Deep Learning-Based Hotspot Detection Method for Integrated Circuit Layout
By integrating deep convolutional semantic features with layout geometric spacing features into a hybrid feature extraction network, combined with optimal transport mapping and rule checking, the problems of insufficient representation of small size deviations and weak cross-process transfer capability in traditional methods are solved, realizing efficient and accurate detection and rapid adaptation of hot spots in integrated circuit layout.
Patent Information
- Application Number
- CN202611189356.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-06
- Publication Date
- 2026-11-10
AI Technical Summary
Traditional integrated circuit layout hotspot detection methods lack consideration for the unique geometric features of the layout, such as spacing and linewidth, resulting in insufficient ability to characterize minute dimensional deviations. Furthermore, they have weak cross-process node transfer capabilities, high sample collection and labeling costs, and difficulty in ensuring model accuracy.
A hybrid feature extraction network is adopted to fuse deep convolutional semantic features and layout geometric spacing features to construct an optimal transport mapping network. A perturbation feature set is generated by perturbing the principal components in the neighborhood. Combined with a pre-defined rule checking engine and a new process adaptation network, cross-process migration adaptation and pixel-level hotspot localization are achieved.
It significantly improves the distinguishability and representation accuracy of layout features, realizes efficient and accurate detection of hot spots in integrated circuit layouts, adapts to rapid iteration of new process nodes, and provides pixel-level accurate positioning of hot spot areas.
Smart Images

Figure CN123070032A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of integrated circuit operation and maintenance technology, and more specifically, to a method for detecting hotspots in integrated circuit layouts based on deep learning. Background Technology
[0002] With the development of integrated circuit process node technology, layout hotspot detection of integrated circuit layout has become an important process affecting the operation and maintenance of integrated circuits. However, traditional layout hotspot detection methods still have certain limitations.
[0003] On the one hand, traditional methods have relatively simple feature representation dimensions, mostly relying on image-like convolutional features, often lacking consideration of geometric explicit features such as spacing and line width unique to the layout, resulting in insufficient representation ability for small size deviations and easy to miss detections; on the other hand, they have weak cross-process node transfer ability, requiring a large number of labeled samples to retrain the model for each new process node, and the cost of sample collection and labeling is extremely high, making it difficult to guarantee model accuracy in scenarios with few samples. Summary of the Invention
[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for detecting hotspots in integrated circuit layouts based on deep learning, the method comprising: S1: Collect existing process normal layout blocks and inject random perturbations to generate an enhanced positive sample set. Extract the normal feature set and the feature set to be tested based on the hybrid feature extraction network, and save the multi-scale intermediate feature map. Simultaneously perform feature quality pre-check on the normal feature set. S2: Divide the normal feature set into a source feature set and a target feature set, perform neighborhood principal component perturbation on the source feature set to generate a perturbed source feature set, and simultaneously combine it with the target feature set to obtain the optimal transport mapping network, and generate the transport cost threshold in combination with a preset mechanism. S3: Cluster the normal feature set to generate a compressed reference set, quantify the transport cost of the features to be tested and perform a first-level screening based on the optimal transport mapping network, and perform a second-level verification in combination with the preset rule checking engine to generate the first hot spot detection result; S4: Extract the new process feature set of the normal layout of the new process based on the hybrid feature extraction network, and perform cross-process migration adaptation on the optimal transport mapping network to generate a new process adaptation network. Perform hot spot detection based on the new process adaptation network to generate a second hot spot detection result. S5: Obtain the nearest neighbor normal features corresponding to the first and second hotspot detection results, perform differential backpropagation based on the optimal transport mapping network and the new process adaptation network to generate feature attribution weights, and combine multi-scale intermediate feature maps to generate hotspot location maps.
[0005] Compared with the prior art, the present invention has the following beneficial effects: Step S1 uses a hybrid feature extraction network to fuse deep convolutional semantic features and map geometric spacing features, which makes up for the deficiency of single image features in representing geometric size deviations, thereby significantly improving the discriminativeness and representation accuracy of map features. Step S2 constructs a mapping distribution model of the normal feature space based on the optimal transport theory, and combines it with neighborhood principal component perturbation to achieve high-precision fitting of the normal map feature manifold. Step S3 significantly improves the efficiency of initial screening by constructing a compressed reference set, and simultaneously combines the preset rule checking engine to exclude violations with clear rules, while retaining suspicious samples for manual review, thus achieving a balance between detection efficiency and accuracy. Step S4 generates a new process adaptation network by performing cross-process migration adaptation on the optimal transport mapping network, thereby achieving rapid iterative adaptation to new process nodes. Step S5 then maps the hotspot detection results at the feature level back to the pixel space of the input layout, achieving precise pixel-level hotspot location. This provides direct location guidance for subsequent layout correction and optimization, ultimately achieving efficient and accurate detection of hotspots in the integrated circuit layout. Attached Figure Description
[0006] Figure 1 This is a flowchart of the steps of the integrated circuit layout hotspot detection method based on deep learning in this invention; Figure 2 This is a schematic diagram of the hybrid feature extraction network in the deep learning-based integrated circuit layout hotspot detection method of the present invention. Detailed Implementation
[0007] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating the steps of the deep learning-based integrated circuit layout hotspot detection method of the present invention. Figure 2 This is a schematic diagram of the hybrid feature extraction network in the deep learning-based integrated circuit layout hotspot detection method of the present invention. The following is a detailed introduction to the deep learning-based integrated circuit layout hotspot detection method.
[0008] Step S1: Collect existing process normal layout blocks and inject random perturbations to generate an enhanced positive sample set. Extract normal feature sets and test feature sets based on a hybrid feature extraction network, and save multi-scale intermediate feature maps. Simultaneously perform feature quality pre-check on the normal feature set.
[0009] Specifically, step S1 includes steps S11-S13: Step S11: Collect existing normal process layout blocks and inject random perturbations to generate an enhanced positive sample set. Specifically: First, normal layout blocks are collected from the mass production layouts of existing process nodes that have passed tape-out verification. All normal layout blocks must pass complete design rule checks and photolithography simulation verification to confirm that there are no process hotspot risks, thus forming the original positive sample set.
[0010] It should be noted that a stratified sampling strategy is adopted during the sample collection process to cover different layout graphic types. The layout graphic types include at least a combination of graphic types of typical layout layers such as metal lines, vias, contact holes, diffusion regions, and polysilicon gates. At the same time, it covers graphic structures with different density regions and different orientations to ensure the diversity and representativeness of the samples and avoid distribution bias caused by a single sample.
[0011] Subsequently, based on the design rule documents of the corresponding process nodes, key geometric constraint parameters such as line tolerance, spacing tolerance, and contact hole position tolerance are extracted. These parameters are used as boundary conditions for the perturbation amplitude. Random perturbations are applied to each layout block in the original positive sample set. The random perturbations include at least three types of operations: applying polygon edge scaling, chamfering, and via center displacement to each layout block in the positive sample set. The amplitude of all perturbations is strictly controlled within the tolerance range allowed by the corresponding design rules to ensure that the layout blocks generated by the perturbations still meet the design rule requirements.
[0012] Finally, all the layout tiles generated through perturbation are merged with the original positive sample set to form an enhanced positive sample set.
[0013] In one possible embodiment, taking the metal layer layout of a 28nm process node as an example, assume that the original positive sample set contains 12,000 normal layout blocks with a size of 10 micrometers × 10 micrometers, each normal layout block corresponding to a 256-pixel × 256-pixel binarized layout image, and the original normal samples cover typical graphic scenarios such as parallel metal lines, via arrays, line end structures, and corner graphics, with the sample proportions of low-density, medium-density, and high-density regions being 30%, 40%, and 30%, respectively; extract key geometric constraint parameters such as the minimum linewidth and allowable linewidth fluctuation tolerance, minimum spacing and spacing tolerance, and contact hole center position tolerance, as well as the corresponding rule tolerances, from the design rules corresponding to this process node; for each original layout block, randomly select 1 to 3 perturbation type combinations, and randomly select the perturbation amplitude within the rule tolerances. For example, when performing polygon edge scaling, the selected polygon edge can be scaled inward or outward along the normal direction by 1n. Random values within the range of m to 2.8nm are scaled and the line width and spacing values are checked point by point to ensure that the line width and spacing at all positions are not less than the minimum rule requirements. When performing chamfering, isosceles right angles with side lengths of 2nm to 5nm are cut off from the right-angle vertices of the polygon. After chamfering, the interior angles of the graphic still meet the minimum angle constraints of the design rules to avoid acute angle violations. When performing via center displacement, the center of the via graphic is shifted by a random distance within the range of 1nm to 5nm along the x or y direction. After the shift, the coverage margin between the via and the underlying metal layer is calculated to ensure that the margin still meets the coverage requirements of the design rules. Each original normal layout block generates multiple augmented samples with different perturbations. Finally, the original positive samples and augmented samples are merged to form an enhanced positive sample set, thereby effectively expanding the distribution boundary of the normal graphic and improving the adaptability of the subsequent model to process fluctuations (this example only illustrates the operation of random perturbations; specific parameters need to be set according to actual needs).
[0014] Step S12: Extract the normal feature set and the feature set to be tested based on the hybrid feature extraction network, and save the multi-scale intermediate feature maps. Specifically: All plots in the enhanced positive sample set are sequentially input into the hybrid feature extraction network, and the corresponding hybrid feature vectors are output block by block. The hybrid feature vectors corresponding to all plots constitute the normal feature set. While extracting features, the intermediate feature maps output by the deep residual network at the four residual stages are saved simultaneously. Each plot corresponds to a set of multi-scale intermediate feature maps, which are then stored one-to-one with the normal feature set.
[0015] Simultaneously, the test layout is sliced into blocks according to a preset size and overlap step to generate a set of test layout blocks. Then, each block is input into the hybrid feature extraction network to extract hybrid features. The hybrid feature vectors corresponding to all test layout blocks together constitute the test feature set. Similarly, the multi-scale intermediate feature map corresponding to each test layout block is saved simultaneously.
[0016] It should be noted that the hybrid feature extraction network comprises two parallel branches: a deep residual network and an explicit spacing encoder. The deep residual network is used to extract deep semantic features of the layout image, while the explicit spacing encoder is used to extract local geometric spacing features of the layout. Finally, the two branches are concatenated to form a hybrid feature vector that integrates the two types of information. Specifically: First, the deep residual network branch uses the classic ResNet architecture as its basic backbone, truncating the network before the global average pooling layer to directly output a high-dimensional convolutional feature map. This convolutional feature map is the first feature map, which is used to preserve the spatial location information of the features.
[0017] Next, the explicit spacing encoder is used to extract the geometric features of the layout. The specific process is as follows: Uniform grid point sampling is performed within the map block area, and multiple equidistant sector areas are divided with each grid point as the center; Calculate the perpendicular distance from the grid point to the nearest polygon edge within each sector, and combine the distance values of all sectors to form the local distance vector of the grid point; The local distance vectors of all grid points are flattened into one-dimensional vectors and encoded and mapped through one or more fully connected layers to output a fixed-length spacing feature vector. The spacing feature vector is used to characterize the local geometric properties of the layout, such as density and spacing distribution.
[0018] Finally, the hybrid feature extraction network performs global average pooling on the convolutional feature map output by the deep residual network, reducing it to a one-dimensional global convolutional feature vector. This vector is then concatenated with the spacing feature vector output by the explicit spacing encoder along the feature dimension to generate the final hybrid feature vector.
[0019] In the pre-training stage, the hybrid feature extraction network employs a self-supervised task of masked image reconstruction, randomly masking a portion of pixel regions in the layout image within the training set. The training set uses historical normal integrated circuit layouts, and features are annotated by domain experts. The hybrid feature extraction network extracts corresponding features from the layout image through an encoder and then connects to a symmetrical reconstruction decoder, aiming to reconstruct the original complete layout image. It is trained using mean squared error loss, and after pre-training, the decoder is removed, while the encoder portion is retained for use as the feature extraction network.
[0020] In one possible implementation, it is assumed that the deep residual network branch uses ResNet-18 as the basic architecture, the network input is a single-channel binary image of 256 pixels × 256 pixels, and the global average pooling layer and fully connected classification layer at the end of the network are removed, leaving only the convolutional output of the four residual stages. Finally, the convolutional feature map output of the fourth residual stage has a size of 16 pixels × 16 pixels and 512 channels. This feature map not only contains high-level semantic information, but also retains spatial resolution information, thereby supporting subsequent multi-scale localization calculations. The explicit spacing encoder uniformly samples 16×16 grid points (256 in total) within a 256-pixel × 256-pixel layout block, meaning the pixel spacing between adjacent grid points is 16, achieving uniform coverage of the entire layout area. Centered on each grid point, it divides the area into 8 equal-angle sector sectors, each covering a 45-degree angle range, continuously covering the entire 360-degree circumference. Within each sector, it calculates the Euclidean distance from the grid point along the sector centerline to the nearest polygon edge, obtaining 8 corresponding distance values, thus forming an 8-dimensional local distance vector for that grid point. This vector can be used to accurately reflect the graphic spacing distribution in eight directions around the grid point. The local distance vectors of the above 256 grid points are concatenated to generate a 2048-dimensional global distance vector. Then, it is encoded through two fully connected layers. The first layer maps the 2048-dimensional vector to 256-dimensional vectors and connects it to the ReLU activation function. The second layer maps the 256-dimensional vector to 64-dimensional vectors, and finally outputs a 64-dimensional distance feature vector. After completing feature extraction from the two parallel branches of the deep residual network and the explicit spacing encoder, global average pooling is performed on the 16×16×512 convolutional feature map output by the ResNet branch. That is, the global average value is calculated for the feature map of each channel to obtain a 512-dimensional global convolutional feature vector. This is then concatenated with the 64-dimensional spacing feature vector to obtain a 576-dimensional hybrid feature vector.
[0021] Following the previous example, suppose a certain enhanced positive sample set contains 48,000 samples. These samples are input into the hybrid feature extraction network through batch inference, with a batch size of 32. After extracting features batch by batch, a normal feature set containing 48,000 576-dimensional hybrid feature vectors is obtained. At the same time, four sets of intermediate feature maps at different scales are saved for each sample. These intermediate feature maps are the 64×64×64 feature map output from the first residual stage, the 32×32×128 feature map output from the second residual stage, the 16×16×256 feature map output from the third residual stage, and the 8×8×512 feature map output from the fourth residual stage. All intermediate feature maps are numbered and stored according to the sample index, forming a multi-scale intermediate feature map set for use in subsequent hotspot localization steps. For the test layout, firstly, a sliding window is used to divide the layout into multiple test layout blocks using preset parameters such as a block size of 10 micrometers × 10 micrometers and an overlap step of 5 micrometers. This ensures that all areas of the layout are covered and that no edge areas are missed. Next, a hybrid feature extraction network is used to extract features from each of the test layout blocks to form a test feature set. Similarly, each test feature also corresponds to a set of multi-scale intermediate feature maps for subsequent hotspot localization. Finally, after all features are extracted, Z-Score normalization is performed on the hybrid feature vectors in the test feature set based on the mean and standard deviation of the normal feature set to eliminate the influence of differences in dimensions.
[0022] Step S13, perform feature quality pre-check on the normal feature set, specifically: First, obtain the mean vector of all mixed feature vectors in the normal feature set, and use it as the center vector of the normal feature space to represent the distribution center of the normal features.
[0023] Subsequently, the Euclidean distances from all mixed feature vectors in the normal feature set to the center vector are obtained, resulting in the distance distribution of all samples. This distance distribution is used to characterize the degree of dispersion of normal samples relative to the center. The 95th percentile of the distance distribution is taken as the feature density score. This score is used to characterize the degree of clustering of normal features as a whole. The lower the score, the better the feature clustering and the higher the discriminative power of the feature space.
[0024] Next, the density score is compared with a preset threshold, which is set by a domain expert based on the actual precision requirements of the process node. If the density score does not exceed the preset threshold, the feature quality detection is deemed to have passed and no correction operation is required. If the density score exceeds a preset threshold, it indicates that the feature space dispersion is too large and the representation ability of the feature extraction network does not meet the requirements. At this time, the optimization and fine-tuning of the hybrid feature extraction network is triggered. The feature distribution is optimized by adjusting the weights of the hybrid feature extraction network, and the feature quality detection is re-performed until the density score drops below the preset threshold before proceeding to the next step.
[0025] In one possible implementation, assuming a density score of 0.72 corresponds to a normal feature set, and the preset density threshold is set to 0.8, 0.72 is less than the threshold, indicating that the feature quality meets the standard and no network fine-tuning is required. However, if the actual calculated density score is 0.82, exceeding the threshold of 0.8, then the feature dispersion is considered too large, and an optimization method can be selected based on the degree of deviation. If the deviation is small, such as the actual density score exceeding the preset threshold by no more than 10%, the domain adaptive fine-tuning method is adopted. 20% of the samples are randomly selected from the normal feature set as the fine-tuning set. The last two layers of the hybrid feature extraction network are retrained through the center loss function. After training for 5 rounds, the density score is recalculated until the density score drops below the preset threshold. If the deviation is large, such as the actual density score exceeding the preset threshold by more than 10%, then the entire network is retrained. The loss weights and training rounds of the self-supervised pre-training are adjusted, the hybrid feature extraction network is retrained, features are extracted again, and feature quality pre-detection is re-executed until the density score drops below the preset threshold.
[0026] Step S2: Divide the normal feature set into a source feature set and a target feature set. Perform neighborhood principal component perturbation on the source feature set to generate a perturbed source feature set. Simultaneously combine the target feature set to obtain the optimal transport mapping network. Combine this with a preset mechanism to generate a transport cost threshold.
[0027] Specifically, step S2 includes steps S21-S23: Step S21: Divide the normal feature set into a source feature set and a target feature set, and perform neighborhood principal component perturbation on the source feature set to generate a perturbed source feature set. Specifically: First, the normal feature set is divided into a source feature set and a target feature set by random uniform partitioning. For example, for a normal feature set containing 48,000 samples, all samples are first randomly shuffled to eliminate the influence of sample order. Then, a uniform binary partitioning method is used to generate a source feature set and a target feature set, each containing 24,000 feature vectors.
[0028] Subsequently, after the segmentation is completed, the feature distributions of the two subsets, the source feature set and the target feature set, are statistically verified. The mean and variance of each dimension of the two subsets are obtained to confirm that the relative difference between the mean and variance of each dimension of the two subsets is less than the preset threshold. At the same time, the KS test is used to confirm that there is no significant difference in the overall distribution of the two subsets. After confirming the consistency of the distribution, the random uniform segmentation process ends. If any of the relative differences in mean or variance exceed the preset threshold or the KS test distribution is significantly different, the random uniform splitting process must be re-executed, and statistical verification and KS test must be performed again until both verification results meet the standards. If the number of times random uniform segmentation is re-executed exceeds the preset threshold, and the results of statistical verification and KS test still fail to meet the standard, other preset segmentation mechanisms are used to divide the normal feature set. For example, a stratified sampling segmentation strategy is adopted, which performs stratified sampling based on the mean and variance of the distribution of each dimension of the complete normal feature set, and evenly distributes the samples to the source and target feature sets according to the distribution proportion.
[0029] Next, for each feature vector in the source feature set, retrieve the k nearest neighbor vectors within the source feature set to form a neighborhood set; Obtain the covariance matrix of the neighborhood set, perform principal component analysis on the covariance matrix, and extract the first principal component direction vector. The first principal component direction vector is used to characterize the direction of maximum variance of the feature distribution of the local region corresponding to the current feature vector. Each feature vector in the source feature set is translated along the direction of the corresponding first principal component, where the translation amount is sampled from a Gaussian distribution with a mean of zero and a variance equal to the variance of the projection of the neighborhood set along the direction of the first principal component, thus generating a perturbation source feature set.
[0030] In one possible embodiment, for each feature vector in the source feature set, the number of nearest neighbors k is set to 10. Using Euclidean distance as the quantification index, the 10 nearest neighbor features of the feature are retrieved within the source feature set, forming a neighborhood set of 10 samples. The covariance matrix of the 10-dimensional neighborhood set is calculated, and the first principal component direction vector is obtained through eigenvalue decomposition. This direction corresponds to the feature vector corresponding to the largest eigenvalue of the covariance matrix. The projection values of all samples in the neighborhood set along the first principal component direction are calculated, and the variance of the projection values is calculated. It is assumed that the projection variance is 0.06. Then, a Gaussian random number is generated with a mean of 0 and a variance of 0.06 as the translation step size. The original feature vector is superimposed with the translation step size along the first principal component direction to generate a perturbed feature vector. The above perturbing operation is performed on all samples in the source feature set one by one to generate multiple perturbed source features, forming a perturbed source feature set.
[0031] Step S22: Combining the feature set of the disturbance source and the feature set of the target, obtain the optimal transport mapping network. Specifically: First, the key process parameters in the preset constraint rules are encoded into fixed-length process constraint condition vectors. The encoded key process parameters include at least three types of core parameters: minimum line width, minimum spacing, and minimum area. These parameters are then encoded into fixed-length vectors through normalization and dimensional expansion. The process constraint vector and each hybrid feature vector in the disturbance source feature set are then directly spliced and fused according to the feature dimension channel to generate an enhanced disturbance source feature set.
[0032] For example, the three types of core parameters are normalized using the maximum-minimum normalization method to generate a normalized vector. The normalized vector is then subjected to a linear transformation by a fully connected layer, and a ReLU activation function is introduced for nonlinear activation to remove invalid negative features, ultimately generating a fixed-length process constraint vector.
[0033] Next, the enhanced perturbation source feature set is input into the original transport mapping network, which outputs the corresponding transport feature set. Using the combined transport distance obtained through a preset optimal transport cost function as the single-sample cost metric, the optimal transport matrix between the transport features and the corresponding samples of the target feature set is solved using the Sinkhorn algorithm. The original transport mapping network is trained with the goal of minimizing the average transport cost, resulting in the optimal transport mapping network. Specifically: Step A1: For the transport features obtained by mapping a single disturbance source feature and the sample features in the target feature set, a preset optimal transport cost function is used to measure the transport cost and generate the corresponding combined transport distance. The preset optimal transport cost function is expressed as the weighted sum of the normalized Euclidean distance and the cosine distance between the transport features and the sample features in the target feature set. The weighting coefficients are preset by domain experts based on historical data.
[0034] Step A2: During the training phase of the original transport mapping network, each time a batch of perturbation source features are taken from the enhanced perturbation source feature set and input into the optimal transport mapping network, the corresponding batch of transport feature sets are output, and an equal number of batch target feature sets are matched from the target feature set to construct the batch transport cost matrix. , For batch sample size, a single element within the matrix The combined transport distance corresponding to the i-th transport feature and the j-th target feature.
[0035] Step A3: Initialize a uniform probability matrix and successively perform row and column normalization corrections. After iterative convergence, output the optimal transport matrix. The optimal transport matrix is represented as a weight coefficient matrix with the same dimension as the batch transport cost matrix. Based on the optimal transport matrix, weight the batch transport cost matrix to generate the corresponding average transport cost, i.e.: ; in For the optimal transport matrix, the second OK The elements of the column.
[0036] Step A4: With the goal of minimizing the average transport cost, the parameters of the original transport mapping network are updated using the mini-batch gradient descent algorithm. That is, after each forward calculation to obtain the average transport cost, backpropagation is performed to solve the gradients of the weights and biases of each layer of the network, and the network parameters are iteratively updated in conjunction with the optimizer to continuously reduce the overall distribution deviation between the transport feature set and the target feature set.
[0037] Before training, some features within the normal feature set are randomly selected as an independent validation set and do not participate in the training parameter update. When each round of full training is completed, the average transport cost is calculated on the validation set. If the average transport cost on the validation set does not decrease for several consecutive rounds or the fluctuation range is less than the preset threshold, it is determined that the model has converged and there is no room for optimization. Training is terminated immediately, and the final original transport mapping network is taken as the optimal transport mapping network.
[0038] It should be noted that the original transport mapping network adopts a three-layer fully connected network as its main structure. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function. This network takes the enhanced perturbation source feature set as input and outputs a set of transport features with the same feature dimension as the target feature set. Furthermore, all weights of the fully connected layers of the original transport mapping network are subject to spectral normalization constraints. By restricting the spectral norm of the weight matrix, the mapping function is guaranteed to satisfy Lipschitz continuity, thereby avoiding drastic distortion in the mapping and ensuring the stability and rationality of the optimal transport solution process.
[0039] In one possible embodiment, assuming the feature vector in the enhanced perturbation source feature set is a 592-dimensional concatenated vector formed by concatenating a 576-dimensional perturbation source feature vector with a 16-dimensional process constraint vector, the dimensions of the three fully connected layers of the original transport mapping network can be set to 256, 128, and 576 dimensions respectively. The second 128-dimensional layer serves as the bottleneck layer of the network, used to extract low-dimensional latent space features. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function. The output layer does not have an activation function and directly outputs a 576-dimensional transport feature vector. Furthermore, spectral normalization constraints are applied to the weight matrices of the three fully connected layers. After each weight update, the spectral norm of the weight matrix is normalized to limit the spectral norm of the weights to 1, thereby limiting the Lipschitz constant of the mapping, ensuring the smoothness of the mapping, and thus guaranteeing the numerical stability of the optimal transport solution.
[0040] Step S23: Generate a transportation cost threshold based on a preset mechanism, specifically: First, a bootstrap resampling method is used to perform M rounds of random sampling with replacement from the original normal feature set, and a resampled dataset is obtained in each round of sampling. For each resampled dataset, the corresponding process in steps S21-S22 is called to train a corresponding optimal transport mapping network variant, and the transport cost corresponding to all normal samples under the optimal transport mapping network variant is obtained by combining the preset optimal transport cost function.
[0041] Next, all transportation cost data from the Mth round are collected to generate a normal cost sample set. The tail distribution of the normal cost sample set is fitted using extreme value theory, and the upper tail part of the cost is fitted using a generalized Pareto distribution to generate the corresponding fitted distribution.
[0042] Finally, the upper preset quantile of the fitted distribution is taken as the final transportation cost threshold.
[0043] In one possible implementation, it is assumed that the bootstrap resampling rounds M are 100 rounds, with 24,000 samples sampled with replacement from 48,000 normal samples in each round, and the sampling results in each round are different. An optimal transport mapping network variant is trained based on the sampling data of each round, resulting in a total of 100 network variants. The transport cost of all 48,000 normal samples under each variant is calculated, resulting in a total of 4.8 million transport cost sample points. The upper tail portion of all cost samples is fitted with a generalized Pareto distribution, and samples above the 95th percentile are selected as tail samples for parameter estimation. The shape and scale parameters of the distribution are obtained through maximum likelihood estimation. After fitting, the upper 99.9th percentile of the distribution is taken as the transport cost threshold.
[0044] Step S3: Cluster the normal feature set to generate a compressed reference set, quantify the transport cost of the features to be tested and perform a first-level screening based on the optimal transport mapping network, and perform a second-level verification in combination with the preset rule checking engine to generate the first hot spot detection result.
[0045] Specifically, step S3 includes steps S31-S33: Step S31: Cluster the normal feature set to generate a compressed reference set. Specifically: The K-Means clustering algorithm is used to perform clustering operations on the normal feature set. The number of clusters is... The optimal silhouette coefficient criterion is determined by iterating through different numbers of clusters, calculating the silhouette coefficient of the corresponding clustering results, and selecting the number of clusters corresponding to the largest silhouette coefficient as the optimal number. Values, after clustering, are retained The center vectors of each cluster together form a compressed reference set, which is used to characterize the overall distribution of the normal feature set.
[0046] In one possible embodiment, for the normal feature set containing 48,000 samples, multiple values of the cluster number between 32 and 256 are tested, with a step size set to 32; K-Means clustering is performed 10 times for each cluster number, and the average silhouette coefficient is taken as the evaluation index for that cluster number; assuming that, after calculation, when the cluster number... When the value is 128, the silhouette coefficient reaches its maximum value, meaning that the intra-cluster distance is small and the inter-cluster distance is large, resulting in the optimal clustering effect. Therefore, 128 is selected as the final number of clusters. After performing K-Means clustering, 128 cluster center vectors are obtained, thus forming a compressed reference set with a dimension of 128×576.
[0047] Step S32: Based on the optimal transport mapping network, perform transport cost quantification and primary screening for each feature in the feature set to be tested. Specifically: First, the feature to be tested is concatenated with the process condition vector in terms of dimension, and then input into the optimal transport mapping network to output the corresponding transport feature.
[0048] Subsequently, based on the preset optimal transport cost function, the combined transport distance between the transport feature and each center vector in the compressed reference set is calculated, and the minimum value among all combined transport distances is taken as the individual cost of the sample to be tested. The individual cost is the corresponding optimal transport cost, which is used to characterize the minimum cost of mapping the sample to the normal distribution.
[0049] Next, the optimal transport cost is compared with the transport cost threshold. If the optimal transport cost exceeds the transport cost threshold, it means that the sample cannot be properly mapped and fits the data, deviating from the normal distribution. The corresponding plot is then marked as a candidate hotspot. If the optimal transport cost does not exceed the transport cost threshold, it is determined to be a normal plot and is directly excluded without needing to proceed to the subsequent verification stage.
[0050] Step S33: Combine the preset rules to check the engine to perform secondary verification and generate the first hotspot detection result. Specifically: The preset rule checking engine is invoked to perform local design rule verification on each candidate hot spot map block, checking whether it triggers preset core design rules such as line width, spacing, area, and alignment. If a candidate hot spot plot triggers a violation of the preset core design rules, that is, if there is a clear geometric rule non-compliance, it will be directly determined as a hot spot; If a candidate hot spot block does not trigger any rule violations, meaning its geometric dimensions all conform to the design rules, but its transport cost is abnormal, it is considered a potential process hot spot, marked as a questionable sample, and transferred to the manual review queue. Engineers will make the final judgment based on lithography simulation or process experience.
[0051] It should be noted that the preset rule checking engine consists of a three-layer architecture: a rule parsing and compilation layer, a local geometry operation layer, and a constraint determination and output layer. Specifically: The rule parsing and compilation layer is used to preload the complete set of design rule files for the corresponding process node. It pre-compiles discrete constraint rules such as line width, spacing, via coverage, graphic area, and edge tolerance in text form into a quantized judgment operator library that can be directly called. It also solidifies the threshold parameters and geometric judgment logic of various rules and sorts them according to the corresponding rule priority. For example, it prioritizes the verification of core rules such as spacing and line width that are prone to hot spots in the layout. The local geometry operation layer is used to receive candidate hotspot blocks from the primary screening output. It only performs graphic parsing and geometric operations on the current local block, without loading or traversing the global layout data. The specific process is as follows: First, the metal polygons, through-hole polygons, and diffusion area polygons within the block are vectorized at the edges, extracted at the vertices, and clipped at the bounding boxes to lock the outlines of all valid graphics within the block. Then, through basic geometric operators such as polygon distance calculation, overlapping area calculation, edge parallelism verification, and vertex spacing matching, the actual geometric parameters of all graphics within the block are output in batches. The constraint judgment output layer is used to match the local actual geometric parameters obtained in real time with the pre-compiled process rule thresholds one by one to complete the violation judgment. If any core geometric parameter does not meet the process design rules, it immediately outputs the clear rule violation type, violation coordinate position and deviation value. If there is no violation, it outputs the compliance judgment result.
[0052] Step S4: Extract the new process feature set of the normal layout of the new process based on the hybrid feature extraction network, and perform cross-process migration adaptation on the optimal transport mapping network to generate a new process adaptation network. Perform hot spot detection based on the new process adaptation network to generate a second hot spot detection result.
[0053] Specifically, step S4 includes steps S41-S43: Step S41: Extract the new process feature set of the normal layout of the new process based on the hybrid feature extraction network. Specifically: First, the normal layout blocks of the new process node are obtained. These layout blocks need to be checked against the design rules of the new process to confirm that there are no rule violations. Then, the hybrid feature extraction network and feature extraction process in step S1 are fully reused to extract hybrid features from the normal layout blocks of the new process node block by block. All feature vectors together constitute the new process feature set.
[0054] Step S42: Perform cross-process migration adaptation on the optimal transport mapping network to generate a new process adaptation network. Specifically: First, determine whether the number of samples in the new process feature set is less than a preset low-sample threshold: If the number of samples is not less than the preset small sample threshold, skip the current step and proceed directly to the subsequent domain adaptation training. If the number of samples is less than the preset small sample threshold, a linear alignment method based on the maximum mean difference is adopted to map the existing process normal feature set and the new process feature set to the regeneration kernel Hilbert space. With minimizing the maximum mean difference loss between the two distributions as the optimization objective, an initial linear alignment matrix is solved. Then, through this alignment matrix, the normal feature set extracted in step S1 to characterize the normal features of the existing process is linearly transformed to the new process feature space to generate a pre-aligned feature set.
[0055] Next, before the last fully connected layer of the optimal transport mapping network, that is, after the bottleneck layer and before the output layer, a single hidden layer fully connected network is inserted to generate the original new process adaptation network, and the original new process adaptation network is subjected to domain adaptation training. The domain adaptation training refers to freezing all the original weights of the current optimal transport mapping network and updating only the parameters of the inserted single hidden layer fully connected network. The loss function used in the domain adaptation training process consists of the following two weighted parts: The first part is the distribution alignment loss, which is to minimize the average transport cost between the transport features output by the single hidden layer fully connected network and the new process feature set; The second part is the covariance alignment loss. The covariance matrix of the output features of the constraint domain adaptation module is made to approximate the covariance matrix of the new process feature set. The F-norm difference between the two covariance matrices is used to quantify this loss.
[0056] Finally, the original new process adaptation network after training is used as the new process adaptation network.
[0057] In one possible embodiment, assuming a new process feature set contains 200 samples and a preset few-sample threshold is 500, since the number of samples in the new process feature set is lower than the preset few-sample threshold, few-sample pre-alignment preprocessing is triggered. This involves selecting a Gaussian kernel as the kernel function of the regenerated kernel Hilbert space, with the kernel bandwidth set to the median distance of the feature space. The normal feature set corresponding to the existing process, extracted in step S1, is then used. Using the new process feature set as input, a maximum mean difference loss function is constructed. The linear alignment matrix is then solved using gradient descent. After solving, the result is... Each eigenvector in the matrix is multiplied by the linear alignment matrix to obtain the pre-aligned feature set; Assume the current optimal transport mapping network has three fully connected layers with dimensions of 592-dimensional input, 256-dimensional input, 128-dimensional bottleneck layer, and 576-dimensional output, respectively. A single hidden-layer fully connected network is inserted after the 128-dimensional bottleneck layer and before the 576-dimensional output layer. That is, the output of the bottleneck layer first passes through the single hidden-layer fully connected network and then enters the output layer. The input dimension of the single hidden-layer fully connected network is 128-dimensional, the hidden layer dimension is set to half of the input dimension, i.e., 64-dimensional, and the output dimension remains 128-dimensional to ensure that the input and output dimensions match, so as not to affect the structure of the original network. A ReLU activation function is added after the hidden layer, and no activation function is set for the output layer, generating the original new process adaptation network. After inserting a single-hidden-layer fully connected network, all original parameters of the original new process adaptation network are frozen, i.e., all original parameters corresponding to the optimal transport mapping network. The pre-aligned feature set is used as the source domain input, and the new process feature set is used as the target domain label. Only the parameters of the single-hidden-layer fully connected network are trained. The Adam optimizer is used, and training is performed in combination with the corresponding loss function. Before training, some features within the pre-aligned feature set are randomly selected as an independent validation set and do not participate in the training parameter update. After each round of full training, the average transport cost is calculated on the validation set. If the average transport cost on the validation set does not decrease for several consecutive rounds or the fluctuation range is less than a preset threshold, it is determined that the model has converged and there is no room for optimization, and training is terminated immediately.
[0058] Step S43: Perform hotspot detection based on the new process-adapted network to generate a second hotspot detection result. That is, completely reuse the same process as step S3 to perform hotspot detection and generate a second hotspot detection result.
[0059] Step S5: Obtain the nearest neighbor normal features corresponding to the first and second hotspot detection results, perform differential backpropagation based on the optimal transport mapping network and the new process adaptation network to generate feature attribution weights, and combine them with multi-scale intermediate feature maps to generate hotspot location maps.
[0060] Specifically, step S5 includes steps S51-S53: Step S51: Obtain the nearest neighbor normal features corresponding to the first and second hotspot detection results. Specifically: First, extract the mixed features corresponding to the map blocks that have been identified as hotspots in the first and second hotspot detection results, and mark them as hotspot features.
[0061] Subsequently, in the corresponding normal feature set, the feature vector closest to the hotspot feature is retrieved by Euclidean distance and used as the baseline normal feature. This baseline normal feature is used to characterize the distribution of normal patterns that are closest to the hotspot pattern, and the difference between the two is the abnormal component that causes the hotspot.
[0062] For example, if a certain map block is identified as a hotspot, its corresponding 576-dimensional hybrid feature vector is extracted. Then, in a normal feature set containing 48,000 samples, the Euclidean distance between the hybrid feature vector and each normal feature is calculated one by one, and the feature vector with the smallest distance is selected as the benchmark normal feature.
[0063] Step S52: Based on the optimal transport mapping network and the new process adaptation network, differential backpropagation is performed to generate feature attribution weights. Specifically: First, the difference vector between the hotspot feature and the baseline normal feature is obtained. This difference vector represents the feature difference corresponding to the anomaly.
[0064] Subsequently, the difference vector is used as the initial gradient signal to perform layer-by-layer backpropagation in the latent space of the optimal transport mapping network or the new process adaptation network, backpropagating from the output layer to the bottleneck layer, and accumulating the gradient magnitude of the output of each fully connected layer along the way.
[0065] Finally, the gradient magnitude corresponding to each dimension of the bottleneck layer features is obtained. This magnitude is the feature attribution weight for that dimension. The larger the magnitude, the higher the contribution of that dimension to the anomaly.
[0066] In one possible embodiment, for a certain map patch identified as a hotspot in the first hotspot detection result, its corresponding hotspot feature and corresponding baseline normal feature are obtained, and the difference vector between the hotspot feature and the baseline normal feature is used as the initial gradient of the output end of the optimal transport mapping network. At the same time, it is input into the optimal transport mapping network for backpropagation. During the backpropagation process, the gradient of the output feature of each layer is calculated. Finally, the gradient magnitude corresponding to the 128-dimensional feature of the bottleneck layer is extracted, and the absolute value is taken to obtain the 128-dimensional feature attribution weight vector.
[0067] Step S53: Combine multi-scale intermediate feature maps to generate a hotspot location map. Specifically: First, the feature attribution weight vector of the bottleneck layer is mapped to the channel weight of the corresponding high-level feature map, and then weighted channel by channel with the intermediate feature map of the same scale.
[0068] Subsequently, an upsampling operation is performed through transposed convolution to improve the resolution of the feature map, and a fusion operation is performed with the intermediate feature map of the same scale at the next level, which is multiplied element-wise and then added together to pass the abnormal weights to a finer scale.
[0069] Next, the upsampling and fusion process is repeated layer by layer until the original resolution of the input image is reached.
[0070] Next, a high-resolution anomaly heatmap with the same size as the input map is generated. In the heatmap, the higher the value, the greater the degree of anomaly.
[0071] Finally, an adaptive threshold segmentation algorithm is used to binarize the abnormal heatmap. After segmentation, a binary mask of the abnormal region is obtained. The region with a pixel value of 1 in the mask is the abnormal region. The minimum bounding rectangle of the binary mask is extracted. This rectangular region is the final hotspot location map, which can intuitively mark the specific location of the hotspot in the layout, so as to facilitate engineers to carry out subsequent layout correction work.
[0072] In one possible implementation, assuming the bottleneck layer corresponds to an 8×8×512 intermediate feature map of the fourth residual stage, firstly, the 128-dimensional feature attribution weights are mapped to 512-dimensional channel weights through a fully connected layer, and then weighted channel-wise with the 8×8×512 fourth-stage intermediate feature map to obtain a weighted high-level feature map. Subsequently, a 2×2 transposed convolution is used to upsample the weighted feature map, resulting in a resolution of 16×16 and a channel number of 256, which is then compared with the 16×16×256 intermediate feature map output from the third stage. After element-wise multiplication and element-wise addition, a fused 16×16 feature map is obtained. This is then upsampled to 32×32 resolution using transposed convolution and fused with the 32×32×128 intermediate feature map from the second stage. Next, it is upsampled to 64×64 resolution and fused with the 64×64×64 intermediate feature map from the first stage. Finally, it is upsampled to the original input resolution of 256×256 to generate a single-channel anomaly heatmap. Each pixel value in this heatmap represents the degree of anomaly at that location; a higher value indicates that the area is more likely to be the source of the hotspot.
[0073] The Otsu method is used to automatically calculate the segmentation threshold for the 256×256 anomaly heatmap. Assuming the calculated threshold is 0.6 times the maximum value of the heatmap, pixels with values greater than the threshold are set to 1, and pixels with values less than or equal to the threshold are set to 0, resulting in a binary anomaly mask. Morphological opening operations are then performed on the binary mask to remove small noise points and make the anomaly region more coherent. Subsequently, the minimum bounding rectangle of all anomaly pixels in the mask is calculated. Assuming that the top-left pixel coordinates of a certain minimum bounding rectangle are (52, 64) and the bottom-right pixel coordinates are (118, 136), the area covered by this rectangle is the core location of the hotspot. By superimposing this rectangle on the original map image, a visualized hotspot location map can be obtained.
[0074] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.
[0075] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0076] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A method for detecting hotspots in integrated circuit layouts based on deep learning, characterized in that, It includes the following steps: S1: Collect existing process normal layout blocks and inject random perturbations to generate an enhanced positive sample set. Extract the normal feature set and the feature set to be tested based on the hybrid feature extraction network, and save the multi-scale intermediate feature map. Simultaneously perform feature quality pre-check on the normal feature set. S2: Divide the normal feature set into a source feature set and a target feature set, perform neighborhood principal component perturbation on the source feature set to generate a perturbed source feature set, and simultaneously combine it with the target feature set to obtain the optimal transport mapping network, and generate the transport cost threshold in combination with a preset mechanism. S3: Cluster the normal feature set to generate a compressed reference set, quantify the transport cost of the features to be tested and perform a first-level screening based on the optimal transport mapping network, and perform a second-level verification in combination with the preset rule checking engine to generate the first hot spot detection result; S4: Extract the new process feature set of the normal layout of the new process based on the hybrid feature extraction network, and perform cross-process migration adaptation on the optimal transport mapping network to generate a new process adaptation network. Perform hot spot detection based on the new process adaptation network to generate a second hot spot detection result. S5: Obtain the nearest neighbor normal features corresponding to the first and second hotspot detection results, perform differential backpropagation based on the optimal transport mapping network and the new process adaptation network to generate feature attribution weights, and combine multi-scale intermediate feature maps to generate hotspot location maps.
2. The integrated circuit layout hotspot detection method based on deep learning according to claim 1, characterized in that, Step S1 includes: Collect normal layout blocks under the existing process to obtain the original positive sample set, and randomly perturb the positive sample set in combination with preset constraint rules to generate an enhanced positive sample set. The random perturbation includes at least applying polygon edge scaling, chamfering, and via center displacement to each block in the positive sample set. The enhanced sample set and the test image set are input into a pre-trained hybrid feature extraction network to extract features, and obtain the normal feature set, the test feature set and the corresponding multi-scale intermediate feature map. The normal feature set is subjected to feature quality detection and correction based on preset quality inspection rules.
3. The integrated circuit layout hotspot detection method based on deep learning according to claim 2, characterized in that, Hybrid feature extraction networks, including: A deep residual network is used to extract deep semantic features of the map and output the first feature map. The explicit spacing encoder extracts the nearest polygon edge distance in multiple sector areas, centered on the grid points uniformly sampled within the layout block, forming a local distance vector, which is then encoded into spacing features by a fully connected layer. The first feature map output by the deep residual network is processed and concatenated with the spacing feature to generate a corresponding hybrid feature vector; The enhanced sample set and the mixed feature vectors corresponding to all samples in the test set together constitute the normal feature set and the test feature set.
4. The integrated circuit layout hotspot detection method based on deep learning according to claim 2, characterized in that, Based on preset quality inspection rules, feature quality detection and correction are performed on the normal feature set, including: The mean of all mixed feature vectors within the normal feature set is defined as the center vector; Obtain the Euclidean distances from all mixed feature vectors in the normal feature set to the center vector, take the 95th percentile of all Euclidean distances as the density score, and perform feature quality detection in combination with preset quality inspection rules; If the density score meets the preset quality inspection compliance conditions, it is determined that the feature quality test has passed and no correction operation is required; If the density score does not meet the preset quality inspection compliance conditions, the hybrid feature extraction network is optimized and fine-tuned and the feature quality detection is performed again until the density score does not exceed the preset threshold.
5. The integrated circuit layout hotspot detection method based on deep learning according to claim 1, characterized in that, Step S2 includes: The normal feature set is evenly divided into the source feature set and the target feature set; For each feature vector in the source feature set, retrieve k nearest neighbor vectors within the source feature set to form a neighborhood set; Obtain the covariance matrix of the neighborhood set, and obtain the direction of the first principal component by performing principal component analysis; Each feature vector in the source feature set is translated along the direction of the corresponding first principal component, wherein the translation amount is sampled from a Gaussian distribution with a mean of zero and a variance equal to the variance of the projection of the neighborhood set along the direction of the first principal component, thereby generating a perturbation source feature set; Based on preset constraint rules, a process constraint condition vector is generated. Each feature vector in the disturbance source feature set is concatenated with the process constraint condition vector to generate an enhanced disturbance source feature set. The enhanced perturbation source feature set is input into the original transport mapping network, and the corresponding transport feature set is output. Based on a preset optimal transport cost function, the original transport mapping network is iteratively trained by solving the Sinkhorn optimal transport problem between the transport feature set and the target feature set to generate the optimal transport mapping network.
6. The integrated circuit layout hotspot detection method based on deep learning according to claim 5, characterized in that, The method further includes: A bootstrap resampling method is used to sample multiple times with replacement from the normal feature set to generate a resampled sample set, wherein each sample in the resampled sample set corresponds to a resampled normal feature set; Using each sample in the resampled sample set as the training set, the same process as in step S2 for training and generating the optimal transport mapping network is adopted to generate multiple variants of the optimal transport mapping network. Obtain the transport feature set corresponding to multiple variants of the optimal transport mapping network, fit the generalized Pareto distribution using extreme value theory, and use the upper preset quantile of the generalized Pareto distribution as the transport cost threshold.
7. The integrated circuit layout hotspot detection method based on deep learning according to claim 1, characterized in that, Step S3 includes: Clustering of the normal feature set generates multiple clusters, and the mean of all mixed feature vectors within the multiple clusters is defined as the center vector. A compressed reference set is generated based on the multiple center vectors. Each feature to be tested in the set of features to be tested is input into the optimal transport mapping network to generate the corresponding transport feature set; Based on a preset optimal transport cost function, the optimal transport cost between the transport feature set and the compressed reference set is obtained; If the optimal transportation cost exceeds the transportation cost threshold, it is determined to be a candidate hotspot; If the optimal transportation cost does not exceed the transportation cost threshold, it is considered normal; For the map blocks marked as candidate hotspots, the preset rule checking engine is invoked to verify whether the preset rule has been triggered; If a preset rule is triggered, it will be identified as a hotspot; If the preset rules are not triggered, the sample is marked as suspicious and transferred to the manual review queue. The results of manual review are combined with the map blocks identified as hotspots to form the first hotspot detection result.
8. The integrated circuit layout hotspot detection method based on deep learning according to claim 1, characterized in that, Step S4 includes: Based on a hybrid feature extraction network, process features corresponding to the normal layout of the new process are obtained, and a new process feature set is generated. If the number of samples in the new process feature set is less than a preset threshold, then a preset small sample domain adaptation process is executed to generate a pre-aligned feature set. Before the last fully connected layer of the optimal transport mapping network, a single hidden layer fully connected network is inserted to generate the original new process adaptation network; Based on the pre-aligned feature set or normal feature set, the original new process adaptation network is trained simultaneously with the new process feature set to generate a new process adaptation network. Based on the new process adaptation network, a corresponding transport cost threshold is generated, and the same process as in step S3 is called to perform hotspot detection, generating a second hotspot detection result.
9. The integrated circuit layout hotspot detection method based on deep learning according to claim 1, characterized in that, Step S5 includes: Obtain the mixed feature vector corresponding to the map block identified as a hotspot in the first and second hotspot detection results, and retrieve the corresponding nearest neighbor normal feature in the normal feature set; Based on the hybrid feature vector and the nearest neighbor normal features, a difference vector is generated; Using the difference vector as the gradient signal, layer-by-layer backpropagation is performed in the latent space of the optimal transport mapping network or the new process adaptation network to generate feature attribution weights. The feature attribution weights are fused with the multi-scale intermediate feature map to generate an anomaly heatmap, and the anomaly heatmap is processed to generate a hotspot location map.
