiency and reducing costs.

[0074] In one embodiment, time feature vectors and image feature vectors are extracted from the multi-source heterogeneous dataset, and cross-process feature alignment is performed on the time feature vectors and image feature vectors according to the wafer identifier and processing timestamp to obtain an enhanced cross-process feature set, which may include the following steps:

[0075] Step S201: Extract the process parameter time series from the multi-source heterogeneous dataset, and use the sliding window method to truncate the process parameter time series of a fixed length to obtain the time feature vector.

[0076] Step S202: Collect semiconductor wafer defect images from a multi-source heterogeneous dataset, input the wafer defect images into a preset convolutional neural network, and extract deep features of wafer defect morphology through the convolutional neural network to obtain image feature vectors.

[0077] Step S203: Using the wafer identifier as the unique matching identifier and the processing timestamp as the time sequence matching basis, the time feature vector and the image feature vector are aligned across processes to obtain the aligned cross-process feature set.

[0078] Step S204: Extract the intersection of time features and image features from the aligned cross-process feature set as a fusion feature subset. Based on the attention mechanism, assign weights to each feature component of the fusion feature subset and obtain a weighted feature vector after weighted calculation.

[0079] Step S205: Input the weighted feature vector into the preset defect classification model to output the defect classification label. If the defect probability value corresponding to the defect classification label exceeds the preset defect judgment threshold, an alarm signal is generated.

[0080] Step S206: Extract the associated semiconductor manufacturing process parameters from the alarm signal as associated process parameters, and use the gradient descent algorithm to iteratively optimize the associated process parameters to obtain the optimized process parameter set.

[0081] Step S207: Adjust the processing timestamps of the corresponding processes according to the optimized set of process parameters, and substitute the adjusted processing timestamps into the truncation process of the sliding window method to obtain the updated time feature vector.

[0082] Step S208: Extract subsequence features from the updated time feature vector, and re-align the extracted subsequence features with the image feature vector according to the cross-process alignment method to obtain an enhanced cross-process feature set.

[0083] Specifically, firstly, the time series of process parameters corresponding to each process step are extracted from the constructed multi-source heterogeneous dataset. A sliding window method is used to set a fixed-length time window to truncate the time series of process parameters, retaining the temporal variation characteristics of the process parameters within each window, thus obtaining a time feature vector that can characterize the local temporal regularity of the process parameters. Simultaneously, wafer defect images after each semiconductor process are acquired. The acquired wafer defect images are uniformly preprocessed and then input into a pre-defined convolutional neural network. Through convolution, pooling, and other layer operations of the convolutional neural network, deep visual features such as wafer defect morphology, defect distribution, and defect size are extracted. Redundant visual information is removed to obtain an image feature vector that can accurately characterize the wafer defect features. Subsequently, the wafer identifier is used as a unique matching identifier to ensure that the temporal features of the same wafer correspond to the image features. The processing timestamp is used as the time sequence matching basis to calibrate the feature time sequence deviation of the same wafer under different processes. The extracted time feature vector and image feature vector are aligned across processes to eliminate the feature misalignment problem between processes, resulting in an aligned cross-process feature set.

[0084] Next, from the aligned cross-process feature set, the intersection of temporal and image features is extracted as a fused feature subset that can simultaneously reflect changes in process parameters and wafer defect states. Based on an attention mechanism, weights are assigned to each feature component of the fused feature subset, assigning higher weights to feature components highly correlated with wafer yield and lower weights to those with low correlation. A weighted feature vector highlighting key features is obtained through weighted calculation. This weighted feature vector is input into a preset defect classification model, which outputs the corresponding defect classification label and the defect probability value. If the defect probability value exceeds a preset defect judgment threshold, it indicates an abnormal risk in the current process parameters, and an alarm signal is immediately generated. From the generated alarm signal, semiconductor manufacturing process parameters directly related to the abnormal risk are extracted as associated process parameters. A gradient descent algorithm is used to iteratively optimize these associated process parameters multiple times, gradually correcting process parameter deviations to obtain an optimized set of process parameters. Based on the optimized set of process parameters, the processing timestamps of the corresponding processes are adjusted to ensure that the processing timestamps match the optimized process parameters. The adjusted processing timestamps are then substituted into the aforementioned sliding window method for truncation, and the process parameter time series is re-trunculated to obtain the updated time feature vector. Subsequence feature extraction is performed on the updated time feature vector to further explore the subtle temporal variation features of the process parameters. The extracted subsequence features are then re-aligned with the previously obtained image feature vectors according to the aforementioned cross-process alignment method, and integrated to form an enhanced cross-process feature set with more complete information and more accurate features.

[0085] This embodiment achieves accurate feature extraction and optimization of multi-source heterogeneous data. By using a sliding window method to extract the time series of process parameters, it ensures that the time feature vector fully preserves the temporal variation patterns of process parameters. Convolutional neural networks are used to extract deep features of wafer defects, solving the problems of insufficient and redundant information in traditional image feature extraction and improving the representation accuracy of image features. Through dual matching of wafer identifiers and processing timestamps, precise alignment of time features and image features across processes is achieved, eliminating the risk of feature misalignment between processes and ensuring the integrity and correlation of cross-process feature sets. Feature weight allocation based on an attention mechanism highlights key features closely related to wafer yield, improving the effectiveness of feature vectors. By triggering alarms and optimizing associated process parameters through a defect classification model, dynamic correction of process parameters is achieved, reducing feature distortion caused by process parameter deviations and further improving the quality of time feature vectors. Finally, through subsequence feature extraction and realignment, an enhanced cross-process feature set was obtained, which integrates the temporal features of process parameters and the visual features of wafer defects. This solves the problems of fragmented multi-source features, low feature utilization, and insufficient feature accuracy in traditional feature extraction, and ensures the accuracy and reliability of subsequent yield prediction.

[0086] In one embodiment, mapping the enhanced cross-process feature set to a spatial network topology based on the machine nodes through which the wafer flows, and generating spatiotemporal graph data with node and edge features, may include the following steps:

[0087] Step S301: Obtain the wafer transfer path data in the production line; the transfer path data includes manufacturing machine nodes and wafer transfer edge information between machines.

[0088] Step S302: Construct a spatial network topology based on the flow path data; the spatial network topology uses manufacturing equipment as nodes and wafer flow relationships as edges.

[0089] Step S303: The enhanced cross-process feature set is mapped to the spatial network topology according to the manufacturing machine nodes through which the wafer flows.

[0090] Step S304: Assign corresponding node features to each manufacturing machine node and corresponding edge features to each wafer transfer edge in the spatial network topology to generate spatiotemporal graph data with node features and edge features.

[0091] Specifically, the process begins by acquiring complete wafer transfer path data within the production line. This data explicitly includes information about each manufacturing machine node, the wafer transfer order between machines, and their interrelationships. In other words, the transfer path data covers core information such as the identification of each manufacturing machine, the order in which wafers transfer between machines, and the corresponding processing steps for each machine. Subsequently, based on the acquired transfer path data, a spatial network topology is constructed with manufacturing machines as nodes and wafer transfer relationships as edges. The association logic and transfer rules of each machine node are clarified to ensure consistency between the topology and the actual production process. Building upon this, the enhanced cross-process feature set obtained earlier is precisely matched according to the machine nodes through which the wafers pass. The feature data of the corresponding machine is mapped to the corresponding nodes in the spatial network topology, and the transfer-related feature data is mapped to the corresponding transfer edges in the topology, completing the binding of features and topology. Finally, combining all the above data, corresponding node features are assigned to each machine node, and corresponding edge features are assigned to each transfer edge. By integrating temporal and spatial relationship information, a spatiotemporal graph with node and edge features is generated.

[0092] This embodiment achieves deep integration of wafer transfer path data, multi-source feature data, and spatial network topology, effectively solving the problems of isolated features and insufficient data utilization in traditional yield prediction. The generated spatiotemporal map data completely retains core information such as the temporal changes of process parameters and machine relationships, ensuring data integrity and consistency, and providing accurate data support for subsequent yield prediction model training and risk localization, avoiding the problems of feature disconnect and data waste. At the same time, by combining feature data with the actual flow logic of the production line, the accuracy and reliability of the prediction results are greatly improved, providing scientific and practical technical support for semiconductor manufacturing process optimization and yield improvement, effectively reducing production risks and ensuring the stable and efficient operation of the production line.

[0093] In one embodiment, after generating spatiotemporal graph data with node and edge features, the following steps may also be included:

[0094] Step S401: The spatiotemporal graph data is processed using a graph neural network to obtain the node state vector corresponding to each manufacturing machine node.

[0095] Preferably, the spatiotemporal graph data carrying manufacturing machine node features, wafer transfer edge features, and process timing information is input into the graph neural network. The model performs fusion learning on the process features of each machine node, the correlation features of adjacent nodes, and the interaction features of transfer edges through inter-node message passing and hierarchical feature aggregation operations. It extracts the topological structure features and timing change features in the spatiotemporal graph data. After the network forward operation, it outputs a node state vector that can completely represent the process status and node correlation of each manufacturing machine node.

[0096] Step S402: Calculate the state transition probability of the wafer on the transfer path based on the node state vector and the edge characteristics of the corresponding wafer transfer edge.

[0097] Step S403: Determine the critical transfer path of the wafer between each manufacturing machine node based on the state transition probability.

[0098] Furthermore, the state transition probabilities between all machine nodes are numerically sorted and threshold-screened, retaining the flow paths with probability values ​​higher than the preset judgment standard and eliminating non-core flow links with low probabilities. Then, combined with the processing sequence of the wafer in the actual production line, the screened high-probability flow paths are connected and integrated according to the processing logic to form a continuous and complete flow link, and finally the key flow paths of the wafer between each manufacturing machine node are determined.

[0099] Specifically, the process begins by acquiring spatiotemporal graph data containing node and edge features. This data is then input into a pre-defined graph neural network model. Through hierarchical operations such as convolution aggregation and message passing, the model mines the spatial topological relationships and process timing features of each manufacturing machine node. It aggregates the edge features of adjacent nodes with their own node features, ultimately outputting a node state vector for each manufacturing machine node. This vector fully represents the corresponding machine node's operating state, process attributes, and the degree of association with surrounding nodes on the production line. Subsequently, using the node state vectors of each manufacturing machine node as the core input, and combining them with the edge features carried by the corresponding wafer transfer edges (which specifically include information such as the inter-machine process coupling strength, data interaction frequency, and wafer transfer relationships), the process integrates the dual information of node state and edge features to perform a quantitative analysis of the wafer transfer process between different manufacturing machine nodes. The probability of state transition of the wafer along the transfer path is calculated, and this probability characterizes the feasibility and priority of the wafer transfer from one manufacturing machine node to another. Finally, based on the calculated state transition probabilities of the wafer between each manufacturing machine node, flow links with probability values ​​higher than the preset judgment threshold are screened out. The node sequence and feature information of each flow link are integrated to determine the key flow path of the wafer between each manufacturing machine node. This path clearly reflects the core node links of wafer processing, defect transmission and process influence transmission in the semiconductor production line.

[0100] This embodiment processes spatiotemporal graph data using a graph neural network, fully exploring the spatial topological relationships and deep-seated process timing features of production line machine nodes, thus improving the information completeness and representation accuracy of node state vectors. By combining node state vectors with wafer flow edge features to calculate state transition probabilities, a quantitative representation of the flow relationships between machines is achieved. This solves the problem that traditional flow path analysis relies solely on qualitative descriptions and lacks quantitative evidence, improving the accuracy and scientific rigor of path analysis. Based on the key flow paths determined by state transition probabilities, the system accurately aligns with the actual production logic of semiconductor production lines, clarifying the core node links for wafer defect generation and propagation. This effectively improves the targeting and effectiveness of production line yield control, reduces ineffective process adjustments and resource waste, and ensures the stability and efficiency of semiconductor production line operation.

[0101] In one embodiment, the state transition probability of the wafer along the transfer path is calculated using the following formula:

[0102]

[0103] in, Indicates machine node To machine node The state transition probability, Represents a node The node state vector, Represents the adaptive projection matrix of edge features. Indicates machine node to The edge features corresponding to the wafer transfer edges between them This indicates the attention weight of process coupling between machines. Represents a small constant for numerical smoothing. Indicates historical wafers from arrive The actual circulation frequency, Indicates the wafer is in , Average turnover time between Indicates the time-series flow weighting coefficient. This represents the Sigmoid activation function.

[0104] This embodiment achieves a multi-dimensional quantitative representation of the wafer transfer path state transition probability by fusing multi-dimensional information such as node state vectors, adaptively projected edge features, process coupling attention weights, historical transfer frequency, and average transfer time. This effectively solves the problem of traditional transfer path analysis relying solely on qualitative descriptions or single data dimensions and lacking comprehensive quantitative evidence, thus improving the scientific rigor and accuracy of state transition probability calculations. The small constant introduced in the formula ensures the stability of the calculation process and avoids calculation anomalies. The key transfer path determined by this formula precisely matches the actual operating logic of the semiconductor production line, reflecting both the process coupling relationship between machines and taking into account historical transfer habits and timing efficiency. It can accurately capture the core node links of wafer defect generation and propagation, significantly improving the targeting and effectiveness of production line yield control, reducing ineffective process adjustments and resource waste, and ensuring the stability and efficiency of semiconductor production line operation.

[0105] In one embodiment, the trained yield prediction model is obtained through the following process, which may include:

[0106] Step S501: Obtain the interaction and association information between each manufacturing machine node in the spatiotemporal graph data, and construct a node dependency graph based on the interaction and association information.

[0107] The interactive association information may include the process coupling strength between nodes, the frequency of data interaction, and the flow association relationship.

[0108] Step S502: Use a time-series graph neural network to extract time-series features from the node dependency graph and key flow paths, and capture the dynamic evolution sequence of the process parameters corresponding to each manufacturing machine node over time.

[0109] The dynamic evolution sequence of process parameters can include process parameter drift trends and time series correlation features.

[0110] The node dependency graph (including information such as machine node dependency logic and process coupling strength) and key flow paths (including wafer machine flow sequence and core link information) are used as inputs to the time-series graph neural network. The model uses a time-series attention mechanism and graph convolution operation to fuse and extract the static features and time-series change information of each manufacturing machine node. The core of the time-series feature aggregation is achieved through the following formula: ,in Indicates machine node At any moment The temporal characteristics, Let represent the set of adjacent nodes of machine node i. The adjacency matrix elements represent the node dependency graph, characterizing the dependency strength between nodes i and j. Through continuous extraction and concatenation of time-series features at multiple time points, the dynamic evolution sequence of the process parameters corresponding to each manufacturing machine node over time is finally captured. The process parameter drift trend characterizes the offset law of process parameters with processing time, and the time-series correlation features characterize the correlation relationship and change law of process parameters at different time points.

[0111] Step S503: Perform collaborative model training based on the dynamic evolution sequence of process parameters, and perform encryption operation on the multi-party local model gradient data generated during the training process to obtain encrypted gradient updates.

[0112] Employing a distributed collaborative training model, each participant (e.g., different factories or production lines) independently trains its yield prediction model based on the dynamically evolving sequence of process parameters acquired locally. The gradient of its local model is then calculated using the backpropagation algorithm. (where k represents the kth participant); to ensure the privacy and security of local data of each participant and to avoid gradient leakage, the local model gradient data generated by each participant is encrypted. Homomorphic encryption algorithm is used to implement gradient encryption, and the encryption process can be represented as: ,in Represents a homomorphic encryption function. This represents the public encryption key, which, after encryption, yields the corresponding encryption gradient updates for each participant. This ensures that gradient data is not leaked during transmission and aggregation.

[0113] Step S504: Globally aggregate the encrypted gradient updates using the federated averaging algorithm to obtain the global gradient update result.

[0114] Furthermore, the global gradient update results are... With preset gradient convergence threshold The comparison is performed, and the judgment condition is as follows: (in (This represents the L2 norm of the gradient vector). If this condition is met, it indicates that the model training has reached a stable state and the model parameters no longer change significantly. The collaborative training process is terminated, and the model parameters at this time are used as the final parameters to obtain the trained yield prediction model. If the preset gradient convergence threshold is not reached, the global gradient update result is distributed back to each participant. Each participant updates its local model parameters based on the global gradient and repeats steps S503 to S504 until the global gradient update result meets the convergence condition. This ensures that the trained yield prediction model has high prediction accuracy and generalization ability and can accurately capture the correlation between process parameter changes and wafer yield.

[0115] Step S505: If the global gradient update result reaches the preset gradient convergence threshold, the trained yield prediction model is obtained.

[0116] Specifically, firstly, interaction and correlation information between each manufacturing machine node is extracted from the generated spatiotemporal graph data. This interaction and correlation information includes the process coupling strength between nodes, the frequency of data interaction, and the flow correlation. Based on the above interaction and correlation information, a node dependency graph that can characterize the dependency relationship between machine nodes is constructed. Then, a temporal graph neural network is used to extract temporal features from the constructed node dependency graph and the previously determined key wafer flow paths. The model learns the spatial dependency relationship and temporal change law between machine nodes, and captures the dynamic evolution sequence of the process parameters corresponding to each manufacturing machine node as the processing progresses. This dynamic evolution sequence of process parameters covers the drift trend of process parameters and the temporal correlation characteristics between nodes. Based on the obtained dynamic evolution sequence of process parameters, collaborative training of the yield prediction model is performed in a distributed data scenario. The local model gradient data generated by each participant during local training is encrypted to obtain the corresponding encrypted gradient update. Then, the encrypted gradient updates of multiple parties are globally aggregated using the federated averaging algorithm to output the global gradient update result. When the global gradient update result meets the preset gradient convergence threshold, the model training process is terminated, and the trained yield prediction model is obtained.

[0117] This embodiment constructs a node dependency graph based on machine interaction association information, which can clearly depict the spatial association and process dependency logic of each manufacturing machine in the production line. Combined with key flow paths, it extracts time-series features, which can completely preserve the time-series change rules of process parameters and the association features between nodes, improving the representation accuracy and completeness of the dynamic evolution sequence of process parameters. By adopting a collaborative training method combined with gradient encryption and federated average aggregation, it can achieve joint modeling of multi-party data while ensuring the privacy, security and compliance of local data of each participant, avoiding the security risks brought about by centralized data aggregation. By controlling the training process through gradient convergence threshold, it can ensure that the yield prediction model reaches a stable performance state, enabling the model to have stronger time-series awareness and generalization ability.

[0118] In one embodiment, the yield prediction model trained by inputting the enhanced cross-process feature set and the current spatial network state can locate the high-risk process machines and time intervals where potential defects may occur, and obtain the defect probability prediction result. This can include the following steps:

[0119] Step S601: Obtain the dynamic process parameters corresponding to the current spatial network status.

[0120] Step S602: The random forest algorithm is used to perform a fusion analysis on the dynamic process parameters and the enhanced cross-process feature set to obtain a preliminary defect risk index.

[0121] Step S603: Input the preliminary defect risk index into the trained yield prediction model to obtain the defect probability prediction results for future batches of wafers.

[0122] Step S604: Based on the defect probability prediction results, perform correlation analysis on the process data of each key flow path covered by the process to determine whether there are abnormal fluctuations in process parameters of each high-risk process machine.

[0123] Step S605: If it is determined that there is abnormal fluctuation in process parameters of a high-risk process machine, then locate the time interval of potential defect occurrence corresponding to the abnormal fluctuation.

[0124] Step S606: Extract historical process log data of the corresponding high-risk process machine from the time interval of potential defect occurrence, and generate process optimization and adjustment instructions based on the historical process log data.

[0125] Step S607: The current spatial network status is dynamically updated using process optimization adjustment instructions to obtain the optimized defect probability prediction result.

[0126] Specifically, the process first acquires the dynamic process parameters corresponding to the current spatial network state. These parameters encompass the real-time operating process parameters of each manufacturing machine, the processing progress of each step, and parameter fluctuation data, comprehensively reflecting the actual operating status of the current production line. Then, a random forest algorithm is used to fuse and analyze the acquired dynamic process parameters with the previously obtained enhanced cross-process feature set. The algorithm mines the correlation between the two types of data, quantifies the correlation between process parameters and wafer defects, and outputs a preliminary defect risk index that reflects the wafer defect risk level. This preliminary defect risk index is input into the trained yield prediction model. The model combines the previously learned dynamic evolution patterns of process parameters, machine correlation relationships, and defect correlation rules to obtain the defect probability prediction results for future batches of wafers through inference calculations. These results include the defect occurrence probability and risk level corresponding to each process machine.

[0127] Based on the defect probability prediction results, correlation analysis is performed on the process data covered by each key workflow path. The standard thresholds of process parameters for each high-risk process machine are compared with the actual operating parameters to determine whether there are abnormal fluctuations in process parameters for each high-risk process machine. If abnormal fluctuations in process parameters are determined for a high-risk process machine, the machine's processing timestamp, the period of parameter fluctuation, and the timing of changes are combined to locate the potential defect occurrence time interval corresponding to the abnormal fluctuation, and to clarify the start and end processing nodes and the specific processing stage of the corresponding process within that time interval. From the located potential defect occurrence time interval, historical process log data of the corresponding high-risk process machine is extracted. This data includes core content such as machine operating parameters, process adjustment records, fault alarm information, and wafer processing status during that period. Based on this historical process log data, the root cause of the abnormal parameter fluctuations is analyzed, and targeted process optimization adjustment instructions are generated. These instructions clearly include the process parameters to be adjusted, the adjustment range, and the adjustment sequence. Finally, the generated process optimization and adjustment instructions are used to dynamically update the machine process parameters, node relationships and flow link information in the current spatial network state. The updated spatial network state and the enhanced cross-process feature set are then re-input into the yield prediction model to obtain the optimized defect probability prediction results.

[0128] This embodiment achieves accurate prediction of wafer defect risk and optimization of process parameters. By fusing dynamic process parameters and enhanced cross-process features through a random forest algorithm, it effectively mines the correlation between multi-source data, improving the accuracy of preliminary defect risk indicators. Inputting the preliminary risk indicators into the trained yield prediction model ensures the accuracy of future batch wafer defect probability predictions, enabling early prediction of defect risks and solving the problem of lagging traditional defect detection. Through correlation analysis of key process flow data, abnormal fluctuations in process parameters of high-risk process equipment and the corresponding potential defect time intervals can be quickly located, achieving precise traceability of defect sources. Based on historical process log data, targeted process optimization adjustment instructions are generated, accurately correcting process parameter deviations. Closed-loop optimization is completed through dynamic updates of the spatial network state, effectively reducing the probability of defects occurring in high-risk process equipment. The entire process ensures both the accuracy and timeliness of defect prediction and achieves dynamic optimization of process parameters, improving the yield control capability and production stability of semiconductor production lines.

[0129] In one embodiment, such as Figure 2 As shown, this application also provides a semiconductor yield collaborative prediction system based on machine learning, which may include:

[0130] The data acquisition module 701 is used to acquire time series data of machine sensor data and wafer defect image data of each process in semiconductor manufacturing, forming a multi-source heterogeneous dataset.

[0131] The feature extraction module 702 is used to extract time feature vectors and image feature vectors from multi-source heterogeneous datasets respectively, and to perform cross-process feature alignment on the time feature vectors and image feature vectors according to the wafer identifier and processing timestamp to obtain an enhanced cross-process feature set.

[0132] The spatiotemporal graph generation module 703 is used to construct a spatial network topology structure with manufacturing equipment as nodes and wafer flow relationships as edges based on the wafer flow path in the production line. It maps the enhanced cross-process feature set to the spatial network topology structure according to the equipment nodes through which the wafer flows, and generates spatiotemporal graph data with node features and edge features.

[0133] The model training module 704 is used to extract time-series features from the spatiotemporal graph data to obtain the dynamic evolution sequence of process parameters. Based on the dynamic evolution sequence, collaborative model training is performed, and the gradients of multiple models are encrypted and aggregated through the federated averaging algorithm to obtain the trained yield prediction model.

[0134] The defect localization module 705 is used to input the enhanced cross-process feature set and the current spatial network status into the trained yield prediction model to locate the high-risk process machine and time interval where potential defects occur, and obtain the defect probability prediction result.

[0135] The aforementioned machine learning-based semiconductor yield collaborative prediction system first acquires time-series data from machine sensor data and wafer defect image data corresponding to each process in semiconductor manufacturing through a data acquisition module. These two types of data are then aggregated and organized to form a multi-source heterogeneous dataset, providing raw data support for subsequent feature processing. Subsequently, the feature extraction module processes this multi-source heterogeneous dataset, extracting the temporal feature vectors corresponding to the machine sensor time-series data and the visual feature vectors corresponding to the wafer defect images. Using wafer identification as the matching criterion and processing timestamps as the time-series standard, the two types of feature vectors are aligned across processes to eliminate time-series biases, forming a complete cross-process feature set, resulting in an enhanced cross-process feature set. The spatiotemporal graph generation module constructs a spatial network topology with manufacturing machines as nodes and flow relationships as edges based on the actual flow path of the wafer on the production line. The enhanced cross-process feature set is mapped to machine nodes to generate spatiotemporal graph data containing node features and edge features. Then, the model training module extracts time-series features from the spatiotemporal graph data to obtain the dynamic evolution law of process parameters. Combined with the federated averaging algorithm, the gradients of multiple models are encrypted and aggregated to complete the model training and obtain the trained yield prediction model. Finally, the defect localization module inputs the enhanced cross-process feature set and the current production line spatial network state into the trained model to locate high-risk process machines and corresponding time intervals, and outputs the defect probability prediction results.

[0136] In this embodiment, the modules work collaboratively, with a clear and logically coherent data flow from acquisition to feature processing, topology construction, model training, and defect localization. This effectively solves the problems of data fragmentation, insufficient feature extraction, privacy data insecurity, and delayed defect localization in traditional yield control. Through the coordinated efforts of multiple modules, efficient utilization of multi-source data is achieved, improving the accuracy of yield prediction and the targeting of process adjustments. At the same time, the federated learning model ensures data privacy and security, avoiding the security risks of centralized data management. By accurately locating high-risk processes and time intervals, process parameters can be adjusted in a timely manner, production processes can be optimized, yield losses caused by process deviations can be reduced, and the stability and efficiency of semiconductor manufacturing can be improved. This provides reliable technical support for the intelligent and refined management of semiconductor production lines, reducing production losses and management costs.

[0137] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0138] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the machine learning-based semiconductor yield co-prediction method as described above.

[0139] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0140] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0141] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for machine learning based semiconductor yield collaborative prediction, characterized in that, The method includes: Acquire time-series data of machine sensor data and wafer defect image data for each process in semiconductor manufacturing to form a multi-source heterogeneous dataset; Time feature vectors and image feature vectors are extracted from the multi-source heterogeneous dataset. Cross-process feature alignment is performed on the time feature vectors and image feature vectors according to the wafer identifier and processing timestamp to obtain an enhanced cross-process feature set. Based on the flow path of wafers in the production line, a spatial network topology is constructed with manufacturing equipment as nodes and wafer flow relationships as edges. The enhanced cross-process feature set is mapped to the spatial network topology according to the equipment nodes through which the wafers flow, generating spatiotemporal graph data with node features and edge features. The time-series features of the spatiotemporal graph data are extracted to obtain the dynamic evolution sequence of process parameters. Based on the dynamic evolution sequence, a collaborative model is trained and the gradients of multiple models are encrypted and aggregated using a federated averaging algorithm to obtain the trained yield prediction model. The enhanced cross-process feature set and the current spatial network state are input into the trained yield prediction model to locate the high-risk process machines and time intervals where potential defects may occur, and the defect probability prediction results are obtained.

2. The method of claim 1, wherein, The process involves extracting time feature vectors and image feature vectors from the multi-source heterogeneous dataset, and then aligning the time feature vectors and image feature vectors across processes based on wafer identifiers and processing timestamps to obtain an enhanced cross-process feature set, including: The process parameter time series is extracted from the multi-source heterogeneous dataset, and a fixed-length process parameter time series is truncated using the sliding window method to obtain a time feature vector; Semiconductor wafer defect images are collected from the multi-source heterogeneous dataset. The wafer defect images are input into a preset convolutional neural network. The deep features of the wafer defect morphology are extracted through the convolutional neural network to obtain image feature vectors. Using the wafer identifier as the unique matching identifier and the processing timestamp as the time sequence matching basis, the time feature vector and the image feature vector are aligned across processes to obtain the aligned cross-process feature set. The intersection of time features and image features is extracted from the aligned cross-process feature set as a fusion feature subset. The weights of each feature component of the fusion feature subset are assigned based on the attention mechanism, and a weighted feature vector is obtained after weighted calculation. The weighted feature vector is input into a preset defect classification model to output a defect classification label. If the defect probability value corresponding to the defect classification label exceeds a preset defect judgment threshold, an alarm signal is generated. The associated semiconductor manufacturing process parameters are extracted from the alarm signal as associated process parameters, and the gradient descent algorithm is used to iteratively optimize the associated process parameters to obtain an optimized set of process parameters. Adjust the processing timestamps of the corresponding processes according to the optimized set of process parameters, and substitute the adjusted processing timestamps into the truncation process of the sliding window method to obtain the updated time feature vector. Subsequence features are extracted from the updated time feature vector. The extracted subsequence features are then re-aligned with the image feature vector according to the cross-process alignment method to obtain an enhanced cross-process feature set.

3. The method of claim 1, wherein, The step of mapping the enhanced cross-process feature set to the spatial network topology based on the machine nodes through which the wafer flows, and generating spatiotemporal graph data with node features and edge features, includes: Acquire wafer transfer path data in the production line; the transfer path data includes manufacturing machine nodes and wafer transfer edge information between machines; A spatial network topology is constructed based on the flow path data; the spatial network topology uses manufacturing equipment as nodes and wafer flow relationships as edges. The enhanced cross-process feature set is mapped to the spatial network topology according to the manufacturing machine nodes through which the wafer flows; In the spatial network topology, each manufacturing machine node is assigned a corresponding node feature, and each wafer transfer edge is assigned a corresponding edge feature, thereby generating spatiotemporal graph data with node features and edge features.

4. The method of claim 3, wherein, After generating the spatiotemporal graph data with node and edge features, the process further includes: A graph neural network is used to process the spatiotemporal graph data to obtain the node state vector corresponding to each manufacturing machine node; Based on the node state vector and the edge characteristics of the corresponding wafer transfer edge, calculate the state transition probability of the wafer on the transfer path; The critical flow path of the wafer between each of the manufacturing machine nodes is determined based on the state transition probability.

5. The method of claim 4, wherein, The state transition probability of the wafer along the transfer path is calculated using the following formula: in, Indicates machine node To machine node The state transition probability, Represents a node The node state vector, Represents the adaptive projection matrix of edge features. Indicates machine node to The edge features corresponding to the wafer transfer edges between them This indicates the attention weight of process coupling between machines. Represents a small constant for numerical smoothing. Indicates historical wafers from arrive The actual circulation frequency, Indicates the wafer is in , Average turnover time between Indicates the time-series flow weighting coefficient. This represents the Sigmoid activation function.

6. The method according to claim 1, characterized in that, The trained yield prediction model is obtained through the following process: Obtain the interaction and association information between each manufacturing machine node in the spatiotemporal graph data, and construct a node dependency graph based on the interaction and association information; The interactive association information includes the process coupling strength between nodes, the frequency of data interaction, and the flow association relationship; A temporal graph neural network is used to extract temporal features from the node dependency graph and key flow paths, capturing the dynamic evolution sequence of the process parameters corresponding to each manufacturing machine node over time. The dynamic evolution sequence of process parameters includes process parameter drift trends and time-series correlation features; Based on the dynamic evolution sequence of the process parameters, a collaborative model is trained, and the multi-party local model gradient data generated during the training process is encrypted to obtain encrypted gradient updates. The encrypted gradient updates are globally aggregated using a federated averaging algorithm to obtain the global gradient update result. If the global gradient update result reaches the preset gradient convergence threshold, the trained yield prediction model is obtained.

7. The method according to claim 1, characterized in that, The step of inputting the enhanced cross-process feature set and the current spatial network state into the trained yield prediction model to locate high-risk process machines and time intervals where potential defects occur, and obtaining defect probability prediction results, includes: Obtain the dynamic process parameters corresponding to the current spatial network status; The random forest algorithm is used to fuse and analyze the dynamic process parameters and the enhanced cross-process feature set to obtain preliminary defect risk indicators; The preliminary defect risk index is input into the trained yield prediction model to obtain the defect probability prediction results for future batches of wafers. Based on the defect probability prediction results, the process data of each key flow path covered by the process are correlated and analyzed to determine whether there are abnormal fluctuations in process parameters of each high-risk process machine. If it is determined that a certain high-risk process machine has abnormal fluctuations in the process parameters, then the time interval of potential defect occurrence corresponding to the abnormal fluctuations is located. Extract historical process log data of the machine tool corresponding to the high-risk process within the time interval of the potential defect occurrence, and generate process optimization and adjustment instructions based on the historical process log data; The current spatial network state is dynamically updated using the process optimization and adjustment instructions to obtain optimized defect probability prediction results.

8. A semiconductor yield collaborative prediction system based on machine learning, characterized in that, The system includes: The data acquisition module is used to acquire time-series data from machine sensors and wafer defect image data for each process in semiconductor manufacturing, forming a multi-source heterogeneous dataset; The feature extraction module is used to extract time feature vectors and image feature vectors from the multi-source heterogeneous dataset, and to perform cross-process feature alignment on the time feature vectors and image feature vectors according to the wafer identifier and processing timestamp to obtain an enhanced cross-process feature set. The spatiotemporal graph generation module is used to construct a spatial network topology structure with manufacturing equipment as nodes and wafer flow relationships as edges based on the wafer flow path in the production line. The enhanced cross-process feature set is mapped to the spatial network topology structure according to the equipment nodes through which the wafer flows, generating spatiotemporal graph data with node features and edge features. The model training module is used to extract time-series features from the spatiotemporal graph data to obtain the dynamic evolution sequence of process parameters, perform collaborative model training based on the dynamic evolution sequence, and encrypt and aggregate the gradients of multiple models through the federated averaging algorithm to obtain the trained yield prediction model. The defect localization module is used to input the enhanced cross-process feature set and the current spatial network status into the trained yield prediction model to locate the high-risk process machines and time intervals where potential defects occur, and obtain the defect probability prediction result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

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

Images