Layout potential hotspot detection method based on visual state space model

Abstract

The embodiment of the invention provides a layout potential hot spot detection method based on a visual state space model, which belongs to the technical field of data processing, and specifically comprises the following steps: step 1, constructing an image data set according to a sample layout file; 2, dividing the image data set into a label pool and a label-free pool; 3, training the initial model by using the sample data in the label pool to obtain a trained model; 4, judging whether the current sampling number meets the given total sampling number or not, if yes, ending training to obtain a trained hotspot detector, and if not, executing one-time active sampling, collecting multiple sample data in the label-free pool, adding the sample data into the label pool, updating the label pool and the sample data in the label-free pool, returning to the step 3, and if not, executing one-time active sampling, collecting multiple sample data in the label-free pool, and returning to the step 3; obtaining a trained hotspot detector until a given total sampling number is met; and 5, inputting the target layout file into a hot spot detector to obtain a detection result. Through the scheme disclosed by the invention, the detection efficiency and adaptability are improved.

Description

Technical Field

[0001] The disclosed embodiments relate to the field of data processing technology, and in particular to a method for detecting potential hotspots in a layout based on a visual state space model. Background Art

[0002] As the result of the underlying physical design, the chip layout uses plane geometric graphics to describe the physical design in the integrated circuit, and contains relevant physical information of various devices. An important part of the physical verification process is to check whether the generated layout file meets the manufacturing process requirements and achieves circuit performance. Among them, the manufacturing process requires that the designed circuit meets specific geometric design rules according to different process characteristics, including limited area, minimum width, minimum spacing between graphics, etc., to avoid circuit failure caused by short circuit and open circuit. The mask is manufactured according to the information provided in the layout. The size of the geometric graphics in the layout is positively correlated with the size of the physical layer on the actual chip. Due to the influence of the optical proximity effect, the circuit pattern will produce certain defects when the geometric graphics in the layout are burned into the silicon wafer, causing the circuit to fail to work properly. Therefore, the purpose of the layout potential hot spot detection is to avoid potential short circuits, open circuits and other adverse effects, find possible violation areas before manufacturing, and ensure the manufacturing yield.

[0003] At present, the problem of potential hotspot detection has been studied in combination with deep learning frameworks such as Convolutional Neural Network (CNN), Vision Transformers (ViT), and Graph Neural Network (GNN). However, there are still the following deficiencies in the problem of hotspot detection:

[0004] Although models based on convolutional neural networks have linear complexity, their ability to extract global information is limited.

[0005] The visual converter has quadratic complexity. The attention mechanism is used to further extract the pattern features of the layout and alleviate the long-distance dependency problem. Although it can effectively capture hotspot patterns, the time complexity is O(n 2 ), which is less efficient.

[0006] The complex geometric figures in the layout are represented only by length, width, relative position, etc., and some feature information is lost, such as angle information, curve information, etc.

[0007] Model training requires a large number of training sets. In the hotspot detection problem, the number of hotspot samples is relatively small, and the training set has an imbalanced sample category problem, which leads to overfitting of the model and affects the generalization of the model.

[0008] It can be seen that there is an urgent need for a potential hotspot detection method for the layout based on the visual state space model with high detection efficiency and adaptability. Summary of the invention

[0009] In view of this, an embodiment of the present disclosure provides a method for detecting potential hotspots in a layout based on a visual state space model, which at least partially solves the problems of poor detection efficiency and adaptability in the prior art.

[0010] In a first aspect, an embodiment of the present disclosure provides a method for detecting potential hot spots in a layout based on a visual state space model, comprising:

[0011] Step 1, construct an image dataset based on the sample layout file;

[0012] Step 2, divide the image dataset into a labeled pool and an unlabeled pool;

[0013] Step 3: Use the sample data in the label pool to train the initial model to obtain the trained model;

[0014] Step 4, determine whether the current sampling number meets the given total sampling number. If so, the training is completed and the trained hotspot detector is obtained. If not, perform active sampling once, collect multiple sample data in the unlabeled pool and add them to the labeled pool, update the sample data in the labeled pool and the unlabeled pool accordingly and return to step 3 until the given total sampling number is met, and the trained hotspot detector is obtained.

[0015] Step 5: Input the target layout file into the hotspot detector to obtain the detection result.

[0016] According to a specific implementation of the embodiment of the present disclosure, step 1 specifically includes:

[0017] Step 1.1, read the encoding information of the sample layout file to obtain the geometric information of the hot spot area and the non-hot spot area of ​​the sample layout file;

[0018] Step 1.2, drawing a grayscale layout image according to the coordinate nodes in the geometric figure information and converting it into a grayscale image;

[0019] Step 1.3, based on the hotspot and non-hotspot marks in the geometric information, potential hotspot areas and non-hotspot areas in the grayscale image are cut out to form an image data set.

[0020] According to a specific implementation of the embodiment of the present disclosure, the initial model includes a block embedding layer, an RVSS block, an adaptive average pooling layer and a linear layer connected in sequence;

[0021] The RVSS block includes a residual network structure and a visual state space module, the residual network structure includes three residual blocks stacked by jump connections, the residual block adopts depth-separable convolution, and the depth-separable convolution includes channel-by-channel convolution and point-by-point convolution;

[0022] The visual state space module is used to perform layer normalization on the input feature map, normalize the features of each sample, input the normalized features into the SS2D structure, and cross-scan the pattern features in four directions of up, down, left, and right through a scanning extension operation.

[0023] According to a specific implementation of an embodiment of the present disclosure, the step of collecting multiple sample data and adding them to a label pool includes:

[0024] The predicted value of the trained model after T forward propagation is calculated based on Monte Carlo Dropout, and the uncertainty of the unlabeled sample for the model is judged by the Shannon entropy H(p) of multiple predicted values. The expression of the Shannon entropy H(p) is:

[0025]

[0026] Among them, the probability Represents the hotspot detection model M i The random output of W t represents the model weight, X and Y represent the labeled samples and unlabeled samples, the labeled samples are sampled from the label pool I, and the unlabeled samples are sampled from the unlabeled pool U, Softmax represents the activation function, p represents the probability that the unlabeled sample Y is predicted to be of category c, c represents the sample category, c=0 represents a non-hotspot, and c=1 represents a hotspot;

[0027] Select a preset number of unlabeled samples before Shannon entropy sorting for labeling and add them to the label pool.

[0028] According to a specific implementation of the embodiment of the present disclosure, the expression for updating the sample data in the label pool and the unlabel pool is:

[0029] I=I+Sample(b)

[0030] U=U-Sample(b)

[0031] Wherein, Sample(b) represents b sample data of active sampling.

[0032] According to a specific implementation of the embodiment of the present disclosure, step 5 specifically includes:

[0033] Step 5.1, use the block embedding layer to transform the input image x∈R H×W×3Divide into N×N non-overlapping blocks, map the image to C channels, and get The feature map F;

[0034] Step 5.2: Input the feature map into the RVSS block to perform four-stage feature extraction operations to obtain feature maps of each stage, where the sizes of the feature maps of each stage are

[0035] Step 5.3, use the adaptive average pooling layer to Average pooling is performed at each position in the feature map;

[0036] In step 5.4, the average pooled features are input into the linear layer for hotspot classification to form the detection results.

[0037] The potential hotspot detection scheme for a layout based on a visual state space model in the disclosed embodiment includes: step 1, constructing an image data set according to a sample layout file; step 2, dividing the image data set into a label pool and an unlabeled pool; step 3, using the sample data in the label pool to train the initial model to obtain a trained model; step 4, judging whether the current sampling number satisfies a given total sampling number, if so, the training is terminated to obtain a trained hotspot detector, if not, performing an active sampling, collecting multiple sample data in the unlabeled pool and adding them to the label pool, and updating the sample data in the label pool and the unlabeled pool accordingly and returning to step 3 until the given total sampling number is met to obtain a trained hotspot detector; step 5, inputting the target layout file into the hotspot detector to obtain a detection result.

[0038] The beneficial effects of the embodiments of the present disclosure are as follows: through the scheme of the present disclosure, potential hotspot areas can be accurately and efficiently detected before actual manufacturing to ensure manufacturing feasibility. From the perspective of the model frame of hotspot detection, the time complexity of the hotspot detection problem is reduced, and the model achieves a global receptive field within linear time complexity, reducing the amount of computation for hotspot detection and improving the accuracy of discovering potential hotspot areas. From the perspective of the data set, in the face of a hotspot detection data set with a serious imbalance in the number of categories, an active learning mechanism is used to alleviate the problem of model overfitting. In the active learning sampling process, the performance of the trained model is considered, and samples with higher uncertainty are screened for sampling to improve the generalization of the model, achieve a higher recall rate and a lower false alarm rate, and improve the efficiency of hotspot detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0040] Figure 1 A schematic diagram of a flow chart of a method for detecting potential hot spots in a layout based on a visual state space model provided in an embodiment of the present disclosure;

[0041] Figure 2 A schematic diagram of a specific implementation process of a layout potential hotspot detection method based on a visual state space model provided in an embodiment of the present disclosure;

[0042] Figure 3 A schematic diagram of the structure flow of a hotspot detection model provided in an embodiment of the present disclosure;

[0043] Figure 4 A structural flow chart of an RVSS block provided in an embodiment of the present disclosure;

[0044] Figure 5 A schematic diagram of a residual block design provided in an embodiment of the present disclosure;

[0045] Figure 6 A schematic diagram of a four-directional scanning strategy of SS2D provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0046] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0047] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0048] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0049] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The drawings only show components related to the present disclosure rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0050] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0051] As the result of the underlying physical design, the chip layout uses plane geometric graphics to describe the physical design in the integrated circuit, and contains relevant physical information of various devices. An important part of the physical verification process is to check whether the generated layout file meets the manufacturing process requirements and achieves circuit performance. Among them, the manufacturing process requires that the designed circuit meets specific geometric design rules according to different process characteristics, including limited area, minimum width, minimum spacing between graphics, etc., to avoid circuit failure caused by short circuit and open circuit. The mask is manufactured according to the information provided in the layout. The size of the geometric graphics in the layout is positively correlated with the size of the physical layer on the actual chip. Due to the influence of the optical proximity effect, the circuit pattern will produce certain defects when the geometric graphics in the layout are burned into the silicon wafer, causing the circuit to fail to work properly. Therefore, the purpose of the layout potential hot spot detection is to avoid potential short circuits, open circuits and other adverse effects, find possible violation areas before manufacturing, and ensure the manufacturing yield.

[0052] In advanced process node manufacturing, the layout file must go through a physical verification process to check whether the design meets the manufacturing process constraints to ensure that the circuit can function normally. Traditionally, the exposure process of the physical simulation lithography machine is used to predict the formation of geometric figures during manufacturing. Although the simulation experiment achieves high-accuracy hotspot detection, it consumes a lot of computing resources and time at advanced process nodes. The pattern matching method is heavily dependent on pattern library information and can only match the patterns of existing hotspots in the pattern library, and cannot detect hotspots with new pattern features. With the advantages of machine learning technology in feature extraction, a potential hotspot detection framework is built based on machine learning technology to efficiently extract the features of hotspot patterns, solve the shortcomings of traditional physical simulation methods and pattern matching methods, and achieve fast, accurate and efficient hotspot detection.

[0053] At present, the problem of potential hotspot detection has been studied in combination with deep learning frameworks such as Convolutional Neural Network (CNN), Vision Transformers (ViT), and Graph Neural Network (GNN). However, there are still the following deficiencies in the problem of hotspot detection:

[0054] Although models based on convolutional neural networks have linear complexity, their ability to extract global information is limited.

[0055] The visual converter has quadratic complexity. The attention mechanism is used to further extract the pattern features of the layout and alleviate the long-distance dependency problem. Although it can effectively capture hotspot patterns, the time complexity is O(n 2 ), which is less efficient.

[0056] The complex geometric figures in the layout are represented only by length, width, relative position, etc., and some feature information is lost, such as angle information, curve information, etc.

[0057] Model training requires a large number of training sets. In the hotspot detection problem, the number of hotspot samples is relatively small, and the training set has an imbalanced sample category problem, which leads to overfitting of the model and affects the generalization of the model.

[0058] Therefore, starting from the shortcomings of existing research, this paper re-examines the process of potential hotspot detection in the map to alleviate the problem that model training requires a large number of data sets, and designs a new potential hotspot detection framework based on the visual state space model to achieve linear complexity while ensuring the performance of model feature extraction.

[0059] The embodiment of the present disclosure provides a method for detecting potential hotspots in a layout based on a visual state space model. The method can be applied to the hotspot detection process of a layout in a chip design scenario.

[0060] See also Figure 1 , is a flow chart of a method for detecting potential hot spots in a layout based on a visual state space model provided by an embodiment of the present disclosure. Figure 1 and Figure 2 As shown, the method mainly comprises the following steps:

[0061] Step 1, construct an image dataset based on the sample layout file;

[0062] Furthermore, the step 1 specifically includes:

[0063] Step 1.1, read the encoding information of the sample layout file to obtain the geometric information of the hot spot area and the non-hot spot area of ​​the sample layout file;

[0064] Step 1.2, drawing a grayscale layout image according to the coordinate nodes in the geometric figure information and converting it into a grayscale image;

[0065] Step 1.3, based on the hotspot and non-hotspot marks in the geometric information, potential hotspot areas and non-hotspot areas in the grayscale image are cut out to form an image data set.

[0066] In the specific implementation, the potential hot spot problem is converted into a classification problem for detecting hot spot areas and non-hot spot areas. First, the encoding information of the GDS layout file is read to obtain the geometric information containing the hot spot and non-hot spot areas of the layout. Secondly, the grayscale layout image is drawn according to the coordinate nodes of the geometric figure and converted into image data. Finally, the potential hot area and non-hot spot area are cut out according to the hot spot and non-hot spot marks to obtain the image data set.

[0067] Step 2, divide the image dataset into a labeled pool and an unlabeled pool;

[0068] In the specific implementation, the image dataset is divided into an initial label pool I and an unlabeled pool U. Specifically, the hotspot samples and non-hotspot samples of the image dataset obtained in step 1 are unbalanced. In order to reduce the negative impact of training models with unbalanced datasets, the number of hotspot and non-hotspot category samples in the initial label pool I is balanced. The number of samples in the initial label pool is α*Dataset, and the number of samples in the initial unlabeled pool U is (1-α)*Dataset. Among them, α∈(0,1), and Dataset represents the size of the original image dataset.

[0069] Step 3: Use the sample data in the label pool to train the initial model to obtain the trained model;

[0070] Further, the initial model includes a block embedding layer, an RVSS block, an adaptive average pooling layer and a linear layer connected in sequence;

[0071] The RVSS block includes a residual network structure and a visual state space module, the residual network structure includes three residual blocks stacked by jump connections, the residual block adopts depth-separable convolution, and the depth-separable convolution includes channel-by-channel convolution and point-by-point convolution;

[0072] The visual state space module is used to perform layer normalization on the input feature map, normalize the features of each sample, input the normalized features into the SS2D structure, and cross-scan the pattern features in four directions of up, down, left, and right through a scanning extension operation.

[0073] In specific implementation, use the samples in the label pool I to train the model M, and obtain the trained model M i The overall framework of model M is as follows Figure 3 shown.

[0074] ① Block embedding layer: Use the block embedding layer to embed the input image x∈R H×W×3 Divide into N×N non-overlapping blocks, map the image to C channels, and get The feature map F.

[0075] ② Input the feature map into the RVSS block (Residual Visual State Space) to perform four-stage feature extraction operations. Among them, the RVSS block does not change the size of the feature map. Therefore, stage 1 does not change the size of the feature map; stages 2 to 4 downsample the output of the RVSS block, and the size of the feature map in each stage is

[0076] ③Adaptive average pooling: receiving the output of step ② The feature map is used to perform average pooling on each position in the feature map to reduce information loss.

[0077] ④Classification: Implemented through the classic linear layer. Input the features obtained in step ③, set the number of output features to 2 (hotspot and non-hotspot categories), and implement the classification of layout samples.

[0078] The RVSS block, as the core component of model M, consists of two parts: a residual network structure and a visual state space module, such as Figure 4 As shown. The feature map F obtained in step ① is divided into F1 and F2 using channel separation technology, and the two parts are used to extract feature information respectively; then the aggregation technology is used to restore the channel dimension size, and the channel rearrangement is used to disrupt the feature map in the channel dimension to avoid the loss of features in the channel dimension after the F1 and F2 features are extracted; finally, the input feature map is added to the feature map after feature extraction. Therefore, the RVSS block does not change the size of the feature map. The detailed introduction of the residual network structure and the visual state space module is as follows:

[0079] Residual network structure: Three residual blocks are stacked to alleviate the gradient vanishing problem of deep neural networks through skip connections. Each residual block is designed to use two depthwise separable convolutions (DSC), as shown in the following figure. Figure 5 As shown, they are DSC1 and DSC2 respectively. The residual connection is introduced to propagate the input feature map of DSC1 directly to the output of DSC2 to alleviate the gradient vanishing problem. The DSC operation includes two parts: channel-by-channel convolution (Depthwise) and point-by-point convolution (Pointwise). Depthwise is used to complete the convolution of each channel to obtain a feature map F with the same number of input channels. Pointwise convolution weights F in the channel direction, while considering the image area and channel information of the layout sample to generate a new feature map. ReLU is used as the activation function. Compared with conventional convolution operations, DSC reduces the amount of calculation and parameters, which is more conducive to the realization of lightweight hotspot detectors.

[0080] Visual state space module: implemented based on VMamaba structure. First, the input feature map is normalized, and the features of each sample are normalized to alleviate the gradient disappearance and speed up network training. Secondly, the normalized samples are input to the SS2D (2D-Selective-Scan) structure, and the pattern features in the four directions of up, down, left, and right are cross-scanned through the scan extension operation to ensure the global receptive field. It has linear complexity, which is much faster than O(n 2 ) time complexity, can improve efficiency while extracting global pattern features and improve computational efficiency. The scanning expansion operation of SS2D is as follows Figure 6 Finally, the DropPath regularization technique is used to prevent overfitting, randomly delete subpaths of the multi-branch structure in the model, and reduce the number of model parameters.

[0081] Step 4, determine whether the current sampling number meets the given total sampling number. If so, the training is completed and the trained hotspot detector is obtained. If not, perform active sampling once, collect multiple sample data in the unlabeled pool and add them to the labeled pool, update the sample data in the labeled pool and the unlabeled pool accordingly and return to step 3 until the given total sampling number is met, and the trained hotspot detector is obtained.

[0082] Based on the above embodiment, the step of collecting multiple sample data and adding them into the label pool includes:

[0083] The predicted value of the trained model after T forward propagation is calculated based on Monte Carlo Dropout, and the uncertainty of the unlabeled sample for the model is judged by the Shannon entropy H(p) of multiple predicted values. The expression of the Shannon entropy H(p) is:

[0084]

[0085] Among them, the probability Represents the hotspot detection model M i The random output of W t represents the model weight, X and Y represent the labeled samples and unlabeled samples, the labeled samples are sampled from the label pool I, and the unlabeled samples are sampled from the unlabeled pool U, Softmax represents the activation function, p represents the probability that the unlabeled sample Y is predicted to be of category c, c represents the sample category, c=0 represents a non-hotspot, and c=1 represents a hotspot;

[0086] Select a preset number of unlabeled samples before Shannon entropy sorting for labeling and add them to the label pool. Further, the expression for updating the sample data in the label pool and the unlabeled pool is:

[0087] I=I+Sample(b)

[0088] U=U-Sample(b)

[0089] Wherein, Sample(b) represents b sample data of active sampling.

[0090] In the specific implementation, it is determined whether the current total number of samples S is met, S∈[0,(1-α)*Dataset]. If it is met, the training is completed and the trained hotspot detector is obtained; if it is not met, an active sampling is performed, that is, b sample data Sample(b) is sampled and added to the label pool I. The active learning sampling strategy of the present invention is specifically designed as follows:

[0091] The sampling strategy of uncertainty is adopted, and the sampling is based on the Monte Carlo Dropout (MC Dropout) method. The purpose is to select the sample information with the richest information from the data set for training. In order to alleviate the problem of model overfitting caused by limited hot sample data, the parameters of the model are not fixed during training and testing, but obey the Bernoulli distribution. Further, samples with high uncertainty are selected for training during the training process. The calculation method of uncertainty depends on the M trained in step 3. i , based on MC Dropout, the predicted value after T forward propagation is calculated, and the uncertainty of the unlabeled sample for the model is judged by the Shannon entropy H(p) of multiple predicted values. The entropy calculation process is as follows:

[0092]

[0093] Among them, the probability Represents the hotspot detection model M i Random output (W t represents the model weight), X and Y represent the existing training samples and unlabeled samples, and the model output is compressed by the Softmax function. c represents the sample category (c=0 represents non-hotspot, c=1 represents hotspot). The larger the entropy value of the unlabeled sample, the higher the uncertainty of the sample in the evaluation result of the current model. Samples with high uncertainty are preferentially selected for labeling and added to the label pool I.

[0094] Then update the sample data in the sample label pool I and the unlabeled pool U.

[0095] I=I+Sample(b) (3)

[0096] U=U-Sample(b) (4)

[0097] Repeat the above steps until the given total sampling number S is met to obtain the trained hotspot detector.

[0098] Step 5: Input the target layout file into the hotspot detector to obtain the detection result.

[0099] Based on the above embodiment, step 5 specifically includes:

[0100] Step 5.1, use the block embedding layer to transform the input image x∈R H×W×3 Divide into N×N non-overlapping blocks, map the image to C channels, and get The feature map F;

[0101] Step 5.2: Input the feature map into the RVSS block to perform four-stage feature extraction operations to obtain feature maps of each stage, where the sizes of the feature maps of each stage are

[0102] Step 5.3, use the adaptive average pooling layer to Average pooling is performed at each position in the feature map;

[0103] In step 5.4, the average pooled features are input into the linear layer for hotspot classification to form the detection results.

[0104] In the specific implementation, a hotspot detector is used to detect potential hotspots, so that potential hotspot areas can be found from the layout data and processed in time before manufacturing to ensure manufacturability. We use the open source dataset ICCAD-2012 competition data to perform hotspot detection tasks. The training set and test set in the dataset are shown in Table 1.

[0105] Table 1

[0106]

[0107] The parameters of the model are set as follows: the initial label data pool size is set to 12% of the original training set, the number of samples b sampled at a time is 500, and the total number of samples S is set to 1500. The above parameter settings are used for experiments. The confusion matrix is ​​used to evaluate the model performance of the classification. The definition of the confusion matrix is ​​shown in Table 2. The confusion matrix defines two performance evaluation indicators of the hotspot detector: recall rate and false alarm rate.

[0108] Table 2

[0109]

[0110] Recall is defined as the ratio of correctly detected hotspot samples to the total number of actual hotspot samples:

[0111]

[0112] The false positive rate is defined as the ratio of non-hotspot samples detected as hotspot samples to the total number of actual non-hotspot samples:

[0113]

[0114] The performance of the model trained on 1500 total samples is compared with the latest research as shown in Table 3:

[0115] Table 3

[0116]

[0117] It can be found from Table 3 that the potential hotspot detection method based on visual state space proposed in the present invention significantly reduces the false alarm rate while improving the recall rate of hotspot detection, which verifies the effectiveness of the model.

[0118] The potential hotspot detection method for a layout based on a visual state space model provided in this embodiment ensures manufacturing feasibility by accurately and efficiently detecting potential hotspot areas before actual manufacturing. From the perspective of the model frame of hotspot detection, the time complexity of the hotspot detection problem is reduced, and the model achieves a global receptive field within a linear time complexity, reducing the amount of computation for hotspot detection and improving the accuracy of discovering potential hotspot areas. From the perspective of the data set, in the face of a hotspot detection data set with a serious imbalance in the number of categories, an active learning mechanism is used to alleviate the problem of model overfitting. In the active learning sampling process, the performance of the trained model is considered, and samples with higher uncertainty are screened for sampling to improve the generalization of the model, achieve a higher recall rate and a lower false alarm rate, and improve the efficiency of hotspot detection.

[0119] It should be understood that various parts of the present disclosure may be implemented in hardware, software, firmware or a combination thereof.

[0120] The above is only a specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present disclosure should be included in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be based on the protection scope of the claims.

Claims

1. A method for detecting potential hot spots in a layout based on a visual state space model, characterized in that: include: Step 1, construct an image dataset based on the sample layout file; Step 2, divide the image dataset into a labeled pool and an unlabeled pool; Step 3: Use the sample data in the label pool to train the initial model to obtain the trained model; Step 4, determine whether the current sampling number meets the given total sampling number. If so, the training is completed and the trained hotspot detector is obtained. If not, perform active sampling once, collect multiple sample data in the unlabeled pool and add them to the labeled pool, update the sample data in the labeled pool and the unlabeled pool accordingly and return to step 3 until the given total sampling number is met, and the trained hotspot detector is obtained. Step 5: Input the target layout file into the hotspot detector to obtain the detection result.

2. The method according to claim 1, characterized in that The step 1 specifically includes: Step 1.1, read the encoding information of the sample layout file to obtain the geometric information of the hot spot area and the non-hot spot area of ​​the sample layout file; Step 1.2, drawing a grayscale layout image according to the coordinate nodes in the geometric figure information and converting it into a grayscale image; Step 1.3, based on the hotspot and non-hotspot marks in the geometric information, potential hotspot areas and non-hotspot areas in the grayscale image are cut out to form an image data set.

3. The method according to claim 2, characterized in that The initial model includes a block embedding layer, an RVSS block, an adaptive average pooling layer and a linear layer connected in sequence; The RVSS block includes a residual network structure and a visual state space module, the residual network structure includes three residual blocks stacked by jump connections, the residual block adopts depth-separable convolution, and the depth-separable convolution includes channel-by-channel convolution and point-by-point convolution; The visual state space module is used to perform layer normalization on the input feature map, normalize the features of each sample, input the normalized features into the SS2D structure, and cross-scan the pattern features in four directions of up, down, left, and right through a scanning extension operation.

4. The method according to claim 3, characterized in that The step of collecting multiple sample data and adding them into the label pool includes: The predicted value of the trained model after T forward propagation is calculated based on Monte Carlo Dropout, and the uncertainty of the unlabeled sample for the model is judged by the Shannon entropy H(p) of multiple predicted values. The expression of the Shannon entropy H(p) is: Among them, the probability Represents the hotspot detection model M i The random output of W t represents the model weight, X and Y represent the labeled samples and unlabeled samples, the labeled samples are sampled from the label pool I, and the unlabeled samples are sampled from the unlabeled pool U, Softmax represents the activation function, p represents the probability that the unlabeled sample Y is predicted to be of category c, c represents the sample category, c=0 represents a non-hotspot, and c=1 represents a hotspot; Select a preset number of unlabeled samples before Shannon entropy sorting for labeling and add them to the label pool.

5. The method according to claim 4, characterized in that The expression for updating the sample data in the label pool and the unlabel pool is: I=I+Sample(b) U=U-Sample(b) Wherein, Sample(b) represents b sample data of active sampling.

6. The method according to claim 5, characterized in that The step 5 specifically includes: Step 5.1, use the block embedding layer to transform the input image x∈R H×W×3 Divide into N×N non-overlapping blocks, map the image to C channels, and get The feature map F; Step 5.2: Input the feature map into the RVSS block to perform four-stage feature extraction operations to obtain feature maps of each stage, where the sizes of the feature maps of each stage are Step 5.3, use the adaptive average pooling layer to Average pooling is performed at each position in the feature map; In step 5.4, the average pooled features are input into the linear layer for hotspot classification to form the detection results.

Images