Integrated circuit process parameter optimization method and system based on machine learning

Abstract

The present invention provides an integrated circuit process parameter optimization method and system based on machine learning, which relates to the field of integrated circuit technology, including: acquiring process history data, constructing a process technology knowledge graph to solve the semantic similarity between entities and relationships, extracting key process knowledge and constructing a priori knowledge base; determining the causal dependency between process parameters and yield, setting causal feature selection criteria, screening key causal feature subsets, expanding the process technology knowledge graph, constructing heterogeneous model integration and determining the dynamic changes of process parameters and yield; constructing a multi-objective optimization model and setting optimization goals, performing process parameter optimization solution, generating optimized process parameters and performing transfer learning, comparing and mapping process nodes, constructing a virtual simulation system and performing causal intervention, generating process parameter optimization decisions and correcting the Bayesian causal model.

Description

Technical Field

[0001] The present invention relates to the field of integrated circuit technology, and in particular to an integrated circuit process parameter optimization method and system based on machine learning. Background Art

[0002] Integrated circuit manufacturing is a highly complex process involving hundreds of process steps and thousands of process parameters. In actual production, the setting of process parameters has a decisive influence on the yield, performance and reliability of the chip. Traditional process parameter optimization mainly relies on engineers' experience and trial and error, and gradually approaches the optimal parameter combination by manually adjusting parameters and conducting a large number of experiments.

[0003] In recent years, machine learning technology has been widely used in the industrial field, providing new ideas for modeling, optimization and control of complex manufacturing processes. However, existing technologies still lack effective representation and integration of process knowledge, and it is difficult to fully utilize multi-source heterogeneous data in design, simulation, testing and other links, and ignore the correlation constraints and interactive effects between process parameters, and it is difficult to take into account multiple quality indicators.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for optimizing integrated circuit process parameters based on machine learning, which can at least solve some of the problems existing in the prior art.

[0006] According to a first aspect of an embodiment of the present invention, there is provided a method for optimizing integrated circuit process parameters based on machine learning, comprising:

[0007] Acquire process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with an ontology mapping method, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph based on pre-acquired expert experience knowledge, extract key process knowledge and construct a priori knowledge base in combination with natural language processing technology;

[0008] Based on the process technology knowledge graph, feature extraction and fusion are performed through a graph convolutional neural network to generate feature representations corresponding to process parameters, the causal dependency between process parameters and yield is determined through causal inference theory, the causal graph is determined and causal feature selection criteria are set in combination with a causal structure learning algorithm based on a graph model, a key causal feature subset is screened, key samples are selected and active learning is performed in combination with causal correlation analysis and semantic similarity calculation, the process technology knowledge graph is expanded, and a heterogeneous model integration is constructed through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space, and according to the update of the process technology knowledge graph, the heterogeneous model integration is updated in real time through an incremental learning algorithm of streaming data and the dynamic changes of process parameters and yield are determined;

[0009] Based on the dynamic changes, a multi-objective optimization model is constructed and multiple optimization objectives are set. The process parameters are optimized and solved by a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset. The optimized process parameters are generated and transfer learning is performed through feature mapping and parameter adaptation based on the semantic mapping relationship of the process nodes in the process technology knowledge graph. The process nodes are compared and mapped according to the pre-constructed Bayesian causal model, the causal pattern is determined, and a virtual simulation system is integrated based on the Bayesian causal model and the heterogeneous model, and causal intervention is performed on key process parameters. Process parameter optimization decisions are generated and the Bayesian causal model is corrected according to the process parameter optimization decisions.

[0010] In an optional embodiment,

[0011] Acquire process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with ontology mapping methods, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph according to pre-acquired expert experience knowledge, extract key process knowledge in combination with natural language processing technology and construct a priori knowledge base, including:

[0012] Acquire initial process history data from multi-source heterogeneous process equipment and test equipment, wherein the initial process history data includes equipment parameters, process monitoring data and test results, pre-process the initial process history data, combine data standardization and normalization methods to perform unified representation, and obtain the process history data, perform association rule mining on the process history data according to a priori algorithms, generate association rules, and perform frequent pattern mining on the process history data through a frequent pattern growth algorithm, determine the co-occurrence patterns between different data sources, and perform significance tests and confidence scores on the association rules and the co-occurrence patterns through statistics, and determine strong association rules and high-frequency patterns;

[0013] Based on the strong association rules and the high-frequency patterns, the process history data from different data sources are mapped as heterogeneous data to a common semantic space through an ontology mapping method, a process technology ontology is constructed, and the semantic similarity between different data sources is determined through a semantic similarity calculation method, semantic fusion and interoperation are performed based on the semantic similarity to obtain fused data, based on the process technology ontology and the fused data, process technology knowledge is represented in a process resource description framework format and a process technology knowledge graph is constructed, wherein entities in the process technology knowledge graph include process equipment, process parameters, test indicators and material properties, and entities and relationships in the process technology knowledge graph are embedded into a low-dimensional continuous vector space through a knowledge graph embedding method to obtain a distributed representation of entities and relationships;

[0014] Based on the distributed representation, the semantic similarity between entities and relationships is solved by representation vector operations in the embedded vector space to obtain process technology domain knowledge, and multi-step relational reasoning is performed on the process technology knowledge graph through a path-based reasoning method to generate implicit associations between entities and generate incremental process technology knowledge in combination with the semantic operation rules in the embedded vector space. Based on the incremental process technology knowledge and pre-acquired expert experience knowledge, the key process knowledge in the expert experience knowledge is extracted through natural language processing technology, the key process knowledge is mapped to the process technology knowledge graph, and knowledge alignment and ontology matching are performed through a knowledge fusion method to obtain an enhanced process technology knowledge graph, and the key process knowledge is used as prior knowledge to construct the prior knowledge base.

[0015] In an optional embodiment,

[0016] The multi-step relational reasoning is performed on the process technology knowledge graph through the path-based reasoning method as shown in the following formula:

[0017] ;

[0018] Among them, P(v i|v 0 ) represents and starts from the starting node v 0 To the target node v i The probability, P(v 0 , v i ) indicates that from the starting node v 0 To the target node v i The set of all paths of , p represents the path index, |p| represents the length of the pth path, α represents the priority weight factor of entities and relations, Represents the slave node v j-1 To node v j The transition probability, Represents node v j-1 and node v j , λ represents the time decay factor, △t represents the time difference between the current moment and the moment when the knowledge is generated, which is used to calculate the time decay.

[0019] In an optional embodiment,

[0020] Based on the process technology knowledge graph, feature extraction and fusion are performed through a graph convolutional neural network to generate feature representations corresponding to process parameters, the causal dependency between process parameters and yield is determined through causal inference theory, the causal graph is determined and causal feature selection criteria are set in combination with a causal structure learning algorithm based on a graph model, a key causal feature subset is screened, key samples are selected and active learning is performed in combination with causal correlation analysis and semantic similarity calculation, the process technology knowledge graph is expanded, and a heterogeneous model integration is constructed through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space. According to the update of the process technology knowledge graph, the heterogeneous model integration is updated in real time through an incremental learning algorithm of streaming data and the dynamic changes of process parameters and yield are determined, including:

[0021] Representing the process technology knowledge graph as an undirected weighted graph and setting a corresponding weight matrix for each edge in the undirected weighted graph, constructing a graph convolutional neural network and learning the hidden layer representation corresponding to each node based on the attribute information and topological structure information of the node in the undirected weighted graph, defining a forward propagation rule of the graph convolutional neural network based on the hidden layer representation, performing forward propagation in the graph convolutional neural network according to the forward propagation rule, stacking multiple graph convolutional neural network layers, and extracting feature representations corresponding to process parameters;

[0022] Based on the feature representation, the Robin causal model is selected to formally describe the potential result framework, the process parameters are used as processing variables, the yield rate is used as the result variable, the potential results corresponding to different processing variables are determined according to the causal effect, the conditional independence test is performed through the causal structure learning algorithm, and a causal graph is constructed according to the test results, the key processing variables are identified, and the average causal effect and the conditional average processing effect are used as causal feature selection criteria, the influence of each process parameter on the yield rate is calculated, and the key causal feature subset is obtained by screening according to the preset influence threshold;

[0023] Based on the key causal feature subset, the causal correlation coefficient between each process parameter and the yield rate is calculated through causal correlation analysis, each pair of data is sorted in descending order according to the causal correlation coefficient, and the data pairs in the top 10% are selected as candidate data pairs. The semantic similarity between the candidate data pairs and the entities in the process technology knowledge graph is calculated through a semantic similarity calculation method, and the key samples are obtained by screening according to a preset similarity threshold. Through an active learning strategy, the amount of information corresponding to each key sample is determined according to an uncertainty sampling criterion and the key sample with the maximum amount of information is selected to generate a knowledge triple, and the process technology knowledge graph is expanded according to the knowledge triple;

[0024] According to the expanded process technology map and the common semantic space, different types of pre-acquired prediction models are combined through a stacked generalized architecture to construct a heterogeneous model integration. Based on the updated process technology knowledge map, the parameters in the heterogeneous model integration are updated through the incremental learning algorithm of streaming data, and the updated heterogeneous model integration is used to predict the dynamic changes between the process parameters and yield.

[0025] In an optional embodiment,

[0026] According to the forward propagation rule, forward propagation is performed in the graph convolutional neural network as shown in the following formula:

[0027] ;

[0028] in, represents the node feature matrix of the l+1th layer, σ represents the activation function, is the degree matrix, which indicates the number of connections of each node. is the adjacency matrix with self-loops added to represent the connection relationship between nodes. represents the node feature matrix of the lth layer, represents the attention weight matrix, Represents the residual connection weight matrix.

[0029] In an optional embodiment,

[0030] Based on the dynamic changes, a multi-objective optimization model is constructed and multiple optimization objectives are set. The process parameters are optimized and solved by a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset, the optimized process parameters are generated, and according to the semantic mapping relationship of the process nodes in the process technology knowledge graph, transfer learning is performed through feature mapping and parameter adaptation, the process nodes are compared and mapped according to the pre-constructed Bayesian causal model, the causal pattern is determined, and a virtual simulation system is constructed based on the Bayesian causal model and the heterogeneous model, and causal intervention is performed on key process parameters, and a process parameter optimization decision is generated and the Bayesian causal model is corrected according to the process parameter optimization decision, including:

[0031] Acquire actual production demand, build a multi-objective optimization model according to the actual production demand and the dynamic change of the current time step, select multiple process parameters as decision variables, maximize the yield rate, minimize the production cost and shorten the production cycle as optimization goals, set constraints and establish the objective function corresponding to the multi-objective optimization model according to the physical meaning and actual limitations of the process parameters in the key causal feature subset;

[0032] Based on a multi-objective evolutionary optimization algorithm, the decision variables are mapped to individual codes in the algorithm, multiple individuals are randomly generated and combined as an initial population, for each individual in the initial population, the fitness value under multiple optimization objectives is calculated and an initial non-dominated solution set is obtained, an initial global optimal solution is selected from the initial non-dominated solution set and the prey position is set according to the position of the initial global optimal solution, for each individual in the initial population, according to the prey position and the hunting behavior of whales, the position is updated by simulating the whales rotating around the prey and approaching, the fitness value of each updated individual is calculated again and the prey position is updated according to the position in the current non-dominated solution set, and the update is repeated until the preset maximum number of iterations is reached to obtain the optimized process parameters;

[0033] Add the optimized process parameter as a new node to the process technology knowledge graph to generate an optimized process parameter node, construct a semantic mapping matrix according to the semantic associations between different nodes in the process technology knowledge graph, determine the source process node with the highest similarity for the optimized process parameter node through the semantic mapping matrix, extract the feature representation corresponding to the source process node and map it to the optimized process parameter node through a feature mapping function to obtain a mapping feature representation, adjust the mapping feature representation through a parameter adaptation function to obtain an optimized model parameter;

[0034] The optimization model parameters are added to the Bayesian causal model pre-constructed according to the process technology knowledge graph through a transfer learning algorithm. Based on the Bayesian causal model, the causal patterns of different optimization process parameter nodes are compared and mapped in combination with a causal reasoning algorithm to identify common causal patterns and special causal patterns. Based on the integration of the Bayesian causal model and the heterogeneous model, causal intervention is performed on the key process parameters according to the common causal patterns and the special causal patterns to generate process parameter optimization decisions and update the Bayesian causal model and the process technology knowledge graph.

[0035] In an optional embodiment,

[0036] The causal intervention of the key process parameters according to the common causal model and the special causal model is shown in the following formula:

[0037] ;

[0038] in, represents the probability of the potential value of the outcome variable Y when setting X to a specific value x, given the actual observed value x' of the key process parameter X and the specific value y of the optimization target Y. x is the counterfactual outcome variable, which indicates the potential value of the optimization target Y when the key process parameter X is set to x, where X represents the key process parameter, x' represents the actual observed value of the key process parameter X, and Y represents the optimization target. It indicates the probability that the optimization target Y takes the value y under the intervention condition. It represents the probability that the key process parameter X actually takes the value x' under the condition that the optimization target takes the value y. It represents the marginal probability that the actual value of the key process parameter X is x'.

[0039] A second aspect of an embodiment of the present invention provides an integrated circuit process parameter optimization system based on machine learning, comprising:

[0040] The first unit is used to obtain process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with an ontology mapping method, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph according to pre-acquired expert experience knowledge, extract key process knowledge and construct a priori knowledge base in combination with natural language processing technology;

[0041] The second unit is used to extract and fuse features based on the process technology knowledge graph through a graph convolutional neural network to generate feature representations corresponding to process parameters, determine the causal dependency between process parameters and yield rate through causal inference theory, determine the causal graph and set causal feature selection criteria in combination with a causal structure learning algorithm based on a graph model, screen out a key causal feature subset, select key samples in combination with causal correlation analysis and semantic similarity calculation and perform active learning, expand the process technology knowledge graph, build a heterogeneous model integration through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space, and update the heterogeneous model integration in real time based on the update of the process technology knowledge graph through an incremental learning algorithm of streaming data to determine the dynamic changes of process parameters and yield rate;

[0042] The third unit is used to build a multi-objective optimization model and set multiple optimization goals based on the dynamic changes, optimize and solve the process parameters through a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset, generate optimized process parameters and perform transfer learning through feature mapping and parameter adaptation according to the semantic mapping relationship of the process nodes in the process technology knowledge graph, compare and map the process nodes according to the pre-built Bayesian causal model, determine the causal pattern and integrate the Bayesian causal model and the heterogeneous model to build a virtual simulation system and perform causal intervention on key process parameters, generate process parameter optimization decisions and correct the Bayesian causal model according to the process parameter optimization decisions.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] According to a fourth aspect of the embodiments of the present invention,

[0049] 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.

[0050] In the present invention, association rules and frequent patterns of different data sources are identified through data mining, heterogeneous data are mapped to a common semantic space in combination with ontology mapping, and entities and nodes in the knowledge graph are embedded in a low-dimensional vector space by using knowledge representation learning, so as to realize the semantic association and similarity calculation of process knowledge, select key samples for active learning through causal correlation analysis and semantic similarity calculation, expand the knowledge graph, and improve the accuracy of causal reasoning, build heterogeneous model integration through stacking generalization architecture, and realize real-time update by combining the incremental learning algorithm of streaming data, accurately capture the dynamic changes of process parameters and yield rate, and provide a reliable decision-making basis for causal intervention, and multi-objective optimization can obtain more balanced and comprehensive optimization results, meeting the diversified needs in actual production, and clarifying the influence mechanism of process parameter adjustment on yield rate through causal intervention, so as to improve the interpretability and reliability of optimization decision, in summary, the present invention makes full use of the advantages of data-driven and knowledge-driven, realizes the intelligent and automated process optimization, and provides important technical support for the transformation and upgrading of the manufacturing industry, which has great theoretical significance and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of a flow chart of an integrated circuit process parameter optimization method based on machine learning according to an embodiment of the present invention;

[0052] Figure 2 Schematic diagram of the structure of an integrated circuit process parameter optimization system based on machine learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] 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.

[0054] 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.

[0055] Figure 1 Schematic diagram of the process of the embodiment of the present invention, such as Figure 1 As shown, the method includes:

[0056] S1. Obtain process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with an ontology mapping method, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve the semantic similarity between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph based on pre-acquired expert experience knowledge, extract key process knowledge and construct a priori knowledge base in combination with natural language processing technology;

[0057] The process equipment is a machine or tool used to perform specific manufacturing or processing tasks, usually including machines, sensors, etc. The process history data is a record of various types of data in the production process, including parameter settings, production results and fault information, which is used to analyze and optimize the production process. The data mining technology is a technology used to extract valuable information and patterns from large amounts of data, and is often used for prediction and decision support. The frequent pattern is a pattern or item set that appears more frequently than a preset threshold in a data set, and is usually used for association rule mining. The ontology mapping method is a technology that matches and transforms concepts and relationships between different knowledge ontologies to achieve data sharing and interoperability. The public semantic space is a shared semantic framework that allows knowledge between different systems or applications. The low-dimensional continuous vector space is a representation method that maps high-dimensional data to low-dimensional space, which is usually used to simplify calculations and improve model performance. The representation vector operation is the process of performing mathematical operations on data or knowledge representation, which usually involves vector addition, subtraction, dot product and other operations to obtain information such as similarity or distance. The semantic similarity is an indicator to measure the degree of semantic similarity between different concepts or objects, which is usually based on word meaning, context and other information. The process technology domain knowledge is professional knowledge related to integrated circuit manufacturing or processing, including best practices, technical specifications and experience. The prior knowledge base is a database that stores existing knowledge and experience, which is usually used for reasoning and decision support to improve the accuracy of analysis.

[0058] In an optional embodiment,

[0059] Acquire process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with ontology mapping methods, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph according to pre-acquired expert experience knowledge, extract key process knowledge in combination with natural language processing technology and construct a priori knowledge base, including:

[0060] Acquire initial process history data from multi-source heterogeneous process equipment and test equipment, wherein the initial process history data includes equipment parameters, process monitoring data and test results, pre-process the initial process history data, combine data standardization and normalization methods to perform unified representation, and obtain the process history data, perform association rule mining on the process history data according to a priori algorithms, generate association rules, and perform frequent pattern mining on the process history data through a frequent pattern growth algorithm, determine the co-occurrence patterns between different data sources, and perform significance tests and confidence scores on the association rules and the co-occurrence patterns through statistics, and determine strong association rules and high-frequency patterns;

[0061] Based on the strong association rules and the high-frequency patterns, the process history data from different data sources are mapped as heterogeneous data to a common semantic space through an ontology mapping method, a process technology ontology is constructed, and the semantic similarity between different data sources is determined through a semantic similarity calculation method, semantic fusion and interoperation are performed based on the semantic similarity to obtain fused data, based on the process technology ontology and the fused data, process technology knowledge is represented in a process resource description framework format and a process technology knowledge graph is constructed, wherein entities in the process technology knowledge graph include process equipment, process parameters, test indicators and material properties, and entities and relationships in the process technology knowledge graph are embedded into a low-dimensional continuous vector space through a knowledge graph embedding method to obtain a distributed representation of entities and relationships;

[0062] Based on the distributed representation, the semantic similarity between entities and relationships is solved by representation vector operations in the embedded vector space to obtain process technology domain knowledge, and multi-step relational reasoning is performed on the process technology knowledge graph through a path-based reasoning method to generate implicit associations between entities and generate incremental process technology knowledge in combination with the semantic operation rules in the embedded vector space. Based on the incremental process technology knowledge and pre-acquired expert experience knowledge, the key process knowledge in the expert experience knowledge is extracted through natural language processing technology, the key process knowledge is mapped to the process technology knowledge graph, and knowledge alignment and ontology matching are performed through a knowledge fusion method to obtain an enhanced process technology knowledge graph, and the key process knowledge is used as prior knowledge to construct the prior knowledge base.

[0063] The frequent pattern growth algorithm is an efficient algorithm for discovering frequent item sets, which shortens the calculation time by recursively generating candidate item sets. The co-occurrence pattern is a pattern of items or events that appear at the same time, which is often used to analyze the correlation and mutual influence between variables. The significance test is a statistical method used to determine whether the observed results are statistically significant at a certain confidence level, and is often used to verify hypotheses. The process resource description framework format is a standardized format used to describe resource information in the manufacturing process, such as equipment, materials and process parameters. The path-based reasoning method is to deduce the relationship between nodes by analyzing the paths in the knowledge graph, thereby discovering potential causal relationships or associations. The semantic operation rules are rules that specify how to perform reasoning and operations at the semantic level, which are used to ensure the consistency and effectiveness of reasoning. The incremental process technology knowledge is new process technology domain knowledge generated as data increases.

[0064] Acquire initial process history data from multi-source heterogeneous process equipment and test equipment, including equipment parameters, process monitoring data and test results, pre-process the initial process history data, combine data standardization and normalization methods to uniformly represent it, obtain process history data, perform association rule mining on process history data according to a priori algorithms, generate association rules, perform frequent pattern mining on process history data through frequent pattern growth algorithms, determine co-occurrence patterns between different data sources, perform significance tests and confidence scores on association rules and co-occurrence patterns through statistical methods, and determine strong association rules and high-frequency patterns;

[0065] Based on strong association rules and high-frequency patterns, the process history data from different data sources are mapped as heterogeneous data to the public semantic space through the ontology mapping method to build the process technology ontology. The semantic similarity between different data sources is determined by the semantic similarity calculation method. Semantic fusion and interoperation are performed based on semantic similarity to obtain fused data. Based on the process technology ontology and fused data, the process technology knowledge is represented in the process resource description framework format to build a process technology knowledge graph, in which entities include process equipment, process parameters, test indicators and material properties. The entities and relationships in the process technology knowledge graph are embedded into a low-dimensional continuous vector space through the knowledge graph embedding method to obtain a distributed representation of entities and relationships.

[0066] Based on distributed representation, the semantic similarity between entities and relations is solved by representation vector operations in the embedded vector space to obtain process technology domain knowledge. Multi-step relational reasoning is performed on the process technology knowledge graph through a path-based reasoning method to generate implicit associations between entities. Incremental process technology knowledge is generated by combining the semantic operation rules in the embedded vector space. Based on the incremental process technology knowledge and the pre-acquired expert experience knowledge, the key process knowledge in the expert experience knowledge is extracted through natural language processing technology, and the key process knowledge is mapped to the process technology knowledge graph. Knowledge alignment and ontology matching are performed through knowledge fusion methods to obtain an enhanced process technology knowledge graph. The key process knowledge is used as prior knowledge to construct a prior knowledge base.

[0067] For example, it is assumed that in a semiconductor manufacturing process, the historical data of multiple process equipment are acquired and preprocessed to obtain standardized and normalized process historical data, and the strong association rules between etching time and etching depth are discovered through association rule mining. The co-occurrence pattern of test results between different test equipment is discovered through frequent pattern mining, and the data of different equipment are mapped to a common semantic space. A process ontology containing entities such as equipment parameters, process parameters and test indicators is constructed, and fused data is obtained through semantic fusion, a semiconductor manufacturing process knowledge graph is constructed, and a distributed representation of entities and relationships is obtained through knowledge graph embedding;

[0068] By using distributed representation, the semantic similarity between different entities in the embedded vector space was calculated, and the close relationship between etching depth and transistor performance was discovered, forming new process technology domain knowledge. By performing multi-step relational reasoning on the knowledge graph, the implicit association between etching conditions and transistor performance was discovered, and incremental process technology knowledge was generated. Combined with the experience and knowledge of experts in the semiconductor field, key process knowledge was extracted through natural language processing technology, such as the influence of etching gas flow and pressure on etching depth, and mapped and integrated into the enhanced semiconductor manufacturing process knowledge graph. Using these key process knowledge as prior knowledge, an a priori knowledge base was constructed to provide knowledge support for subsequent process optimization and decision-making.

[0069] In this embodiment, by preprocessing, mining association rules and frequent patterns of multi-source heterogeneous data, data fusion across devices and processes is achieved, key knowledge in the process technology is extracted, and the foundation is laid for subsequent knowledge representation and reasoning. By organizing the process technology knowledge into a knowledge graph and applying the knowledge graph embedding technology, the structured and semantic representation of knowledge is achieved, the interpretability and computability of knowledge are enhanced, and convenience is provided for knowledge reasoning and application. By performing path-based multi-step relational reasoning on the knowledge graph, the implicit association between process technology parameters and product performance is mined, new process technology knowledge is generated, and the original knowledge base is expanded. In summary, this embodiment realizes the automatic extraction, representation and application of process technology knowledge through the combination of data-driven and knowledge-driven, constructs a high-quality prior knowledge base, provides strong decision-making support for process optimization and product quality improvement, and has significant technical effects and application value.

[0070] In an optional embodiment,

[0071] The multi-step relational reasoning is performed on the process technology knowledge graph through the path-based reasoning method as shown in the following formula:

[0072] ;

[0073] Among them, P(v i |v 0 ) represents and starts from the starting node v 0 To the target node v i The probability, P(v 0 , v i ) indicates that from the starting node v 0 To the target node v i The set of all paths of , p represents the path index, |p| represents the length of the pth path, α represents the priority weight factor of entities and relations, Represents the slave node v j-1 To node vj The transition probability, Represents node v j-1 and node v j , λ represents the time decay factor, △t represents the time difference between the current moment and the moment when the knowledge is generated, which is used to calculate the time decay.

[0074] In this embodiment, by considering the multi-step relationship path between entity nodes in the knowledge graph, entity pairs that are not directly connected but have implicit associations can be found, thereby expanding the coverage of associated knowledge. When calculating the transfer probability between nodes, a semantic similarity factor based on vector representation is introduced, so that semantically similar node pairs have a higher transfer probability, thereby improving the semantic rationality of reasoning. The impact of the timeliness of knowledge generation on the credibility of reasoning is taken into account. The influence of knowledge that is far away from the current moment will decay over time, making the reasoning results more dynamic and real-time. In summary, this embodiment improves the comprehensiveness and accuracy of knowledge reasoning, makes the reasoning process more in line with process reality, has stronger interpretability and credibility, and provides more comprehensive, accurate and real-time knowledge support for process decisions.

[0075] S2. Based on the process technology knowledge graph, feature extraction and fusion are performed through a graph convolutional neural network to generate feature representations corresponding to process parameters, the causal dependency between process parameters and yield is determined through causal inference theory, the causal graph is determined and causal feature selection criteria are set in combination with a causal structure learning algorithm based on a graph model, a key causal feature subset is screened, key samples are selected and active learning is performed in combination with causal correlation analysis and semantic similarity calculation, the process technology knowledge graph is expanded, and a heterogeneous model integration is constructed through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space, and according to the update of the process technology knowledge graph, the heterogeneous model integration is updated in real time through an incremental learning algorithm of streaming data and the dynamic changes of process parameters and yield are determined;

[0076] The graph convolutional neural network is a deep learning model for processing graph structure data. By propagating node features in the graph, it can achieve effective representation and classification of nodes and graphs. The causal inference theory is a methodology for studying causal relationships, which aims to infer causal relationships between variables through observation data or experimental design to help decision-making and policy making. The causal dependency refers to the fact that in a system, the change of a certain variable will directly affect the state of another variable. The causal structure learning algorithm based on the graph model aims to learn causal relationships from data and construct a causal graph. The causal graph is a representation of the graph model, in which nodes represent variables and edges represent causal relationships between variables. The causal feature selection criteria are standards for screening important causal features, including average causal effect and Conditional average treatment effect, etc. The key causal feature subset is a set of features that are screened from all possible features and have the most significant impact on the target variable. The causal correlation analysis is the process of evaluating the strength of causal relationships between variables, aiming to identify potential causal relationships by analyzing patterns in the data. The stacked generalization architecture is an integrated learning method that combines the prediction results of multiple base learners to improve the generalization ability and prediction performance of the model. The heterogeneous model integration refers to combining different types of learning models together to fully utilize the advantages of each model and improve overall performance and accuracy. The incremental learning algorithm for streaming data is used for real-time processing and learning of dynamically generated data streams, and usually needs to be efficient and adaptable in order to update the model in a timely manner and make accurate predictions.

[0077] In an optional embodiment,

[0078] Based on the process technology knowledge graph, feature extraction and fusion are performed through a graph convolutional neural network to generate feature representations corresponding to process parameters, the causal dependency between process parameters and yield is determined through causal inference theory, the causal graph is determined and causal feature selection criteria are set in combination with a causal structure learning algorithm based on a graph model, a key causal feature subset is screened, key samples are selected and active learning is performed in combination with causal correlation analysis and semantic similarity calculation, the process technology knowledge graph is expanded, and a heterogeneous model integration is constructed through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space. According to the update of the process technology knowledge graph, the heterogeneous model integration is updated in real time through an incremental learning algorithm of streaming data and the dynamic changes of process parameters and yield are determined, including:

[0079] Representing the process technology knowledge graph as an undirected weighted graph and setting a corresponding weight matrix for each edge in the undirected weighted graph, constructing a graph convolutional neural network and learning the hidden layer representation corresponding to each node based on the attribute information and topological structure information of the node in the undirected weighted graph, defining a forward propagation rule of the graph convolutional neural network based on the hidden layer representation, performing forward propagation in the graph convolutional neural network according to the forward propagation rule, stacking multiple graph convolutional neural network layers, and extracting feature representations corresponding to process parameters;

[0080] Based on the feature representation, the Robin causal model is selected to formally describe the potential result framework, the process parameters are used as processing variables, the yield rate is used as the result variable, the potential results corresponding to different processing variables are determined according to the causal effect, the conditional independence test is performed through the causal structure learning algorithm, and a causal graph is constructed according to the test results, the key processing variables are identified, and the average causal effect and the conditional average processing effect are used as causal feature selection criteria, the influence of each process parameter on the yield rate is calculated, and the key causal feature subset is obtained by screening according to the preset influence threshold;

[0081] Based on the key causal feature subset, the causal correlation coefficient between each process parameter and the yield rate is calculated through causal correlation analysis, each pair of data is sorted in descending order according to the causal correlation coefficient, and the data pairs in the top 10% are selected as candidate data pairs. The semantic similarity between the candidate data pairs and the entities in the process technology knowledge graph is calculated through a semantic similarity calculation method, and the key samples are obtained by screening according to a preset similarity threshold. Through an active learning strategy, the amount of information corresponding to each key sample is determined according to an uncertainty sampling criterion and the key sample with the maximum amount of information is selected to generate a knowledge triple, and the process technology knowledge graph is expanded according to the knowledge triple;

[0082] According to the expanded process technology map and the common semantic space, different types of pre-acquired prediction models are combined through a stacked generalized architecture to construct a heterogeneous model integration. Based on the updated process technology knowledge map, the parameters in the heterogeneous model integration are updated through the incremental learning algorithm of streaming data, and the updated heterogeneous model integration is used to predict the dynamic changes between the process parameters and yield.

[0083] The undirected weighted graph is a graph structure in which the edges have no direction and are assigned weights to represent the strength of the relationship between nodes. The hidden layer representation is the process of transforming input data and extracting features through the hidden layer in the neural network. The forward propagation rule is the process of calculating the output in the neural network, passing information layer by layer from the input layer to the output layer. The formal description is a precise definition of the system or model, and mathematical symbols and language are usually used to express its structure and behavior. The Robin causal model is a causal inference framework that emphasizes the concept of potential results and allows researchers to evaluate the treatment effect under intervention. The potential result framework is a theoretical basis for causal inference, emphasizing the potential results of each individual under different interventions. The treatment variable refers to a variable that can be used in experiments or observations. For the manipulated independent variable, the researcher observes the impact of the treatment variable on the outcome variable by changing the treatment variable. The outcome variable is the variable affected by the treatment variable, reflecting the effect or result of the treatment. The causal structure learning algorithm is used to identify causal relationships from the data and establish a causal graph structure. The conditional independence test is a statistical method used to determine whether two variables are independent given other variables. The average causal effect refers to the average difference in the outcome variable between the treatment group and the control group under a specific intervention. The conditional average treatment effect refers to the difference between the average result of the treatment group and the average result of the control group under specific conditions, emphasizing the analysis of specific subsets. The causal correlation coefficient is a measure that reflects the strength and direction of the causal relationship between two variables.

[0084] The process technology knowledge graph is represented as an undirected weighted graph, and a corresponding weight matrix is ​​set for each edge to construct a graph convolutional neural network. The hidden layer representation corresponding to each node is learned based on the attribute information and topological structure information of the nodes in the undirected weighted graph. Based on the hidden layer representation, the forward propagation rule of the graph convolutional neural network is defined, and forward propagation is performed in the graph convolutional neural network according to the forward propagation rule. Multiple graph convolutional neural network layers are stacked to extract the feature representation corresponding to the process parameters. Based on the feature representation of the process parameters, the Robin causal model is selected to formally describe the potential result framework. The process parameters are used as processing variables and the yield rate is used as the result variable. The potential results corresponding to different processing variables are determined according to the causal effect. The conditional independence test is performed through the causal structure learning algorithm. The causal graph is constructed according to the test results, and the key processing variables are identified. The average causal effect and the conditional average processing effect are used as the causal feature selection criteria. The impact of each process parameter on the yield rate is calculated, and the key causal feature subset is obtained by screening according to the pre-set impact threshold.

[0085] Based on the key causal feature subset, the causal correlation coefficient between each process parameter and the yield is calculated through causal correlation analysis, and each pair of data is sorted in descending order according to the causal correlation coefficient. The data pairs in the top 10% are selected as candidate data pairs. The semantic similarity between the candidate data pairs and the entities in the process technology knowledge graph is calculated through the semantic similarity calculation method. The key samples are screened according to the pre-set similarity threshold. The amount of information corresponding to each key sample is determined according to the uncertainty sampling criterion through the active learning strategy. The key samples with the maximum amount of information are selected to generate knowledge triples. The process technology knowledge graph is expanded according to the knowledge triples. According to the expanded process technology knowledge graph and the public semantic space, the different types of pre-acquired prediction models are combined through the stacked generalization architecture to construct a heterogeneous model integration. Based on the updated process technology knowledge graph, the parameters in the heterogeneous model integration are updated through the incremental learning algorithm of streaming data. The updated heterogeneous model integration is used to predict the dynamic changes between process parameters and yield.

[0086] Exemplarily, by collecting historical data, a process technology knowledge graph containing information such as equipment parameters, process parameters and product yield is constructed. The process technology knowledge graph is represented as an undirected weighted graph, a weight matrix is ​​set for each edge, and a graph convolutional neural network containing 3 layers is constructed. The hidden layer representation of each process parameter node is obtained through node attribute and topological structure learning. The Robin causal model is selected to formally describe the potential result framework. Process parameters such as etching time and oxygen flow rate are used as processing variables, and wafer yield is used as the result variable. The conditional independence test is performed through the causal structure learning algorithm, the causal graph is constructed, and the average causal effect of each process parameter on the yield is calculated. According to the impact threshold, etching time, oxygen flow rate and RF power are screened out as the key causal feature subsets. Based on the key causal feature subsets, the effect of each process parameter on the yield is calculated. The causal correlation coefficient between the number and the yield rate is calculated, and the data pairs in the top 10% are selected as candidate data pairs. The semantic similarity between the candidate data pairs and the entities in the knowledge graph is calculated through the Word2Vec model, and key samples such as "etching time-10s" and "oxygen flow-50sccm" are screened out. According to the uncertainty sampling criterion, the key samples with the largest amount of information are selected to generate knowledge triples, such as (etching time, impact, yield rate), and added to the knowledge graph. According to the expanded knowledge graph and public semantic space, the decision tree, support vector machine and other prediction models are combined through the stacking generalization method to construct a heterogeneous model integration. When new process parameter data appears on the production line, the model parameters are updated in real time through the incremental learning algorithm of the streaming data. The updated model integration is used to predict the dynamic impact of process parameter adjustment on the yield rate.

[0087] In this embodiment, by representing the process technology knowledge graph as an undirected weighted graph and using a graph convolutional neural network to learn the hidden layer representation of the nodes, the structured information and node attribute information of the knowledge graph are fully utilized to obtain high-quality feature representation of the process parameters, which provides a good feature basis for subsequent causal reasoning and key sample selection. Through indicators such as average causal effect and conditional average treatment effect, the key process parameters with the greatest impact on the yield rate are screened out, the automatic selection of causal features is realized, and the pertinence of process optimization is improved. An active learning strategy is adopted to select the samples with the largest amount of information according to the uncertainty sampling criterion to generate knowledge triples, expand the knowledge graph, and continuously enrich and improve it. The process technology knowledge base improves the coverage and accuracy of knowledge, and adopts the incremental learning algorithm of streaming data to update the model parameters in real time, so that the prediction model can adapt to the dynamic changes of process parameters, realize the dynamic prediction of yield rate, and provide strong support for real-time process optimization. Through technologies such as causal reasoning and active learning, the whole process optimization from knowledge representation to knowledge expansion to knowledge application is realized, and the intelligence level of process technology optimization is improved. In summary, this embodiment realizes the comprehensive integration of process technology data, automatic extraction and updating of knowledge, and dynamic optimization of process parameters, which significantly improves the intelligence level and real-time performance of process optimization and provides a strong guarantee for the continuous improvement of process quality.

[0088] In an optional embodiment,

[0089] According to the forward propagation rule, forward propagation is performed in the graph convolutional neural network as shown in the following formula:

[0090] ;

[0091] in, represents the node feature matrix of the l+1th layer, σ represents the activation function, is the degree matrix, which indicates the number of connections of each node. is the adjacency matrix with self-loops added to represent the connection relationship between nodes. represents the node feature matrix of the lth layer, represents the attention weight matrix, Represents the residual connection weight matrix.

[0092] In this embodiment, by introducing an adjacency matrix with self-loops, not only the direct connection relationship between nodes is considered, but also the features of the node itself are included in the calculation through self-loops, so that the feature representation of each node contains both local neighborhood information and global topological structure information, thereby improving the expressiveness of the node representation. By introducing an attention weight matrix, the features of different neighbor nodes are weighted and aggregated, and the importance of neighbor nodes is adaptively adjusted. Through multi-level feature propagation and fusion, a more abstract and high-level node representation can be obtained. Residual connections can alleviate the gradient vanishing problem in deep model training, promote the back propagation of gradients, and improve the training efficiency and convergence speed of the model. In summary, this embodiment effectively captures the local and global information of the node, realizes multi-level feature extraction and nonlinear transformation, improves the quality and expressiveness of the node representation, and provides strong technical support for the representation and optimization of process technology knowledge based on graph convolutional neural networks.

[0093] S3. Based on the dynamic changes, a multi-objective optimization model is constructed and multiple optimization objectives are set. The process parameters are optimized and solved by a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset. The optimized process parameters are generated and transfer learning is performed through feature mapping and parameter adaptation according to the semantic mapping relationship of the process nodes in the process technology knowledge graph. The process nodes are compared and mapped according to the pre-constructed Bayesian causal model, the causal pattern is determined, and a virtual simulation system is integrated based on the Bayesian causal model and the heterogeneous model, and causal intervention is performed on key process parameters. A process parameter optimization decision is generated and the Bayesian causal model is corrected according to the process parameter optimization decision.

[0094] The multi-objective evolutionary optimization algorithm is a type of algorithm used to simultaneously optimize multiple conflicting objective functions, usually based on the principles of genetic algorithms. The semantic mapping relationship refers to the process of establishing connections between concepts and entities in different fields or different representation forms. The feature mapping is the process of transforming input data from the original space to a new feature space, usually used to improve the performance of the model. The parameter adaptation refers to adjusting the parameter settings of the model according to different tasks and data characteristics to optimize the performance and adaptability of the model. The Bayesian causal model is a causal analysis framework based on Bayesian inference, which allows researchers to infer causal relationships between variables and evaluate intervention effects through prior knowledge and data updates. The causal intervention refers to applying operations to a variable in the system to observe its impact on other variables.

[0095] In an optional embodiment,

[0096] Based on the dynamic changes, a multi-objective optimization model is constructed and multiple optimization objectives are set. The process parameters are optimized and solved by a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset, the optimized process parameters are generated, and according to the semantic mapping relationship of the process nodes in the process technology knowledge graph, transfer learning is performed through feature mapping and parameter adaptation, the process nodes are compared and mapped according to the pre-constructed Bayesian causal model, the causal pattern is determined, and a virtual simulation system is constructed based on the Bayesian causal model and the heterogeneous model, and causal intervention is performed on key process parameters, and a process parameter optimization decision is generated and the Bayesian causal model is corrected according to the process parameter optimization decision, including:

[0097] Acquire actual production demand, build a multi-objective optimization model according to the actual production demand and the dynamic change of the current time step, select multiple process parameters as decision variables, maximize the yield rate, minimize the production cost and shorten the production cycle as optimization goals, set constraints and establish the objective function corresponding to the multi-objective optimization model according to the physical meaning and actual limitations of the process parameters in the key causal feature subset;

[0098] Based on a multi-objective evolutionary optimization algorithm, the decision variables are mapped to individual codes in the algorithm, multiple individuals are randomly generated and combined as an initial population, for each individual in the initial population, the fitness value under multiple optimization objectives is calculated and an initial non-dominated solution set is obtained, an initial global optimal solution is selected from the initial non-dominated solution set and the prey position is set according to the position of the initial global optimal solution, for each individual in the initial population, according to the prey position and the hunting behavior of whales, the position is updated by simulating the whales rotating around the prey and approaching, the fitness value of each updated individual is calculated again and the prey position is updated according to the position in the current non-dominated solution set, and the update is repeated until the preset maximum number of iterations is reached to obtain the optimized process parameters;

[0099] Add the optimized process parameter as a new node to the process technology knowledge graph to generate an optimized process parameter node, construct a semantic mapping matrix according to the semantic associations between different nodes in the process technology knowledge graph, determine the source process node with the highest similarity for the optimized process parameter node through the semantic mapping matrix, extract the feature representation corresponding to the source process node and map it to the optimized process parameter node through a feature mapping function to obtain a mapping feature representation, adjust the mapping feature representation through a parameter adaptation function to obtain an optimized model parameter;

[0100] The optimization model parameters are added to the Bayesian causal model pre-constructed according to the process technology knowledge graph through a transfer learning algorithm. Based on the Bayesian causal model, the causal patterns of different optimization process parameter nodes are compared and mapped in combination with a causal reasoning algorithm to identify common causal patterns and special causal patterns. Based on the integration of the Bayesian causal model and the heterogeneous model, causal intervention is performed on the key process parameters according to the common causal patterns and the special causal patterns to generate process parameter optimization decisions and update the Bayesian causal model and the process technology knowledge graph.

[0101] The yield rate refers to the ratio of qualified products to total production products in the production process. It is an important indicator for measuring production efficiency and quality, and is usually used to optimize process parameters and decisions. The individual encoding refers to the assignment of a specific representation form to each individual or solution in the evolutionary algorithm so as to perform crossover, mutation and selection in genetic operations. The non-dominated solution set refers to a set of solutions in which no other solution is better than the current solution in all objectives in multi-objective optimization. The semantic mapping matrix is ​​a matrix that represents the semantic relationship between different concepts or entities. It is usually used in fields such as knowledge graphs and information retrieval to help identify similarities and correlations. The source process node refers to the starting or key node in the process flow, which is usually used to describe the starting point or important link of the process flow. The feature mapping function is a function used to convert input data into feature space, which helps to extract the key features of the data for subsequent analysis and model training. The common causal pattern refers to the causal relationship pattern that is prevalent in multiple systems or situations, which helps to identify causal laws in different environments and facilitates reasoning and intervention. The special causal pattern refers to the causal relationship observed under specific conditions or scenarios, which has uniqueness and limitations and is usually used for in-depth analysis of specific problems or phenomena.

[0102] Obtain actual production demand, build a multi-objective optimization model based on the actual production demand and the dynamic changes of the current time step, select multiple process parameters as decision variables, and take maximizing yield, minimizing production cost and shortening production cycle as optimization goals. According to the physical meaning and actual limitations of process parameters in the key causal feature subset, set constraints and establish the objective function corresponding to the multi-objective optimization model;

[0103] Based on a multi-objective evolutionary optimization algorithm, the decision variables are mapped to individual codes in the algorithm, multiple individuals are randomly generated and combined as the initial population, for each individual in the initial population, the fitness value under multiple optimization objectives is calculated and the initial non-dominated solution set is obtained, the initial global optimal solution is selected from the initial non-dominated solution set, and the prey position is set according to the position of the initial global optimal solution, for each individual in the initial population, according to the prey position and the hunting behavior of the whale, the position is updated by simulating the whale's rotation and approach around the prey, the fitness value of each updated individual is calculated again, and the prey position is updated according to the position in the current non-dominated solution set, and the update is repeated until the preset maximum number of iterations is reached to obtain the optimized process parameters;

[0104] Add the optimized process parameters as new nodes to the process technology knowledge graph to generate optimized process parameter nodes, build a semantic mapping matrix based on the semantic associations between different nodes in the process technology knowledge graph, determine the source process node with the highest similarity through the semantic mapping matrix for the optimized process parameter nodes, extract the feature representation corresponding to the source process node, and map it to the optimized process parameter node through the feature mapping function to obtain the mapping feature representation, adjust the mapping feature representation through the parameter adaptation function to obtain the optimized model parameters, add the optimized model parameters to the Bayesian causal model pre-built according to the process technology knowledge graph through the transfer learning algorithm, compare and map the causal patterns of different optimized process parameter nodes based on the Bayesian causal model and the causal reasoning algorithm, identify the common causal patterns and special causal patterns, perform causal intervention on key process parameters based on the common causal patterns and special causal patterns based on the Bayesian causal model and heterogeneous model integration, generate process parameter optimization decisions, and update the Bayesian causal model and the process technology knowledge graph;

[0105] For example, when an integrated circuit manufacturing company produces 0.18μm CMOS process chips, it faces problems such as unstable yield, high production cost, and difficulty in shortening the production cycle. According to the real-time changes in customer orders and equipment performance, a multi-objective optimization model is constructed, and key process parameters such as lithography exposure time, development time, and etching depth are selected as decision variables. The optimization objectives are to maximize the yield, minimize the production cost, and minimize the production cycle. At the same time, according to the physical meaning of the process parameters and equipment limitations, corresponding constraints are set, such as exposure time between 0.5-2 seconds, development time not exceeding 60 seconds, etching depth controlled between 200-500nm, etc. Based on the multi-objective whale optimization algorithm (MOWOA), the decision variables are mapped to whale individual codes. 100 whale individuals are randomly generated as the initial population, and the fitness value of each individual under the three optimization objectives is calculated to obtain the initial non-dominated solution set, and the individual with the highest comprehensive fitness value is selected as the initial prey position. Then, by simulating the behavior of whales rotating and approaching their prey, the position of each individual was updated, the fitness value was repeatedly calculated, the non-dominated solution set and the prey position were updated, and 500 iterations were reached. A set of optimized process parameter combinations was obtained, such as exposure time of 1.2 seconds, development time of 45 seconds, etching depth of 350nm, etc. The optimized process parameter combination was added to the integrated circuit process knowledge graph to generate a new optimized process parameter node. By calculating the semantic similarity between nodes, a semantic mapping matrix was constructed. For the new optimized process parameter node, a source process node with the highest similarity was found through the semantic mapping matrix, and its feature representation was extracted and mapped to the optimized process parameter node. Then, the mapping features were fine-tuned through the parameter adaptation function to obtain the optimized model parameters, and the optimized model parameters were added to the pre-built Bayesian causal model through the transfer learning algorithm. Combined with the causal reasoning algorithm, the common causal patterns between different optimization process parameter nodes are identified, such as "increased exposure time → reduced lithography size → improved yield", "too long development time → increased sidewall roughness → decreased yield", etc., as well as special causal patterns, such as "increased etching depth → increased dielectric damage → increased production costs", etc., and causal intervention is performed on key process parameters, generating a series of process optimization decisions, such as appropriately increasing exposure time, reducing development time, controlling etching depth, etc. According to the feedback of the optimization decisions, the Bayesian causal model and the integrated circuit process technology knowledge graph are dynamically updated.

[0106] In this embodiment, a multi-objective optimization model is constructed, and multiple optimization objectives such as yield rate, production cost and production cycle are considered at the same time, and the whale optimization algorithm is introduced for solving. A combination of process parameters that performs well under multiple objectives is obtained, which can obtain a more balanced and comprehensive optimization result and meet the diversified needs in actual production. Through the knowledge graph, the intrinsic connection between the process parameters can be revealed, which provides a basis for subsequent causal analysis. The introduction of causal reasoning makes process optimization no longer a black box process, but is based on the understanding and control of the process mechanism, which improves the pertinence and reliability of optimization. Through model integration, the optimization system can enhance its adaptability to uncertainty and dynamic changes, and improve the robustness of optimization decisions. In summary, this embodiment establishes a closed-loop optimization mechanism from data to knowledge to decision-making, which significantly improves the efficiency, accuracy and interpretability of process optimization, provides new ideas and methods for the development of intelligent manufacturing, and has broad application prospects.

[0107] In an optional embodiment,

[0108] The causal intervention of the key process parameters according to the common causal model and the special causal model is shown in the following formula:

[0109] ;

[0110] in, represents the probability of the potential value of the outcome variable Y when setting X to a specific value x, given the actual observed value x' of the key process parameter X and the specific value y of the optimization target Y. x is the counterfactual outcome variable, which indicates the potential value of the optimization target Y when the key process parameter X is set to x, where X represents the key process parameter, x' represents the actual observed value of the key process parameter X, and Y represents the optimization target. It indicates the probability that the optimization target Y takes the value y under the intervention condition. It represents the probability that the key process parameter X actually takes the value x' under the condition that the optimization target takes the value y. It represents the marginal probability that the actual value of the key process parameter X is x'.

[0111] In this embodiment, by expressing the effectiveness of causal intervention, a quantitative decision-making basis is provided for process optimization. By reflecting the impact of the actual state of process parameters on the optimization target, the causal intervention is closer to the actual production situation. By incorporating marginal probability, the formula takes into account the overall distribution of the optimization target, making the causal intervention strategy more globally optimal. Precise regulation based on causal intervention can effectively improve the efficiency and quality of process optimization. In summary, this embodiment realizes precise regulation of key process parameters, improves the efficiency, quality and interpretability of process optimization, and provides a new idea and means for the optimization of complex process systems. At the same time, it also lays a theoretical foundation for the application of causal reasoning in the field of intelligent manufacturing, and promotes the deep integration of causal analysis technology and industrial practice.

[0112] Figure 2 Schematic diagram of the structure of the system of the embodiment of the present invention, as shown in Figure 2 As shown, the system comprises:

[0113] The first unit is used to obtain process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with an ontology mapping method, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph according to pre-acquired expert experience knowledge, extract key process knowledge and construct a priori knowledge base in combination with natural language processing technology;

[0114] The second unit is used to extract and fuse features based on the process technology knowledge graph through a graph convolutional neural network to generate feature representations corresponding to process parameters, determine the causal dependency between process parameters and yield rate through causal inference theory, determine the causal graph and set causal feature selection criteria in combination with a causal structure learning algorithm based on a graph model, screen out a key causal feature subset, select key samples in combination with causal correlation analysis and semantic similarity calculation and perform active learning, expand the process technology knowledge graph, build a heterogeneous model integration through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space, and update the heterogeneous model integration in real time based on the update of the process technology knowledge graph through an incremental learning algorithm of streaming data to determine the dynamic changes of process parameters and yield rate;

[0115] The third unit is used to build a multi-objective optimization model and set multiple optimization goals based on the dynamic changes, optimize and solve the process parameters through a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset, generate optimized process parameters and perform transfer learning through feature mapping and parameter adaptation according to the semantic mapping relationship of the process nodes in the process technology knowledge graph, compare and map the process nodes according to the pre-built Bayesian causal model, determine the causal pattern and integrate the Bayesian causal model and the heterogeneous model to build a virtual simulation system and perform causal intervention on key process parameters, generate process parameter optimization decisions and correct the Bayesian causal model according to the process parameter optimization decisions.

[0116] According to a third aspect of the embodiments of the present invention,

[0117] An electronic device is provided, comprising:

[0118] processor;

[0119] a memory for storing processor-executable instructions;

[0120] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0121] According to a fourth aspect of the embodiments of the present invention,

[0122] 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.

[0123] 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.

[0124] 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. An integrated circuit process parameter optimization method based on machine learning, characterized in that: include: Acquire process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with an ontology mapping method, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph based on pre-acquired expert experience knowledge, extract key process knowledge and construct a priori knowledge base in combination with natural language processing technology; Based on the process technology knowledge graph, feature extraction and fusion are performed through a graph convolutional neural network to generate feature representations corresponding to process parameters, the causal dependency between process parameters and yield is determined through causal inference theory, the causal graph is determined and causal feature selection criteria are set in combination with a causal structure learning algorithm based on a graph model, a key causal feature subset is screened, key samples are selected and active learning is performed in combination with causal correlation analysis and semantic similarity calculation, the process technology knowledge graph is expanded, and a heterogeneous model integration is constructed through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space, and according to the update of the process technology knowledge graph, the heterogeneous model integration is updated in real time through an incremental learning algorithm of streaming data and the dynamic changes of process parameters and yield are determined; Based on the dynamic changes, a multi-objective optimization model is constructed and multiple optimization objectives are set. The process parameters are optimized and solved by a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset, the optimized process parameters are generated, and according to the semantic mapping relationship of the process nodes in the process technology knowledge graph, transfer learning is performed through feature mapping and parameter adaptation. The process nodes are compared and mapped according to the pre-constructed Bayesian causal model, the causal pattern is determined, and a virtual simulation system is integrated based on the Bayesian causal model and the heterogeneous model, and causal intervention is performed on key process parameters, and a process parameter optimization decision is generated and the Bayesian causal model is corrected according to the process parameter optimization decision; Acquire process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with ontology mapping methods, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph according to pre-acquired expert experience knowledge, extract key process knowledge in combination with natural language processing technology and construct a priori knowledge base, including: Acquire initial process history data from multi-source heterogeneous process equipment and test equipment, wherein the initial process history data includes equipment parameters, process monitoring data and test results, pre-process the initial process history data, combine data standardization and normalization methods to perform unified representation, and obtain the process history data, perform association rule mining on the process history data according to a priori algorithms, generate association rules, and perform frequent pattern mining on the process history data through a frequent pattern growth algorithm, determine the co-occurrence patterns between different data sources, and perform significance tests and confidence scores on the association rules and the co-occurrence patterns through statistics, and determine strong association rules and high-frequency patterns; Based on the strong association rules and the high-frequency patterns, the process history data from different data sources are mapped as heterogeneous data to a common semantic space through an ontology mapping method, a process technology ontology is constructed, and the semantic similarity between different data sources is determined through a semantic similarity calculation method, semantic fusion and interoperation are performed based on the semantic similarity to obtain fused data, based on the process technology ontology and the fused data, process technology knowledge is represented in a process resource description framework format and a process technology knowledge graph is constructed, wherein entities in the process technology knowledge graph include process equipment, process parameters, test indicators and material properties, and entities and relationships in the process technology knowledge graph are embedded into a low-dimensional continuous vector space through a knowledge graph embedding method to obtain a distributed representation of entities and relationships; Based on the distributed representation, the semantic similarity between entities and relationships is solved by representation vector operations in the embedded vector space to obtain process technology domain knowledge, and multi-step relational reasoning is performed on the process technology knowledge graph through a path-based reasoning method to generate implicit associations between entities and generate incremental process technology knowledge in combination with the semantic operation rules in the embedded vector space. Based on the incremental process technology knowledge and pre-acquired expert experience knowledge, the key process knowledge in the expert experience knowledge is extracted through natural language processing technology, the key process knowledge is mapped to the process technology knowledge graph, and knowledge alignment and ontology matching are performed through a knowledge fusion method to obtain an enhanced process technology knowledge graph, and the key process knowledge is used as prior knowledge to construct the prior knowledge base.

2. The method according to claim 1, characterized in that The multi-step relational reasoning is performed on the process technology knowledge graph through the path-based reasoning method as shown in the following formula: Among them, P(v i |v0) represents and goes from the starting node v0 to the target node v i The probability, P(v0,v i ) represents the transition from the starting node v0 to the target node v i , p represents the path index, |p| represents the length of the pth path, α represents the priority weight factor of entities and relations, P(v j |v j-1 ) represents the slave node v j-1 To node v j The transition probability, φ(v j-1 , v j ) represents node v j-1 and node v j , λ represents the time decay factor, △t represents the time difference between the current moment and the moment when the knowledge is generated, which is used to calculate the time decay.

3. The method according to claim 1, characterized in that Based on the process technology knowledge graph, feature extraction and fusion are performed through a graph convolutional neural network to generate feature representations corresponding to process parameters, the causal dependency between process parameters and yield is determined through causal inference theory, the causal graph is determined and causal feature selection criteria are set in combination with a causal structure learning algorithm based on a graph model, a key causal feature subset is screened, key samples are selected and active learning is performed in combination with causal correlation analysis and semantic similarity calculation, the process technology knowledge graph is expanded, and a heterogeneous model integration is constructed through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space. According to the update of the process technology knowledge graph, the heterogeneous model integration is updated in real time through an incremental learning algorithm of streaming data and the dynamic changes of process parameters and yield are determined, including: Representing the process technology knowledge graph as an undirected weighted graph and setting a corresponding weight matrix for each edge in the undirected weighted graph, constructing a graph convolutional neural network and learning the hidden layer representation corresponding to each node based on the attribute information and topological structure information of the node in the undirected weighted graph, defining a forward propagation rule of the graph convolutional neural network based on the hidden layer representation, performing forward propagation in the graph convolutional neural network according to the forward propagation rule, stacking multiple graph convolutional neural network layers, and extracting feature representations corresponding to process parameters; Based on the feature representation, the Robin causal model is selected to formally describe the potential result framework, the process parameters are used as processing variables, the yield rate is used as the result variable, the potential results corresponding to different processing variables are determined according to the causal effect, the conditional independence test is performed through the causal structure learning algorithm, and a causal graph is constructed according to the test results, the key processing variables are identified, and the average causal effect and the conditional average processing effect are used as causal feature selection criteria, the influence of each process parameter on the yield rate is calculated, and the key causal feature subset is obtained by screening according to the preset influence threshold; Based on the key causal feature subset, the causal correlation coefficient between each process parameter and the yield rate is calculated through causal correlation analysis, each pair of data is sorted in descending order according to the causal correlation coefficient, and the data pairs in the top 10% are selected as candidate data pairs. The semantic similarity between the candidate data pairs and the entities in the process technology knowledge graph is calculated through a semantic similarity calculation method, and the key samples are obtained by screening according to a preset similarity threshold. Through an active learning strategy, the amount of information corresponding to each key sample is determined according to an uncertainty sampling criterion and the key sample with the maximum amount of information is selected to generate a knowledge triple, and the process technology knowledge graph is expanded according to the knowledge triple; According to the expanded process technology map and the common semantic space, different types of pre-acquired prediction models are combined through a stacked generalized architecture to construct a heterogeneous model integration. Based on the updated process technology knowledge map, the parameters in the heterogeneous model integration are updated through the incremental learning algorithm of streaming data, and the updated heterogeneous model integration is used to predict the dynamic changes between the process parameters and yield.

4. The method according to claim 3, characterized in that According to the forward propagation rule, forward propagation is performed in the graph convolutional neural network as shown in the following formula: Among them, H (l+1) represents the node feature matrix of the l+1th layer, σ represents the activation function, is the degree matrix, which indicates the number of connections of each node. is the adjacency matrix with self-loops added to represent the connection relationship between nodes, H (l) represents the node feature matrix of the lth layer, represents the attention weight matrix, Represents the residual connection weight matrix.

5. The method according to claim 1, characterized in that Based on the dynamic changes, a multi-objective optimization model is constructed and multiple optimization objectives are set. The process parameters are optimized and solved by a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset, the optimized process parameters are generated, and according to the semantic mapping relationship of the process nodes in the process technology knowledge graph, transfer learning is performed through feature mapping and parameter adaptation, the process nodes are compared and mapped according to the pre-constructed Bayesian causal model, the causal pattern is determined, and a virtual simulation system is constructed based on the Bayesian causal model and the heterogeneous model, and causal intervention is performed on key process parameters, and a process parameter optimization decision is generated and the Bayesian causal model is corrected according to the process parameter optimization decision, including: Acquire actual production demand, build a multi-objective optimization model according to the actual production demand and the dynamic change of the current time step, select multiple process parameters as decision variables, maximize the yield rate, minimize the production cost and shorten the production cycle as optimization goals, set constraints and establish the objective function corresponding to the multi-objective optimization model according to the physical meaning and actual limitations of the process parameters in the key causal feature subset; Based on a multi-objective evolutionary optimization algorithm, the decision variables are mapped to individual codes in the algorithm, multiple individuals are randomly generated and combined as an initial population, for each individual in the initial population, the fitness value under multiple optimization objectives is calculated and an initial non-dominated solution set is obtained, an initial global optimal solution is selected from the initial non-dominated solution set and the prey position is set according to the position of the initial global optimal solution, for each individual in the initial population, according to the prey position and the hunting behavior of whales, the position is updated by simulating the whales rotating around the prey and approaching, the fitness value of each updated individual is calculated again and the prey position is updated according to the position in the current non-dominated solution set, and the update is repeated until the preset maximum number of iterations is reached to obtain the optimized process parameters; Add the optimized process parameter as a new node to the process technology knowledge graph to generate an optimized process parameter node, construct a semantic mapping matrix according to the semantic associations between different nodes in the process technology knowledge graph, determine the source process node with the highest similarity for the optimized process parameter node through the semantic mapping matrix, extract the feature representation corresponding to the source process node and map it to the optimized process parameter node through a feature mapping function to obtain a mapping feature representation, adjust the mapping feature representation through a parameter adaptation function to obtain an optimized model parameter; The optimization model parameters are added to the Bayesian causal model pre-constructed according to the process technology knowledge graph through a transfer learning algorithm. Based on the Bayesian causal model, the causal patterns of different optimization process parameter nodes are compared and mapped in combination with a causal reasoning algorithm to identify common causal patterns and special causal patterns. Based on the integration of the Bayesian causal model and the heterogeneous model, causal intervention is performed on the key process parameters according to the common causal patterns and the special causal patterns to generate process parameter optimization decisions and update the Bayesian causal model and the process technology knowledge graph.

6. The method according to claim 5, characterized in that The causal intervention of the key process parameters according to the common causal model and the special causal model is shown in the following formula: Among them, P(Y x |X=x′,Y=y) represents the probability of the potential value of the outcome variable Y when setting X to a specific value x, given the actual observed value x' of the key process parameter X and the specific value y of the optimization target Y, Y x is the counterfactual outcome variable, which represents the potential value of the optimization target Y when the key process parameter X is set to x. X represents the key process parameter, x' represents the actual observed value of the key process parameter X, and Y represents the optimization target. P(Y=y|do(X=x)) represents the probability that the optimization target Y takes the value of y under the condition that the optimization target takes the value of y. P(X=x′|Y=y) represents the probability that the key process parameter X actually takes the value of x' under the condition that the optimization target takes the value of y. P(X=x′) represents the marginal probability that the key process parameter X actually takes the value of x'.

7. An integrated circuit process parameter optimization system based on machine learning, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain process history data from process equipment and test equipment, identify association rules and frequent patterns between different data sources in the process history data through data mining technology, map heterogeneous data to a common semantic space in combination with an ontology mapping method, construct a process technology knowledge graph according to a knowledge graph construction method, embed entities and nodes in the process technology knowledge graph into a low-dimensional continuous vector space through a knowledge representation learning algorithm, solve semantic similarities between entities and relationships through representation vector operations, obtain process technology domain knowledge, enhance the process technology knowledge graph according to pre-acquired expert experience knowledge, extract key process knowledge and construct a priori knowledge base in combination with natural language processing technology; The second unit is used to extract and fuse features based on the process technology knowledge graph through a graph convolutional neural network to generate feature representations corresponding to process parameters, determine the causal dependency between process parameters and yield rate through causal inference theory, determine the causal graph and set causal feature selection criteria in combination with a causal structure learning algorithm based on a graph model, screen out a key causal feature subset, select key samples in combination with causal correlation analysis and semantic similarity calculation and perform active learning, expand the process technology knowledge graph, build a heterogeneous model integration through a stacked generalization architecture based on the expanded process technology knowledge graph and the public semantic space, and update the heterogeneous model integration in real time based on the update of the process technology knowledge graph through an incremental learning algorithm of streaming data to determine the dynamic changes of process parameters and yield rate; The third unit is used to build a multi-objective optimization model and set multiple optimization goals based on the dynamic changes, optimize and solve the process parameters through a multi-objective evolutionary optimization algorithm in combination with the key causal feature subset, generate optimized process parameters and perform transfer learning through feature mapping and parameter adaptation according to the semantic mapping relationship of the process nodes in the process technology knowledge graph, compare and map the process nodes according to the pre-built Bayesian causal model, determine the causal pattern and integrate the Bayesian causal model and the heterogeneous model to build a virtual simulation system and perform causal intervention on key process parameters, generate process parameter optimization decisions and correct the Bayesian causal model according to the process parameter optimization decisions.

8. 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 according to any one of claims 1 to 6.

9. 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 6 is implemented.

Images