Integrated circuit hot spot detection method based on multi-modal feature fusion and dynamic optimization

Abstract

The invention provides an integrated circuit hot spot detection method based on multi-modal feature fusion and dynamic optimization, which relates to the technical field of integrated circuits, and comprises the following steps: acquiring multi-modal feature data, executing progressive block clustering to extract multi-scale features, fusing the features by using a multi-head attention mechanism, and constructing a hot spot feature map. A hot spot area is determined by combining spatial autocorrelation analysis, and a detection result is optimized based on synchronization analysis of a morphological evolution and energy evolution sequence. The hot spot detection accuracy and robustness can be improved, the false detection rate can be reduced, and the potential hot spot area in the complex layout can be effectively identified.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit technology, and in particular to an integrated circuit hotspot detection method based on multimodal feature fusion and dynamic optimization. Background Technology

[0002] As integrated circuit manufacturing processes continue to evolve towards smaller feature sizes, the impact of optical proximity effects and other manufacturing process variations on chip manufacturing yield is becoming increasingly prominent. Hotspot regions refer to locations in the integrated circuit layout where circuit functionality may fail due to manufacturing process variations. Hotspot detection technology, as a key technology for improving integrated circuit manufacturing yield, has become an indispensable part of the design verification process. Traditional hotspot detection methods mainly rely on rule checking and simulation. With the increasing complexity of integrated circuits, these methods face severe challenges in terms of efficiency and accuracy. In recent years, machine learning, especially deep learning technology, has shown great potential in the field of hotspot detection. By learning hotspot feature patterns from historical data, hotspot prediction can be performed more efficiently.

[0003] However, existing hotspot detection technologies have significant shortcomings. Most hotspot detection methods focus only on single-modal features, such as geometric shape or electrical parameters, ignoring the complex interactions between different types of features, resulting in one-sided detection results and a lack of comprehensive judgment ability. Existing methods lack effective mechanisms for processing multi-scale features, making it difficult to simultaneously capture the influence of local microstructure and global layout features, and prone to missed detections and false detections in complex layout environments. Traditional hotspot detection methods usually use static detection models, which cannot dynamically adjust the detection strategy according to the features of different regions, making it difficult to adapt to the increasingly complex layout changes and hotspot pattern evolution in modern integrated circuit design. Summary of the Invention

[0004] This invention provides a method for detecting hotspots in integrated circuits based on multimodal feature fusion and dynamic optimization, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for detecting hotspots in integrated circuits based on multimodal feature fusion and dynamic optimization, comprising: Acquire multimodal feature data of integrated circuit design layout and generate layout feature vector sets; Progressive block clustering is performed on the feature vector group of the map to obtain multi-scale feature components. Local map feature descriptions are extracted at different scale layers. Feature fusion is performed through a multi-head attention mechanism to generate a fused feature matrix. The fused feature matrix is ​​input into the residual network, and hotspot regions are determined based on the gradient change magnitude of the feature regions to generate a target semantic feature set. A hotspot feature map is constructed based on the target semantic feature set. The spatial distribution characteristics of hotspot regions are determined by clustering methods to generate original hotspot detection data. A hotspot density distribution map is constructed based on the original hotspot detection data, and the correlation strength matrix between hotspot regions is calculated using spatial autocorrelation analysis. Based on the correlation strength matrix and the hotspot density distribution map, a region expansion operation is performed to generate a set of candidate hotspot regions. Geometric and electrical features are extracted from candidate hotspot regions, and iterative optimization is performed based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data. Based on the target hotspot detection data, output the hotspot location information in the integrated circuit design layout.

[0006] In one optional embodiment, progressive block clustering is performed on the map feature vector group to obtain multi-scale feature components. Local map feature descriptions are extracted at different scale layers, and feature fusion is performed through a multi-head attention mechanism to generate a fused feature matrix, including: Progressive block clustering is performed based on the map feature vector group. By dynamically adjusting the clustering radius, structural feature components and topological feature components of multiple scale layers are obtained. The structural feature components and the topological feature components are converted into a map feature coefficient matrix through sparse coding. The map feature coefficient matrix is ​​then used to reconstruct the map feature representation of different scale layers. The spatial distribution matrix is ​​calculated based on the layout feature representation, and the feature vector set is obtained by eigenvalue decomposition. After mapping the feature vector set to a high-dimensional feature space using a kernel function, a low-dimensional feature subspace that preserves the topological structure of the layout is obtained by manifold learning. A device connection graph is constructed from the low-dimensional feature subspace, and a similarity matrix between device nodes is calculated. Based on the similarity matrix, the layout feature combination pattern is identified, the information gain value of the layout features is calculated, and a local layout feature description with significant weights is generated. The hierarchical fusion network with multi-head attention mechanism is trained using the local map feature descriptions to obtain the weight coefficients of map features at different scales. The local map feature descriptions with different weight coefficients are fused by residual connections and optimized through backpropagation to obtain the fused feature matrix.

[0007] In one optional embodiment, the fused feature matrix is ​​input into the residual network, and hotspot regions are determined based on the gradient change magnitude of the feature regions to generate a target semantic feature set, including: Perform residual network operations on the fused feature matrix to obtain the initial feature map; Calculate the difference between the directional gradient value and the texture complexity value of each feature region in the initial feature map and the feature regions in the corresponding adaptive neighborhood set, and generate a feature difference evaluation matrix; Perform iterative comparison operation on the feature difference evaluation matrix, take the position with value greater than the first preset value as hot spot feature point, take the position with value less than the second preset value as non-hot spot feature point, keep the feature value of the hot spot feature point position and set the feature value of the non-hot spot feature point position to zero, and obtain the target area feature map. The target region feature map is divided into multiple feature reconstruction blocks. Reconstruction weight coefficients are assigned according to the gradient change magnitude of each feature reconstruction block. Weighted reconstruction is performed according to the reconstruction weight coefficients to generate a target semantic feature set.

[0008] In an optional embodiment, calculating the difference between the directional gradient value and texture complexity value of each feature region in the initial feature map and the corresponding feature regions in the adaptive neighborhood set, and generating a feature difference evaluation matrix includes: Device connection information is extracted from the initial feature map, an adaptive neighborhood set for each feature region is constructed, and the number of feature regions in the adaptive neighborhood set is determined based on the topological and electrical similarity between feature regions. Perform multi-directional rotation filtering on the feature regions in the initial feature map to obtain multi-scale directional response vectors, and determine the directional gradient value based on the directional response amplitude; Calculate the local spectral entropy and local intrinsic dimension of the feature region to generate a texture complexity value, and construct a multi-scale feature vector by combining the directional gradient value and the texture complexity value; Calculate the feature weights of each scale component of the multi-scale feature vector to generate a comprehensive feature vector of the feature region. Calculate the local covariance matrix between the feature region and the feature region in the adaptive neighborhood set, and calculate the Mahalanobis distance difference metric of the comprehensive feature vector based on the local covariance matrix; The confidence level is calculated based on the distribution of Mahalanobis distance difference metric values. The Mahalanobis distance difference metric values ​​and the confidence level are combined to generate a feature difference evaluation matrix. Each value of the feature difference evaluation matrix corresponds to the gradient change magnitude of the feature region at the same position in the initial feature map.

[0009] In one optional embodiment, constructing a hotspot density distribution map based on the original hotspot detection data and calculating the correlation strength matrix between hotspot regions using spatial autocorrelation analysis includes: Normalization is performed on the raw hotspot detection data to obtain standard values ​​of hotspot data. These standard values ​​are then used to construct a hotspot location matrix, and spatial coordinate information and hotspot attribute information are extracted. Substitute spatial coordinate information and hotspot attribute information into the kernel density function to obtain the initial density estimate. Set the bandwidth adjustment parameter based on the initial density estimate and calculate the kernel density value for the preset grid points to obtain the hotspot density distribution map. Based on the hotspot density distribution map, the spatial autocorrelation coefficient is calculated to obtain the proximity relationship matrix. The spatial weight function is constructed using the proximity relationship matrix to calculate the correlation degree of hotspot areas. Hotspot areas that meet the preset correlation threshold are grouped into a hotspot association set. Based on the hotspot association set, a spatial network topology is constructed, node connection relationships are extracted to obtain an association path set, the spatial distance between adjacent nodes in the association path set is calculated to determine the distance matrix, and the distance matrix and the association degree of the hotspot area are combined to generate an association strength matrix.

[0010] In one optional embodiment, performing a region expansion operation based on the correlation strength matrix and the hotspot density distribution map to generate a candidate hotspot region set includes: The correlation strength matrix is ​​input into the spatial distance attenuation function to obtain the attenuation coefficient matrix. Based on the attenuation coefficient matrix, the region growth constraint condition is constructed. Based on the region growth constraint condition and the hotspot density distribution map, the initial expansion boundary of the hotspot region is determined. A partitioning operation is performed on the initial expansion boundary to obtain the core expansion region and the transition expansion region. The density gradient value of the core expansion region is calculated using the hotspot density distribution map to delineate the expansion range of the core region. A buffer zone is constructed around the expansion range of the core region to obtain the expansion range of the transition region. Based on the spatial distance decay function, the expansion thresholds of the core expansion region and the transition expansion region are calculated. The expansion thresholds are substituted into the adaptive weight function to obtain the boundary adjustment parameters. The optimization operation is performed on the expansion region according to the boundary adjustment parameters to obtain the optimized boundary. The overlapping regions in the optimization boundary are detected, and the overlapping regions are merged according to the correlation strength matrix. The spatial continuity index and regional integrity index of the merged region are calculated, and the regions that meet the preset index thresholds are formed into a candidate hotspot region set.

[0011] In one optional embodiment, geometric and electrical features are extracted from candidate hotspot regions, and iterative optimization is performed based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data, including: Geometric features are extracted from candidate hotspot regions, the trend of boundary curvature change of geometric features is calculated, and Fourier spectral analysis is performed to obtain morphological evolution sequences. Electrical features are extracted from candidate hotspot areas, the power loss distribution of the electrical features is calculated, and wavelet transform is performed to obtain the energy evolution sequence. Align the morphological evolution sequence with the energy evolution sequence in the time-frequency domain, calculate the phase coherence spectrum between the sequences, and extract the frequency component with the largest amplitude in the phase coherence spectrum to construct a feature co-evolution matrix. Based on the feature co-evolution matrix, the synchronization degree between the morphological evolution sequence and the energy evolution sequence is calculated, a feature enhancement function is constructed, and the feature enhancement function is used to iteratively optimize the hotspot region; Calculate the difference between the morphological evolution sequence and the difference between the energy evolution sequence of two adjacent iterations of optimization. When both the difference between the morphological evolution sequence and the difference between the energy evolution sequence are less than a preset threshold, output the optimized hotspot region as the target hotspot detection data.

[0012] A second aspect of this invention provides an integrated circuit hotspot detection system based on multimodal feature fusion and dynamic optimization, comprising: The first unit is used to acquire multimodal feature data of integrated circuit design layout and generate layout feature vector groups; The second unit is used to perform progressive block clustering on the map feature vector group to obtain multi-scale feature components, extract local map feature descriptions at different scale layers, and perform feature fusion through a multi-head attention mechanism to generate a fused feature matrix. The third unit is used to input the fused feature matrix into the residual network, determine the hotspot regions based on the gradient change magnitude of the feature regions, and generate the target semantic feature set. The fourth unit is used to construct a hotspot feature map based on the target semantic feature set, determine the spatial distribution characteristics of hotspot regions through clustering methods, and generate original hotspot detection data. The fifth unit is used to construct a hotspot density distribution map based on the original hotspot detection data, and to calculate the correlation strength matrix between hotspot regions using spatial autocorrelation analysis; based on the correlation strength matrix and the hotspot density distribution map, it performs region expansion operations to generate a set of candidate hotspot regions; The sixth unit is used to extract geometric and electrical features from candidate hotspot regions, and to perform iterative optimization based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data. The seventh unit is used to output hotspot location information in the integrated circuit design layout based on the target hotspot detection data.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] In this embodiment of the invention, multi-dimensional information of integrated circuit layout is extracted through a multi-modal feature fusion mechanism. Combined with progressive block clustering and multi-head attention mechanisms, layout features at different scales can be effectively captured, significantly improving the sensitivity and accuracy of hotspot detection and reducing the false alarm and false positive rates. A feature extraction and target semantic feature construction method based on residual networks is adopted, combined with hotspot feature maps and spatial autocorrelation analysis, to achieve precise localization of hotspot regions and assessment of correlation strength, improving the detection method's ability to identify potential hotspots in complex layout structures. An iterative optimization method for synchronizing morphological evolution sequences and energy evolution sequences is introduced. Through comprehensive consideration of geometric and electrical features, dynamic optimization and adjustment of hotspot regions are achieved, enhancing the stability and reliability of detection results and providing more accurate guidance for integrated circuit design optimization. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the integrated circuit hotspot detection method based on multimodal feature fusion and dynamic optimization according to an embodiment of the present invention. Figure 2 This is a data flow graph for adaptive boundary identification of hotspot regions based on the correlation strength matrix. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the integrated circuit hotspot detection method based on multimodal feature fusion and dynamic optimization according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire multimodal feature data of integrated circuit design layout and generate layout feature vector sets; Progressive block clustering is performed on the feature vector group of the map to obtain multi-scale feature components. Local map feature descriptions are extracted at different scale layers. Feature fusion is performed through a multi-head attention mechanism to generate a fused feature matrix. The fused feature matrix is ​​input into the residual network, and hotspot regions are determined based on the gradient change magnitude of the feature regions to generate a target semantic feature set. A hotspot feature map is constructed based on the target semantic feature set. The spatial distribution characteristics of hotspot regions are determined by clustering methods to generate original hotspot detection data. A hotspot density distribution map is constructed based on the original hotspot detection data, and the correlation strength matrix between hotspot regions is calculated using spatial autocorrelation analysis. Based on the correlation strength matrix and the hotspot density distribution map, a region expansion operation is performed to generate a set of candidate hotspot regions. Geometric and electrical features are extracted from candidate hotspot regions, and iterative optimization is performed based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data. Based on the target hotspot detection data, output the hotspot location information in the integrated circuit design layout.

[0020] In one specific implementation, the hotspot detection technology for integrated circuit design layouts achieves accurate identification and location of hotspot regions through multimodal feature extraction and fusion analysis. After acquiring integrated circuit design layout data, multimodal feature data, including geometric features, electrical features, and process features, are extracted and converted into numerical vectors to form a layout feature vector set. The layout feature vector set contains feature information such as the topological relationships, linewidths, spacing, and density of the circuit layout, with each feature vector representing a set of attributes for a specific region in the layout.

[0021] Progressive block clustering is performed on the feature vector group of the map, using an adaptive window size for multi-level region partitioning. The lowest level uses fine-grained partitioning with a window size of 10μm × 10μm; the middle layer uses medium-grained partitioning with a window size of 50μm × 50μm; and the top level uses coarse-grained partitioning with a window size of 200μm × 200μm. At each level, local map feature descriptions are extracted, including edge complexity, density distribution, and connectivity patterns. Feature fusion is performed using a multi-head attention mechanism to integrate feature information from different levels, with each attention head focusing on different types of feature association patterns. The multi-head attention mechanism uses eight attention heads, each with an output dimension of 64, ultimately generating a 512-dimensional fused feature matrix.

[0022] The fused feature matrix is ​​input into a residual network containing 16 convolutional layers for deep feature extraction, with each convolutional layer having 128 output channels. During feature extraction, potential hotspot regions are identified based on the gradient change magnitude of the feature regions. Regions with gradient change magnitudes exceeding a preset threshold of 0.75 are marked as candidate hotspots. Feature aggregation and filtering are performed on candidate hotspot regions to extract semantic features, including hotspot type, severity, and impact range, forming a target semantic feature set.

[0023] A hotspot feature map is constructed based on the target semantic feature set. Nodes in the map represent hotspot regions, and edges represent the relationships between hotspots. The spatial distribution characteristics of hotspot regions are determined using density clustering, with a density threshold of 5 hotspots per square millimeter. Clustering analysis results show that hotspot regions are mainly concentrated in high-density interconnect areas of circuits and power distribution networks. Raw hotspot detection data is generated based on the clustering results, including hotspot location coordinates, type identifiers, and severity scores.

[0024] A hotspot density distribution map is constructed based on the original hotspot detection data, using a kernel density estimation method with a kernel function bandwidth of 25 μm. The density distribution map visually displays the concentrated areas of hotspots, with density values ​​normalized from 0 to 1. Spatial autocorrelation analysis is used to calculate the correlation strength matrix between hotspot regions, with elements ranging from -1 to 1, representing the degree of correlation between hotspots. Hotspot regions with a correlation strength exceeding 0.6 are considered significantly correlated. Based on the correlation strength matrix and the hotspot density distribution map, a region expansion operation is performed, merging and expanding regions with strong correlation and high density values ​​to generate a set of candidate hotspot regions.

[0025] Geometric and electrical features are extracted from candidate hotspot regions. Geometric features include region area, perimeter, and shape complexity, while electrical features include current density, potential gradient, and resistance distribution. For each candidate hotspot region, a morphological evolution sequence and an energy evolution sequence are generated. The morphological evolution sequence describes the trend of geometric properties changing with parameters, while the energy evolution sequence describes the trend of electrical properties changing with parameters. The two sequences are synchronized, and their similarity is calculated. Regions with a similarity exceeding 0.85 are identified as true hotspots. Through an iterative optimization process, the boundaries of hotspot regions are adjusted, the core regions of hotspots are precisely located, and finally, target hotspot detection data is generated.

[0026] Based on the target hotspot detection data, this method outputs hotspot location information in the integrated circuit design layout, including hotspot center coordinates, hotspot boundary contours, hotspot type, severity score, and suggested remediation solutions. The hotspot location accuracy reaches ±2μm, accurately identifying areas that may lead to manufacturing defects. In a practical application, in a chip design containing 50 million transistors, this method successfully detected 127 potential hotspot regions. Compared with traditional methods, the detection accuracy is improved by 18%, and the false negative rate is reduced by 23%, providing important data for integrated circuit design verification and optimization.

[0027] Through the above-described technical implementation methods, high-precision detection of hot spots in integrated circuit design layouts is achieved, providing effective support for improving chip manufacturing yield and significantly reducing the number of design iterations and product launch time.

[0028] In one optional implementation, progressive block clustering is performed on the map feature vector group to obtain multi-scale feature components. Local map feature descriptions are extracted at different scale layers, and feature fusion is performed through a multi-head attention mechanism to generate a fused feature matrix, including: Progressive block clustering is performed based on the map feature vector group. By dynamically adjusting the clustering radius, structural feature components and topological feature components of multiple scale layers are obtained. The structural feature components and the topological feature components are converted into a map feature coefficient matrix through sparse coding. The map feature coefficient matrix is ​​then used to reconstruct the map feature representation of different scale layers. The spatial distribution matrix is ​​calculated based on the layout feature representation, and the feature vector set is obtained by eigenvalue decomposition. After mapping the feature vector set to a high-dimensional feature space using a kernel function, a low-dimensional feature subspace that preserves the topological structure of the layout is obtained by manifold learning. A device connection graph is constructed from the low-dimensional feature subspace, and a similarity matrix between device nodes is calculated. Based on the similarity matrix, the layout feature combination pattern is identified, the information gain value of the layout features is calculated, and a local layout feature description with significant weights is generated. The hierarchical fusion network with multi-head attention mechanism is trained using the local map feature descriptions to obtain the weight coefficients of map features at different scales. The local map feature descriptions with different weight coefficients are fused by residual connections and optimized through backpropagation to obtain the fused feature matrix.

[0029] In one specific implementation, an integrated circuit layout feature vector set is obtained. This feature vector set contains data such as coordinate information, area information, shape features, and hierarchical relationships of each device in the layout. Taking an integrated circuit layout with 1000 device nodes as an example, each device node has an 8-dimensional feature vector, forming an initial feature matrix of 1000×8.

[0030] When implementing progressive block clustering, the initial cluster radius r0 = 5 μm, the decrease coefficient α = 0.8, and the minimum radius threshold rmin = 0.5 μm are set. In the first clustering iteration, using r0 as the search radius, device nodes within 5 μm are identified and grouped into one class, resulting in the initial clustering result C1. At this point, 125 first-level clusters are obtained, with each cluster containing an average of 8 device nodes. For each cluster, structural feature components (including density, average device size, orientation, etc.) and topological feature components (including connectivity, path complexity, etc.) are calculated to form the feature description L1 of the first scale layer.

[0031] The second clustering iteration adjusts the radius to r1 = r0 × α = 4 μm, and the 125 clusters obtained from the first clustering are clustered again to obtain clustering result C2, which contains 52 secondary clusters. The structural feature components and topological feature components of these 52 clusters are calculated to form the feature description L2 of the second scale layer.

[0032] Following the iterative process described above, the third and fourth clustering iterations are performed, with radii of 3.2 μm and 2.56 μm respectively, forming feature descriptions at scales L3 and L4. The iteration terminates when the cluster radius is less than rmin. Finally, a multi-scale feature component set {L1, L2, L3, L4} containing four scale layers is obtained.

[0033] For the feature components of each scale layer, a sparse coding method is used for transformation. An overcomplete dictionary D containing 256 basis vectors is constructed. Sparse coding is performed on the structural and topological feature components of each scale layer to solve for the optimal sparse coefficients, minimizing the reconstruction error and ensuring that the L1 norm of the coefficients does not exceed a set threshold of 0.1. Taking the L1 layer as an example, the feature coefficient matrix S1∈R125×256 is obtained through sparse coding, where each row represents the sparse representation of the corresponding cluster in dictionary D.

[0034] The obtained feature coefficient matrix is ​​used to reconstruct the map feature representation of each scale layer. The reconstruction process involves multiplying the feature coefficient matrix with the dictionary; for example, the reconstruction representation of L1 is R1 = S1 × D. In this way, the map feature representation of each scale layer {R1, R2, R3, R4} is obtained.

[0035] Based on the reconstructed graph feature representation, the spatial distribution matrix is ​​calculated. Taking L2 layer as an example, an adjacency graph G is constructed, where the nodes are 52 second-level clusters, and the edge weights are determined based on the Euclidean distance and connection strength between nodes. The Laplacian matrix of G is calculated, and eigenvalue decomposition is performed. The eigenvectors corresponding to the first 20 smallest non-zero eigenvalues ​​are taken to form the eigenvector group V2.

[0036] A radial basis function kernel is used to map the feature vector set to a high-dimensional feature space. The kernel parameter σ is set to 0.5 times the average distance between nodes. After mapping, the t-SNE manifold learning method is applied to reduce the high-dimensional features to 3 dimensions, resulting in a low-dimensional feature subspace T2 that preserves the topological structure.

[0037] A device connectivity graph H is constructed from a low-dimensional feature subspace, with edge weights calculated based on the cosine similarity of nodes in the T2 space. A connection is established when the similarity between two nodes exceeds 0.75. The similarity between all node pairs is calculated to form a similarity matrix M2∈R52×52.

[0038] Based on the similarity matrix, a frequent pattern mining algorithm was used to identify feature combination patterns in the local map. With a support threshold of 0.3 and a confidence threshold of 0.8, 15 significant feature combination patterns were identified. For each pattern, an information gain value was calculated, and the eight patterns with an information gain exceeding 0.5 were selected as key patterns. A saliency weight was calculated for each key pattern, with the weight value proportional to the pattern's information gain. Combining these weights with the corresponding local map features, a local map feature description F2 with saliency weights was generated.

[0039] To describe the local map features of each scale layer {F1, F2, F3, F4}, a multi-head attention fusion network with 8 attention heads was trained. Each attention head is responsible for capturing feature relevance from different perspectives. In attention calculation, the query matrix, key matrix, and value matrix all have a dimension of 64. The network was trained with a batch size of 32, a learning rate of 0.001, and 200 training epochs.

[0040] After training, the weight coefficient matrix W∈R4×8 of the feature distribution at different scales is obtained, representing the weight distribution of the four scale layers under eight attention heads. Weighted residual connection fusion is then performed, fusing the features at each scale according to the weight coefficients. The residual connections are set to direct addition with a stride of 1 to ensure no information is lost during the fusion process.

[0041] Finally, backpropagation optimization was performed using the cross-entropy loss function and the Adam optimizer with a weight decay coefficient of 0.0001. After optimization, a fusion feature matrix with dimensions of 1000×128 was obtained, where each row corresponds to an original device node, and the 128-dimensional features of each node represent the comprehensive layout feature information at multiple scales.

[0042] In one optional implementation, the fused feature matrix is ​​input into the residual network, and hotspot regions are determined based on the gradient change magnitude of the feature regions to generate a target semantic feature set, including: Perform residual network operations on the fused feature matrix to obtain the initial feature map; Calculate the difference between the directional gradient value and the texture complexity value of each feature region in the initial feature map and the feature regions in the corresponding adaptive neighborhood set, and generate a feature difference evaluation matrix; Perform iterative comparison operation on the feature difference evaluation matrix, take the position with value greater than the first preset value as hot spot feature point, take the position with value less than the second preset value as non-hot spot feature point, keep the feature value of the hot spot feature point position and set the feature value of the non-hot spot feature point position to zero, and obtain the target area feature map. The target region feature map is divided into multiple feature reconstruction blocks. Reconstruction weight coefficients are assigned according to the gradient change magnitude of each feature reconstruction block. Weighted reconstruction is performed according to the reconstruction weight coefficients to generate a target semantic feature set.

[0043] In one specific implementation, in the integrated circuit hotspot detection method based on multimodal feature fusion and dynamic optimization, when performing residual network operations on the fused feature matrix, an improved residual network structure is specifically adopted. This structure contains multiple residual blocks, each consisting of two convolutional layers and a short-circuit connection. After the fused feature matrix is ​​input into the residual network, it first passes through an initial convolutional layer for feature extraction, with a kernel size of 3×3, a stride of 1, and 64 output channels. Subsequently, the feature map passes through five residual blocks sequentially, each containing two convolutional layers, both with a kernel size of 3×3, and output channels of 64, 128, 256, 512, and 512 respectively. Within each residual block, the input feature map is activated by the ReLU function after passing through the first convolutional layer, then through the second convolutional layer, and finally the result is added element-wise with the input feature map to form a short-circuit connection. To accommodate different channel numbers, a 1×1 convolutional layer is added to the short-circuit connection for adjustment. After processing by the residual network, an initial feature map of size W×H×C is obtained, where W and H represent the width and height of the feature map, respectively, and C represents the number of channels.

[0044] When calculating the difference between each feature region in the initial feature map and the corresponding feature regions in the adaptive neighborhood set, for each position (i, j), the adaptive neighborhood set is defined as a K×K region centered at (i, j). The value of K is dynamically adjusted according to the size of the feature map; in practical applications, K is set to 7. For each feature region, its directional gradient value is calculated. The directional gradient value is obtained by taking the square root of the sum of the squares of the horizontal and vertical gradients. The horizontal gradient is calculated by applying the Sobel operator to the feature map, and the vertical gradient also uses the Sobel operator, but in a different direction. The texture complexity value is obtained by calculating the entropy value of the gray-level co-occurrence matrix within the feature region. The gray-level co-occurrence matrix statistically represents the co-occurrence frequency of gray values ​​of adjacent pixels within the feature region. For a feature region at position (i, j), calculate the difference in directional gradient value Δg(i, j, m, n) and texture complexity value Δt(i, j, m, n) between it and each position (m, n) in its neighborhood set. Then, weightedly combine these two difference values ​​to obtain the overall difference value D(i, j, m, n), with weight parameters set to 0.6 and 0.4, respectively. Based on the calculated difference value, construct a feature difference evaluation matrix F, where each element F(i, j) of matrix F represents the value with the largest difference between position (i, j) and its neighborhood set.

[0045] When performing iterative comparison operations on the feature difference evaluation matrix, during the processing, set the first preset value to the mean of the difference matrix plus 1.5 times the standard deviation, and the second preset value to the mean of the difference matrix minus 0.8 times the standard deviation. Specifically, calculate the mean μ and the standard deviation σ of the feature difference evaluation matrix F, and then determine the high threshold T_high = μ + 1.5σ and the low threshold T_low = μ - 0.8σ. For each element F(i, j) in the matrix F, if F(i, j) > T_high, mark the position (i, j) as a hot spot region feature point; if F(i, j) < T_low, mark the position (i, j) as a non-hot spot region feature point; if T_low ≤ F(i, j) ≤ T_high, keep the original value unchanged. After marking, keep the eigenvalue at the position of the hot spot region feature point unchanged, and set the eigenvalue at the position of the non-hot spot region feature point to 0, so as to obtain the target region feature map G. In a practical case, for an initial feature map with a size of 256×256×64, the calculated mean μ is 0.42 and the standard deviation σ is 0.18. Therefore, the high threshold T_high is 0.69 and the low threshold T_low is 0.28. After processing, approximately 17% of the positions are marked as hot spot region feature points, and approximately 32% of the positions are marked as non-hot spot region feature points.

[0046] When dividing the target region feature map into multiple feature reconstruction blocks, specifically, the target region feature map G is divided into N×N blocks of equal size. The value of N is determined according to the size of the feature map, and in practice, it is usually set to 16, i.e., divided into 16×16 blocks. For each block B(p, q), its gradient change magnitude M(p, q) is calculated. The gradient change magnitude is obtained by averaging the directional gradient values ​​of all non-zero elements within the block. Based on the calculated gradient change magnitude, a reconstruction weight coefficient W(p, q) is assigned to each block. The calculation of the weight coefficient considers the ratio of the gradient change magnitude of the block to the global average gradient change magnitude, specifically W(p, q) = M(p, q) / M_avg, where M_avg is the average gradient change magnitude of all blocks. To avoid the weights being too large or too small, W(p, q) is normalized to ensure that its range is between [0.5, 2.0]. Then, each block is reconstructed using weighted reconstruction coefficients. The reconstruction process involves multiplying the value of each element within the block by its corresponding weight coefficient, i.e., G'(i,j) = G(i,j) × W(p,q), where (i,j) is the pixel position within block B(p,q). After weighted reconstruction, a target semantic feature set S is generated. This set contains salient features of hotspot regions while suppressing interference information from non-hotspot regions. In a practical case, for a 16×16 block partition, the calculated average gradient change magnitude M_avg of the blocks is 0.58, with the gradient change magnitude of each block ranging from 0.21 to 1.42. The normalized weight coefficients range from 0.52 to 1.85. After weighted reconstruction, the feature values ​​of hotspot regions are enhanced, making these regions more prominent in subsequent detection and analysis.

[0047] By following the steps above, hotspot areas in integrated circuit design can be effectively identified, improving the accuracy and efficiency of hotspot detection.

[0048] In one optional implementation, calculating the difference between the directional gradient value and texture complexity value of each feature region in the initial feature map and the corresponding feature regions in the adaptive neighborhood set, and generating a feature difference evaluation matrix includes: Device connection information is extracted from the initial feature map, an adaptive neighborhood set for each feature region is constructed, and the number of feature regions in the adaptive neighborhood set is determined based on the topological and electrical similarity between feature regions. Perform multi-directional rotation filtering on the feature regions in the initial feature map to obtain multi-scale directional response vectors, and determine the directional gradient value based on the directional response amplitude; Calculate the local spectral entropy and local intrinsic dimension of the feature region to generate a texture complexity value, and construct a multi-scale feature vector by combining the directional gradient value and the texture complexity value; Calculate the feature weights of each scale component of the multi-scale feature vector to generate a comprehensive feature vector of the feature region. Calculate the local covariance matrix between the feature region and the feature region in the adaptive neighborhood set, and calculate the Mahalanobis distance difference metric of the comprehensive feature vector based on the local covariance matrix; The confidence level is calculated based on the distribution of Mahalanobis distance difference metric values. The Mahalanobis distance difference metric values ​​and the confidence level are combined to generate a feature difference evaluation matrix. Each value of the feature difference evaluation matrix corresponds to the gradient change magnitude of the feature region at the same position in the initial feature map.

[0049] In one specific implementation, when extracting device connection information from the initial feature map, a graph-based connection extraction method is used. Each feature region in the initial feature map is considered a node in the graph, and the connections between feature regions are considered edges. Specifically, an edge detection algorithm is used to extract device boundaries in the initial feature map. An improved Canny algorithm is used for edge detection, with a low threshold of 0.1, a high threshold of 0.3, and a standard deviation of 1.5. After obtaining the edge information, morphological operations such as erosion and dilation are used to extract device contours. The erosion operation uses a 3×3 structuring element and iterates twice; the dilation operation uses a 5×5 structuring element and iterates once. Based on the extracted contours, a region growing algorithm is used to identify different device regions, and a connection matrix is ​​established to record the connections between devices. For each feature region in the initial feature map, when constructing an adaptive neighborhood set, the number of feature regions in the neighborhood set is determined based on topological similarity and electrical similarity. Topological similarity is calculated based on the connectivity of the devices to which the feature region belongs; the connectivity is the number of other regions directly connected to that region. Electrical similarity is calculated based on the electrical characteristics of the feature region, such as resistance and capacitance. In practical applications, for regions with high connectivity and similar electrical properties, the number of feature regions in the adaptive neighborhood set is relatively large, typically 15-25; while for regions with low connectivity and large differences in electrical properties, the number of feature regions is relatively small, typically 5-10.

[0050] When performing multi-directional rotation filtering on the feature regions in the initial feature map, a Gabor filter bank is used to extract texture features in different directions. The Gabor filter bank contains filters in eight different directions with angles of 0°, 22.5°, 45°, 67.5°, 90°, 112.5°, 135°, and 157.5°. Each direction uses three filters at three different scales with scale parameters of 2, 4, and 8 pixels, and wavelength parameters of 4, 8, and 16 pixels, respectively. For each feature region in the feature map, these 24 Gabor filters (8 directions × 3 scales) are applied, resulting in 24 response values, forming a multi-scale directional response vector. Based on the directional response vector, the response amplitude for each direction is calculated. The response amplitude is the absolute value of the filter response for that direction. Then, the direction with the largest response amplitude is selected as the dominant direction for that feature region; the response amplitude in that direction is the directional gradient value.

[0051] When calculating the texture complexity value by calculating the local spectral entropy and local intrinsic dimension of a feature region, the local spectral entropy of the feature region is calculated. This is done by performing a two-dimensional discrete Fourier transform on the feature region to calculate the probability distribution of the spectrum, and then calculating the entropy value based on this distribution. Specifically, for each feature region, an 11×11 local window is extracted, and a two-dimensional discrete Fourier transform is performed on this window to obtain a spectrum. The spectrum is then normalized to obtain a spectral probability distribution, and the entropy value of this distribution is calculated as the local spectral entropy. The local intrinsic dimension is calculated using the box counting method. The box counting method covers the feature region with grids of different scales and statistically analyzes the relationship between the number of non-empty grids and the scale, thereby estimating the fractal dimension of the region. Using five different grid scales (1, 2, 3, 4, and 5 pixels), the number of non-empty grids at each scale is calculated, and the local intrinsic dimension is obtained through regression analysis. The texture complexity value is obtained by weighting the local spectral entropy and the local intrinsic dimension, with weights of 0.7 and 0.3, respectively. Subsequently, the directional gradient values ​​and texture complexity values ​​are combined to construct a multi-scale feature vector, which contains directional gradient values ​​and texture complexity values ​​at different scales.

[0052] When calculating the feature weights of each scale component in a multi-scale feature vector, a weight allocation method based on information entropy is adopted. For each scale component in the multi-scale feature vector, its distribution entropy in the entire feature map is calculated. The higher the distribution entropy, the richer the information contained in that scale component, and the higher the weight should be assigned. Specifically, the value range of each scale component is divided into 10 equally wide intervals, and the frequency of the scale component value falling into each interval in each feature region of the feature map is statistically analyzed, and the entropy value of this frequency distribution is calculated. Then, the entropy values ​​of each scale component are normalized to obtain the feature weights of each scale component. In practical applications, for feature vectors with three scales, the calculated weights are 0.28, 0.42, and 0.30, respectively. Based on these weights, the components of the multi-scale feature vector are weighted and combined to generate a comprehensive feature vector.

[0053] When calculating the local covariance matrix between the feature region and the feature regions in the adaptive neighborhood set, for each feature region in the feature map, its composite eigenvector and the composite eigenvector of all feature regions in the adaptive neighborhood set are extracted, and the covariance matrix of the vectors is calculated. The calculation of the covariance matrix includes: calculating the mean vector of the composite eigenvectors; calculating the outer product of the differences between each eigenvector and the mean vector; and averaging all the outer product results to obtain the covariance matrix. After obtaining the local covariance matrix, the Mahalanobis distance difference metric between the feature region and each feature region in the adaptive neighborhood set is calculated based on the local covariance matrix. The Mahalanobis distance considers the correlation between the components of the eigenvectors and can more accurately measure the differences between feature regions. For the feature region and the feature regions in the adaptive neighborhood set, the difference vector of their composite eigenvectors is calculated, multiplied by the inverse of the local covariance matrix, and then multiplied by the transpose of the difference vector to obtain the Mahalanobis distance difference metric.

[0054] When calculating confidence scores based on the distribution of Mahalanobis distance dissimilarity metrics, for each feature region in the feature map, the mean and standard deviation of its Mahalanobis distance dissimilarity metrics compared to all feature regions in the adaptive neighborhood set are calculated. Then, based on the normal distribution assumption, the confidence score corresponding to each Mahalanobis distance dissimilarity metric is calculated. The confidence score is calculated using the z-score method: the mean of the Mahalanobis distance dissimilarity metric is subtracted, and then divided by the standard deviation to obtain the z-value. The corresponding confidence score is then obtained by looking up a table or calculating it. For example, for a z-value of 2.0, the corresponding confidence score is approximately 0.9544; for a z-value of 1.5, the corresponding confidence score is approximately 0.8664. The Mahalanobis distance dissimilarity metric and the confidence score are combined to generate a feature dissimilarity evaluation matrix, which is generated by multiplying the Mahalanobis distance dissimilarity metric by the confidence score. Each value in the feature dissimilarity evaluation matrix corresponds to the gradient change magnitude of feature regions at the same location in the initial feature map. The larger the value, the more significant the difference between the feature region and the feature regions in its adaptive neighborhood set, and the more likely it is to be a hotspot region.

[0055] In one optional implementation, a hotspot density distribution map is constructed based on the original hotspot detection data, and the correlation strength matrix between hotspot regions is calculated using spatial autocorrelation analysis, including: Normalization is performed on the raw hotspot detection data to obtain standard values ​​of hotspot data. These standard values ​​are then used to construct a hotspot location matrix, and spatial coordinate information and hotspot attribute information are extracted. Substitute spatial coordinate information and hotspot attribute information into the kernel density function to obtain the initial density estimate. Set the bandwidth adjustment parameter based on the initial density estimate and calculate the kernel density value for the preset grid points to obtain the hotspot density distribution map. Based on the hotspot density distribution map, the spatial autocorrelation coefficient is calculated to obtain the proximity relationship matrix. The spatial weight function is constructed using the proximity relationship matrix to calculate the correlation degree of hotspot areas. Hotspot areas that meet the preset correlation threshold are grouped into a hotspot association set. Based on the hotspot association set, a spatial network topology is constructed, node connection relationships are extracted to obtain an association path set, the spatial distance between adjacent nodes in the association path set is calculated to determine the distance matrix, and the distance matrix and the association degree of the hotspot area are combined to generate an association strength matrix.

[0056] In one specific implementation, when performing normalization on the raw hotspot detection data to obtain standard values ​​for the hotspot data, a maximum-minimum normalization method is used to process the raw hotspot detection data. The raw hotspot detection data includes hotspot location coordinates and hotspot intensity values. The hotspot location coordinates describe the location of the hotspot on the chip in micrometers, and the hotspot intensity value represents the severity of the hotspot. Normalization maps the hotspot intensity values ​​to a range of 0 to 1. The calculation method is to subtract the minimum value of all hotspot intensity values ​​from each hotspot intensity value, and then divide by the difference between the maximum and minimum values. For example, for a set of raw hotspot intensity values ​​{28.5, 42.7, 15.6, 89.3, 37.2}, the minimum value is 15.6 and the maximum value is 89.3. After normalization, the values ​​are {0.175, 0.367, 0, 1, 0.292}. The hotspot location coordinates are mapped to a normalized grid with a grid size of 512×512, covering the entire chip area. The normalized hotspot data is used to construct a hotspot location matrix. Each element in the matrix corresponds to a location in the grid, and its value is the hotspot intensity at that location. If there is no hotspot at that location, the value is 0. Spatial coordinate information and hotspot attribute information are extracted from the hotspot location matrix. The spatial coordinate information includes the row and column indices of the hotspot in the normalized grid, and the hotspot attribute information includes the hotspot intensity value, hotspot area, and hotspot shape characteristics.

[0057] When substituting spatial coordinate information and hotspot attribute information into the kernel density function to obtain the initial density estimate, a Gaussian kernel function is used for kernel density estimation. The Gaussian kernel function assigns weights based on distance decay, centered on the hotspot location; the decay rate of the kernel function is controlled by the bandwidth parameter. The initial bandwidth parameter is set to 5% of the grid side length, i.e., 25.6 grid units. For each hotspot, the contributed density value is determined by the product of the hotspot intensity value and the Gaussian kernel function value. The density contributions of all hotspots are superimposed to obtain the initial kernel density estimate for the entire grid. A bandwidth adjustment parameter is set based on the initial density estimate, which is based on the coefficient of variation of the initial density distribution. The coefficient of variation is the standard deviation of the initial density estimate divided by the mean. When the coefficient of variation is greater than 1.5, the bandwidth adjustment parameter is set to 0.8; when the coefficient of variation is between 0.8 and 1.5, the bandwidth adjustment parameter is set to 1.0; when the coefficient of variation is less than 0.8, the bandwidth adjustment parameter is set to 1.2. The final bandwidth is the initial bandwidth multiplied by the bandwidth adjustment parameter. For example, the coefficient of variation for the initial density estimate is 1.28, the bandwidth adjustment parameter is set to 1.0, and the final bandwidth is maintained at 25.6 grid units. When calculating the kernel density value for the preset grid points, the entire 512×512 grid is divided into smaller cells, each cell being 4×4 grid points, resulting in a total of 128×128 cells. The kernel density value is calculated for the center point of each cell, and the kernel density value is obtained by summing the density values ​​contributed by all hotspots according to the Gaussian kernel function. The calculated kernel density values ​​are used to construct a hotspot density distribution map, which is represented in pseudo-color, with red indicating high-density areas and blue indicating low-density areas.

[0058] When calculating the spatial autocorrelation coefficient to obtain the proximity matrix based on the hotspot density distribution map, the Moran's index method is used to measure spatial autocorrelation. Specifically, for each cell in the hotspot density distribution map, the correlation between its density value and that of surrounding cells is calculated. The average density value of all cells is calculated; the difference between each cell's density value and the average density value is calculated; the product of the density differences between cell pairs is calculated; weights are determined based on the distance between cells; the sum of the weighted products of density differences between cell pairs is calculated; and finally, the Moran's index is obtained by dividing by the density variance. The Moran's index value is between -1 and 1, with values ​​close to 1 indicating a high positive correlation, close to -1 indicating a high negative correlation, and close to 0 indicating a random distribution. For example, a calculated Moran's index of 0.72 indicates strong spatial clustering of the hotspot distribution. Based on the Moran's index calculation results, a proximity matrix is ​​constructed, with a size equal to the number of cells multiplied by the number of cells. The matrix element values ​​represent the spatial correlation between corresponding cell pairs. If the Moran's local index of two cells is greater than 0.5, they are considered to be adjacent, and the corresponding matrix element value is set to 1; otherwise, it is set to 0.

[0059] A spatial weighting function is constructed using a proximity matrix to calculate the correlation between hotspot regions. The spatial weights between hotspot regions are defined based on the proximity matrix. The spatial weighting function considers the distance decay effect, with the weight value decreasing as distance increases. The weighting function uses an exponential decay form, where the weight value equals the negative distance to base 2 divided by a power of the feature distance, which is set to 30 grid units. Based on the spatial weighting function, the correlation between each hotspot region and other hotspot regions is calculated. The correlation calculation considers three factors: hotspot strength, distance, and spatial weight. The correlation value equals the product of the strengths of two hotspot regions multiplied by their spatial weight. For each hotspot region, the correlation with all other hotspot regions is calculated, and the correlation values ​​are sorted. A preset correlation threshold is set to 30% of the maximum correlation value. Hotspot regions with correlation values ​​greater than this threshold are grouped into a hotspot correlation set. For example, if the maximum correlation value between a hotspot region and other regions is 0.85, and the preset correlation threshold is 0.255, then 8 hotspot regions have correlation values ​​exceeding the threshold with this region. These 9 regions together form a hotspot correlation set.

[0060] When constructing a spatial network topology based on hotspot association sets, each hotspot region is considered a node in the network, and connections are established between hotspot regions with association relationships. The spatial network topology is represented as an undirected graph, where nodes are hotspot regions and edges represent the association relationships between them. For each pair of hotspot regions in the hotspot association set, if their association degree exceeds a preset association threshold, an edge is established between them. The node connection relationships are extracted from the constructed spatial network topology to obtain an association path set. The association path set contains all possible paths in the network, and each path consists of a series of connected nodes. Specifically, for a hotspot association set containing 9 nodes, the constructed spatial network topology contains 14 edges, and the extracted association path set contains 23 different paths. When calculating the spatial distance matrix between adjacent nodes in the association path set, for each pair of adjacent nodes in the association path set, their Euclidean distance in the normalized grid is calculated. The size of the distance matrix is ​​the number of nodes multiplied by the number of nodes, and the matrix element values ​​are the distances between corresponding node pairs. If two nodes are not adjacent, the corresponding matrix element value is set to infinity. The association strength matrix is ​​generated by combining the distance matrix and the association degree of hotspot regions. The calculation method is to divide the association degree by the square of the distance and then multiply by a scaling factor of 10000 to ensure the result value is within the range of 0 to 1. The association strength matrix reflects the strength of the association between hotspot regions; a larger value indicates a stronger association. In the example, for a hotspot association set consisting of 9 nodes, the calculated association strength matrix shows a maximum value of 0.92, a minimum value of 0.05, and an average value of 0.37.

[0061] The above method can effectively identify hotspot distribution patterns in integrated circuits and reveal the spatial correlation structure between hotspot regions, providing a basis for subsequent hotspot prediction and layout optimization. This method not only considers the spatial distribution characteristics of hotspots but also combines hotspot attribute information for comprehensive analysis, enabling it to more accurately capture the complex relationships between hotspot regions.

[0062] In one optional implementation, performing a region expansion operation based on the correlation strength matrix and the hotspot density distribution map to generate a candidate hotspot region set includes: The correlation strength matrix is ​​input into the spatial distance attenuation function to obtain the attenuation coefficient matrix. Based on the attenuation coefficient matrix, the region growth constraint condition is constructed. Based on the region growth constraint condition and the hotspot density distribution map, the initial expansion boundary of the hotspot region is determined. A partitioning operation is performed on the initial expansion boundary to obtain the core expansion region and the transition expansion region. The density gradient value of the core expansion region is calculated using the hotspot density distribution map to delineate the expansion range of the core region. A buffer zone is constructed around the expansion range of the core region to obtain the expansion range of the transition region. Based on the spatial distance decay function, the expansion thresholds of the core expansion region and the transition expansion region are calculated. The expansion thresholds are substituted into the adaptive weight function to obtain the boundary adjustment parameters. The optimization operation is performed on the expansion region according to the boundary adjustment parameters to obtain the optimized boundary. The overlapping regions in the optimization boundary are detected, and the overlapping regions are merged according to the correlation strength matrix. The spatial continuity index and regional integrity index of the merged region are calculated, and the regions that meet the preset index thresholds are formed into a candidate hotspot region set.

[0063] In one specific implementation, when inputting the association strength matrix into the spatial distance attenuation function to obtain the attenuation coefficient matrix, an exponential spatial distance attenuation function is used. The exponential spatial distance attenuation function defines the law that association strength decreases with increasing distance, and its functional form is an exponential negative attenuation form. Specifically, for each element in the association strength matrix, the standardized distance between the corresponding two hotspot regions is calculated. The standardized distance is the maximum value of the actual distance divided by the grid size. The standardized distance is substituted into the spatial distance attenuation function to calculate the attenuation coefficient. The formula for calculating the attenuation coefficient is the negative standardized distance of e multiplied by the power of the attenuation control parameter, which is set to 2.5. For example, for a pair of hotspot regions with a standardized distance of 0.3, the calculated attenuation coefficient is 0.47; for a pair of hotspot regions with a standardized distance of 0.8, the attenuation coefficient is 0.14. The attenuation coefficients of all hotspot region pairs are combined into an attenuation coefficient matrix, with the same size as the association strength matrix. Region growth constraints are constructed based on the attenuation coefficient matrix, including the maximum expansion radius and direction weights. The maximum expansion radius is determined by the attenuation coefficient and is calculated by multiplying the base expansion radius by the attenuation coefficient; the base expansion radius is set to 50 grid units. The directional weights are determined based on the relative positions between hotspot regions. The weights along the connection direction of the hotspot region are higher, while the weights perpendicular to the connection direction are lower.

[0064] When determining the initial expansion boundary of hotspot regions based on region growth constraints and hotspot density distribution maps, a region growth algorithm is employed. Starting from the center point of each hotspot region, expansion proceeds outward according to the region growth constraints. During expansion, it is determined whether the hotspot density value of the point to be added exceeds a threshold, which is set at 30% of the center point's hotspot density value. If it exceeds the threshold, it is added to the expansion region; otherwise, expansion in that direction stops. Simultaneously, it is ensured that the expansion does not exceed the maximum expansion radius limit. This process is repeated in all directions until the expansion in all directions reaches the stopping condition, thus obtaining the initial expansion boundary of the hotspot region.

[0065] When performing partitioning operations on the initial expansion boundary to obtain the core expansion region and the transition expansion region, region segmentation is performed based on hotspot density gradients. The hotspot density gradient is calculated for each point within the initial expansion boundary; the gradient value reflects the rate of change of hotspot density near that point. The Sobel operator is used for gradient calculation, calculating the horizontal and vertical gradients within a 3×3 neighborhood, and calculating the gradient magnitude. Based on the distribution of gradient magnitudes, an adaptive thresholding method is used to determine the segmentation threshold, which is set to the average gradient magnitude plus 0.5 times the standard deviation. Regions with gradient magnitudes below the threshold are classified as core expansion regions, and regions with gradient magnitudes above the threshold are classified as transition expansion regions. The density gradient value of the core expansion region is calculated using the hotspot density distribution map; the density gradient value reflects the spatial trend of hotspot density, calculated as the difference between the highest and lowest density values ​​within the region divided by the distance between them. The expansion range of the core region is determined based on the density gradient values. The expansion range is controlled by an expansion factor, which is inversely proportional to the density gradient value; the larger the density gradient value, the smaller the expansion factor. The expansion factor is calculated by dividing the base expansion factor by the square root of the density gradient value; the base expansion factor is set to 2.0.

[0066] When constructing a buffer zone around the core area to obtain the transition area, the core area's expansion boundary is extended outward by a certain distance to form the buffer zone. The width of the buffer zone is determined by the size and shape of the core area, calculated as 30% of the core area's equivalent radius, which is the radius of a circle with the same area as the core area. The union of the buffer zone and the original transition area forms the transition area's expansion range. Expansion thresholds for both the core and transition areas are calculated based on a spatial distance attenuation function. These thresholds are determined by the average hotspot density and the distance attenuation coefficient of the area. The expansion threshold for the core area is calculated as the average hotspot density multiplied by 0.8, while the expansion threshold for the transition area is calculated as the average hotspot density multiplied by 0.5 multiplied by the distance attenuation coefficient.

[0067] When substituting the extended threshold into the adaptive weighting function to obtain the boundary adjustment parameters, a sigmoid adaptive weighting function is used. This function maps the extended threshold to the boundary adjustment parameters, which control the movement distance and direction of the boundary points. The adaptive weighting function is in the form of a sigmoid function, taking the difference between the extended threshold and the reference threshold as input and outputting the adjustment parameters between 0 and 1. The reference threshold is set to the average global hotspot density, which is 0.35 in this example. Optimization operations are performed on the extended region based on the boundary adjustment parameters to obtain the optimized boundary. The optimization process includes two steps: boundary smoothing and convex hull generation. Boundary smoothing is achieved by moving the boundary points; the movement distance is controlled by the boundary adjustment parameters, and the movement direction is determined by the local curvature—concave regions move outwards, and convex regions move inwards to reduce boundary irregularities. Convex hull generation is based on the optimized set of boundary points, using the Graham scan algorithm to generate the convex hull, which serves as the final optimized boundary.

[0068] When detecting overlapping regions in the optimized boundaries, the optimized boundaries of all hotspot regions are compared pairwise, and their intersection area is calculated. If the proportion of the intersection area of ​​two regions to the area of ​​the smaller region exceeds a preset overlap threshold, the two regions are considered to overlap. The overlap threshold is set to 30%. Overlapping regions are merged based on the association strength matrix, with the merging decision based on association strength and overlap degree. If the association strength between two overlapping regions is greater than the association threshold (set to 0.4), they are merged into one region; otherwise, the region with the smaller overlap is retained, and the boundary of the other region is reduced to eliminate the overlap. The merging operation is achieved by calculating the union of the two regions, and the boundary of the new region is the convex hull of the union of the original two region boundaries. The spatial continuity index and the region integrity index of the merged region are calculated. The spatial continuity index reflects the connectivity within the region and is calculated by dividing the average length of the shortest path between any two points within the region by the equivalent diameter of the region; a value closer to 1 indicates better continuity. The region integrity index reflects the regularity of the region boundary and is calculated by dividing the region area by its convex hull area; a value closer to 1 indicates higher integrity.

[0069] When forming a candidate hotspot region set by regions that meet preset threshold indicators, the spatial continuity threshold is set to 0.75, and the region integrity threshold is set to 0.8. For each merged region, the spatial continuity and region integrity indices are calculated. If both indices exceed the corresponding thresholds, the region is added to the candidate hotspot region set; otherwise, further boundary optimization is performed on the region until the threshold requirements are met or the maximum number of iterations (set to 5) is reached. In the example, starting from 20 initial hotspot regions, after boundary expansion, optimization, and merging, 15 merged regions are obtained, of which 12 regions meet the threshold requirements and are added to the candidate hotspot region set. The regions in the candidate hotspot region set will serve as the final result of integrated circuit hotspot detection. These regions have high hotspot density, good spatial continuity, and regular boundary shapes, accurately reflecting the potential hotspot distribution in integrated circuits.

[0070] The aforementioned methods for determining and optimizing hotspot region boundaries can effectively improve the accuracy and robustness of hotspot detection. Compared to traditional region growing methods based on fixed thresholds, these methods consider the spatial correlation between hotspot regions and the spatial distribution characteristics of hotspot density, enabling more accurate determination of hotspot region boundaries.

[0071] like Figure 2 As shown, a data flow diagram for adaptive boundary recognition of hotspot regions based on the correlation strength matrix is ​​presented.

[0072] In one optional implementation, geometric and electrical features are extracted from candidate hotspot regions, and iterative optimization is performed based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data, including: Geometric features are extracted from candidate hotspot regions, the trend of boundary curvature change of geometric features is calculated, and Fourier spectral analysis is performed to obtain morphological evolution sequences. Electrical features are extracted from candidate hotspot areas, the power loss distribution of the electrical features is calculated, and wavelet transform is performed to obtain the energy evolution sequence. Align the morphological evolution sequence with the energy evolution sequence in the time-frequency domain, calculate the phase coherence spectrum between the sequences, and extract the frequency component with the largest amplitude in the phase coherence spectrum to construct a feature co-evolution matrix. Based on the feature co-evolution matrix, the synchronization degree between the morphological evolution sequence and the energy evolution sequence is calculated, a feature enhancement function is constructed, and the feature enhancement function is used to iteratively optimize the hotspot region; Calculate the difference between the morphological evolution sequence and the difference between the energy evolution sequence of two adjacent iterations of optimization. When both the difference between the morphological evolution sequence and the difference between the energy evolution sequence are less than a preset threshold, output the optimized hotspot region as the target hotspot detection data.

[0073] In one specific implementation, a boundary contour-based feature extraction method is used when extracting geometric features from a set of candidate hotspot regions. For each hotspot region in the set, a corresponding sequence of boundary contour points is extracted. The boundary contour points are arranged clockwise, and each point contains row and column coordinates in a standardized grid. Based on the boundary contour point sequence, the geometric features of the region are calculated, including area, perimeter, compactness, major axis length, minor axis length, and eccentricity. Compactness is calculated by dividing the square of the perimeter by 4π times the area; a value closer to 1 indicates a shape closer to a circle. Eccentricity is calculated by subtracting 1 from the ratio of the major axis length to the minor axis length; a smaller value indicates a shape closer to a circle.

[0074] When calculating the boundary curvature variation trend of geometric features, the local curvature is calculated for the boundary contour point sequence. The local curvature is calculated using the curvature of the circle determined by three adjacent points. For each point in the boundary contour point sequence, the curvature of that point is calculated as the reciprocal of the radius of the circle determined by that point and its corresponding two adjacent points. To reduce the influence of noise, the calculated curvature sequence is Gaussian smoothed, with a smoothing window size of 5% of the number of boundary points. The smoothed curvature sequence reflects the changing trend of the boundary shape. Fourier spectral analysis is performed on the curvature sequence to obtain the morphological evolution sequence. Specifically, a discrete Fourier transform is performed on the curvature sequence to obtain a frequency domain representation, which includes the amplitude spectrum and phase spectrum. The top 10 frequency components with the largest amplitudes are selected, and their amplitudes and phases are combined to reconstruct the morphological evolution sequence. The morphological evolution sequence reflects the trend of the boundary shape of the hotspot region changing over time or space. In the example, after Fourier transform, the main frequency components of the boundary curvature sequence of a hotspot region are 0.05, 0.12, 0.24, 0.33 and 0.46, respectively, with corresponding amplitudes of 0.72, 0.54, 0.38, 0.25 and 0.17.

[0075] When extracting electrical features from candidate hotspot regions, relevant parameters are calculated based on the electrical models of these hotspot regions. For each hotspot region, its electrical features are extracted, including current density, resistivity, and power density. Current density is estimated based on wiring density and current flow direction within the region; resistivity is calculated based on material properties and temperature distribution; and power density is obtained by multiplying current density and resistivity. When calculating the power loss distribution of the electrical features, the hotspot region is divided into multiple small units, and the power loss value is calculated for each unit. The power loss calculation considers the product of the square of the current density and the resistivity, as well as the effect of temperature on resistivity. The power loss values ​​of all units are combined to form a power loss distribution map. Wavelet transform is performed on the power loss distribution map to obtain an energy evolution sequence. The wavelet transform uses the Daubechies 4 wavelet and performs a 5-level decomposition to obtain wavelet coefficients at different scales. The energy at each scale is calculated based on the wavelet coefficients, using the square root of the sum of the squares of the wavelet coefficients at that scale. The energies at different scales are arranged in ascending order to form an energy evolution sequence. The energy evolution sequence reflects the distribution characteristics of power loss in hotspot regions at different scales. In the example, after wavelet transform, the energy values ​​of power loss distribution in a certain hotspot region at the five scales are 0.85, 0.63, 0.47, 0.32, and 0.21, respectively.

[0076] Aligning morphological evolution sequences with energy evolution sequences in the time-frequency domain requires addressing the issues of different sequence lengths and scale inconsistencies. Interpolation and normalization methods are used to align the two sequences. Specifically, both sequences are normalized to the [0, 1] interval, and then the shorter sequence is interpolated to ensure both sequences have the same length. After alignment, the phase coherence spectrum between the sequences is calculated, which measures the degree of phase synchronization between the two sequences in the time-frequency domain. The phase coherence spectrum is calculated based on continuous wavelet transform. Continuous wavelet transforms are performed on both sequences to obtain time-frequency representations. The cross spectrum between the two time-frequency representations is calculated, and then normalized to obtain the phase coherence spectrum. The value of the phase coherence spectrum is between 0 and 1; the closer the value is to 1, the higher the degree of phase synchronization of the corresponding frequency components. The frequency components with the largest amplitudes in the phase coherence spectrum are extracted; these frequency components correspond to the most synchronized parts of morphological and energy evolution. Based on these frequency components and coherence values, a characteristic co-evolution matrix is ​​constructed. Each element of the feature co-evolution matrix represents the degree of co-change between the morphological evolution sequence and the energy evolution sequence at a specific frequency and time point.

[0077] When calculating the synchronization degree between morphological evolution sequences and energy evolution sequences based on the feature co-evolution matrix, a global synchronization index method is used. The global synchronization index is calculated based on the eigenvalue decomposition of the feature co-evolution matrix, specifically the ratio of the largest eigenvalue to the sum of all eigenvalues. The synchronization degree value ranges from 0 to 1; a value closer to 1 indicates a higher degree of synchronization between the two sequences. When constructing the feature enhancement function, an adaptive enhancement strategy is designed based on the synchronization degree. The feature enhancement function is an exponential function, with eigenvalues ​​as input and enhanced eigenvalues ​​as output. The enhancement degree is controlled by the synchronization degree; a higher synchronization degree results in a greater enhancement degree. Specifically, the calculation involves multiplying the original eigenvalue by a power of e, where the synchronization degree is multiplied by the enhancement factor, and the enhancement factor is set to 2.5. When iteratively optimizing hotspot regions using the feature enhancement function, the boundaries and internal features of the hotspot regions are adjusted. Boundary adjustments are based on the main frequency components in the morphological evolution sequence, while internal feature adjustments are based on the main scale components in the energy evolution sequence. The adjustment process is iterative. After each iteration, the morphological evolution sequence and energy evolution sequence are recalculated, and the optimization effect is evaluated.

[0078] When calculating the difference between the morphological evolution sequence and the energy evolution sequence between two adjacent iterations, the Euclidean distance between the sequences is used to measure the difference. The morphological evolution sequence difference is calculated as the Euclidean distance between the morphological evolution sequence obtained in the current iteration and the morphological evolution sequence obtained in the previous iteration; the energy evolution sequence difference is calculated similarly. When both the morphological evolution sequence difference and the energy evolution sequence difference are less than a preset threshold, the optimization process is considered to have converged, and the optimized hotspot region is output as the target hotspot detection data. The preset threshold for the morphological evolution sequence is set to 0.05, and the preset threshold for the energy evolution sequence is set to 0.08. In practical applications, the optimization process typically converges after 3 to 7 iterations. The final output target hotspot detection data includes the boundary contour, geometric features, and electrical features of the optimized hotspot region, which can be used for subsequent integrated circuit design optimization and hotspot avoidance.

[0079] The method described above, based on the synergistic optimization of geometric and electrical features, effectively improves the accuracy and reliability of hotspot region detection. This method not only considers the morphological characteristics of hotspot regions but also integrates electrical properties for comprehensive analysis, enabling a more accurate description of the essential characteristics and evolutionary patterns of hotspot regions. The integrated circuit hotspot detection system based on multimodal feature fusion and dynamic optimization in this invention includes: The first unit is used to acquire multimodal feature data of integrated circuit design layout and generate layout feature vector groups; The second unit is used to perform progressive block clustering on the map feature vector group to obtain multi-scale feature components, extract local map feature descriptions at different scale layers, and perform feature fusion through a multi-head attention mechanism to generate a fused feature matrix. The third unit is used to input the fused feature matrix into the residual network, determine the hotspot regions based on the gradient change magnitude of the feature regions, and generate the target semantic feature set. The fourth unit is used to construct a hotspot feature map based on the target semantic feature set, determine the spatial distribution characteristics of hotspot regions through clustering methods, and generate original hotspot detection data. The fifth unit is used to construct a hotspot density distribution map based on the original hotspot detection data, and to calculate the correlation strength matrix between hotspot regions using spatial autocorrelation analysis; based on the correlation strength matrix and the hotspot density distribution map, it performs region expansion operations to generate a set of candidate hotspot regions; The sixth unit is used to extract geometric and electrical features from candidate hotspot regions, and to perform iterative optimization based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data. The seventh unit is used to output hotspot location information in the integrated circuit design layout based on the target hotspot detection data.

[0080] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0081] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0082] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hotspot detection method for integrated circuits based on multimodal feature fusion and dynamic optimization, characterized in that, include: Acquire multimodal feature data of integrated circuit design layout and generate layout feature vector sets; Progressive block clustering is performed on the feature vector group of the map to obtain multi-scale feature components. Local map feature descriptions are extracted at different scale layers. Feature fusion is performed through a multi-head attention mechanism to generate a fused feature matrix. The fused feature matrix is ​​input into the residual network, and hotspot regions are determined based on the gradient change magnitude of the feature regions to generate a target semantic feature set. A hotspot feature map is constructed based on the target semantic feature set. The spatial distribution characteristics of hotspot regions are determined by clustering methods to generate original hotspot detection data. A hotspot density distribution map is constructed based on the original hotspot detection data, and the correlation strength matrix between hotspot regions is calculated using spatial autocorrelation analysis. Based on the correlation strength matrix and the hotspot density distribution map, a region expansion operation is performed to generate a set of candidate hotspot regions. Geometric and electrical features are extracted from candidate hotspot regions, and iterative optimization is performed based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data. Based on the target hotspot detection data, output the hotspot location information in the integrated circuit design layout.

2. The method according to claim 1, characterized in that, Progressive block clustering is performed on the feature vector group of the map to obtain multi-scale feature components. Local map feature descriptions are extracted at different scale layers. Feature fusion is performed through a multi-head attention mechanism to generate a fused feature matrix, including: Progressive block clustering is performed based on the map feature vector group. By dynamically adjusting the clustering radius, structural feature components and topological feature components of multiple scale layers are obtained. The structural feature components and the topological feature components are converted into a map feature coefficient matrix through sparse coding. The map feature coefficient matrix is ​​then used to reconstruct the map feature representation of different scale layers. The spatial distribution matrix is ​​calculated based on the layout feature representation, and the feature vector set is obtained by eigenvalue decomposition. After mapping the feature vector set to a high-dimensional feature space using a kernel function, a low-dimensional feature subspace that preserves the topological structure of the layout is obtained by manifold learning. A device connection graph is constructed from the low-dimensional feature subspace, and a similarity matrix between device nodes is calculated. Based on the similarity matrix, the layout feature combination pattern is identified, the information gain value of the layout features is calculated, and a local layout feature description with significant weights is generated. The hierarchical fusion network with multi-head attention mechanism is trained using the local map feature descriptions to obtain the weight coefficients of map features at different scales. The local map feature descriptions with different weight coefficients are fused by residual connections and optimized through backpropagation to obtain the fused feature matrix.

3. The method according to claim 1, characterized in that, The fused feature matrix is ​​input into the residual network, and hotspot regions are determined based on the gradient change magnitude of the feature regions to generate a target semantic feature set, including: Perform residual network operations on the fused feature matrix to obtain the initial feature map; Calculate the difference between the directional gradient value and the texture complexity value of each feature region in the initial feature map and the feature regions in the corresponding adaptive neighborhood set, and generate a feature difference evaluation matrix; Perform iterative comparison operation on the feature difference evaluation matrix, take the position with value greater than the first preset value as hot spot feature point, take the position with value less than the second preset value as non-hot spot feature point, keep the feature value of the hot spot feature point position and set the feature value of the non-hot spot feature point position to zero, and obtain the target area feature map. The target region feature map is divided into multiple feature reconstruction blocks. Reconstruction weight coefficients are assigned according to the gradient change magnitude of each feature reconstruction block. Weighted reconstruction is performed according to the reconstruction weight coefficients to generate a target semantic feature set.

4. The method according to claim 3, characterized in that, Calculate the difference between the directional gradient value and texture complexity value of each feature region in the initial feature map and the corresponding feature regions in the adaptive neighborhood set, and generate a feature difference evaluation matrix including: Device connection information is extracted from the initial feature map, an adaptive neighborhood set for each feature region is constructed, and the number of feature regions in the adaptive neighborhood set is determined based on the topological and electrical similarity between feature regions. Perform multi-directional rotation filtering on the feature regions in the initial feature map to obtain multi-scale directional response vectors, and determine the directional gradient value based on the directional response amplitude; Calculate the local spectral entropy and local intrinsic dimension of the feature region to generate a texture complexity value, and construct a multi-scale feature vector by combining the directional gradient value and the texture complexity value; Calculate the feature weights of each scale component of the multi-scale feature vector to generate a comprehensive feature vector of the feature region. Calculate the local covariance matrix between the feature region and the feature region in the adaptive neighborhood set, and calculate the Mahalanobis distance difference metric of the comprehensive feature vector based on the local covariance matrix; The confidence level is calculated based on the distribution of Mahalanobis distance difference metric values. The Mahalanobis distance difference metric values ​​and the confidence level are combined to generate a feature difference evaluation matrix. Each value of the feature difference evaluation matrix corresponds to the gradient change magnitude of the feature region at the same position in the initial feature map.

5. The method according to claim 1, characterized in that, A hotspot density distribution map is constructed based on the original hotspot detection data, and the correlation strength matrix between hotspot regions is calculated using spatial autocorrelation analysis, including: Normalization is performed on the raw hotspot detection data to obtain standard values ​​of hotspot data. These standard values ​​are then used to construct a hotspot location matrix, and spatial coordinate information and hotspot attribute information are extracted. Substitute spatial coordinate information and hotspot attribute information into the kernel density function to obtain the initial density estimate. Set the bandwidth adjustment parameter based on the initial density estimate and calculate the kernel density value for the preset grid points to obtain the hotspot density distribution map. Based on the hotspot density distribution map, the spatial autocorrelation coefficient is calculated to obtain the proximity relationship matrix. The spatial weight function is constructed using the proximity relationship matrix to calculate the correlation degree of hotspot areas. Hotspot areas that meet the preset correlation threshold are grouped into a hotspot association set. Based on the hotspot association set, a spatial network topology is constructed, node connection relationships are extracted to obtain an association path set, the spatial distance between adjacent nodes in the association path set is calculated to determine the distance matrix, and the distance matrix and the association degree of the hotspot area are combined to generate an association strength matrix.

6. The method according to claim 1, characterized in that, Based on the correlation strength matrix and hotspot density distribution map, a region expansion operation is performed to generate a set of candidate hotspot regions, including: The correlation strength matrix is ​​input into the spatial distance attenuation function to obtain the attenuation coefficient matrix. Based on the attenuation coefficient matrix, the region growth constraint condition is constructed. Based on the region growth constraint condition and the hotspot density distribution map, the initial expansion boundary of the hotspot region is determined. A partitioning operation is performed on the initial expansion boundary to obtain the core expansion region and the transition expansion region. The density gradient value of the core expansion region is calculated using the hotspot density distribution map to delineate the expansion range of the core region. A buffer zone is constructed around the expansion range of the core region to obtain the expansion range of the transition region. Based on the spatial distance decay function, the expansion thresholds of the core expansion region and the transition expansion region are calculated. The expansion thresholds are substituted into the adaptive weight function to obtain the boundary adjustment parameters. The optimization operation is performed on the expansion region according to the boundary adjustment parameters to obtain the optimized boundary. The overlapping regions in the optimization boundary are detected, and the overlapping regions are merged according to the correlation strength matrix. The spatial continuity index and regional integrity index of the merged region are calculated, and the regions that meet the preset index thresholds are formed into a candidate hotspot region set.

7. The method according to claim 1, characterized in that, Geometric and electrical features are extracted from candidate hotspot regions. Iterative optimization is then performed based on the synchronization analysis of morphological and energy evolution sequences to generate target hotspot detection data, including: Geometric features are extracted from candidate hotspot regions, the trend of boundary curvature change of geometric features is calculated, and Fourier spectral analysis is performed to obtain morphological evolution sequences. Electrical features are extracted from candidate hotspot areas, the power loss distribution of the electrical features is calculated, and wavelet transform is performed to obtain the energy evolution sequence. Align the morphological evolution sequence with the energy evolution sequence in the time-frequency domain, calculate the phase coherence spectrum between the sequences, and extract the frequency component with the largest amplitude in the phase coherence spectrum to construct a feature co-evolution matrix. Based on the feature co-evolution matrix, the synchronization degree between the morphological evolution sequence and the energy evolution sequence is calculated, a feature enhancement function is constructed, and the feature enhancement function is used to iteratively optimize the hotspot region; Calculate the difference between the morphological evolution sequence and the difference between the energy evolution sequence of two adjacent iterations of optimization. When both the difference between the morphological evolution sequence and the difference between the energy evolution sequence are less than a preset threshold, output the optimized hotspot region as the target hotspot detection data.

8. An integrated circuit hotspot detection system based on multimodal feature fusion and dynamic optimization, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire multimodal feature data of integrated circuit design layout and generate layout feature vector groups; The second unit is used to perform progressive block clustering on the map feature vector group to obtain multi-scale feature components, extract local map feature descriptions at different scale layers, and perform feature fusion through a multi-head attention mechanism to generate a fused feature matrix. The third unit is used to input the fused feature matrix into the residual network, determine the hotspot regions based on the gradient change magnitude of the feature regions, and generate the target semantic feature set. The fourth unit is used to construct a hotspot feature map based on the target semantic feature set, determine the spatial distribution characteristics of hotspot regions through clustering methods, and generate original hotspot detection data. The fifth unit is used to construct a hotspot density distribution map based on the original hotspot detection data, and to calculate the correlation strength matrix between hotspot regions using spatial autocorrelation analysis; based on the correlation strength matrix and the hotspot density distribution map, it performs region expansion operations to generate a set of candidate hotspot regions; The sixth unit is used to extract geometric and electrical features from candidate hotspot regions, and to perform iterative optimization based on the synchronization analysis of morphological evolution sequences and energy evolution sequences to generate target hotspot detection data. The seventh unit is used to output hotspot location information in the integrated circuit design layout based on the target hotspot detection data.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images