Photoetching hot spot detection method and system based on non-fixed convolution kernel, and medium

Abstract

The invention provides a photoetching hot spot detection method and system based on a non-fixed convolution kernel and a medium, and the method comprises the steps: obtaining a data set which comprises a photoetching hot spot sample and a photoetching non-hot spot sample; a photoetching hot spot detection model based on a non-fixed convolution kernel is established, the model comprises a KAN feature extraction network and a classifier, and the KAN feature extraction network is constructed through residual connection of a plurality of KAN layers; carrying out model iteration training: inputting the data set into the photoetching hot spot detection model for learning, carrying out back propagation optimization by adopting a loss function, and carrying out multiple iterations until the training is completed; and classifying the input photoetching images by using the trained photoetching hot spot detection model to generate a photoetching hot spot detection result. According to the method, the KAN layer is introduced to reconstruct the convolutional neural network, the photoetching hot spot detection model based on the non-fixed convolution kernel is designed, the feature extraction capability can be effectively improved, the model prediction precision can be improved, and the false alarm rate can be reduced.

Description

Technical Field

[0001] The present application relates to the field of integrated circuit technology, and in particular to a lithography hotspot detection method, system and medium based on a non-fixed convolution kernel. Background Art

[0002] As the process technology nodes of integrated circuits continue to advance, the size of integrated circuit layout patterns continues to shrink, resulting in many lithography patterns that meet design rules but have poor actual process windows. These are called lithography hotspot patterns, which refer to patterns or combinations of patterns in the layout that are prone to cause lithography defects.

[0003] In advanced integrated circuit manufacturing, photolithography hotspot detection directly impacts device yield and reliability. Existing photolithography hotspot detection methods primarily fall into four categories: photolithography simulation, pattern matching, machine learning, and deep learning. However, while photolithography simulation-based methods offer high accuracy, they are computationally expensive, requiring hours to days for a single simulation, significantly increasing chip turnaround time. Pattern matching-based methods rely too heavily on a pre-set hotspot library and are unable to identify new defects. While machine learning is widely used in photolithography hotspot detection, it requires manual extraction of image features, making it difficult to adapt to increasingly complex photolithography layouts.

[0004] Deep learning methods can use convolutional neural networks to automatically extract image features, giving them significant advantages when dealing with complex, high-dimensional, and nonlinear problems. However, existing deep learning-based methods still have some problems: for example, the invention patent application with application publication number CN116051887A discloses a lithography hotspot classification method based on a residual network, and the invention patent application with application publication number CN117314828A discloses a lithography hotspot detection method and system based on a deep learning LHD model. Both of these detection methods are based on fixed convolution kernels, which are prone to causing a "semantic gap" and a high rate of missed detection. At the same time, datasets such as ICCAD 2012, which serve as the basis for model training, suffer from severe class imbalance, with a positive-to-negative sample ratio of up to 1:104, which can also seriously affect the model training effect. In addition, the unreasonable choice of loss function can also easily lead to model bias.

[0005] It should be noted that the above introduction to the technical background is merely intended to provide a clear and complete description of the technical solutions of this application and facilitate understanding by those skilled in the art. Simply because these solutions are described in the background technology section of this application, it should not be assumed that the above technical solutions are well known to those skilled in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide a lithography hotspot detection method, system and medium based on a non-fixed convolution kernel to solve the problem of insufficient feature extraction capability caused by the fixed convolution kernel in the existing method; at the same time, it is also necessary to solve the model bias problem caused by sample class imbalance and improper loss function selection.

[0007] In order to solve the above problems, in a first aspect, a lithography hotspot detection method based on a non-fixed convolution kernel is provided, comprising the following steps:

[0008] S1. Acquire a data set, wherein the data set includes a lithography hotspot sample and a lithography non-hotspot sample;

[0009] S2. Establishing a lithography hotspot detection model based on a non-fixed convolution kernel, wherein the lithography hotspot detection model includes a KAN feature extraction network and a classifier, wherein the KAN feature extraction network is constructed by residual connections of multiple KAN layers;

[0010] S3, performing iterative model training: inputting the data set into the lithography hotspot detection model for learning, and performing back propagation optimization using a loss function, and iterating multiple times until the training is completed;

[0011] S4. Use the trained lithography hotspot detection model to classify the input lithography image and generate a lithography hotspot detection result.

[0012] This application reconstructs the convolutional neural network by introducing the KAN layer and designs a lithography hotspot detection model based on non-fixed convolution kernels, which can effectively improve the feature extraction capability, improve the high accuracy of model prediction, and reduce the false alarm rate.

[0013] Before step S3, the hotspot samples are augmented with rotations of 0°, 90°, 180°, and 270°, and the non-hotspot samples are randomly undersampled, ensuring that the ratio of the processed hotspot samples to the processed non-hotspot samples is within 1:30. Physically constrained augmentation effectively balances the dataset and reduces model bias.

[0014] The KAN feature extraction network includes an initial KAN layer, an average pooling layer, a first KAN residual block group, a second KAN residual block group, a third KAN residual block group, a fourth KAN residual block group, a global average pooling layer and a fully connected layer.

[0015] The first KAN residual block group includes three KAN residual blocks, the second KAN residual block group includes four KAN residual blocks, the third KAN residual block group includes six KAN residual blocks, and the fourth KAN residual block group includes three KAN residual blocks; the KAN residual block includes two 3×3 KAN layers.

[0016] The KAN layer is represented as follows:

[0017]

[0018] Where n(x) is a univariate nonlinear learnable function; b(x) is a basis function; SiLU(x) is an activation function; Spline(x) is a spline function; ω b and ω s is the corresponding weight coefficient. We utilize the universal approximation property of KAN to strengthen high-order nonlinear mapping, optimize cross-channel feature interaction through tensor decomposition, and realize dynamic weight generation.

[0019] In the KAN residual block, 1×1 convolution is used to align the feature dimensions. The features of the KAN layer and the original skip connection features are added element-by-element and then output through the activation function. By optimizing the residual block structure, the gradient decay problem of deep networks is further alleviated.

[0020] The spline function is a B-spline function.

[0021] In step S3, the loss function is a dynamic weighted binary cross entropy loss function, which is expressed as:

[0022]

[0023] Among them, y i is the true label of the input sample, p i is the hotspot probability of the input sample, and ω is a dynamically adjustable weight coefficient. Through the cross-loss mechanism, the hotspot missed detection rate can be reduced, and false positives can also be reduced.

[0024] On the other hand, the present invention further provides a lithography hotspot detection system based on a non-fixed convolution kernel, wherein the lithography hotspot detection system is used to implement the method described in any one of the first aspects, including:

[0025] KAN feature extraction network, used to receive input images and perform feature extraction to generate feature vectors;

[0026] The classifier is used to receive the feature vector generated by the KAN feature extraction network and perform classification prediction, and output the lithography hotspot detection result.

[0027] In another aspect, the present invention further provides a computer storage medium, wherein the computer storage medium stores a computer program; when the computer program is run on a computer, the computer executes the method as described in any one of the first aspects.

[0028] Compared with the existing technology, the beneficial effects of the present invention mainly include the following: the present invention reconstructs the ResNet34 network by introducing the KAN layer to replace the traditional fixed convolution kernel, thereby improving the feature extraction ability of the model; optimizes the residual block structure, and realizes cross-layer dimension alignment of the KAN layer and the original input through 1×1 convolution, thereby alleviating the gradient attenuation problem of deep networks; by performing physical constraint enhancement processing on the samples of the data set, balancing the data set distribution, and combining with the dynamic weighted cross-entropy loss function, it effectively suppresses the sample imbalance bias, improves the high accuracy of model prediction, and reduces the false alarm rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the specific embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 This is the traditional ResNet34 network framework diagram.

[0031] Figure 2 This is a KAN feature extraction network framework diagram provided by the present invention.

[0032] Figure 3 A comparison chart of feature extraction between traditional convolutional layers and KAN layers.

[0033] Figure 4 Schematic diagram of the feature extraction process of the KAN layer.

[0034] Figure 5 This is the layout graphic after data enhancement. DETAILED DESCRIPTION

[0035] The foregoing and other technical aspects, features, and functions of the present invention are clearly presented in the following detailed description of a preferred embodiment with reference to the accompanying drawings. Directional terms such as up, down, left, right, front, and back, used in the following embodiments, are intended solely to refer to the directions in the accompanying drawings. Therefore, the directional terms used are for illustrative purposes only and are not intended to limit the present invention.

[0036] The following detailed description of the various embodiments of the present application is provided in conjunction with the accompanying drawings. However, those skilled in the art will appreciate that many technical details are provided in the various embodiments of the present application to facilitate a better understanding of the present application. However, even without these technical details and the various variations and modifications based on the following embodiments, the technical solutions claimed in the present application can still be implemented.

[0037] The steps in the following embodiments do not correspond one-to-one to the steps in the summary of the invention.

[0038] Example 1

[0039] The present application provides a method, system, and medium for detecting lithography hotspots based on a non-fixed convolution kernel. It should be understood that the lithography hotspot detection method of the present application is implemented based on a deep learning approach, i.e., by constructing a lithography hotspot detection model with a specific network architecture and iteratively training it using a suitable dataset, ultimately obtaining a trained lithography hotspot detection model. This lithography hotspot detection model can then be used to perform classification predictions on input lithography images.

[0040] Specifically, refer to Figures 2 to 5 As shown, this embodiment provides a lithography hotspot detection method based on a non-fixed convolution kernel, and the lithography hotspot detection method includes the following steps:

[0041] Step 1: Build a dataset for model training;

[0042] The sample sources for constructing the dataset can be obtained independently, such as by obtaining actual data from a wafer fab, or by leveraging existing datasets, such as ICCAD 2012 and ICCAD 2019. In this application, model training and testing are based on the ICCAD 2012 dataset. As shown in Table 1 below, the ICCAD 2012 dataset includes five benchmark cases, each of which includes corresponding training and test data. The training data is used to support model training, while the test data is used to test the performance of the trained model.

[0043] Table 1 Sample composition of ICCAD 2012 dataset

[0044]

[0045] Step 2: Perform physical constraint data enhancement on the dataset;

[0046] From the sample composition of the dataset shown in Table 1, it can be found that there is a serious class imbalance between hotspot samples (or hotspot images, lithography hotspot samples, etc.) and non-hotspot samples (or non-hotspot images, lithography non-hotspot samples, etc.). In order to balance the distribution of the dataset, it is necessary to preprocess the hotspot samples and non-hotspot samples separately. First, each hotspot sample is enhanced by rotations of 0°, 90°, 180°, and 270°, which can increase the number of hotspot samples while maintaining the authenticity of the optical features of the sample. The preprocessing results of the hotspot samples are shown in Figure 1. Figure 5As shown in Figure 2, random undersampling of non-hotspot samples can also be performed to further optimize the positive-negative sample ratio (i.e., the ratio of hotspot samples to non-hotspot samples) to within 1:30. Random undersampling is a method for balancing a dataset by reducing the number of majority class samples. It is easy to implement and does not require complex algorithms or additional computing resources.

[0047] After performing the above physical constraint data enhancement processing on the original data set, the data set can be effectively balanced, thereby reducing the model's preference for the majority class in the subsequent training process and improving the recognition ability of the minority class (i.e., hot samples).

[0048] Step 3: Establish a lithography hotspot detection model based on non-fixed convolution kernel;

[0049] It is understood that obtaining a complete lithography hotspot detection model includes two processes: preliminary model construction and model training. First, the lithography hotspot detection model architecture of the present invention is introduced. The lithography hotspot detection model generally includes two parts: a feature extraction network and a classifier. The feature extraction network is used to extract features of the input image and transmit them to the classifier, while the classifier is used to provide prediction results. The feature extraction network in the model of the present invention is obtained by improving the ResNet34 network.

[0050] The ResNet34 network is composed of multiple convolutional layers, residual modules and fully connected layers. It builds a deep network by reusing convolutional layers composed of 3×3 convolution kernels and residual connections. Figure 1 As shown, a ResNet34 network typically consists of an initial convolutional layer, a maximum pooling layer, a first residual block group, a second residual block group, a third residual block group, a fourth residual block group, a global average pooling layer, and a fully connected (FC) layer. The first residual block group, the second residual block group, the third residual block group, and the fourth residual block group are composed of 3, 4, 6, and 3 residual blocks, respectively. Each residual block typically includes two 3×3 convolutional layers. The residual module design allows the network to directly pass input to output via skip connections, enabling the network to learn the residual between input and output, thereby alleviating the vanishing gradient problem. However, the convolutional layers in the ResNet34 network use fixed convolution kernels, whose weights remain unchanged during training. That is, the weights are predefined and cannot be updated during training. The use of fixed convolution kernels can easily create a "semantic gap," resulting in a high miss detection rate for lithography hotspot detection. In addition, the feature interaction between traditional residual blocks is limited to linear addition, which cannot adaptively adjust the information flow. Hot spot features are prone to signal attenuation when propagating in deep networks.

[0051] The lithography hotspot detection model of the present invention uses the traditional ResNet34 as the backbone network, integrating the KAN module and reconstructing the residual module. By replacing the traditional convolutional layer with a fixed convolution kernel with the KAN (Kolmogorov-Arnold Network) layer, its universal approximation property is used to enhance high-order nonlinear mapping, and cross-channel feature interaction is optimized through tensor decomposition, achieving dynamic weight generation. The residual block structure is also optimized to further alleviate the gradient decay problem of deep networks.

[0052] The first step is to integrate the KAN module. Figure 2 As shown, the lithography hotspot detection model in the present application includes two parts: a KAN feature extraction network and a classifier (not shown in the figure), wherein the KAN feature extraction network is formed by replacing the convolution layer composed of the traditional 3×3 fixed convolution kernels in the ResNet34 network with a KAN layer. The KAN feature extraction network of the present invention is constructed by residual connections of multiple KAN layers, specifically including an initial KAN layer, an average pooling layer, a first KAN residual block group, a second KAN residual block group, a third KAN residual block group, a fourth KAN residual block group, a global average pooling layer and a fully connected layer (FC) connected in sequence; wherein, the first KAN residual block group / the second KAN residual block group / the third KAN residual block group / the fourth KAN residual block group are respectively composed of 3 / 4 / 6 / 3 residual blocks, corresponding to each KAN residual block, including two 3×3 KAN layers.

[0053] like Figure 2 As shown in the lithography hotspot detection model of this application, a series of KAN layers, pooling layers, and fully connected layers are used to extract features from the input image, and then the extracted feature information is input to the classification layer for classification prediction of the input image. This model replaces the traditional fixed convolution layer with the KAN layer, and uses the characteristics of the KAN layer to extract local features of the input image and generate a feature map. The feature extraction comparison between the KAN layer and the traditional convolution layer is as follows: Figure 3 As shown in the figure, the feature extraction process of the KAN layer is as follows: Figure 4 As shown in the figure; the dimensionality of the input feature map is then reduced through the pooling layer (Avg pooling) to retain the main feature information; finally, the fully connected layer (FC) maps the extracted features to the classification layer, and the activation function is used to obtain the predicted probability that the input image has lithography hotspots (i.e., the lithography detection result).

[0054] Among them, the mathematical expression of the KAN layer is as follows:

[0055]

[0056] in, n(x) is a univariate nonlinear learnable function that can be adjusted through training data to optimize model performance; b(x) is the basis function; the activation function is SiLU(x); Spline(x) is the spline function; ω b and ω s By adjusting the weights and biases, the above KAN layer can efficiently extract complex image features.

[0057] In this embodiment, the weight ω b and ω s It is dynamically generated through the B-spline function. Specifically, for the 3×3 local area of ​​the input feature map, an independent single-variable function fCout, Cin(x) is defined for each input channel Cin and output channel Cout, and an adaptive weight matrix is ​​generated through the single-variable function fCout, Cin(x). Since the weights are dynamically generated based on the local information of the input feature map, they can better capture the high-order features of hot spots and automatically adjust the contributions of the two types of features: process-sensitive features and geometric basic features. Compared with traditional fixed-weight convolution kernels, the dynamic weight generation method adopted in this application can adapt to different input modes more flexibly.

[0058] The second is the reconstruction of the residual module. The gradient vanishing problem is a common problem in deep learning. It refers to the phenomenon that during the backpropagation process, the gradient gradually becomes smaller, resulting in slow updates of the network weights, and the training process becomes very slow or even stagnant. In the ResNet34 network, the gradient vanishing problem is reduced by skip connections; in this application, the feature dimensions are further aligned through 1×1 convolution, and the features extracted by the KAN layer are added to the original skip connection features element by element and then output through the activation function. The dynamic and static features are fused (i.e., the controlled fusion of process-sensitive features and geometric basic features) to achieve gradient gain, which can further solve the problem of hot spot signal attenuation in deep networks.

[0059] The above completes the initial model establishment. However, model training requires a suitable dataset. In step 2 above, the dataset was preprocessed for model training. The model training process is as follows: preprocessed samples are fed into the model as input images. The model predicts the hotspot probability for each input image. Based on the difference between the predicted hotspot probability and the input image's true label, an appropriate loss function combined with a backpropagation algorithm is used to optimize the model's feature extraction and prediction capabilities. After multiple iterations of training, the model can detect lithography hotspots in input lithography images.

[0060] In this embodiment, a dynamic weighted binary cross entropy loss function is used. The dynamic weighted binary cross entropy loss function expression is as follows:

[0061]

[0062] Among them, y i is the true label of the input sample, p i is the hotspot probability predicted by the model, and ω is a dynamically adjustable weight coefficient. Binary cross entropy directly optimizes the model output for hotspot detection by maximizing the probability of correct prediction. When the true label is a hotspot (y i Value is 1): If the model predicts that the probability pi is close to 0, log(pi) approaches negative infinity, causing the loss to increase sharply. This forces the model to correct the misjudgment of the hotspot area to avoid missed detection. When the true label is non-hotspot (y i If the model predicts a probability of pi close to 1, log(1-pi) also approaches negative infinity, penalizing the model for misjudging normal areas and avoiding false positives. This crossover loss mechanism is crucial for lithography manufacturing, as missed hotspot detection can lead to device failure, while false positives increase subsequent processing costs.

[0063] In this embodiment, the training parameters of the model are: 80 iterations, batch size is set to 64, initial learning rate is 0.0001, and optimizer is Adam (β=0.99).

[0064] Afterwards, the trained lithography hotspot detection model was verified based on the ICCAD 2012 dataset and compared with the following existing methods. In this embodiment, recall rate, precision rate and F1 score (F1-score) were used as evaluation indicators.

[0065] It should be noted that in this application, four lithography hotspot detection methods based on deep learning are provided as comparative examples, among which Comparative Example 1 represents a detection method using the ResNet34 network, and Comparative Examples 2 to 4 are respectively methods recorded in the following three documents: Comparative Example 2: Shin M, Lee J H. Accurate lithography hotspot detection using deep convolutional neural networks[J]. Journal of Micro / Nanolithography, MEMS, and MOEMS, 2016, 15(4): 043507; Comparative Example 3: Zhou KB, Zhang KF, Liu J, et al. An imbalance aware lithography hotspot detection method based on HDAM and pre-trained GoogLeNet[J]. Measurement Science and Technology, 2021, 32(12): 125008; Comparative Example 4: Chen YN, Li YJ, Wu B, et al. Lightweight Hotspot Detection Model Fusing SE and ECA Mechanisms[J].Micromachines,2024,15(10):1217.

[0066] The recall rate is the ratio of the number of hotspot samples correctly identified by the model to all actual hotspot samples. The precision rate is the ratio of the number of hotspot samples correctly identified by the model to the total number of hotspot samples identified by the model. The F1 score is the harmonic mean of the recall rate and the precision rate. The values ​​of the above three indicators are all between 0 and 1. The larger the value, the better the model performance. As shown in Table 2 below, in the five benchmark cases of ICCAD 2012, this method has achieved a significant improvement compared with other methods in five different case cases. The average recall rate of the method of the present invention is 95.24%, the precision rate is 98.98%, and the F1-score is 97.07%, which is 40.55% higher than the F1-score of the traditional ResNet34 method.

[0067] Table 2 Performance comparison of hotspot detection models of different methods

[0068]

[0069] Example 2:

[0070] The present application also provides a lithography hotspot detection system based on a non-fixed convolution kernel. The lithography hotspot detection system is the lithography hotspot detection model after training in Example 1, which is used to implement the lithography hotspot method described in Example 1. The lithography hotspot detection system includes: a KAN feature extraction network, which is used to receive an input image and perform feature extraction to generate a feature vector; a classifier, which is used to receive the feature vector generated by the KAN feature extraction network and perform classification prediction, and output a lithography hotspot detection result.

[0071] Example 3:

[0072] The embodiments of the present application provide a computer device for implementing the above-mentioned method for detecting a lithographic hotspot based on a non-fixed convolution kernel, or the steps of the above-mentioned method for detecting a lithographic hotspot based on a non-fixed convolution kernel. The computer device may include: a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other via the communication bus; the processor may call logic instructions in the memory, and the logic instructions in the memory may be implemented in the form of software functional units and stored in a computer-readable storage medium when sold or used as an independent product.

[0073] Example 4:

[0074] An embodiment of the present application provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the above-mentioned lithography hotspot detection method based on a non-fixed convolution kernel, or the steps of the above-mentioned lithography hotspot detection method based on a non-fixed convolution kernel.

[0075] Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or the portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0076] In order to facilitate the description of the present invention, some common English nouns or letters are used for illustrative reference only and are not intended to be restrictive or specific. The scope of protection of the present invention should not be limited by their possible Chinese translations or specific letters.

[0077] It should also be noted that, in this article, relational terms such as "first" and "second" are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

Claims

1. A lithography hotspot detection method based on a non-fixed convolution kernel, characterized in that: The steps include: S1. Acquire a data set, wherein the data set includes a lithography hotspot sample and a lithography non-hotspot sample; S2. Establishing a lithography hotspot detection model based on a non-fixed convolution kernel, wherein the lithography hotspot detection model includes a KAN feature extraction network and a classifier, wherein the KAN feature extraction network is constructed by residual connections of multiple KAN layers; S3, performing iterative model training: inputting the data set into the lithography hotspot detection model for learning, and performing back propagation optimization using a loss function, and iterating multiple times until the training is completed; S4. Use the trained lithography hotspot detection model to classify the input lithography image and generate a lithography hotspot detection result.

2. The method for detecting lithography hotspots based on a non-fixed convolution kernel according to claim 1, wherein: Before step S3, the photolithography hotspot samples are subjected to rotation enhancement at 0°, 90°, 180° and 270° respectively, and the photolithography non-hotspot samples are randomly undersampled, so that the number ratio of the processed photolithography hotspot samples to the processed photolithography non-hotspot samples is within 1:

30.

3. The method for detecting lithography hotspots based on a non-fixed convolution kernel according to claim 1, wherein: The KAN feature extraction network includes an initial KAN layer, an average pooling layer, a first KAN residual block group, a second KAN residual block group, a third KAN residual block group, a fourth KAN residual block group, a global average pooling layer and a fully connected layer.

4. The method for detecting hotspots in lithography based on a non-fixed convolution kernel according to claim 3, wherein: The first KAN residual block group includes three KAN residual blocks, the second KAN residual block group includes four KAN residual blocks, the third KAN residual block group includes six KAN residual blocks, and the fourth KAN residual block group includes three KAN residual blocks; The KAN residual block includes two 3×3 KAN layers.

5. The method for detecting lithography hotspots based on a non-fixed convolution kernel according to claim 4, wherein: The KAN layer is represented as follows: Where n(x) is a univariate nonlinear learnable function; b(x) is a basis function; SiLU(x) is an activation function; Spline(x) is a spline function; ω b and ω s is the corresponding weight coefficient.

6. The method for detecting lithography hotspots based on a non-fixed convolution kernel according to claim 5, characterized in that: In the KAN residual block, the feature dimensions are aligned through 1×1 convolution, and the features of the KAN layer and the original skip connection features are added element by element and then output through the activation function.

7. The method for detecting hotspots in lithography based on a non-fixed convolution kernel according to claim 5, wherein: The spline function is a B-spline function.

8. The method for detecting hotspots in lithography based on a non-fixed convolution kernel according to claim 1, wherein: In step S3, the loss function is a dynamic weighted binary cross entropy loss function, which is expressed as: Among them, y i is the true label of the input sample, p i is the hotspot probability of the input sample, and ω is a dynamically adjustable weight coefficient.

9. A lithography hotspot detection system based on a non-fixed convolution kernel, characterized in that: The photolithography hotspot detection system is used to implement the method according to any one of claims 1 to 8, comprising: KAN feature extraction network, used to receive input images and perform feature extraction to generate feature vectors; The classifier is used to receive the feature vector generated by the KAN feature extraction network and perform classification prediction, and output the lithography hotspot detection result.

10. A computer storage medium, characterized in that The computer storage medium stores a computer program; when the computer program is run on a computer, the computer executes the method according to any one of claims 1 to 8.

Images