Graph neural network interpretability method, system and equipment for defect detection and medium

Abstract

The invention relates to the technical field of graph neural networks, in particular to a graph neural network interpretability method, system and device for defect detection and a medium. The method comprises the following steps: firstly, constructing a semiconductor graph according to acquired industrial semiconductor data; the features are embedded into different types of nodes, and a semiconductor graph neural network is constructed; secondly, initializing interpretable model parameters of the semiconductor graph neural network; then, a random approximate annealing algorithm is called to generate candidate nodes, marginal gains are calculated, and an approximate global optimal explanatory subgraph is obtained; and finally, calling a learning pruning algorithm, predicting explanatory scores of the nodes, and filtering until the loss function converges to obtain a trained global model. And the interpretability is improved, and meanwhile, the approximation of a global optimal solution and a higher convergence speed are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of graph neural networks, and in particular, to an interpretable method, system, device and medium of a graph neural network for defect detection. Background Art

[0002] Currently, Graph Neural Networks (GNNs) have become an important deep learning tool in industrial semiconductor analysis, and are widely used in fields such as semiconductor defect analysis, semiconductor defect detection, and semiconductor performance prediction. Graph neural networks can effectively capture the topological relationships between nodes in the semiconductor material structure, such as atoms, defects, and doping elements, and identify complex interaction patterns in the graph, demonstrating powerful data representation capabilities. However, with the wide application of graph neural networks in industrial semiconductors, their "black box" characteristics have attracted increasing attention. Especially in fields such as industrial semiconductor manufacturing and defect analysis, which have high requirements for model interpretability, how to improve the interpretability of graph neural networks has become a key issue.

[0003] To address the limitations of graph neural network interpretability methods in complex tasks, researchers have begun to explore methods that combine heuristic search and machine learning, attempting to find a balance between interpretability and efficiency. Heuristic search methods, such as Monte Carlo Tree Search (MCTS) and greedy algorithms, find explanatory subgraphs by simulating random paths, which can avoid local optimum problems to a certain extent. However, these methods often face problems of high computational complexity and slow convergence speed when dealing with large-scale semiconductor graph tasks, and it is difficult to adapt to diverse semiconductor graph data. In contrast, machine learning-based methods convert discrete problems into continuous and differentiable optimization problems, usually solved using gradient descent. However, since the problem space is usually non-convex, gradient descent is prone to getting stuck in local optima and it is difficult to guarantee finding the global optimal explanation. However, in the specific application of semiconductor graph neural network interpretability problems, heuristic search can provide a good global approximation, but it is inefficient on large-scale graphs; machine learning methods, although computationally efficient, are limited in their performance for non-convex problems.

[0004] With the rapid development of industrial big data technology, the existing GNN interpretability methods have gradually revealed defects. For example, when dealing with large-scale and complex semiconductor defect data and the relationship between material microstructures, it is difficult for existing methods to efficiently reveal the specific impact of key defect types or doping elements on material properties. In addition, these methods are insufficient in dealing with data heterogeneity and diversity, resulting in the interpretability results being less intuitive and reliable in practical applications such as semiconductor manufacturing and reliability analysis. Generally speaking, limited by the complexity of semiconductor material structures and the large amount of data, existing methods are difficult to guarantee the global optimal solution and cannot obtain ideal interpretability effects in complex tasks of semiconductor GNN models.

[0005] To address the limitations of semiconductor graph neural network interpretability methods in complex tasks, researchers have started to explore more advanced interpretability techniques. For example: combinations of multi-level instance-level and model-level methods, as well as structured interpretation methods for specific semiconductor material structures and defect mechanisms. These new methods attempt to provide more detailed explanations by identifying key subgraph structures and visualizing complex material micro interactions. However, these methods are prone to falling into the local optimal dilemma brought about by non-convex optimization problems in practice. In addition, when dealing with large-scale semiconductor device structure diagrams, there are also computational efficiency issues and it is difficult to effectively scale to industrial-level applications. Therefore, how to improve computational efficiency while ensuring interpretability effects has become an important challenge in semiconductor GNN interpretability research. Summary of the Invention

[0006] The present invention aims at the problems that the existing interpretability methods are prone to falling into local optimal solutions and have a slow convergence speed, and proposes a graph neural network interpretability method, system, device and medium for defect detection; first, construct a semiconductor graph according to the obtained industrial semiconductor data; and embed features into different types of nodes to construct a semiconductor graph neural network; secondly, initialize the interpretability model parameters of the semiconductor graph neural network; then call the stochastic approximation annealing algorithm to generate candidate nodes and calculate the marginal gain to obtain an approximately globally optimal explanatory subgraph; finally, call the learning pruning algorithm to predict and filter the interpretability scores of the nodes until the loss function converges to obtain a trained global model; while improving interpretability, ensure the approximation of the global optimal solution and a higher convergence speed.

[0007] The specific implementation content of the present invention is as follows: A graph neural network interpretability method for defect detection specifically includes the following steps: Step S1: Construct a semiconductor graph according to the type of the obtained semiconductor data and the information affecting the energy band structure; Step S2: Train a graph neural network according to the constructed semiconductor graph, embed features into different types of nodes, and construct a semiconductor graph neural network; Step S3: Initialize the interpretability model parameters of the semiconductor graph neural network; Step S4: Call the stochastic approximation annealing algorithm to generate candidate nodes and calculate the marginal gain to obtain an approximately globally optimal interpretive subgraph; Step S5: Call the learning pruning algorithm to predict the interpretive scores of nodes and filter them; Step S6: Repeat Step S2 - Step S5 until the loss function converges to obtain a trained global model.

[0008] To better implement the present invention, further, Step S1 specifically includes the following steps: Step S11: Obtain industrial semiconductor data and obtain the data type of the industrial semiconductor data and the information affecting the energy band structure; Step S12: Represent industrial semiconductor data of different data types as different types of nodes, and use the information affecting the energy band structure as edges to construct a semiconductor graph structure.

[0009] To better implement the present invention, further, Step S2 specifically includes the following steps: Step S21: Call the graph neural network to embed the features of different types of nodes and capture the structural relationships between nodes; Step S22: Call the cross-entropy loss function to calculate the classification error and call the consistency regularization output to construct a semiconductor graph neural network.

[0010] To better implement the present invention, further, the interpretability model parameters in Step S3 include the parameters of stochastic approximation annealing and the parameters of reinforcement learning pruning; The parameters of the stochastic approximation annealing include the initial temperature, the cooling factor, and the initial acceptance probability constant; The parameters of the reinforcement learning pruning include the learning rate of the policy network, the size of the action space, and the initial policy network parameters.

[0011] To better implement the present invention, further, Step S4 specifically includes the following steps: Step S41: Call the approximate annealing function to generate candidate nodes; Step S42: Calculate the marginal gain of the candidate nodes according to the candidate nodes, the interpretive score of the current subgraph, and the interpretive score after adding the candidate nodes; Step S43: Calculate the selection probability of the candidate nodes according to the marginal gain and the number of candidate nodes; Step S44: Converge to the global optimal solution according to the set temperature control factor to obtain an approximately globally optimal interpretive subgraph.

[0012] To better implement the present invention, further, the step S41 specifically includes the following steps: Step S411: Select basic nodes as candidate nodes according to a set ratio; Step S412: Call the approximate annealing function to anneal the elements in the candidate nodes to generate new candidate nodes.

[0013] To better implement the present invention, further, the step S5 specifically includes the following steps: Step S51: Call the reinforcement learning pruning algorithm, and predict the interpretability score of the candidate nodes according to the set training policy network to obtain a subgraph with high interpretability; Step S52: Model the pruning process as a Markov decision process, and select the optimal action according to the marginal gain to filter the nodes with interpretability scores lower than the set threshold.

[0014] Based on the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection, to better implement the present invention, further, a graph neural network interpretability system for defect detection is proposed, which is used to execute the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection; it includes a construction unit, an initialization unit, an annealing unit, a pruning unit, and a training unit; The construction unit is used to construct a semiconductor graph by taking the atoms of the industrial semiconductor data as nodes and the chemical bonds or adjacency relationships as edges according to the obtained industrial semiconductor data; and train a graph neural network according to the constructed semiconductor graph, embed features into different types of nodes, and construct a semiconductor graph neural network; The initialization unit is used to initialize the interpretability model parameters of the semiconductor graph neural network; The annealing unit is used to call the stochastic approximation annealing algorithm to generate candidate nodes and calculate the marginal gain to obtain an approximately globally optimal interpretive subgraph; The pruning unit is used to call the learning pruning algorithm to predict and filter the interpretability scores of the nodes; The training unit is used to iterate until the loss function converges to obtain a globally trained model.

[0015] Based on the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection, to better implement the present invention, further, an electronic device is proposed, including a memory and a processor; a computer program is stored on the memory; when the computer program is executed on the processor, the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection is implemented.

[0016] Based on the proposed graph neural network interpretability method for industrial semiconductor defect detection above, to better implement the present invention, further, a computer-readable storage medium is proposed, on which computer instructions are stored; when the computer instructions are executed on the above-mentioned electronic device, the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection is implemented.

[0017] The present invention has the following beneficial effects: (1) Based on the random approximation annealing and reinforcement learning pruning strategies, the present invention significantly improves the efficiency of the graph neural network GNN interpretability task and the approximation ability of the global optimal solution. In each iteration, by calculating the marginal gain probability, i.e., the selection probability, of each candidate element, a feasible solution is randomly generated; during the process of gradually decreasing the temperature, the Metropolis criterion is applied to achieve the transformation from extensive search to gradual convergence. At the same time, by training the policy network through the reinforcement learning pruning strategy, subgraphs with lower interpretability scores are filtered, so as to converge to an approximate global optimal solution in a shorter time.

[0018] (2) The present invention shows better effects than the prior art in both small-scale graph tasks such as molecular structure classification and large-scale graph tasks such as industrial semiconductor defect detection, significantly improving the accuracy and efficiency of the GNN interpretability task.

[0019] (3) In the experiment, the present invention shows significant advantages over the existing state-of-the-art baseline models in terms of interpretation quality and time efficiency, making it applicable to complex graph data analysis tasks and having good practicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flowchart of the interpretability method of the graph neural network for semiconductor defect analysis provided by the present invention.

[0021] Figure 2 It is a schematic diagram of the interpretability methods of different types of graph neural networks.

[0022] Figure 3 It is a schematic diagram of random approximation annealing, reinforcement learning pruning, and reward mechanism. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. It should be understood that the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments, and therefore should not be regarded as a limitation on the protection scope. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0024] In the description of the present invention, it should be noted that unless otherwise clearly specified and limited, the terms "set", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can also be directly connected, or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0025] Embodiment 1: This embodiment proposes an interpretability method for graph neural networks for defect detection, which specifically includes the following steps: Step S1: Construct a semiconductor graph according to the type of semiconductor data obtained and the information affecting the energy band structure.

[0026] The specific steps of step S1 are as follows: Step S11: Obtain industrial semiconductor data, and obtain the data type of the industrial semiconductor data and the information affecting the energy band structure; Step S12: Represent industrial semiconductor data of different data types as different types of nodes, and use the information affecting the energy band structure as edges to construct a semiconductor graph structure.

[0027] Step S2: Train a graph neural network according to the constructed semiconductor graph, embed features into different types of nodes, and construct a semiconductor graph neural network.

[0028] The specific steps of step S2 are as follows: Step S21: Call the graph neural network to embed the features of different types of nodes and capture the structural relationship between the nodes; Step S22: Call the cross-entropy loss function to calculate the classification error, and call the consistency regularization output to construct a semiconductor graph neural network.

[0029] Step S3: Initialize the interpretability model parameters of the semiconductor graph neural network.

[0030] The interpretability model parameters in step S3 include the parameters of stochastic approximation annealing and the parameters of reinforcement learning pruning; The parameters of stochastic approximation annealing include the initial temperature, the cooling factor, and the initial acceptance probability constant; The parameters of reinforcement learning pruning include the learning rate of the policy network, the size of the action space, and the initial policy network parameters.

[0031] Step S4: Call the stochastic approximation annealing algorithm to generate candidate nodes, and calculate the marginal gain to obtain an approximately globally optimal explanatory subgraph.

[0032] Step S4 specifically includes the following steps: Step S41: Call the approximate annealing function to generate candidate nodes; Step S41 specifically includes the following steps: Step S411: Select basic nodes as candidate nodes according to a set ratio; Step S412: Call the approximate annealing function to anneal the elements in the candidate nodes to generate new candidate nodes.

[0033] Step S42: Calculate the marginal gain of the candidate nodes according to the candidate nodes, the interpretability score of the current subgraph, and the interpretability score after adding the candidate nodes; Step S43: Calculate the selection probability of the candidate nodes according to the marginal gain and the number of candidate nodes; Step S44: Converge to the global optimal solution according to the set temperature control factor to obtain an approximately globally optimal interpretive subgraph.

[0034] Step S5: Call the learning pruning algorithm to predict and filter the interpretability scores of the nodes.

[0035] Step S5 specifically includes the following steps: Step S51: Call the reinforcement learning pruning algorithm and predict the interpretability scores of the candidate nodes according to the set training policy network to obtain a subgraph with high interpretability; Step S52: Model the pruning process as a Markov decision process and select the optimal action according to the marginal gain to filter the nodes with interpretability scores lower than the set threshold.

[0036] Step S6: Repeat Step S2 - Step S5 until the loss function converges to obtain a trained global model.

[0037] Working principle: In this embodiment, first, a semiconductor graph neural network is obtained by training a graph neural network based on a semiconductor graph; then, the random approximate annealing algorithm is used to control the selection probability of candidate nodes by gradually decreasing the temperature, gradually transitioning from a wide search in the high-temperature stage to convergence in the low-temperature stage to obtain an approximately globally optimal interpretive subgraph; the policy network is trained based on the reinforcement learning pruning strategy, and the interpretability scores of the nodes are predicted by the policy network and the nodes with scores lower than the threshold are pruned, so as to screen out subgraphs with higher interpretability in the solution space. This embodiment shows significant advantages in interpretability quality and time efficiency, significantly improving the accuracy and efficiency of the GNN interpretability task, making it better applicable to complex graph data analysis tasks, while ensuring the approximation of the global optimal solution and a higher convergence speed.

[0038] Embodiment 2: Based on the above Embodiment 1, as Figure 1 , Figure 2 , Figure 3 shown, a specific embodiment will be described in detail.

[0039] Step S1: Construct a semiconductor graph, taking atoms in industrial semiconductor data as nodes, chemical bonds or adjacency relationships as edges, and combining lattice parameters and material properties as features of nodes and edges; The specific operation of Step S1 is: Construct a semiconductor graph, select the main data types in the data, such as: material atoms, doping elements, defect types, crystal structure units, etc., and represent them as different node types. Then, based on information such as the bonding relationship between atoms, the correlation between defects and material properties, and the influence of doping elements on the energy band structure, construct the edge relationship between nodes, thereby forming a complete semiconductor graph structure.

[0040] Step S2: Train a graph neural network based on the semiconductor graph, use the graph neural network to embed the features of different types of nodes, capture the structural relationship between nodes for defect detection, and construct a semiconductor graph neural network; The specific operation of Step S2 is: Train to obtain a semiconductor graph neural network; after constructing the semiconductor graph, use the graph neural network GNN to embed the features of different types of nodes, capture the structural relationship between nodes, calculate the classification error using the cross-entropy loss function, and adopt consistency regularization to ensure the consistency of the output of the model on perturbed data. Update the model parameters through multiple rounds of iteration, so that the constructed semiconductor graph neural network can effectively capture and express the complex relationships in semiconductor data.

[0041] Step S3: Initialize the interpretability model parameters of the semiconductor graph neural network, where the interpretability model parameters include parameters of stochastic approximation annealing and parameters of reinforcement learning pruning.

[0042] Specifically, according to the semiconductor graph neural network obtained in Step S2, initialize the interpretability model parameters of the semiconductor graph neural network, including the parameters of stochastic approximation annealing, namely the initial temperature, cooling factor, and initial acceptance probability constant, and the parameters of reinforcement learning pruning, the learning rate of the policy network, the size of the action space, and the initial policy network parameters; In Step S3, the parameters of stochastic approximation annealing include the initial temperature, cooling factor, and initial acceptance probability constant, and the parameters of reinforcement learning pruning include the learning rate of the policy network, the size of the action space, and the initial policy network parameters.

[0043] Step S4: Using the stochastic approximation annealing algorithm, by gradually reducing the temperature, control the selection probability of candidate nodes, and gradually transition from the extensive search in the high-temperature stage to the convergence in the low-temperature stage to obtain an approximately globally optimal explanatory subgraph.

[0044] In the step S4, nodes with stronger interpretability are screened out based on the selection probability, and nodes are gradually selected according to the Metropolis criterion in the high-temperature stage.

[0045] In the step S4, basic nodes are selected proportionally to reach the number of candidate nodes required for the current optimization process; then, using the marginal gain of the candidate nodes as the input, perform annealing calculation on it through the temperature control function to obtain the selection probability of the new candidate nodes; the temperature control function is: where: T i is the temperature after the i-th iteration; T i-1 is the temperature after the (i - 1)-th iteration; α is the temperature control factor.

[0046] The calculation formula for the marginal gain of the candidate nodes is: where: r i,j is the marginal gain of the j-th candidate node; G i,j-1 represents the current subgraph; e j represents the candidate node; is the interpretability score of the current subgraph; is the new interpretability score after adding the node; The calculation formula for the selection probability of the candidate nodes is: where: B is the total number of candidate nodes.

[0047] Step S5: Train a policy network based on the reinforcement learning pruning strategy. During the reinforcement learning process, gradually optimize the parameters of the policy network to increase the expected value of the cumulative reward, using the marginal gain and interpretability score of the nodes as the reward signals; predict the interpretability score of the nodes through the policy network and perform pruning operations on the nodes with scores lower than the threshold, so as to screen out subgraphs with higher interpretability in the solution space.

[0048] In the step S5, model the pruning process as a Markov decision process, select the optimal action according to the marginal gain of each node, and gradually filter out the interpretability nodes with scores lower than the threshold.

[0049] Step S6: Repeat Step S2 - Step S5 until the loss function converges, and obtain the global model of the trained semiconductor graph neural network to process the semiconductor graph defect analysis task.

[0050] In this embodiment, the randomness of the search process is controlled by the annealing temperature, gradually transitioning from a wide search in the high-temperature stage to convergence in the low-temperature stage. This design reduces randomness by dynamically adjusting the temperature, enabling the model to comprehensively explore the possible solution space in the early stage and focusing more on the fine-tuning of the current best solution in the later stage, thus ensuring the approximation of the global optimal solution. In addition, the introduction of the reinforcement learning pruning strategy reduces the redundant nodes in the search space, further improving the overall convergence speed and computational efficiency.

[0051] Preferably, for the approximate annealing function, Random Approximation Function (RAF), the marginal gain of the candidate nodes is used as the input, and it is annealed through the temperature control function to obtain the new candidate node selection probability. The process of generating candidate nodes by the random approximation annealing function Random Approximation Function (RAF) is as follows: 1) Select basic nodes proportionally to reach the number of candidate nodes required for the current optimization process; 2) Perform annealing operations on each element in the candidate nodes through RAF to obtain the new candidate node selection probability. During the training process, the feature map that conforms to each input is calculated through the generated candidate nodes in the forward propagation, while in the backpropagation process, only the basic nodes existing in the initial graph neural network are updated, and the generated candidate nodes are not updated.

[0052] Preferably, the calculation method of the marginal gain is described by the random approximation annealing algorithm formula of this embodiment. In each iteration, the probability of selecting a node is defined based on its marginal gain, and the specific formula is as follows: where r i,j is the marginal gain of the j th candidate node. This marginal gain value is calculated from the difference between the interpretability score of the current subgraph (fidelity, a common scoring criterion in GNN interpretability) and the new interpretability score after adding this node. B is the total number of candidate nodes. Specifically, the marginal gain r i,j is calculated by the formula: where G i,j-1Represents the current sub - graph, e j Represents a candidate node. Through a probability selection mechanism based on marginal gain, in each iteration of this embodiment, nodes with higher improvement in model interpretability are effectively selected, further improving the interpretability quality.

[0053] Preferably, the reinforcement learning pruning strategy in this embodiment uses a neural network to predict the interpretability score of the current candidate node as the pruning basis. By using a pre - trained policy network to predict the importance score of nodes, nodes with scores lower than the threshold are pruned, thereby reducing the search space. Specifically, in each round of iteration, the policy network uses the marginal gain and interpretability score of the node as the reward signal to guide the model to only retain candidate nodes with potentially high interpretability, making the subsequent annealing process more efficient.

[0054] Preferably, the stochastic approximation annealing strategy of this embodiment controls the search by gradually reducing the temperature. In each iteration, the temperature T is calculated according to the formula: Where: T i is the temperature after the i - th iteration; T i-1 is the temperature after the i (i - 1)-th iteration; α is the temperature control factor, which is used to reduce the temperature after each iteration, so that the search gradually converges from extensive exploration to an approximate global optimal solution to prevent the exponential operation from causing the selection probability of some nodes to be zero. This process of gradually reducing the temperature allows the model to conduct a large - scale search in the high - temperature stage to avoid falling into local optima. In the low - temperature stage, the search gradually converges to achieve a more refined adjustment effect. This temperature control method effectively combines the exploration and convergence characteristics, thus ensuring the approximation of the global optimal solution.

[0055] Through the fusion of reinforcement learning and annealing strategy, this embodiment significantly improves the interpretability and efficiency in large - scale graph tasks. During the model construction process, in each round of iteration, the temperature and pruning threshold are adjusted according to the interpretability score, and through multiple iterations of optimization, the global model is finally obtained when the loss function converges.

[0056] This embodiment not only overcomes the convergence speed problem of traditional heuristic methods but also avoids the local - optimum dilemma of gradient - based machine learning methods, especially showing excellent computational efficiency and interpretability effects when dealing with large - scale graph data.

[0057] In each iteration of this embodiment, the marginal gain of candidate nodes is calculated by the stochastic approximation annealing algorithm, nodes are gradually selected according to the Metropolis criterion, and at the same time, a reinforcement learning pruning strategy is used to predict the interpretability score of nodes and filter out low-score nodes. This embodiment can effectively solve the local optimum problem existing in traditional GNN interpretability methods in large-scale graph tasks and improve the convergence speed. In the initial stage, this embodiment explores the global space at a high temperature, gradually reduces the temperature to achieve convergence, and combines reinforcement learning for effective pruning to reduce the search space, further improving the model calculation efficiency and the quality of interpretability results. In this way, this embodiment not only ensures the approximation of the global optimal solution but also significantly improves the interpretability and processing efficiency of the model in large-scale graph tasks.

[0058] In this embodiment, an open semiconductor dataset is selected, and experiments are conducted on two tasks: semiconductor self-defect detection and semiconductor foreign-impurity defect detection. The former defect types include: vacancies, self-interstitials, and antisites, and five different charge states are calculated for each defect. The latter includes 77 possible impurity atoms, occupying five different point defect positions, for a total of 12,474 possible defects. The selection of these datasets is aimed at evaluating the test performance of the method proposed in this embodiment in large-scale graph neural network interpretability tasks.

[0059] In the two tasks of semiconductor self-defect detection and semiconductor foreign-impurity defect detection, this embodiment combines the stochastic approximation annealing and the reinforcement learning pruning strategy, and selects the Crystal Graph Convolutional Neural Network (CGCNN) model to adapt to the characteristics of industrial semiconductor defect detection tasks.

[0060] To verify the effectiveness of this embodiment in semiconductor graph neural networks, as shown in Table 1, compared with six existing benchmark GNN interpretability methods, this embodiment shows significant advantages in forward fidelity and stability, and at the same time has a short convergence time. As shown in Table 2, in the semiconductor foreign-impurity defect task, this embodiment is superior to other methods in terms of forward fidelity, negative fidelity, and stability, and maintains high efficiency in terms of time. Generally speaking, this embodiment shows higher forward fidelity and stability under different tasks, while significantly reducing the calculation time, highlighting its superiority in industrial semiconductor-related graph neural network tasks.

[0061] Table 1 Experimental results of GNN interpretability in the semiconductor self-defect detection task Table 2 Experimental results of GNN interpretability in the semiconductor foreign-impurity defect task Working principle: The random approximation annealing algorithm of this embodiment selects candidate elements based on the marginal gain probability, i.e., the selection probability, and gradually reduces the temperature to achieve the convergence process from extensive exploration to approximate global optimum. Compared with the existing method that only relies on machine learning to transform discrete problems into continuous optimization problems, this embodiment can better avoid falling into local optimal solutions and provide an explanation closer to the global optimum. The reinforcement learning pruning strategy of this embodiment uses a neural network model to predict the interpretability score, i.e., the fidelity, of each candidate subgraph, a common scoring criterion in GNN interpretability, and filters out the explanations with lower scores based on the expected reward. During the search process, pruning is achieved through reinforcement learning, reducing the solution space of the combinatorial optimization problem and accelerating the convergence process, thereby improving the efficiency of the explanation task. Different from the traditional greedy pruning method, this embodiment ensures the quality and accuracy of the explanation while improving the time efficiency. As Figure 2 shown, Figure 2 in (a), (b), and (c) of [Figure] respectively represent the schematic diagrams of the GNN interpretability method based on the machine learning method, the GNN interpretability method based on the heuristic search method, and the interpretability method of this embodiment.

[0062] As Figure 2 shown in (a) of [Figure], the GNN interpretability method based on machine learning solves the problem by transforming the discrete combinatorial optimization problem into a continuous differentiable problem. However, since such problems are usually non-convex, the commonly used gradient descent algorithm is prone to falling into local optima. The part circled by the dashed line in the figure indicates that these methods can only approximate the optimal solution within a certain range and cannot guarantee to be close to the global optimum.

[0063] As Figure 2 shown in (b) of [Figure], the GNN interpretability method based on heuristic search, such as Monte Carlo tree search (MCTS), etc., can find the optimal subgraph by repeatedly exploring local random paths, thus effectively avoiding falling into local optima. However, as the graph scale increases, the number of branches of the search tree grows exponentially, making these methods show low convergence efficiency in large-scale graph tasks. The search process of the heuristic search method is represented by the extended tree structure in the figure. Although it can approach the global optimum, its efficiency is affected.

[0064] As Figure 2As shown in (c), this embodiment combines the advantages of machine learning and heuristic search based on the random annealing method, and has significant advantages in global optimal approximation and convergence efficiency, improving the efficiency and global optimal approximation ability of the graph neural network interpretability task. First, the Random Approximation Annealing algorithm is used to select candidate elements, and the selection probability is calculated based on the marginal gain in each iteration. By gradually reducing the temperature and introducing the Metropolis criterion, random annealing realizes the process from extensive exploration to gradual convergence, and finally approaches the global optimum. In addition, random annealing also introduces a reinforcement learning pruning strategy, predicting the interpretability score of the subgraph as a reward, and reducing the solution space of the combinatorial optimization problem through pruning to accelerate convergence. The policy network of random annealing proposed in this embodiment shows effective pruning of the search space through reinforcement learning, thus improving the efficiency of the algorithm.

[0065] As Figure 3 shown, Figure 3 (a) in shows the process of random approximation annealing, demonstrating the selection process of candidate nodes at different temperatures. As the temperature gradually decreases, the system gradually transitions from random exploration to the convergence stage, and finally approaches the global optimal solution. Figure 3 (b) in shows the process of predicting the importance of nodes through the policy network, using reinforcement learning for pruning to screen out the subgraph nodes most likely to approach the global optimum, thereby effectively reducing the search space. Figure 3 (c) in summarizes the reward mechanism of the entire process of reinforcement learning pruning. The reinforcement learning policy network will evaluate and predict which nodes contribute the most to interpretability based on the reward scores of the nodes. By combining the annealing strategy and reinforcement learning, this embodiment shows higher efficiency and accuracy in seeking the optimal solution of graph neural network (GNN) interpretability.

[0066] Specifically, in the semiconductor self-defect detection task, this embodiment can be used to explain which material compositions and microstructural features have important impacts on semiconductor performance. In this task, this embodiment first uses the random approximation annealing algorithm to randomly select material features with higher marginal gain in each iteration. For example, specific doping elements or point defect types such as vacancies, interstitial atoms, and anti-site defects. Subsequently, through the pruning process, the material features that contribute the most to the electrical or optical performance of the semiconductor are further screened out. For example, the critical doping concentration, energy level distribution, or defect state density. This embodiment can effectively reduce the search space and only retain the key features crucial for semiconductor performance, thus providing a clear explanation and pointing out the core role of these features in semiconductor device performance optimization and reliability analysis.

[0067] Other parts of this embodiment are the same as those of the above Embodiment 1, so they will not be elaborated here.

[0068] Example 3: Based on any one of the above-mentioned Embodiment 1 - Embodiment 2, this embodiment proposes a graph neural network interpretability system for defect detection, which is used to execute the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection; it includes a construction unit, an initialization unit, an annealing unit, a pruning unit, and a training unit; The construction unit is used to construct a semiconductor graph by taking the atoms of the industrial semiconductor data as nodes and the chemical bonds or adjacency relationships as edges according to the obtained industrial semiconductor data; and train a graph neural network based on the constructed semiconductor graph, embed features into different types of nodes, and construct a semiconductor graph neural network; The initialization unit is used to initialize the interpretability model parameters of the semiconductor graph neural network; The annealing unit is used to call the stochastic approximation annealing algorithm to generate candidate nodes and calculate the marginal gain to obtain an approximately globally optimal explanatory subgraph; The pruning unit is used to call the learning pruning algorithm to predict and filter the interpretability scores of the nodes; The training unit is used to iterate until the loss function converges to obtain a globally trained model.

[0069] This embodiment also proposes an electronic device, including a memory and a processor; a computer program is stored on the memory; when the computer program is executed on the processor, the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection is implemented.

[0070] This embodiment also proposes a computer-readable storage medium, on which a computer instruction is stored; when the computer instruction is executed on the above-mentioned electronic device, the above-mentioned graph neural network interpretability method for industrial semiconductor defect detection is implemented.

[0071] Other parts of this embodiment are the same as any one of the above-mentioned Embodiment 1 - Embodiment 2, so they will not be elaborated here.

[0072] The above is only a preferred embodiment of the present invention, and does not impose any form of limitation on the present invention. Any simple modification or equivalent change made to the above embodiments based on the technical essence of the present invention falls within the protection scope of the present invention.

Claims

1. A graph neural network interpretability method for defect detection, characterized in that: The specific steps include: Step S1: constructing a semiconductor map according to the type of semiconductor data obtained and information affecting the energy band structure; Step S2: training a graph neural network based on the constructed semiconductor graph, embedding features into different types of nodes, and constructing a semiconductor graph neural network; Step S3: Initialize the interpretable model parameters of the semiconductor graph neural network; Step S4: calling the random approximate annealing algorithm to generate candidate nodes and calculate the marginal gain to obtain an explanatory subgraph that is approximately globally optimal; Step S5: Call the learning pruning algorithm to predict the explanatory score of the node and filter it; Step S6: Repeat steps S2 to S5 until the loss function converges to obtain a trained global model.

2. A graph neural network interpretability method for defect detection according to claim 1, characterized in that: The step S1 specifically includes the following steps: Step S11: Acquire industrial semiconductor data, and acquire the data type of the industrial semiconductor data and information affecting the energy band structure; Step S12: Represent industrial semiconductor data of different data types as different types of nodes, and use information that affects the energy band structure as edges to construct a semiconductor graph structure.

3. A graph neural network interpretability method for defect detection according to claim 1, characterized in that: The step S2 specifically includes the following steps: Step S21: calling the graph neural network to embed features of different types of nodes and capture the structural relationship between nodes; Step S22: Call the cross entropy loss function to calculate the classification error, and call the consistency regularization output to build a semiconductor graph neural network.

4. The graph neural network interpretability method for defect detection according to claim 1, characterized in that: The interpretability model parameters in step S3 include parameters of random approximate annealing and parameters of reinforcement learning pruning; The parameters of the random approximate annealing include initial temperature, cooling factor and initial acceptance probability constant; The parameters of the reinforcement learning pruning include the learning rate of the policy network, the size of the action space and the initial policy network parameters.

5. The graph neural network interpretability method for defect detection according to claim 1, characterized in that: The step S4 specifically comprises the following steps: Step S41: calling the approximate annealing function to generate candidate nodes; Step S42: Calculate the marginal gain of the candidate node according to the candidate node, the explanatory score of the current subgraph, and the explanatory score after adding the candidate node; Step S43: Calculate the selection probability of the candidate node according to the marginal gain and the number of candidate nodes; Step S44: Converging to the global optimal solution according to the set temperature control factor, and obtaining an explanatory subgraph that is approximately global optimal.

6. A graph neural network interpretability method for defect detection according to claim 5, characterized in that: The step S41 specifically includes the following steps: Step S411: selecting basic nodes as candidate nodes according to a set ratio; Step S412: calling an approximate annealing function to anneal the elements in the candidate node to generate a new candidate node.

7. The graph neural network interpretability method for defect detection according to claim 1, characterized in that: The step S5 specifically comprises the following steps: Step S51: calling the reinforcement learning pruning algorithm, and predicting the explanatory scores of the candidate nodes according to the set training strategy network, to obtain a subgraph with high explanatory power; Step S52: Model the pruning process as a Markov decision process, select the optimal action based on the marginal gain, and filter out nodes whose explanatory scores are lower than the set threshold.

8. A graph neural network interpretability system for defect detection, used to execute the graph neural network interpretability method for industrial semiconductor defect detection as claimed in claim 1; characterized in that, Including construction unit, initialization unit, annealing unit, pruning unit, and training unit; The construction unit is used to construct a semiconductor graph based on the acquired industrial semiconductor data, using atoms of the industrial semiconductor data as nodes and chemical bonds or adjacency relationships as edges; and to train a graph neural network based on the constructed semiconductor graph, embed features into different types of nodes, and construct a semiconductor graph neural network; The initialization unit is used to initialize the interpretable model parameters of the semiconductor graph neural network; The annealing unit is used to call a random approximate annealing algorithm to generate candidate nodes and calculate marginal gains to obtain an explanatory subgraph that is approximately globally optimal; The pruning unit is used to call the learning pruning algorithm to predict the explanatory score of the node and filter it; The training unit is used to iterate until the loss function converges to obtain a trained global model.

9. An electronic device, characterized in that: It comprises a memory and a processor; a computer program is stored on the memory; when the computer program is executed on the processor, the graph neural network interpretability method for industrial semiconductor defect detection as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 9, the graph neural network interpretability method for industrial semiconductor defect detection as described in any one of claims 1-7 is implemented.

Images