Capsule network-based hotspot detection method

Abstract

The embodiment of the invention provides a hot spot detection method based on a capsule network, and belongs to the technical field of data identification, and the method specifically comprises the steps: 1, obtaining a sample data set of a hot spot detection problem; step 2, performing hybrid enhancement on the hotspot samples in the sample data set to obtain a new sample data set; 3, designing a hotspot detector based on the capsule network; 4, training a hot spot detector by using the new sample data set to obtain a hot spot detection model; and 5, inputting the target layout into the hot spot detection model to obtain a hot spot detection result. According to the scheme provided by the invention, the hybrid hotspot enhancement method is provided from the level of the data set, so that the quality of the data set is improved, and the generalization of the data set is ensured; from the perspective of a hot spot detection model, a hot spot detection network structure is established based on a framework of a capsule network, a hot spot detector sensitive to pose and detail information is realized, and the detection efficiency, accuracy and adaptability are improved.

Description

Technical Field

[0001] The disclosed embodiments relate to the field of data identification technology, and in particular to a hotspot detection method based on a capsule network. Background Art

[0002] At present, the hotspot detection method based on machine learning has gradually transitioned from statistical machine learning to deep learning based on neural networks. Statistical models extract and select features of layout data based on the mathematical principles of statistical probability, which are easily affected by data bias and data sparsity, and are not suitable for solving the problem of extreme imbalance between hotspot and non-hotspot data. With the development of deep learning, deep neural network models are used to extract pattern features of hotspots. They have powerful nonlinear modeling capabilities, can process complex pattern features, and extract details and global information from samples, but rely on large-scale data volumes. At the same time, the model has poor interpretability and limited adaptability to specific fields. The quality of the training data set directly affects the performance of the model. In the problem of hotspot detection in chip layout, the model's ability to extract features from sample data is crucial. After investigation and analysis, the existing hotspot detection methods based on deep learning have the following shortcomings:

[0003] The impact of the quality of the layout dataset on model training is ignored. Due to the characteristics of the field of integrated circuit physical design, there is an imbalance between the number of hot spots and non-hot spots in the layout samples. However, current research only uses a simple and single hot spot enhancement method to balance the number gap between different categories, which makes it difficult to ensure the quality of the enhanced dataset.

[0004] The existing hotspot detector framework design is not sensitive to the pose information of the feature. The architecture based only on CNN is not capable of analyzing the spatial hierarchy of the sample, and does not consider the relative position relationship between the features. If the sample data undergoes geometric changes such as rotation, the output of the model will be unstable.

[0005] It can be seen that there is an urgent need for a hotspot detection method based on capsule networks with high detection efficiency, accuracy and adaptability. Summary of the invention

[0006] In view of this, the embodiments of the present disclosure provide a hotspot detection method based on a capsule network, which at least partially solves the problems of poor detection efficiency, accuracy and adaptability in the prior art.

[0007] The present disclosure provides a hotspot detection method based on a capsule network, including:

[0008] Step 1, obtain a sample data set for the hotspot detection problem;

[0009] Step 2, performing mixed enhancement on the hotspot samples in the sample data set to obtain a new sample data set;

[0010] Step 3, design a hotspot detector based on capsule network;

[0011] Step 4, using the new sample data set to train the hotspot detector to obtain a hotspot detection model;

[0012] Step 5: Input the target layout into the hotspot detection model to obtain the hotspot detection result.

[0013] According to a specific implementation of the embodiment of the present disclosure, step 1 specifically includes:

[0014] Step 1.1, obtain the GDS layout in the hotspot detection problem;

[0015] Step 1.2, extract hotspot samples and non-hotspot samples from the GDS map and convert them into binary images to obtain multiple map images containing the hotspot center and surrounding areas;

[0016] Step 1.3, crop and scale each layout image to obtain a sample data set.

[0017] According to a specific implementation of the embodiment of the present disclosure, step 2 specifically includes:

[0018] Step 2.1, randomly add rectangular blocks to the hotspot samples to expand the hotspot samples, and constrain the upper bounds of the length and width of the added rectangular blocks, the upper bound of the number of added rectangular blocks, and the upper bound of the ratio of the added area to the area of ​​the hotspot samples;

[0019] Step 2.2, when the number of hotspot samples in the sample data set is greater than the threshold, the hotspot samples are expanded based on the generative adversarial network of Wassertein distance;

[0020] Step 2.3, expand the hotspot samples by geometric enhancement;

[0021] Step 2.4, based on the hotspot samples expanded from step 2.1 to step 2.3, hybrid enhancement is performed according to a preset formula to obtain the final generated hotspot samples, which are combined with non-hotspot samples to form a new sample data set.

[0022] According to a specific implementation of an embodiment of the present disclosure, the expression of the generative adversarial network based on Wassertein distance is:

[0023]

[0024] Among them, P r represents the original sample distribution, P g represents the sample distribution generated by the generator, is the marginal distribution P r and P gThe set of all joint distributions γ(x,y), inf represents the infimum of the expectation E under the joint distribution γ.

[0025] According to a specific implementation method of the embodiment of the present disclosure, the expression of the preset formula is

[0026] G(h)=α·R(h)+β·W(h)+δ·H(h)

[0027] (α+β+δ=1)

[0028] Among them, R(h), W(h), and H(h) are the number of hotspot samples generated by randomly adding rectangular blocks, the number of hotspot samples generated by the generative adversarial network based on Wassertein distance, and the number of hotspot samples generated by geometric enhancement. α, β, and δ represent the proportion of hotspot samples generated by each method to the final generated hotspot samples.

[0029] According to a specific implementation of the embodiment of the present disclosure, the hotspot detector includes a three-stage feature extractor and an improved capsule network structure;

[0030] The three-stage feature extractor includes a first convolutional layer, a residual connection layer, and a multi-head self-attention mechanism layer;

[0031] The improved capsule network structure includes a second convolutional layer, an initial capsule layer and a hotspot capsule layer.

[0032] According to a specific implementation of the embodiment of the present disclosure, step 4 specifically includes:

[0033] Step 4.1, input the new sample data set into the three-stage feature extractor to obtain the feature map;

[0034] Step 4.2, convolve the feature map through the second convolutional layer to obtain a tensor structure;

[0035] Step 4.3, input the tensor structure into the initial capsule layer to obtain the vector information corresponding to the GDS layout;

[0036] Step 4.4, input the vector information into the hotspot capsule layer, update the output vector of the capsule and the coupling coefficient between capsules based on the dynamic routing mechanism, and finally obtain the probability that the feature represented by the current capsule is identified as a certain category and calculate the interval loss L1 accordingly, where the expression of the interval loss L1 is:

[0037] L1=Tmax(0,h + -||v||) 2 +λ(1-T)max(0,||v||-h - ) 2

[0038] Where T indicates whether a hotspot or non-hotspot exists (1 for existence, 0 for non-existence); h + and h - is a predefined threshold used to penalize false positive and false negative prediction results, and λ is a regularization parameter that adjusts the proportion of false positive and false negative;

[0039] Step 4.5: The decoder decodes the features represented by the current capsule according to the probability that the features represented by the current capsule are identified as a certain category, and obtains the predicted image R ij ;

[0040] Step 4.6, according to the predicted image R ij and the original GDS layout sample I ij Calculate the reconstruction loss L2, combine the interval loss L1 and the reconstruction loss L2 to get the network loss function Loss, and based on the network loss function Loss, guide the parameter update of the hotspot detector by minimizing Loss to obtain a trained hotspot detection model;

[0041] Among them, the expression of the reconstruction loss L2 is

[0042]

[0043] Among them, I ij and R ij Represent the pixel values ​​of the original image and the reconstructed image in the i-th row and j-th column respectively, and α represents the weight of the reconstruction loss in the total loss;

[0044] The expression of the network loss function Loss is:

[0045] Loss=L1+L2.

[0046] The hotspot detection scheme based on capsule network in the embodiment of the present disclosure includes: step 1, obtaining a sample data set for the hotspot detection problem; step 2, performing mixed enhancement on the hotspot samples in the sample data set to obtain a new sample data set; step 3, designing a hotspot detector based on capsule network; step 4, training the hotspot detector with the new sample data set to obtain a hotspot detection model; step 5, inputting the target layout into the hotspot detection model to obtain a hotspot detection result.

[0047] The beneficial effects of the embodiments of the present disclosure are as follows: through the scheme of the present disclosure, starting from the level of the data set, based on the serious imbalance in the number of hotspot data and non-hotspot data in the layout, a hybrid hotspot enhancement method is proposed to improve the quality of the data set and ensure the generalization of the data set. Starting from the perspective of the hotspot detection model, the importance of the directions and positional relationships of various geometric shapes in the layout data in the hotspot detection problem is analyzed, and a network structure for hotspot detection is built based on the framework of the capsule network to realize a hotspot detector that is sensitive to posture and detail information, thereby improving detection efficiency, accuracy and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] 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.

[0049] Figure 1 A flowchart of a hotspot detection method based on a capsule network provided in an embodiment of the present disclosure;

[0050] Figure 2 A framework diagram of an implementation of a hotspot detection method based on a capsule network provided in an embodiment of the present disclosure;

[0051] Figure 3 A schematic diagram of a process for obtaining a sample data set provided in an embodiment of the present disclosure;

[0052] Figure 4 A schematic diagram of a framework of a three-stage feature extractor provided in an embodiment of the present disclosure;

[0053] Figure 5 A schematic diagram of a capsule network hierarchical structure provided in an embodiment of the present disclosure;

[0054] Figure 6 A schematic diagram of a dynamic routing mechanism provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0055] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0056] 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.

[0057] 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.

[0058] 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 illustrations 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 form, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0059] 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.

[0060] With the development of Moore's Law, the number of transistors in integrated circuits has increased exponentially, and the size of transistors has continued to decrease. At present, the process nodes of integrated circuits have gradually shrunk to 7nm and below, which brings challenges to semiconductor chip manufacturing. Photolithography is considered to be the most complex process in chip manufacturing, but the development of light source wavelength does not match the rapid development of process nodes, increasing the interference and diffraction effects between adjacent graphics during the exposure process, resulting in deviations between the final photolithography pattern and the design layout. Due to the influence of the above-mentioned optical proximity effect, the pattern on the silicon wafer is distorted and a low-fidelity pattern appears. This type of circuit failure is called a "hotspot".

[0061] In order to ensure "zero defects" in the chip manufacturing process, the hot spots in the chip lithography layout must be accurately detected. With the development of artificial intelligence, machine learning technology has been widely used in hot spot detection of chip layouts to solve the problems of inefficiency, time-consuming, and inability to detect new pattern hot spots that have not appeared in the process library. The traditional methods are shown in Table 1, mainly including lithography simulation and pattern matching.

[0062] Table 1

[0063]

[0064]

[0065] Existing hotspot detection methods based on deep learning have the following disadvantages:

[0066] The impact of the quality of the layout dataset on model training is ignored. Due to the characteristics of the field of integrated circuit physical design, there is an imbalance between the number of hot spots and non-hot spots in the layout samples. Taking the ICCAD-2012 dataset as an example, as shown in Table 2, the imbalance rate is defined as:

[0067]

[0068] However, current research only uses a simple and single hotspot enhancement method to balance the quantity gap between different categories, which makes it difficult to ensure the quality of the enhanced dataset;

[0069] The existing hotspot detector framework design is not sensitive to the pose information of the feature. The architecture based only on CNN is not capable of analyzing the spatial hierarchy of the sample, and does not consider the relative position relationship between the features. If the sample data undergoes geometric changes such as rotation, the output of the model will be unstable.

[0070] Table 2

[0071]

[0072] The embodiment of the present disclosure provides a hotspot detection method based on a capsule network, which can be applied to the hotspot detection process in chip design scenarios.

[0073] See also Figure 1 , is a flow chart of a hotspot detection method based on a capsule network provided by an embodiment of the present disclosure. Figure 1 and Figure 2 As shown, the method mainly comprises the following steps:

[0074] Step 1, obtain a sample data set for the hotspot detection problem;

[0075] Furthermore, the step 1 specifically includes:

[0076] Step 1.1, obtain the GDS layout in the hotspot detection problem;

[0077] Step 1.2, extract hotspot samples and non-hotspot samples from the GDS map and convert them into binary images to obtain multiple map images containing the hotspot center and surrounding areas;

[0078] Step 1.3, crop and scale each layout image to obtain a sample data set.

[0079] Get the original dataset for the hotspot detection problem: Use the open source dataset of the ICCAD competition to train and evaluate the hotspot detector. The original competition data is the GDS layout. Extract the hotspot and non-hotspot areas from the layout and convert them into an image dataset. The specific steps are as follows: Figure 3 shown.

[0080] Step 2, performing mixed enhancement on the hotspot samples in the sample data set to obtain a new sample data set;

[0081] Based on the above embodiment, step 2 specifically includes:

[0082] Step 2.1, randomly add rectangular blocks to the hotspot samples to expand the hotspot samples, and constrain the upper bounds of the length and width of the added rectangular blocks, the upper bound of the number of added rectangular blocks, and the upper bound of the ratio of the added area to the area of ​​the hotspot samples;

[0083] Step 2.2, when the number of hotspot samples in the sample data set is greater than the threshold, the hotspot samples are expanded based on the generative adversarial network of Wassertein distance;

[0084] Step 2.3, expand the hotspot samples by geometric enhancement;

[0085] Step 2.4, based on the hotspot samples expanded from step 2.1 to step 2.3, hybrid enhancement is performed according to a preset formula to obtain the final generated hotspot samples, which are combined with non-hotspot samples to form a new sample data set.

[0086] Furthermore, the expression of the generative adversarial network based on Wassertein distance is:

[0087]

[0088] Among them, P r represents the original sample distribution, P g represents the sample distribution generated by the generator, is the marginal distribution P r and P gThe set of all joint distributions γ(x,y), inf represents the infimum of the expectation E under the joint distribution γ.

[0089] Furthermore, the expression of the preset formula is

[0090] G(h)=α·R(h)+β·W(h)+δ·H(h)

[0091] (α+β+δ=1)

[0092] Among them, R(h), W(h), and H(h) are the number of hotspot samples generated by randomly adding rectangular blocks, the number of hotspot samples generated by the generative adversarial network based on Wassertein distance, and the number of hotspot samples generated by geometric enhancement. α, β, and δ represent the proportion of hotspot samples generated by each method to the final generated hotspot samples.

[0093] In the specific implementation, a mixed hotspot sample data enhancement method is proposed to solve the problem of imbalanced categories in the map dataset, improve the generalization of the dataset, and ensure the balance of sample numbers between different categories.

[0094] Analyze the image dataset obtained in step 1. The number of hotspot samples in the map is seriously unbalanced compared to the number of non-hotspot samples, which affects the quality of the training set. The hotspot samples are expanded by hybrid enhancement. The specific strategy is as follows:

[0095] ① Randomly add rectangles to expand hotspot samples: randomly add rectangular blocks to the hotspot image, and constrain the upper bounds of the length and width of the added rectangular blocks, the upper bound of the number of added rectangular blocks, and the upper bound of the ratio of the added area to the original sample area. Among them, randomness is reflected in many aspects, including the added position, the length of the rectangle, the width of the rectangle, the number of added rectangular blocks, etc.

[0096] ② Generative adversarial network based on Wassertein distance to expand hotspot samples: Since data enhancement using machine learning depends on the generalization of the original dataset, it is only when the number of hotspot samples in the original training set is greater than a given threshold (T h ), this method is used for data enhancement. Compared with other adversarial generative networks, the generative adversarial network based on Wassertein distance can effectively solve the gradient vanishing problem when performing data enhancement, while avoiding mode collapse and ensuring the diversity of hot spot samples. The generator and discriminator are better measured by calculating the Wassertein distance. The calculation formula is as follows:

[0097]

[0098] Among them, P r represents the original sample distribution, Pg represents the sample distribution generated by the generator, is the marginal distribution P r and P g The set of all joint distributions γ(x,y), inf represents the infimum of the expectation E under the joint distribution γ.

[0099] ③Geometric enhancement to expand hotspot samples: Geometric enhancement, as a basic image enhancement method, is also applicable to hotspot sample enhancement. Geometric enhancement includes geometric operations such as translation, rotation, flipping, and mirroring.

[0100] For the hotspot sample enhancement methods ①, ② and ③ mentioned above, in the present invention, they are mixed and used to expand the hotspot sample data, and G(h) is used to represent the number of hotspot samples finally generated. The generation method is as follows:

[0101] G(h)=α·R(h)+β·W(h)+δ·H(h)

[0102] (α+β+δ=1)

[0103] Among them, R(h), W(h), and H(h) are the number of hotspot samples generated by methods ①, ②, and ③, respectively. α, β, and δ represent the proportion of hotspot samples generated by each method to G(h).

[0104] Step 3, design a hotspot detector based on capsule network;

[0105] Furthermore, the hotspot detector includes a three-stage feature extractor and an improved capsule network structure;

[0106] The three-stage feature extractor includes a first convolutional layer, a residual connection layer, and a multi-head self-attention mechanism layer;

[0107] The improved capsule network structure includes a second convolutional layer, an initial capsule layer and a hotspot capsule layer.

[0108] Step 4, using the new sample data set to train the hotspot detector to obtain a hotspot detection model;

[0109] Based on the above embodiment, step 4 specifically includes:

[0110] Step 4.1, input the new sample data set into the three-stage feature extractor to obtain the feature map;

[0111] Step 4.2, convolve the feature map through the second convolutional layer to obtain a tensor structure;

[0112] Step 4.3, input the tensor structure into the initial capsule layer to obtain the vector information corresponding to the GDS layout;

[0113] Step 4.4, input the vector information into the hotspot capsule layer, update the output vector of the capsule and the coupling coefficient between capsules based on the dynamic routing mechanism, and finally obtain the probability that the feature represented by the current capsule is identified as a certain category and calculate the interval loss L1 accordingly, where the expression of the interval loss L1 is:

[0114] L1=Tmax(0,h + -||v||) 2 +λ(1-T)max(0,||v||-h - ) 2

[0115] Where T indicates whether a hotspot or non-hotspot exists (1 for existence, 0 for non-existence); h + and h - is a predefined threshold used to penalize false positive and false negative prediction results, and λ is a regularization parameter that adjusts the proportion of false positive and false negative;

[0116] Step 4.5: The decoder decodes the features represented by the current capsule according to the probability that the features represented by the current capsule are identified as a certain category, and obtains the predicted image R ij ;

[0117] Step 4.6, according to the predicted image R ij and the original GDS layout sample I ij Calculate the reconstruction loss L2, combine the interval loss L1 and the reconstruction loss L2 to get the network loss function Loss, and based on the network loss function Loss, guide the parameter update of the hotspot detector by minimizing Loss to obtain a trained hotspot detection model;

[0118] Among them, the expression of the reconstruction loss L2 is

[0119]

[0120] Among them, I ij and R ij Represent the pixel values ​​of the original image and the reconstructed image in the i-th row and j-th column respectively, and α represents the weight of the reconstruction loss in the total loss;

[0121] The expression of the network loss function Loss is:

[0122] Loss=L1+L2.

[0123] In the specific implementation, given that the existing hotspot detector framework design is insensitive to the orientation and other posture information of the geometric shapes in the layout, we designed an efficient hotspot detection method based on the capsule network architecture, which improves the sensitivity to image attributes such as various posture information of geometric shapes. The details are as follows:

[0124] 1) Three-stage feature extractor: Considering that the size of the model is proportional to the number of model parameters, when designing the feature extractor, a small kernel convolution method is adopted as a whole to reduce the model's computational and storage requirements and reduce the complexity of the model. Residual connections directly connect residual blocks in a cross-layer manner to alleviate problems such as gradient explosion or gradient vanishing, which is conducive to training deep network structures and exploring complex hot spot pattern features. The design of the multi-head self-attention mechanism can extract the dependencies between different positions of geometric shapes in the layout, and through the features learned by multiple heads, the feature subspace is represented from different levels of layout geometric shape attributes to improve the performance of the feature extractor. The specific design is as follows: Figure 4 shown.

[0125] 2) Improved capsule network framework: The capsule network combines posture information and can achieve performance comparable to that of the CNN network through training with a small amount of data. In the layout hotspot detection task, due to the limitations of design rules, the hotspots in the layout may undergo various forms of changes, which correspond to the position and morphological relationship between the various geometric shapes in the layout image in terms of the image data set. The hotspot detector designed based on the posture-sensitive network architecture can better capture the hotspot and non-hotspot patterns in the layout and implement an efficient hotspot detection algorithm. The input and output of neurons in the capsule network are both in the form of vectors. The length of the vector represents the probability of the existence of the entity, and the direction of the vector is the attribute information of each geometric figure in the layout. Each capsule represents an attribute, which enriches the expression of features compared to the CNN network. The hierarchical structure of the improved capsule network suitable for hotspot detection tasks proposed in the present invention is as follows: Figure 5 As shown:

[0126] ①7x7 convolution: After extracting the features of the layout through the three-stage feature extractor, a feature map is obtained. A 7x7 convolution operation is performed on the feature map to reduce the size of the feature map. Compared with the 3x3 convolution in the feature extractor, it contains features of a larger spatial range;

[0127] ② Initial capsule layer: The tensor structure required by the capsule network is constructed through step ①. The output of each calculation of the traditional convolution operation is a scalar, while the output of the initial capsule layer is a vector, which enriches the feature expression of the hotspot pattern and uses information such as the direction of the vector to represent the attributes of each geometric shape in the layout;

[0128] ③ Hotspot capsule layer: Based on the output vector of step ②, the capsule output vector and coupling coefficient and other parameters are updated based on the dynamic routing mechanism to ensure that the high-level capsule correctly receives the feature vector output by the low-level capsule. The parameter update based on the dynamic routing process is as follows:

[0129] U j|i =W ij U i

[0130] S j =∑ i c ij U j|i

[0131]

[0132] b ij =b ij +U j|i V j

[0133]

[0134] Among them U i and V j Represents the output vectors of low-level capsule i and high-level capsule j respectively; weight matrix W ij Encodes the spatial relationship between the features of low-level and high-level capsules; the prediction vector U j|i The output vector U of the lower layer capsule i After the weight matrix W ij Affine transformation: The total input S of the capsule j For all U j|i The weighted sum of ij is a temporary variable that measures the similarity score between two capsules i and j and is initialized to 0; c ij is the coupling coefficient, ensuring that the sum of the coupling coefficients between the high-level capsule j and all the low-level capsules i is 1. The dynamic routing process is as follows Figure 6 shown.

[0135] Since the hotspot detection problem is essentially a binary classification problem, after the above process, the hotspot capsule layer finally obtains the vector of two capsules, and uses the normalized vector modulus to represent the probability value of being identified as a certain category, and finally obtains the probability that the feature represented by the current capsule is identified as a certain category. Then the decoder decodes the feature represented by the current capsule according to the probability that the feature represented by the current capsule is identified as a certain category, obtains the predicted image, calculates the loss value based on the predicted image and the GDS layout, and updates the parameters of the hotspot detector based on the loss value to obtain a trained hotspot detection model.

[0136] The open source ICCAD-2012 and ICAD-2019 competition training set training model is used to obtain the preprocessed enhanced training set as described in steps 1 and 2, and input it into the framework built in step 3 for training. In order to give full play to the dominant role of the original hotspot dataset, only the original hotspot dataset is used in the initial stage of model training, and the enhanced training set is gradually introduced in the later stage to obtain a trained hotspot detection model.

[0137] At the same time, the test sets of ICCAD-2012 and ICCAD-2019 can be processed in step 1 and input into the trained network in step 4 to obtain the results of hotspot detection. The recall rate (Recall) and false alarm rate (FA) are used to evaluate the performance of hotspot detection. The definitions of recall rate and false alarm rate are as follows:

[0138]

[0139]

[0140] Since the four ICCAD-2012 datasets are all at the 28nm node, the four datasets are combined for training and testing. The performance of the hotspot detector is compared with recent studies as shown in Table 3:

[0141] Table 3

[0142]

[0143] The ICCAD-2019 dataset has two sub-datasets. The test set in ICCAD2019-1 contains hotspot patterns that have never appeared in the training set. The test set in ICCAD2019-2 contains only slight differences between hotspot samples and non-hotspot samples, which belongs to fine-grained image recognition. The results of the evaluation using the ICCAD-2019 dataset are shown in Table 4:

[0144] Table 4

[0145]

[0146]

[0147] From the test results of two open source data sets, the hotspot detection method based on capsule network proposed in this paper is accurate and has a low false alarm rate, which can improve the efficiency of hotspot detection.

[0148] Step 5: Input the target layout into the hotspot detection model to obtain the hotspot detection result.

[0149] In specific implementation, after obtaining the trained hotspot detection model, during actual chip design, the target layout can be input into the hotspot detection model to obtain the hotspot detection result.

[0150] The hotspot detection method based on capsule network provided in this embodiment starts from the level of data set and proposes a hybrid hotspot enhancement method based on the serious imbalance in the number of hotspot data and non-hotspot data in the layout, thereby improving the quality of the data set and ensuring the generalization of the data set. From the perspective of the hotspot detection model, the importance of the directions and positional relationships of various geometric shapes in the layout data in the hotspot detection problem is analyzed, and a network structure for hotspot detection is built based on the framework of capsule network to realize a hotspot detector that is sensitive to posture and detail information, thereby improving detection efficiency, accuracy and adaptability.

[0151] It should be understood that various parts of the present disclosure may be implemented in hardware, software, firmware or a combination thereof.

[0152] 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 hotspot detection method based on capsule network, characterized in that: include: Step 1, obtain a sample data set for the hotspot detection problem; Step 2, performing mixed enhancement on the hotspot samples in the sample data set to obtain a new sample data set; Step 3, design a hotspot detector based on capsule network; Step 4, using the new sample data set to train the hotspot detector to obtain a hotspot detection model; Step 5: Input the target layout into the hotspot detection model to obtain the hotspot detection result.

2. The method according to claim 1, characterized in that , the step 1 specifically includes: Step 1.1, obtain the GDS layout in the hotspot detection problem; Step 1.2, extract hotspot samples and non-hotspot samples from the GDS map and convert them into binary images to obtain multiple map images containing the hotspot center and surrounding areas; Step 1.3, crop and scale each layout image to obtain a sample data set.

3. The method according to claim 2, characterized in that , the step 2 specifically includes: Step 2.1, randomly add rectangular blocks to the hotspot samples to expand the hotspot samples, and constrain the upper bounds of the length and width of the added rectangular blocks, the upper bound of the number of added rectangular blocks, and the upper bound of the ratio of the added area to the area of ​​the hotspot samples; Step 2.2, when the number of hotspot samples in the sample data set is greater than the threshold, the hotspot samples are expanded based on the generative adversarial network of Wassertein distance; Step 2.3, expand the hotspot samples by geometric enhancement; Step 2.4, based on the hotspot samples expanded from step 2.1 to step 2.3, hybrid enhancement is performed according to a preset formula to obtain the final generated hotspot samples, which are combined with non-hotspot samples to form a new sample data set.

4. The method according to claim 3, characterized in that ,The expression of the generative adversarial network based on Wassertein distance is Among them, P r represents the original sample distribution, P g represents the sample distribution generated by the generator, is the marginal distribution P r and P g The set of all joint distributions γ(x,y), inf represents the infimum of the expectation E under the joint distribution γ.

5. The method according to claim 4, characterized in that , the expression of the preset formula is G(h)=α·R(h)+β·W(h)+δ·H(h) (α+β+δ=1) Among them, R(h), W(h), and H(h) are the number of hotspot samples generated by randomly adding rectangular blocks, the number of hotspot samples generated by the generative adversarial network based on Wassertein distance, and the number of hotspot samples generated by geometric enhancement. α, β, and δ represent the proportion of hotspot samples generated by each method to the final generated hotspot samples.

6. The method according to claim 5, characterized in that ,The hotspot detector includes a three-stage feature extractor and an improved capsule ,network structure; The three-stage feature extractor includes a first convolutional layer, a residual connection layer, and a multi-head self-attention mechanism layer; The improved capsule network structure includes a second convolutional layer, an initial capsule layer and a hotspot capsule layer.

7. The method according to claim 6, characterized in that , the step 4 specifically includes: Step 4.1, input the new sample data set into the three-stage feature extractor to obtain the feature map; Step 4.2, convolve the feature map through the second convolutional layer to obtain a tensor structure; Step 4.3, input the tensor structure into the initial capsule layer to obtain the vector information corresponding to the GDS layout; Step 4.4, input the vector information into the hotspot capsule layer, update the output vector of the capsule and the coupling coefficient between capsules based on the dynamic routing mechanism, and finally obtain the probability that the feature represented by the current capsule is identified as a certain category and calculate the interval loss L1 accordingly, where the expression of the interval loss L1 is L1 = Tmax(0,h + -||v||) 2 +λ(1-T)max(0,||v||-h - ) 2 Where T indicates whether a hotspot or non-hotspot exists (1 for existence, 0 for non-existence); h + and h - is a predefined threshold used to penalize false positive and false negative prediction results, and λ is a regularization parameter that adjusts the proportion of false positive and false negative; Step 4.5: The decoder decodes the features represented by the current capsule according to the probability that the features represented by the current capsule are identified as a certain category, and obtains the predicted image R ij ; Step 4.6, according to the predicted image R ij and the original GDS layout sample I ij Calculate the reconstruction loss L2, combine the interval loss L1 and the reconstruction loss L2 to get the network loss function Loss, and based on the network loss function Loss, guide the parameter update of the hotspot detector by minimizing Loss to obtain a trained hotspot detection model; Among them, the expression of the reconstruction loss L2 is Among them, I ij and R ij Represent the pixel values ​​of the original image and the reconstructed image in the i-th row and j-th column respectively, and α represents the weight of the reconstruction loss in the total loss; The expression of the network loss function Loss is: Loss=L1+L2.

Images