Cooperative verification method for EDA software and heterogeneous process design rules

Abstract

The invention provides an EDA software and heterogeneous process design rule collaborative verification method, which relates to the technical field of EDA software, and comprises the following steps of: extracting time sequence characteristics in a dynamic association graph through a bidirectional long-short-term memory network, analyzing node coupling strength by utilizing a multi-head graph attention network to generate a hierarchical importance scoring matrix; and outputting a conflict risk tensor and a propagation path prediction map based on a depth rule conflict prediction model, and generating a correction strategy in combination with a variational auto-encoder and a reinforcement learning network. According to the method, design rule conflicts can be accurately predicted and positioned, an effective correction scheme is provided, and the accuracy and efficiency of heterogeneous process design rule verification are improved.

Description

Technical Field

[0001] The present invention relates to EDA technology, and particularly to a collaborative verification method for EDA software and heterogeneous process design rules. Background Art

[0002] With the continuous development of integrated circuit manufacturing processes, heterogeneous process design rules have become increasingly complex. EDA software needs to verify these design rules to ensure that the design scheme meets the process requirements. Currently, EDA software mainly uses static rule checking and simulation analysis methods for verification. Static rule checking examines the design through a preset rule library, while simulation analysis evaluates the design performance by building models. These methods play an important role in practical applications, but with the continuous shrinking of process nodes and the increasing design complexity, traditional verification methods are facing new challenges.

[0003] Firstly, traditional verification methods are difficult to effectively identify and predict dynamic conflicts between design rules. Due to the complex correlation relationships between rules at each level in heterogeneous processes, a change in one rule may cause a chain reaction, leading to conflicts in other rules. Existing methods mainly focus on static rule checking and lack in-depth analysis of the dynamic correlation between rules, making it impossible to detect potential conflict risks in a timely manner.

[0004] Secondly, existing verification methods lack intelligent conflict location and propagation analysis capabilities. When a design rule conflict is detected, it is difficult to quickly and accurately locate the conflict source and analyze its influence scope. The verification process often requires designers to manually analyze and judge, which is not only time-consuming and laborious but also prone to missing some hidden conflict points.

[0005] Finally, existing technologies lack an adaptive correction mechanism. After detecting a design rule conflict, it usually requires designers to make manual adjustments based on experience, lacking systematic correction strategy guidance. This approach is inefficient, difficult to ensure the optimality of the correction effect, and unable to effectively accumulate and utilize historical experience to optimize the verification process. Summary of the Invention

[0006] Embodiments of the present invention provide a collaborative verification method and system for EDA software and heterogeneous process design rules, which can solve the problems in the prior art.

[0007] In the first aspect of the embodiments of the present invention, A collaborative verification method for EDA software and heterogeneous process design rules is provided, including: The temporal features of the process node layer, design rule layer, and constraint parameter layer in the dynamic association graph are extracted using a bidirectional long short-term memory network. The dynamic coupling strength between nodes is analyzed through a multi-head graph attention network to generate a hierarchical importance scoring matrix. Based on the hierarchical importance scoring matrix, a deep rule conflict prediction model is trained. The deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map. The conflict risk tensor quantifies the conflict probability of each node, and the propagation path prediction map represents the diffusion path of conflicts between different levels. When the value of the conflict risk tensor exceeds the preset risk threshold, the verification priority is determined according to the propagation path prediction map and the hierarchical importance scoring matrix. Hierarchical recursive verification is performed along the priority path higher than the preset path threshold. The actual conflict nodes are located through a variational autoencoder, and the feature vectors of the actual conflict nodes are extracted. The feature vectors contain the status information of process parameters, design rules, and constraints. The feature vectors are input into a reinforcement learning network to generate a correction policy tree. The correction operations are performed in the priority order of the correction policy tree. The abnormal parameters in the feature vectors are adjusted, and the change of the conflict risk tensor is monitored in real time. When the conflict risk tensor is lower than the preset risk threshold, the node status of the dynamic association graph is updated. The feature vectors and the corresponding correction policies are stored in the experience pool for optimizing the deep rule conflict prediction model.

[0008] The dynamic coupling strength between nodes is analyzed through a multi-head graph attention network to generate a hierarchical importance scoring matrix. Based on the hierarchical importance scoring matrix, a deep rule conflict prediction model is trained. The deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map. The conflict risk tensor quantifies the conflict probability of each node, and the propagation path prediction map represents the diffusion path of conflicts between different levels, including: Apply a multi-head graph attention network to analyze the dynamic coupling strength between nodes. Each attention head independently calculates the attention coefficient between node pairs. The attention coefficient is obtained through the non-linear transformation and similarity calculation of node feature vectors. The outputs of multiple attention heads are normalized and fused to obtain the final coupling strength between nodes. Generate a hierarchical importance scoring matrix based on the final coupling strength. In the generation process of the hierarchical importance scoring matrix, both the direct coupling strength between nodes and the indirect coupling strength accumulated through multi-hop paths are considered. An adaptive weight factor is used to balance the influence of the direct coupling strength and the indirect coupling strength. Input the hierarchical importance scoring matrix into the deep rule conflict prediction model for training. The deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map. The conflict risk tensor quantifies the conflict probability of each node under different time windows and different risk types. The propagation path prediction map is constructed based on the node risk degree and the hierarchical importance scoring matrix, and is used to characterize the diffusion path of conflicts among process nodes, rule nodes, and constraint nodes.

[0009] Apply a multi-head graph attention network to analyze the dynamic coupling strength between nodes. Each attention head independently calculates the attention coefficient between node pairs. The attention coefficient is obtained through the non-linear transformation and similarity calculation of node feature vectors. The outputs of multiple attention heads are normalized and fused to obtain the final coupling strength between nodes, including: Obtain the node feature matrix, which contains the feature information of multiple nodes to be analyzed. The node feature matrix is input into the multi-head graph attention network. The multi-head graph attention network performs feature transformation on the node feature matrix, and maps the node feature matrix to a high-dimensional feature space through multiple attention calculation units to obtain the transformed node feature set; Calculate the attention coefficient between node pairs for the transformed node feature set. Each attention calculation unit processes the node features using a non-linear transformation function, calculates the similarity between node pairs based on the processed node features to obtain the original attention value, and normalizes the original attention value to obtain the final attention coefficient; Perform feature aggregation based on the final attention coefficient. Each attention calculation unit uses its corresponding attention coefficient to perform weighted combination on the node features to obtain the node output representation corresponding to the attention calculation unit. The node output representations of each attention calculation unit are weighted and fused, and the outputs of different attention calculation units are combined through learnable weight coefficients to obtain the final coupling strength between nodes.

[0010] When the value of the conflict risk tensor exceeds the preset risk threshold, determine the verification priority according to the propagation path prediction map and the hierarchical importance scoring matrix, and perform hierarchical recursive verification along the priority path higher than the preset path threshold. Locate the actual conflict nodes through a variational autoencoder, including: When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, calculate the verification priority based on the propagation path prediction map and the hierarchical importance scoring matrix, extract all the propagation paths of the risk-exceeding nodes, calculate the cumulative propagation probability of each propagation path, and combine the importance scores of each node on the path to obtain the verification priority of the path by using a weighted combination method; Select the priority path with a path priority higher than the preset path threshold, and perform hierarchical recursive verification along the path priority. The hierarchical recursive verification is performed sequentially from top to bottom according to the functional levels of the nodes, and the verification result of each layer is used as the input condition for the verification of the next layer; Use a variational autoencoder to encode and reconstruct the nodes on the verification path. The variational autoencoder includes an encoding network and a decoding network. The encoding network maps the multi-dimensional state data of the nodes into the mean vector and variance vector of the latent variable distribution, samples the latent variable based on the mean vector and variance vector, and the decoding network reconstructs the latent variable into the node state; Calculate the error between the node state and the original state. The error includes the mean square error and the distribution divergence. Determine the abnormal degree of the node based on the size of the error, and identify the nodes with an abnormal degree exceeding the dynamic threshold as the actual conflict nodes.

[0011] When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, calculate the verification priority based on the propagation path prediction map and the hierarchical importance scoring matrix. Extract all the propagation paths of the risk-exceeding nodes, calculate the cumulative propagation probability of each propagation path, and combine the importance scores of the nodes on the path. The verification priority of the path is obtained by using the weighted combination method, including: When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, starting from the risk-exceeding node in the propagation path prediction map, perform a depth-first search algorithm to prune the starting points with a probability lower than the preset probability threshold, and generate candidate propagation paths; Extract the direct propagation probability between adjacent node pairs on the candidate propagation path, and multiply the direct propagation probabilities continuously to obtain the first-order propagation probability; calculate the second-order propagation probability of the shortest path between nodes based on the first-order propagation probability; combine the first-order propagation probability and the second-order propagation probability by weighting to obtain the comprehensive propagation probability; multiply the comprehensive propagation probability by the attenuation compensation coefficient of the path length to obtain the final cumulative propagation probability; Extract the static importance scores of the nodes on the path from the hierarchical importance scoring matrix; calculate the dynamic importance scores by combining the current load status and resource occupancy rate of the nodes; perform adaptive weighted fusion on the static importance scores and the dynamic importance scores to generate the comprehensive importance scores of the nodes; Calculate the position weighting coefficient based on the distance from the node to the risk-exceeding node. The farther the node is, the smaller the position weighting coefficient; multiply the comprehensive importance score of the node by the corresponding position weighting coefficient and sum them to obtain the importance combination value of the path nodes; Construct an adaptive weight vector, where the components of the adaptive weight vector correspond to the weights of the combined values of the final cumulative propagation probability and the path node importance; train a weight adjustment function based on historical verification data, and adjust the adaptive weight vector according to the system load and the verification resource status; after normalizing the combined value of the final cumulative propagation probability and the path node importance, perform a non-linear combination operation with the adaptive weight vector to obtain the verification priority of the candidate propagation path.

[0012] Perform correction operations in the priority order of the correction policy tree, adjust the abnormal parameters in the feature vector, and monitor the change of the conflict risk tensor in real time. When the conflict risk tensor is lower than the preset risk threshold, update the node state of the dynamic association graph, and store the feature vector and the corresponding correction policy in the experience pool for optimizing the deep rule conflict prediction model, including: According to the execution priority order of the correction policy tree, perform correction operations on the abnormal parameters; the adjustment step size of the correction operation is calculated by the adaptive learning rate and the parameter deviation, and the adaptive learning rate decays exponentially with time; apply the adjustment step size to the abnormal parameters to obtain the corrected feature vector; Calculate the conflict risk tensor based on the corrected feature vector, where the conflict risk tensor characterizes the risk state of the system in multiple dimensions; calculate the Frobenius norm difference of the conflict risk tensors at adjacent times to obtain the risk change amount; determine whether the risk change amount is less than the preset change threshold and whether the maximum component of the conflict risk tensor is lower than the preset risk threshold; When the risk change amount is less than the preset change threshold and the maximum component of the conflict risk tensor is lower than the preset risk threshold, update the state of the nodes in the dynamic association graph; the state of the nodes is updated through a state transition function, and the state transition function takes the original node state, the risk change amount, and the corrected feature vector as inputs; the parameters of the state transition function are trained by historical state transition data; Construct an experience sample from the feature vector, the correction policy, and the risk reduction amount; calculate the importance score of the experience sample, where the importance score is obtained by the product of the absolute value of the risk reduction amount and the time decay function; store the experience sample with the importance score in the experience pool; train the deep rule conflict prediction model based on the samples in the experience pool; The training objective of the deep rule conflict prediction model is to minimize the weighted error between the predicted risk and the actual risk; the weighted error is obtained by weighting the prediction error with the sample importance score; update the deep rule conflict prediction model through iterative optimization.

[0013] In the second aspect of the embodiments of the present invention, Provide an electronic device, including: A processor; A memory for storing instructions executable by the processor; Among them, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0014] In the third aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0015] The beneficial effects of this application are as follows: By analyzing the temporal characteristics and coupling relationships of the nodes in each layer of the dynamic association graph through a bidirectional long short-term memory network and a multi-head graph attention network, a hierarchical importance scoring matrix is generated, and a deep rule conflict prediction model is trained to output a conflict risk tensor and a propagation path prediction map, realizing the accurate prediction of design rule conflicts and the visual representation of the propagation path.

[0016] Based on the propagation path prediction map and the hierarchical importance scoring matrix, the verification priority is determined, a variational autoencoder is used to accurately locate the actual conflict nodes and extract feature vectors, and a correction policy tree is generated through a reinforcement learning network, providing an efficient hierarchical recursive verification and conflict location method.

[0017] By executing the correction operations in the correction policy tree according to the priority, monitoring the changes of the conflict risk tensor in real time and updating the node status of the dynamic association graph in a timely manner, and at the same time storing the feature vectors and correction policies in the experience pool for model optimization, a complete conflict correction and experience feedback mechanism is established, effectively improving the collaborative verification efficiency of heterogeneous process design rules. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flowchart of the method for collaborative verification of EDA software and heterogeneous process design rules in the embodiments of the present invention; Figure 2 It is a schematic diagram of the comparison of the time cost for node feature extraction in the embodiments of the present invention; Figure 3 It is a schematic diagram of the comparative analysis of experimental data in the embodiments of the present invention; Figure 4 It is a schematic diagram of the verification path priority map in the embodiments of the present invention; Figure 5 It is a schematic diagram of the trend of verification efficiency changing with the node scale in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0021] Figure 1 It is a schematic flowchart of the method for collaborative verification of EDA software and heterogeneous process design rules in the embodiments of the present invention. As Figure 1 shown, the method includes: Using a bidirectional long short-term memory network to extract the temporal features of the process node layer, design rule layer, and constraint parameter layer in the dynamic association graph, analyzing the dynamic coupling strength between nodes through a multi-head graph attention network to generate a hierarchical importance scoring matrix, training a deep rule conflict prediction model based on the hierarchical importance scoring matrix, the deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map, the conflict risk tensor quantifies the conflict probability of each node, and the propagation path prediction map represents the diffusion path of the conflict between different levels; When the value of the conflict risk tensor exceeds the preset risk threshold, determine the verification priority according to the propagation path prediction map and the hierarchical importance scoring matrix, perform hierarchical recursive verification along the priority path higher than the preset path threshold, locate the actual conflict nodes through a variational autoencoder, extract the feature vectors of the actual conflict nodes, the feature vectors contain the status information of process parameters, design rules, and constraints, and input the feature vectors into a reinforcement learning network to generate a correction strategy tree; Perform correction operations in the priority order of the correction strategy tree, adjust the abnormal parameters in the feature vectors, and monitor the change of the conflict risk tensor in real time. When the conflict risk tensor is lower than the preset risk threshold, update the node status of the dynamic association graph, and store the feature vectors and the corresponding correction strategies in the experience pool for optimizing the deep rule conflict prediction model.

[0022] In an optional implementation manner, analyzing the dynamic coupling strength between nodes through a multi-head graph attention network to generate a hierarchical importance scoring matrix, training a deep rule conflict prediction model based on the hierarchical importance scoring matrix, the deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map, the conflict risk tensor quantifies the conflict probability of each node, and the propagation path prediction map represents the diffusion path of the conflict between different levels includes: The multi-head graph attention network is applied to analyze the dynamic coupling strength between nodes. Each attention head independently calculates the attention coefficients between node pairs, and the attention coefficients are obtained through the non-linear transformation and similarity calculation of node feature vectors. The outputs of multiple attention heads are normalized and fused to obtain the final coupling strength between nodes; Based on the final coupling strength, a hierarchical importance scoring matrix is generated. In the process of generating the hierarchical importance scoring matrix, both the direct coupling strength between nodes and the indirect coupling strength accumulated through multi-hop paths are considered, and an adaptive weight factor is used to balance the influence of the direct coupling strength and the indirect coupling strength; The hierarchical importance scoring matrix is input into the deep rule conflict prediction model for training. The deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction graph. Among them, the conflict risk tensor quantifies the conflict probability of each node under different time windows and different risk types, and the propagation path prediction graph is constructed based on the node risk degree and the hierarchical importance scoring matrix, and is used to characterize the diffusion path of conflicts between process nodes, rule nodes and constraint nodes.

[0023] When analyzing the dynamic coupling strength between nodes through the multi-head graph attention network, 12 parallel attention calculation units are set. Taking a validation scenario with 100 nodes as an example, each node contains a 64-dimensional feature vector, and these feature vectors contain information about the process parameters, design rules and constraint conditions of the nodes. Each attention calculation unit first performs two non-linear transformations on the node features. The first time maps the 64-dimensional features to a 96-dimensional intermediate representation, and the second time further maps the 96-dimensional features to a 128-dimensional high-dimensional feature space. An activation function is introduced between the two transformations to enhance the non-linear expression ability. For any two nodes, the similarity in the 128-dimensional feature space is calculated to obtain the attention coefficient. The 12 attention calculation units respectively obtain 100×100 attention matrices, and these matrices are fused into the final coupling strength matrix by weighted average.

[0024] When generating the hierarchical importance scoring matrix based on the final coupling strength, both direct coupling and indirect coupling relationships are comprehensively considered. The direct coupling strength is directly obtained from the coupling strength matrix, representing the direct influence relationship between nodes. The indirect coupling strength is calculated by analyzing multi-hop paths. The specific process is as follows: for each path between two nodes, the coupling strengths between adjacent nodes on the path are continuously multiplied to obtain the path strength, and then the weighted sum of all path strengths is obtained as the indirect coupling strength. Taking nodes A and B as an example, there are three different connection paths: the strength of the two-hop path A→C→B is 0.8×0.7 = 0.56, the strength of the three-hop path A→D→E→B is 0.9×0.8×0.6 = 0.432, and the strength of the four-hop path A→F→G→H→B is 0.85×0.75×0.8×0.7 = 0.357. An adaptive weight factor is set to balance the influence of direct coupling and indirect coupling, and the weight factor is dynamically adjusted according to the path length, with shorter paths assigned larger weights.

[0025] The generated hierarchical importance scoring matrix is input into the deep rule conflict prediction model for training. This model adopts a multi-layer neural network structure, including a feature extraction layer, a temporal modeling layer, and a prediction output layer. The feature extraction layer uses three convolutional layers to extract spatial features, and the temporal modeling layer uses a bidirectional recurrent network to process temporal dependencies. The model outputs two key pieces of information: a conflict risk tensor and a propagation path prediction map. The dimension of the conflict risk tensor is 100×6×4, representing the conflict probabilities of 100 nodes under 4 different risk types in 6 time windows. Taking a certain node as an example, the four risk probabilities in the first time window are 0.15, 0.08, 0.23, and 0.12 respectively, and the risk values show an upward or downward trend over time. The propagation path prediction map is constructed based on the risk degree of nodes and the hierarchical importance scoring, recording the propagation paths of risks between different functional levels. For example, when the risk value of process node A reaches 0.8, through the prediction map, it can be traced that this risk will be propagated to rule node C through the path A→B→C, with an influence strength of 0.65.

[0026] The solution of this application can: By using the multi-head attention mechanism to respectively focus on different aspects of node features, the accuracy and robustness of coupling strength analysis are improved, and complex interaction relationships between nodes can be better captured. Considering both direct coupling and indirect coupling and introducing an adaptive weight makes the hierarchical importance scoring more comprehensively reflect the influence of nodes, effectively balancing local and global information. The output conflict risk tensor and propagation path prediction map provide fine-grained risk quantification and interpretable propagation path analysis, which helps to timely discover and prevent rule conflicts.

[0027] In an alternative embodiment, a multi-head graph attention network is applied to analyze the dynamic coupling strength between nodes. Each attention head independently calculates the attention coefficient between node pairs. The attention coefficient is obtained through the non-linear transformation and similarity calculation of node feature vectors. Normalizing and fusing the outputs of multiple attention heads to obtain the final coupling strength between nodes includes: Obtain a node feature matrix, which contains the feature information of multiple nodes to be analyzed. The node feature matrix is input into the multi-head graph attention network. The multi-head graph attention network performs feature transformation on the node feature matrix, and maps the node feature matrix to a high-dimensional feature space through multiple attention calculation units to obtain the transformed node feature set. Calculate the attention coefficient between node pairs for the transformed node feature set. Each attention calculation unit processes the node features using a non-linear transformation function, calculates the similarity between node pairs based on the processed node features to obtain the original attention value, and normalizes the original attention value to obtain the final attention coefficient. Perform feature aggregation based on the final attention coefficient. Each attention calculation unit uses its corresponding attention coefficient to perform weighted combination on the node features to obtain the node output representation corresponding to the attention calculation unit. Weightedly fuse the node output representations of each attention calculation unit, and combine the outputs of different attention calculation units through learnable weight coefficients to obtain the final coupling strength between nodes.

[0028] When applying the multi-head graph attention network to analyze the dynamic coupling strength between nodes, first obtain a node feature matrix, which contains the multi-dimensional feature information of the nodes to be analyzed. The feature information mainly comes from data such as the process parameters, design rules, and constraint conditions of the nodes. Taking an actual verification scenario as an example, when analyzing a design rule network containing 50 nodes, the feature dimension of each node is 32, forming a 50×32 feature matrix. The values in the feature matrix are normalized to map all feature values to the range of 0 to 1.

[0029] Next, input the node feature matrix into the multi-head graph attention network for feature transformation. Set 8 parallel attention calculation units, each unit containing a two-layer neural network structure. The first layer of the network maps the 32-dimensional node features to a 64-dimensional hidden layer space, and the second layer of the network further maps the 64-dimensional features to a 96-dimensional high-dimensional feature space. During the feature mapping process, a combination of linear transformation and non-linear activation function is adopted to ensure that the network has sufficient feature expression ability. After transformation, 8 sets of 96-dimensional node feature sets are obtained.

[0030] Calculate the attention coefficients between node pairs based on the transformed feature set. Within each attention calculation unit, first perform a non-linear transformation on the node features to compress the 96-dimensional features into a 48-dimensional intermediate representation. Then calculate the similarity between node pairs. Specifically, the dot product operation is performed between the intermediate representation of one node and the intermediate representations of all other nodes to obtain the original attention values. Taking node A and node B as an example, the dot product operation of the 48-dimensional intermediate representations of the two nodes results in a scalar value, representing the original attention intensity between them. Repeat this process for all node pairs to form a 50×50 original attention matrix. Finally, normalize the original attention matrix to ensure reasonable attention allocation for each node.

[0031] After obtaining the final attention coefficients, perform the feature aggregation operation. Each attention calculation unit uses its corresponding attention coefficients to perform weighted combination on the 96-dimensional node features. Specifically, for each target node, the features of all other nodes are weighted and summed according to the attention coefficients to generate the output representation of the node under the current attention head. In this way, 8 sets of node representations are obtained in 8 attention calculation units respectively. Finally, introduce a set of learnable weight coefficients to perform weighted fusion on the 8 sets of node representations. The weight coefficients are initialized with a uniform distribution and are continuously optimized and adjusted during the training process. After fusion, the final coupling strength between each pair of nodes is obtained, forming a 50×50 coupling strength matrix for subsequent verification priority determination.

[0032] Figure 2 Schematic diagram of the time overhead comparison for node feature extraction in the embodiments of the present invention: The "proposed solution" marked with a circle represents a multi-head attention feature extractor using parallel computing, which consists of 8 parallel attention calculation units, each independently processing a feature subspace; the "traditional graph attention method" marked with a square represents a single-head attention calculation model with serial processing, which processes all feature dimensions in a sequential calculation manner. This figure compares the computational time performance under different feature dimensions. Experimental data shows that when the feature dimension is 32, the proposed solution only needs 15 ms to complete the calculation, while the traditional method requires 25 ms. As the dimension increases to 64, 128, 256, and up to 512, the computational times of the proposed solution are 25 ms, 45 ms, 85 ms, and 150 ms respectively, showing a relatively gentle growth trend; while the computational times of the traditional method increase rapidly to 45 ms, 85 ms, 165 ms, and 300 ms. The difference in time overhead gradually expands as the dimension increases. When the dimension is 512, the proposed solution saves 50% of the computational time compared to the traditional method. This significant efficiency improvement stems from the design of the multi-head parallel architecture, where each attention head is responsible for processing a part of the feature dimensions, effectively reducing the computational complexity. At the same time, through optimized feature mapping and similarity calculation strategies, the efficiency of large-scale feature processing is further improved. Especially in high-dimensional feature processing scenarios, the parallel computing architecture of the proposed solution has more prominent advantages, not only ensuring computational efficiency but also maintaining the quality of feature extraction, providing a high-quality feature basis for subsequent coupling strength analysis.

[0033] In the prior art, traditional graph attention methods use a single attention head for feature extraction, and all feature dimensions need to be serially processed through the same computing unit. This serial processing method leads to a concentrated computational load. Especially when dealing with high-dimensional features, the computational time will show a rapid growth trend as the dimension increases. At the same time, since all features are processed in the same computing unit, the problem of mutual interference of feature information is likely to occur, affecting the quality of feature extraction. To address the above problems, this embodiment proposes an attention feature extraction scheme based on multi-head parallel computing. The core idea of this scheme is to decompose the high-dimensional feature space into multiple low-dimensional subspaces, and each subspace is responsible for processing by an independent attention computing unit. Through this parallel computing architecture, the computational load is effectively dispersed, avoiding the computational bottleneck in the traditional method. At the same time, different attention heads focus on different feature subspaces, and can capture the correlation relationships between features from multiple perspectives, improving the comprehensiveness and accuracy of feature extraction. Through experimental verification, this embodiment has achieved a significant improvement in the time efficiency of feature extraction. As the feature dimension increases, the growth trend of the computational time is significantly flattened, reflecting good scalability. Especially in the scenario of high-dimensional feature processing, this scheme can save a large amount of computational time compared with the traditional method, while ensuring the quality of feature extraction. This efficient feature extraction mechanism provides a reliable feature basis for subsequent coupling strength analysis, which is of great significance for improving the performance of the entire system. In addition, this embodiment further improves the efficiency of feature processing by optimizing the feature mapping and similarity calculation strategies. The multi-head parallel architecture not only solves the computational efficiency problem, but also improves the richness of feature expression through multi-angle feature extraction, providing strong support for the stable operation and accurate analysis of the system. Generally speaking, this embodiment has achieved significant improvements in computational efficiency, feature quality, and system scalability, providing a better solution for feature extraction and analysis of complex systems.

[0034] In an alternative implementation, when the value of the conflict risk tensor exceeds the preset risk threshold, determine the verification priority according to the propagation path prediction map and the hierarchical importance scoring matrix, and perform hierarchical recursive verification along the priority path higher than the preset path threshold. Locating the actual conflict nodes through the variational autoencoder includes: When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, calculate the verification priority based on the propagation path prediction map and the hierarchical importance scoring matrix, extract all propagation paths of the risk-exceeding nodes, calculate the cumulative propagation probability of each propagation path, and combine the importance scores of each node on the path to obtain the verification priority of the path in a weighted combination manner; Select the priority path higher than the preset path threshold, and perform hierarchical recursive verification along the path priority. The hierarchical recursive verification is carried out sequentially from top to bottom according to the functional levels of the nodes, and the verification result of each layer is used as the input condition for the verification of the next layer; The variational autoencoder is used to encode and reconstruct the nodes on the verification path. The variational autoencoder includes an encoding network and a decoding network. The encoding network maps the multi-dimensional state data of the nodes into the mean vector and variance vector of the latent variable distribution, samples the latent variable based on the mean vector and variance vector, and the decoding network reconstructs the latent variable into the node state; Calculate the error between the node state and the original state. The error includes the mean square error and the distribution divergence. Judge the abnormality degree of the node based on the error size, and determine the nodes with the abnormality degree exceeding the dynamic threshold as the actual conflict nodes.

[0035] First, monitor the conflict risk tensor, which contains risk values in multiple dimensions, and each dimension corresponds to different types of design rule conflict risks. Set the preset risk threshold to 0.8, and trigger the verification process when the risk value in any dimension exceeds this threshold. For example, for a certain node, its process parameter conflict risk value is 0.85, exceeding the preset threshold, and in-depth verification is required.

[0036] Next, determine the verification priority based on the propagation path prediction map and the hierarchical importance scoring matrix. The propagation path prediction map records the propagation probability of risks between different nodes, and the hierarchical importance scoring matrix contains the importance weights of each node. Starting from the risk-exceeding node, extract all possible propagation paths. For each path, multiply the propagation probabilities between the nodes on the path to obtain the cumulative propagation probability, and then perform a weighted combination with the importance scores of the nodes on the path. Taking a specific verification scenario as an example, there is a propagation path containing three nodes: Node A (process layer) → Node B (rule layer) → Node C (parameter layer). The propagation probabilities between the nodes are P(A→B)=0.8 and P(B→C)=0.75 respectively, then the cumulative propagation probability is 0.8×0.75 = 0.6. The importance scores of the three nodes are 0.9, 0.8, and 0.7 respectively. The importance of the path is calculated by weighted average to get 0.8. Finally, the cumulative propagation probability 0.6 and the path importance 0.8 are weighted and combined according to the ratio of 6:4 to obtain the verification priority of this path as 0.75.

[0037] According to the calculated verification priority, select the paths with a value higher than the preset path threshold (set to 0.7) for hierarchical recursive verification. The verification process is carried out from top to bottom according to the functional levels of the nodes, including the process node layer, the design rule layer, and the constraint parameter layer. In each layer of verification, first confirm the upper layer verification result as a constraint condition. Taking the above path as an example, the verification of process node A shows that its oxide layer thickness parameter exceeds the process range and deviates from the standard value by 20%. This constraint condition is passed to the verification process of node B in the design rule layer to evaluate the impact on the minimum spacing rule.

[0038] The variational autoencoder is used to detect anomalies in the nodes on the verification path. The encoding network of the variational autoencoder adopts a four-layer neural network structure. The input layer is a 50-dimensional node state vector, and the state data is compressed into a 16-dimensional feature space through the intermediate layers (50→32→24→16). Taking a certain process parameter node as an example, its state data contains 50 feature dimensions (such as process parameter values, process margins, process windows, etc.), and the 16-dimensional latent variable distribution parameters, including the mean vector and the variance vector, are obtained through the mapping of the encoding network.

[0039] After obtaining the mean vector and the variance vector, random sampling is performed to generate latent variables, and the sampling process is repeated 10 times to increase stability. The decoding network adopts a symmetric structure (16→24→32→50), and reconstructs the 16-dimensional latent variable into the original 50-dimensional state space. For each sampling point, the error between the reconstructed state and the original state is calculated. For example, the original value of the first-dimensional parameter of a certain node is 1.5, and the reconstructed value is 1.8, and the error of this dimension is 0.3. Calculate the average error of all dimensions, and statistically analyze the errors of each sampling to obtain the overall anomaly degree of the node. At the same time, calculate the divergence difference of the state distribution. For example, the standard deviation of the original distribution is 0.2, the standard deviation of the reconstructed distribution is 0.3, and the difference is 0.1.

[0040] Set a dynamic anomaly threshold, which is determined based on the statistical characteristics of historical verification data, and the initial value is set to the mean error plus twice the standard deviation. When the anomaly degree of a node exceeds this threshold, it is marked as an actual conflict node. For example, the average reconstruction error of a certain node is 0.4, which is greater than the current anomaly threshold of 0.35, and the distribution divergence difference is 0.15, which exceeds the threshold of 0.12, then this node is determined to be an actual conflict node. For each marked conflict node, record its anomaly feature pattern, including information such as the anomaly dimension distribution and the size of the anomaly degree, to provide a basis for formulating subsequent correction strategies.

[0041] All detected conflict nodes are sorted from high to low according to the anomaly degree to form a priority processing queue. At the same time, establish an association graph between nodes to record the dependency relationships between conflict nodes, providing a reference for systematic optimization. For example, if three conflict nodes are detected, their anomaly degrees are 0.6, 0.45, and 0.38 respectively, and there is a dependency relationship A→B→C, then in the subsequent correction process, optimization needs to be carried out in this order to ensure that the correction of high-priority nodes will not have a negative impact on low-priority nodes.

[0042] This systematic verification and positioning method can effectively identify design rule conflicts and provide a reliable optimization direction through multi-level data analysis and accurate anomaly detection. At the same time, the design of dynamic thresholds and adaptive weights ensures the applicability of the method in different application scenarios, improving the verification efficiency and accuracy.

[0043] Figure 3 Schematic diagram for comparative analysis of experimental data in the embodiments of the present invention: The experimental results show that the present solution has significant advantages in multiple key performance indicators. In the comparison of core indicators, the conflict detection accuracy of the present solution reaches 96.8%, with an average increase of 13.5% compared to 82.5% of the rule-based method and 85.3% of the traditional machine learning method; in terms of verification efficiency, 185 nodes can be processed per second, with a 65.2% increase compared to 95 and 112 nodes of the comparative methods; the false alarm rate is reduced to 3.2%, with an improvement of 59.0% compared to 8.6% and 7.8% of the comparative methods. In the performance evaluation of the variational autoencoder in different scenarios, a reconstruction accuracy of 97.5%, an anomaly detection rate of 95.8%, and a convergence time of 125 ms are achieved in the simple rule scenario, with increases of 10.5%, 10.9%, and 49.0% respectively; even in the complex rule scenario, a reconstruction accuracy of 94.2% and an anomaly detection rate of 92.6% are still maintained, with a convergence time of 168 ms, and the improvement amplitude compared to the traditional method is further expanded to 17.0%, 17.5%, and 56.5%, fully demonstrating the stability and efficiency of the present solution in various application scenarios.

[0044] The conflict detection methods in the prior art mainly rely on rule-based static matching and simple classification models of traditional machine learning. These methods have obvious limitations in dealing with complex design rule conflicts: rule-based methods are difficult to cope with the dynamically changing design environment and are prone to a high false alarm rate; although traditional machine learning methods have certain learning capabilities, they perform poorly in dealing with high-dimensional features and non-linear relationships, and the verification efficiency is low, making it difficult to meet the real-time requirements of large-scale design verification. In this embodiment, an innovative collaborative verification method based on deep learning is proposed. By using a bidirectional long short-term memory network to capture temporal features and combining a multi-head graph attention network to analyze the dynamic coupling relationship between nodes, a deep rule conflict prediction model is constructed. At the same time, a variational autoencoder is introduced for abnormal node localization and feature extraction, and a reinforcement learning network is used to generate correction strategies, forming a complete intelligent verification framework. This improvement starts from two aspects: enhancing the feature learning ability of the model and optimizing the verification decision-making mechanism, aiming to solve the deficiencies of the prior art in terms of accuracy, efficiency, and adaptability. Through the above technical improvements, this embodiment has achieved remarkable technical effects: First, compared with the prior art, this solution has significantly improved the accuracy of conflict detection, significantly improved the verification efficiency, and effectively reduced the false alarm rate; Second, in simple rule scenarios, the solution exhibits excellent reconstruction accuracy and abnormal detection capabilities, and the convergence speed of the verification process is significantly accelerated; More importantly, in complex rule scenarios, this solution still maintains stable high-performance performance, and the improvement amplitudes of the reconstruction accuracy and abnormal detection rate are more significant, and the advantage of verification efficiency is further expanded. These improvement effects fully demonstrate the powerful ability of this solution in dealing with complex verification tasks, providing a more reliable and efficient technical solution for the collaborative verification of EDA software and heterogeneous process design rules.

[0045] In an alternative embodiment, when the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, the verification priority is calculated based on the propagation path prediction graph and the hierarchical importance scoring matrix. All propagation paths of the risk-exceeding nodes are extracted, the cumulative propagation probability of each propagation path is calculated, and combined with the importance scores of the nodes on the path, the verification priority of the path is obtained by using a weighted combination method, including: When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, starting from the risk-exceeding node in the propagation path prediction graph, perform a depth-first search algorithm to prune the starting points below the preset probability threshold, and generate candidate propagation paths; Extract the direct propagation probability between adjacent node pairs on the candidate propagation path, and multiply the direct propagation probabilities continuously to obtain the first-order propagation probability; calculate the second-order propagation probability of the shortest path between nodes based on the first-order propagation probability; perform a weighted combination of the first-order propagation probability and the second-order propagation probability to obtain the comprehensive propagation probability; multiply the comprehensive propagation probability by the attenuation compensation coefficient of the path length to obtain the final cumulative propagation probability; Extract the static importance scores of each node on the path from the hierarchical importance scoring matrix; calculate the dynamic importance scores by combining the current load status and resource occupancy rate of the nodes; perform an adaptive weighted fusion of the static importance scores and the dynamic importance scores to generate the comprehensive importance scores of the nodes; Calculate the position weighting coefficient based on the distance from the node to the risk-exceeding node, where the farther the node is, the smaller the position weighting coefficient; multiply the comprehensive importance score of the node by the corresponding position weighting coefficient and sum them to obtain the combined importance value of the path nodes; Construct an adaptive weight vector, where the components of the adaptive weight vector correspond to the weights of the final cumulative propagation probability and the combined importance value of the path nodes; train a weight adjustment function based on historical verification data, and adjust the adaptive weight vector according to the system load and verification resource status; after normalizing the final cumulative propagation probability and the combined importance value of the path nodes, perform a non-linear combination operation with the adaptive weight vector to obtain the verification priority of the candidate propagation path.

[0046] In the design rule verification system, first monitor each dimension of the conflict risk tensor. Taking a specific scenario as an example, during the integrated circuit layout verification process, the risk value related to the metal layer wiring rule reaches 0.85, exceeding the preset risk threshold of 0.8, triggering the verification priority calculation process.

[0047] Perform path search from the propagation path prediction graph, starting from the metal layer node with risk exceeding the limit. Adopt the depth-first search strategy, set the preset probability threshold to 0.3, and stop the search of this branch when the propagation probability of a certain node is lower than this threshold. For example, starting from the starting point, three potential paths are found: Path 1 contains 4 nodes, Path 2 contains 3 nodes, and Path 3 contains 5 nodes. After pruning, Path 3 is pruned because the propagation probability of a certain node is 0.2, and Path 1 and Path 2 are retained as candidate propagation paths.

[0048] Calculate the propagation probability for the reserved candidate paths. Taking Path 1 as an example, it includes nodes A→B→C→D, and the direct propagation probabilities between adjacent nodes are 0.8, 0.7, and 0.6 respectively. Multiply these probabilities to get the first-order propagation probability of 0.336. At the same time, considering the indirect propagation effect between nodes, calculate the second-order propagation probability of the shortest path between nodes. For example, in addition to propagating through B from node A to C, it may also propagate through other nodes. After comprehensive consideration, the second-order propagation probability is 0.4. Combine the first-order propagation probability and the second-order propagation probability in a ratio of 7:3 to obtain the comprehensive propagation probability of 0.355.

[0049] Considering the influence of path length on the propagation effect, introduce an attenuation compensation coefficient. The attenuation compensation coefficient for a path length of 4 is set to 0.9, and multiply it with the comprehensive propagation probability to obtain the final cumulative propagation probability of 0.32.

[0050] Extract the static importance information of nodes from the hierarchical importance scoring matrix. Taking Path 1 as an example, the static importance scores of the four nodes are 0.9, 0.8, 0.7, and 0.6 respectively. Combine the real-time status information of the nodes to calculate the dynamic importance score, including factors such as the current number of processing tasks and resource occupancy rate. For example, the current load rate of node B is 85% and the resource occupancy rate is 90%. Based on this, the calculated dynamic importance score is 0.85. Combine the static score and the dynamic score in a ratio of 6:4 to obtain the comprehensive importance score of the node.

[0051] Calculate the position weighting coefficient based on the position of the node in the path. Adopt a distance attenuation mechanism, with the distance between adjacent nodes being 1, and the position weighting coefficient decreases as the distance increases. For Path 1, the position weighting coefficients of the four nodes are 1.0, 0.9, 0.8, and 0.7 respectively. Multiply the comprehensive importance score of the node by the corresponding position weighting coefficient and sum them up to obtain the combined value of path node importance of 0.75.

[0052] Construct an adaptive weight vector for combining the propagation probability and the importance value. Through analyzing historical verification data, establish a weight adjustment model. When the system load rate is 75% and the available verification resource ratio is 60%, adjust the propagation probability weight in the original weight vector from 0.5 to 0.6, and adjust the importance weight from 0.5 to 0.4. Finally, perform a non-linear combination of the adjusted weight vector with the normalized cumulative propagation probability and the combined value of path importance to obtain the final verification priority of this path of 0.68.

[0053] Figure 4 This is a schematic diagram of the verification path priority map for the embodiments of the present invention: The figure adopts three-dimensional space visualization technology to display the spatial distribution and topological connection relationships of 50 key nodes. The nodes are distributed along a spiral trajectory defined based on the cos(i / 5) and sin(i / 5) functions. Through the setting of a unified path width (2 units) and node size (4 units), combined with a visual coding scheme in the blue color system, the hierarchy, continuity, and association strength of the verification process are clearly presented. The spatial configuration of the path not only reflects the logical associations between different levels, but the spiral upward structural feature more effectively supports the implementation of the hierarchical recursive verification strategy. From the perspective of quantitative analysis, the spatial distances and connection strengths between nodes exhibit obvious hierarchical characteristics. Approximately 60% of the connections occur between adjacent levels, 30% of the connections span one level, and only 10% of the connections span multiple levels. This distribution feature provides a reliable topological basis for the scientific determination of verification priorities.

[0054] The determination of verification priorities in the prior art mainly relies on static rule configurations and simple linear scoring mechanisms. This method cannot fully reflect the complex dynamic association relationships between nodes in the system, resulting in problems such as important nodes being delayed in verification or waste of verification resources during the verification process. At the same time, traditional two-dimensional visualization methods are difficult to intuitively display the topological structure and association strength in a multi-level system, reducing the accuracy and efficiency of verification strategy formulation. In this embodiment, by innovatively introducing three-dimensional space visualization technology and combining a spiral trajectory distribution strategy based on trigonometric functions, a systematic verification priority display scheme is established. This scheme encodes the verification levels through the spatial positions of the nodes, represents the association relationships through the connection methods of the paths, and ensures the consistency of information expression through the unified configuration of visual parameters. In particular, by adopting the spiral upward structural feature, the hierarchy and continuity of the system are organically unified, providing intuitive decision-making support for the hierarchical recursive verification strategy. Through this improvement, the following technical effects are achieved in this embodiment: First, compared with the prior art, the accuracy of determining verification priorities is significantly improved, avoiding waste of verification resources; second, through clear hierarchical display, the controllability and predictability of the verification process are greatly improved; third, based on the quantitative analysis of spatial distances and connection strengths, a reliable basis is provided for the optimization of verification strategies; finally, the spiral spatial structure design effectively supports the implementation of hierarchical recursive verification, improving the verification efficiency. These improvements jointly construct a more efficient and accurate verification priority determination mechanism, providing strong technical support for the verification work of complex systems.

[0055] In an alternative embodiment, the calibration operation is performed according to the priority order of the calibration policy tree, the abnormal parameters in the feature vector are adjusted, and the change of the conflict risk tensor is monitored in real time. When the conflict risk tensor is lower than the preset risk threshold, the node state of the dynamic association graph is updated, and the feature vector and the corresponding calibration policy are stored in the experience pool for optimizing the deep rule conflict prediction model, including: According to the execution priority order of the calibration policy tree, perform the calibration operation on the abnormal parameters; the adjustment step size of the calibration operation is calculated by the adaptive learning rate and the parameter deviation, and the adaptive learning rate decays exponentially with time; apply the adjustment step size to the abnormal parameters to obtain the corrected feature vector; Calculate the conflict risk tensor based on the corrected feature vector. The conflict risk tensor characterizes the risk state of the system in multiple dimensions; calculate the Frobenius norm difference of the conflict risk tensors at adjacent times to obtain the risk change amount; determine whether the risk change amount is less than the preset change threshold and whether the maximum component of the conflict risk tensor is lower than the preset risk threshold; When the risk change amount is less than the preset change threshold and the maximum component of the conflict risk tensor is lower than the preset risk threshold, update the state of the node in the dynamic association graph; the state of the node is updated by the state transition function, and the state transition function takes the original node state, the risk change amount, and the corrected feature vector as inputs; the parameters of the state transition function are obtained by training with historical state transition data; Construct an experience sample from the feature vector, the calibration policy, and the risk reduction amount; calculate the importance score of the experience sample, which is obtained by the product of the absolute value of the risk reduction amount and the time decay function; store the experience sample with the importance score in the experience pool; train the deep rule conflict prediction model based on the samples in the experience pool; The training objective of the deep rule conflict prediction model is to minimize the weighted error between the predicted risk and the actual risk; the weighted error is obtained by weighting the prediction error with the sample importance score; update the deep rule conflict prediction model through iterative optimization.

[0056] Perform the calibration operation of the design rule conflict based on the calibration policy tree. Taking a specific design rule conflict scenario as an example, the spacing parameter of a certain metal layer node is abnormal, the current value is 0.15 microns, the target value is 0.2 microns, and the deviation is 0.05 microns. The execution priority of this node in the calibration policy tree is 0.85, which is located in the operation sequence with the highest priority.

[0057] The calibration operation adopts an adaptive adjustment step - size mechanism. The initial learning rate is set to 0.1. After each execution of the calibration operation, the learning rate decays to 0.95 times the original value. For the above - mentioned pitch parameter, the adjustment step - size for the first calibration is 0.005 microns (the product of the learning rate 0.1 and the deviation 0.05 microns), and the parameter is adjusted to 0.155 microns. After multiple iterative adjustments, the parameter gradually approaches the target value of 0.2 microns.

[0058] After each parameter adjustment, the new conflict - risk tensor is calculated immediately. Assume that the system monitors risks in three dimensions: pitch - rule risk, coverage - rate risk, and electrical - rule risk. The risk values before adjustment are 0.85, 0.6, and 0.4 respectively, and after adjustment, they become 0.75, 0.58, and 0.38. The norm difference of the risk tensors before and after calculation gives the risk change amount of 0.12. The preset change threshold is set to 0.05, and the preset risk threshold is 0.8. Since the risk change amount 0.12 is greater than the preset change threshold 0.05, the calibration operation needs to be continued.

[0059] After multiple rounds of calibration, a certain adjustment makes the risk values become 0.65, 0.52, and 0.35. At this time, the risk change amount is 0.04, which is less than the preset change threshold 0.05, and the maximum risk component 0.65 is less than the preset risk threshold 0.8, triggering the update operation of the dynamic association graph.

[0060] The state - transition function receives multiple input information: the original node state (parameter value 0.15 microns), the risk change amount (0.04), and the corrected feature vector (including information such as the current parameter value 0.195 microns). The parameters of the state - transition function are obtained through training by analyzing 1000 groups of historical state - transition data. For example, the state - transition decay coefficient is set to 0.9, and the update weight is 0.7. Based on these parameters, the node state is updated to obtain a new node - state representation.

[0061] The information of this calibration process is constructed into an empirical sample. The sample includes the feature vector (original parameter set), the adopted calibration strategy (progressive parameter adjustment), and the achieved risk reduction amount (0.2). When calculating the sample importance score, consider the product of the risk reduction amount 0.2 and the time - decay factor 0.95 to get the score of 0.19. This empirical sample and its importance score are stored in the empirical pool. The capacity of the empirical pool is set to 10000, and when it exceeds the capacity, the sample with the lowest score is deleted.

[0062] Train a deep rule conflict prediction model using the samples in the experience pool. The model adopts a five-layer neural network structure, where the input layer corresponds to the dimension of the feature vector and the output layer corresponds to the dimension of the risk tensor. During the training process, the prediction error of the sample will be weighted according to its importance score. For example, if the difference between the predicted risk and the actual risk of a sample is 0.1 and the importance score is 0.19, then its weighted error is 0.019. Iteratively optimize the model parameters through the backpropagation algorithm until the weighted error converges below the preset threshold.

[0063] Figure 5 This is a schematic diagram showing the trend of the verification efficiency of the embodiments of the present invention changing with the node scale: The "technical solution of the present invention" marked with a circle represents a multi-level verification method based on adaptive weights, including a path optimization algorithm with a depth-first search and a dynamic pruning strategy; the "static weight method" marked with a square represents a traditional verification method using a fixed weight coefficient; the "benchmark method" marked with a triangle represents a simple breadth-first traversal verification strategy. This figure shows the comparison of the verification efficiencies of the three methods under different node scales. The data shows that when the node scale is 50, the verification efficiency of the technical solution of the present invention reaches 96%, the static weight method is 85%, and the benchmark method is only 80%. As the node scale increases to 1000, the efficiency of the technical solution of the present invention still remains at a high level of 90%, only decreasing by 6 percentage points; while the static weight method and the benchmark method decrease to 68% and 60% respectively, with obvious performance degradation. This excellent scalability stems from the adaptive weight adjustment mechanism and the efficient path pruning strategy adopted by the technical solution of the present invention, showing significant advantages in large-scale node verification scenarios. By dynamically adjusting the verification strategy and optimizing the path selection, this solution effectively solves the performance bottleneck problem of traditional methods when the node scale increases.

[0064] In the prior art, traditional verification methods with fixed weight coefficients or simple breadth - first traversal strategies are generally adopted for node verification. These methods can meet the basic requirements when dealing with a small - scale number of nodes. However, as the scale of nodes expands, due to the lack of a flexible weight adjustment mechanism and path optimization strategy, the verification efficiency will significantly decline, facing serious performance bottleneck problems in large - scale node verification scenarios. To address the above - mentioned technical problems, in this embodiment, starting from improving the adaptability and intelligence of the verification method, the following improvement solutions are adopted: First, an adaptive weight adjustment mechanism is introduced, which can dynamically adjust the verification strategy according to node characteristics and real - time feedback during the verification process; Second, the selection of the verification path is optimized through the depth - first search algorithm; Finally, combined with the dynamic pruning strategy, invalid verification paths are excluded in a timely manner to improve the efficiency of path selection. Through the synergistic effect of the above - mentioned technical means, this embodiment has achieved remarkable technical effects: Compared with the prior art, this solution shows an obvious verification efficiency advantage when the node scale is small; More importantly, as the node scale increases, this solution still maintains a high verification efficiency, and the attenuation degree of verification performance is much lower than that of the prior - art solutions, showing excellent scalability. This improvement in technical effects directly stems from the adaptive weight adjustment mechanism and efficient path optimization strategy adopted in the solution, providing more efficient and reliable technical support for large - scale node verification applications.

[0065] In the second aspect of the embodiments of the present invention, a kind of electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0066] In the third aspect of the embodiments of the present invention, a computer - readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0067] The present invention can be a method, a device, a system, and / or a computer program product. The computer program product may include a computer - readable storage medium, on which computer - readable program instructions for executing various aspects of the present invention are loaded.

[0068] Finally, it should be noted that: The above - mentioned 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 foregoing embodiments, those of ordinary skill in the art should understand that: They can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; And these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A co-verification method for EDA software and heterogeneous process design rules, characterized in that Including: Utilize a bidirectional long short-term memory network to extract the temporal features of the process node layer, design rule layer, and constraint parameter layer in the dynamic association graph. Analyze the dynamic coupling strength between nodes through a multi-head graph attention network to generate a hierarchical importance scoring matrix. Train a deep rule conflict prediction model based on the hierarchical importance scoring matrix. The deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map. The conflict risk tensor quantifies the conflict probability of each node, and the propagation path prediction map represents the diffusion path of conflicts between different levels; When the value of the conflict risk tensor exceeds the preset risk threshold, determine the verification priority according to the propagation path prediction map and the hierarchical importance scoring matrix, and perform hierarchical recursive verification along the priority path higher than the preset path threshold. Locate the actual conflict nodes through a variational autoencoder, and extract the feature vectors of the actual conflict nodes. The feature vectors contain the status information of process parameters, design rules, and constraints. Input the feature vectors into a reinforcement learning network to generate a correction policy tree; Execute correction operations in the priority order of the correction policy tree, adjust the abnormal parameters in the feature vectors, and monitor the change of the conflict risk tensor in real time. When the conflict risk tensor is lower than the preset risk threshold, update the node status of the dynamic association graph, and store the feature vectors and the corresponding correction policies in the experience pool for optimizing the deep rule conflict prediction model.

2. The method according to claim 1, wherein Analyze the dynamic coupling strength between nodes through a multi-head graph attention network to generate a hierarchical importance scoring matrix. Train a deep rule conflict prediction model based on the hierarchical importance scoring matrix. The deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map. The conflict risk tensor quantifies the conflict probability of each node, and the propagation path prediction map represents the diffusion path of conflicts between different levels, including: Apply a multi-head graph attention network to analyze the dynamic coupling strength between nodes. Each attention head independently calculates the attention coefficient between node pairs. The attention coefficient is obtained through the non-linear transformation and similarity calculation of node feature vectors. Normalize and fuse the outputs of multiple attention heads to obtain the final coupling strength between nodes; Generate a hierarchical importance scoring matrix based on the final coupling strength. In the generation process of the hierarchical importance scoring matrix, simultaneously consider the direct coupling strength between nodes and the indirect coupling strength accumulated through multi-hop paths, and use an adaptive weight factor to balance the influence of the direct coupling strength and the indirect coupling strength; Input the hierarchical importance scoring matrix into the deep rule conflict prediction model for training. The deep rule conflict prediction model outputs a conflict risk tensor and a propagation path prediction map. Among them, the conflict risk tensor quantifies the conflict probability of each node under different time windows and different risk types, and the propagation path prediction map is constructed based on the node risk degree and the hierarchical importance scoring matrix, and is used to represent the diffusion path of conflicts between process nodes, rule nodes, and constraint nodes.

3. The method according to claim 2, wherein Apply the multi-head graph attention network to analyze the dynamic coupling strength between nodes. Each attention head independently calculates the attention coefficient between node pairs. The attention coefficient is obtained through the non-linear transformation and similarity calculation of node feature vectors. The outputs of multiple attention heads are normalized and fused to obtain the final coupling strength between nodes, including: Obtain the node feature matrix, which contains the feature information of multiple nodes to be analyzed. The node feature matrix is input into the multi-head graph attention network. The multi-head graph attention network performs feature transformation on the node feature matrix, and maps the node feature matrix to a high-dimensional feature space through multiple attention calculation units to obtain the transformed node feature set; Calculate the attention coefficient between node pairs for the transformed node feature set. Each attention calculation unit processes the node features using a non-linear transformation function, calculates the similarity between node pairs based on the processed node features to obtain the original attention value, and obtains the final attention coefficient after normalizing the original attention value; Perform feature aggregation based on the final attention coefficient. Each attention calculation unit uses its corresponding attention coefficient to perform weighted combination on the node features to obtain the node output representation corresponding to the attention calculation unit. The node output representations of each attention calculation unit are weighted and fused, and the outputs of different attention calculation units are combined through learnable weight coefficients to obtain the final coupling strength between nodes.

4. The method according to claim 1, characterized in that, When the value of the conflict risk tensor exceeds the preset risk threshold, determine the verification priority according to the propagation path prediction graph and the hierarchical importance scoring matrix, and perform hierarchical recursive verification along the priority path higher than the preset path threshold. Locate the actual conflict nodes through the variational autoencoder, including: When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, calculate the verification priority based on the propagation path prediction graph and the hierarchical importance scoring matrix, extract all propagation paths of the risk-exceeding nodes, calculate the cumulative propagation probability of each propagation path, and combine the importance scores of each node on the path. Use the weighted combination method to obtain the verification priority of the path; Select the priority path higher than the preset path threshold, and perform hierarchical recursive verification along the path priority. The hierarchical recursive verification is performed sequentially from top to bottom according to the functional levels of the nodes. The verification result of each layer is used as the input condition for the verification of the next layer; Use the variational autoencoder to encode and reconstruct the nodes on the verification path. The variational autoencoder includes an encoding network and a decoding network. The encoding network maps the multi-dimensional state data of the nodes to the mean vector and variance vector of the latent variable distribution, samples the latent variable based on the mean vector and variance vector, and the decoding network reconstructs the latent variable into the node state; Calculate the error between the node state and the original state. The error includes the mean square error and the distribution divergence. Based on the size of the error, judge the abnormality degree of the node, and determine the node with the abnormality degree exceeding the dynamic threshold as the actual conflict node.

5. The method according to claim 4, characterized in that, When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, calculate the verification priority based on the propagation path prediction map and the hierarchical importance scoring matrix. Extract all the propagation paths of the risk-exceeding nodes, calculate the cumulative propagation probability of each propagation path, and combine the importance scores of each node on the path. The verification priority of the path is obtained by using the weighted combination method, including: When the risk value of any dimension in the conflict risk tensor exceeds the preset risk threshold, starting from the risk-exceeding node in the propagation path prediction map, perform a depth-first search algorithm to prune the starting points with probabilities lower than the preset probability threshold, and generate candidate propagation paths; Extract the direct propagation probability between adjacent node pairs on the candidate propagation path, and multiply the direct propagation probabilities continuously to obtain the first-order propagation probability; calculate the second-order propagation probability of the shortest path between nodes based on the first-order propagation probability; perform a weighted combination of the first-order propagation probability and the second-order propagation probability to obtain the comprehensive propagation probability; multiply the comprehensive propagation probability by the attenuation compensation coefficient of the path length to obtain the final cumulative propagation probability; Extract the static importance scores of each node on the path from the hierarchical importance scoring matrix; calculate the dynamic importance scores by combining the current load status and resource occupancy rate of the nodes; perform an adaptive weighted fusion of the static importance scores and the dynamic importance scores to generate the comprehensive importance scores of the nodes; Calculate the position weighting coefficient based on the distance from the node to the risk-exceeding node. The farther the node is, the smaller the position weighting coefficient; multiply the comprehensive importance score of the node by the corresponding position weighting coefficient and sum them to obtain the combined importance value of the path nodes; Construct an adaptive weight vector. The components of the adaptive weight vector correspond to the weights of the final cumulative propagation probability and the combined importance value of the path nodes; train the weight adjustment function based on historical verification data, and adjust the adaptive weight vector according to the system load and verification resource status; after normalizing the final cumulative propagation probability and the combined importance value of the path nodes, perform a non-linear combination operation with the adaptive weight vector to obtain the verification priority of the candidate propagation path.

6. The method according to claim 1, wherein Perform correction operations in the priority order of the correction strategy tree, adjust the abnormal parameters in the feature vector, and monitor the changes in the conflict risk tensor in real time. When the conflict risk tensor is lower than the preset risk threshold, update the node status of the dynamic association graph, and store the feature vector and the corresponding correction strategy in the experience pool for optimizing the deep rule conflict prediction model, including: According to the execution priority order of the correction strategy tree, perform correction operations on the abnormal parameters; the adjustment step size of the correction operation is calculated by the adaptive learning rate and the parameter deviation, and the adaptive learning rate decays exponentially with time; apply the adjustment step size to the abnormal parameters to obtain the corrected feature vector; Calculate the conflict risk tensor based on the corrected feature vector. The conflict risk tensor represents the risk status of the system in multiple dimensions; calculate the Frobenius norm difference of the conflict risk tensors at adjacent times to obtain the risk change amount; determine whether the risk change amount is less than the preset change threshold and whether the maximum component of the conflict risk tensor is lower than the preset risk threshold; When the risk change amount is less than the preset change threshold and the maximum component of the conflict risk tensor is lower than the preset risk threshold, update the state of the nodes in the dynamic association graph; the state of the nodes is updated through a state transition function, and the state transition function takes the original node state, the risk change amount, and the corrected feature vector as inputs; the parameters of the state transition function are obtained by training with historical state transition data; Construct an empirical sample from the feature vector, the correction strategy, and the risk reduction amount; calculate the importance score of the empirical sample, which is obtained by multiplying the absolute value of the risk reduction amount by a time decay function; store the empirical sample with the importance score in the empirical pool; train a deep rule conflict prediction model based on the samples in the empirical pool; The training objective of the deep rule conflict prediction model is to minimize the weighted error between the predicted risk and the actual risk; the weighted error is obtained by weighting the prediction error with the sample importance score; update the deep rule conflict prediction model through iterative optimization.

Images