Geometric topology perception photoetching hot spot detection method based on graph contrast learning

Abstract

The invention provides a geometric topology perception photoetching hot spot detection method based on graph contrast learning. For an input layout sample, a model predicts a category label of the input layout sample, specifically, firstly, hierarchical geometry-topology perception graph representation is constructed through a new multi-level graph representation construction and feature coding method, and multi-scale spatial dependence is captured to distinguish a hot spot mode and a non-hot spot mode; then, node embedding is optimized by adopting a graph contrast learning method, and feature separability is enhanced; and finally, introducing a graphical complex detection mechanism, and carrying out refined judgment on the predicted hot spots. Complex polygon units are modeled into graph nodes, the local geometric attributes and the global topology interaction relation are jointly coded, the structural integrity of the layout is kept, and the recognition capacity of the model for complex shapes and tiny differences is effectively improved.

Description

Technical Field

[0001] This disclosure relates to the field of chips, and more specifically, to a geometric topology-aware lithographic hotspot detection method based on graph contrast learning. Background Technology

[0002] In the field of very large-scale integrated circuit (VLSI) manufacturing, photolithography is a crucial step in transistor structure formation, and its precision directly determines chip yield and performance. As process nodes continue to shrink, the gap between design complexity and manufacturing processes widens significantly, leading to increasingly prominent issues related to lithography hotspots caused by optical imaging errors, process disturbances, and equipment limitations. Lithography hotspots manifest as localized critical areas in the layout that cause printing distortion and are a major factor affecting chip manufacturing yield and electrical reliability. Therefore, achieving efficient and accurate lithography hotspot detection during the design phase has become an indispensable technical step in advanced manufacturing processes.

[0003] Existing research on lithography hotspot detection mainly focuses on three types of methods: lithography simulation, pattern matching, and machine learning. Lithography simulation methods can achieve high-precision results by accurately simulating the optical imaging and development processes, but their computational complexity is extremely high, making it difficult to meet the timeliness requirements of full-chip-level detection. Pattern matching methods utilize predefined hotspot template libraries for detection, including string matching, graph matching, and design rule-based (DRC) matching methods. These methods are highly efficient in pattern recognition within the library, but their detection accuracy drops significantly in novel layouts due to a lack of representation and generalization ability for unknown patterns. Machine learning methods improve the automation level of detection by extracting geometric features and training classifiers (such as support vector machines, random forests, and ensemble learning); however, their feature representation capabilities are limited, making it difficult to effectively capture complex geometric shapes and multi-scale topological dependencies in the layout.

[0004] In recent years, the development of deep learning technology has provided new ideas for lithography hotspot detection. Graph-based deep learning methods model the layout as a graph structure, where nodes and edges represent polygons and their spatial relationships in the chip layout, thereby mining topological dependencies. Image-based methods treat the chip layout as an image and learn spatial texture features through convolutional networks. However, these methods have the following drawbacks: First, current topological graph construction methods have defects. Past graph representation methods typically decompose polygon layout units into rectangular nodes to facilitate feature propagation and neighborhood aggregation within the graph neural network framework. However, this simplification destroys the geometric integrity of the polygons and introduces a large number of redundant edges into the graph structure, leading to topological distortion. The destruction of geometric boundaries causes the model to lose key shape descriptions during the feature learning stage, directly affecting the spatial separability of hotspot regions. Second, imbalanced data distribution affects the model's detection capability. In chip layout data, hotspot samples account for only a very small proportion, resulting in highly imbalanced training data. Although existing studies have used data augmentation strategies (such as rotation, translation, and copying) to expand the data, these methods only form approximate samples in local geometric spaces and cannot effectively cover the diverse layout feature spaces. Training models on imbalanced data can easily lead to class bias, thereby reducing their ability to identify unknown patterns and their generalization performance. Third, there is a lack of dynamic and adaptive hotspot re-evaluation mechanisms. Most deep learning methods lack effective re-identification and refinement mechanisms for false positive samples after training. For ambiguous samples near the discrimination boundary, the model typically cannot perform targeted feature reconstruction and discrimination correction, resulting in a persistent high false positive rate. This lack of dynamic feedback and updating makes it difficult for the model to maintain stable detection performance under constantly evolving design and technological conditions.

[0005] Current lithographic hotspot detection methods suffer from several problems. First, existing graph representation methods based on geometric or topological features often decompose polygonal structures into rectangles or over-discrete them during construction, disrupting the geometric integrity and topological relationships of the layout, resulting in distorted feature representation. This further reduces the separability between hotspot and non-hotspot patterns, leading to a high false positive rate. Second, existing methods generally face the problem of imbalanced datasets. In actual lithographic layouts, the number of hotspot samples is far less than that of non-hotspot samples, causing the model to be prone to overfitting or class bias during training, making it difficult to effectively learn the discriminative features of hotspots. Traditional data augmentation techniques can only alleviate the sample scarcity problem at the level of local geometric transformations, but cannot significantly improve the model's generalization ability under unseen patterns or long-tailed distributions. Third, existing hotspot detection models lack re-evaluation and dynamic optimization mechanisms for false positive samples. For hotspot and non-hotspot layouts that are difficult for a one-stage model to distinguish, existing methods have not proposed a multi-stage training framework, resulting in a consistently high false positive rate and difficulty in effectively maintaining stable and accurate detection results. Summary of the Invention

[0006] The purpose of this disclosure is to provide a geometric topology-aware lithographic hotspot detection method based on graph contrastive learning, which aims to solve the problems of existing graph representation methods based on geometric or topological features, such as rectangular decomposition or excessive discretization of polygonal structures during the construction process, imbalanced datasets, and the lack of re-evaluation and dynamic optimization mechanisms for false positive samples in existing hotspot detection models.

[0007] In general, a geometric topology-aware lithographic hotspot detection method based on graph contrastive learning is provided for input layout samples. ,Model Predict its category label Specifically, firstly, a novel multi-level graph representation construction and feature encoding method is used to construct a hierarchical geometry-topology-aware graph representation to capture multi-scale spatial dependencies and distinguish between hotspot and non-hotspot patterns. Then, a graph contrastive learning method is used to optimize node embedding and enhance feature separability. Finally, an image-based complex detection mechanism is introduced to refine the determination of predicted hotspots.

[0008] The specific implementation of the multi-level graph representation construction and feature encoding method is as follows: the graph structure construction provides the foundation for feature extraction and embedding generation; the node feature extraction and external topological features describe the structural dependencies between polygons; and the graph embedding generation method enhances the node feature expression capability.

[0009] The specific method for constructing the graph structure is as follows: representing the chip layout as a graph structure. Where V is the set of nodes, each node corresponding to a polygon in the map; A is the set of node attributes, including the internal geometric information and external topological information of the nodes; E is the set of edges, representing the adjacency relationship between nodes, where an edge is considered only if the horizontal distance between two polygons is less than a threshold. And the vertical distance is less than the threshold. When, establish edges between corresponding nodes, threshold. and Determined based on the feature dimensions of the photolithography process.

[0010] The features extracted from the node features specifically include: Shape dimensions: the maximum width and length of the polygon; Shape complexity: Measured by the number of rectangles inside the polygon and the shape type; The external topological feature extraction features specifically include: Shape interaction: Count the shape type and quantity of adjacent polygons; Distance interaction: Measure the horizontal, vertical, and minimum Euclidean distance between adjacent polygons; Diagonal relationship: When the vertices of two polygons overlap in the horizontal or vertical direction but do not share an edge, a diagonal relationship is defined.

[0011] The graph embedding generation method is specifically a graph embedding generation method based on graph transformers and graph convolutional networks, and its steps include: Representing the map as a graph structure ; A graph converter is used to perform weighted aggregation of node neighbors through an attention mechanism to achieve node attribute encoding; A graph convolutional network is used to aggregate the features of neighboring nodes to generate node embedding vectors; Multi-layer graph converters are overlaid to achieve the fusion of local geometric dependencies and global topological interactions.

[0012] Where the node represents the first The layer update formula is: in, For attention aggregation function, It is a multilayer perceptron.

[0013] The specific method of the graph comparison learning method is as follows: Given a layout embedding representation as ,in For the number of nodes, For node feature dimensions; By performing max pooling on the nodes, a graph-level representation is obtained. This vector gathers the geometric and topological information of the layout, and can effectively express subtle structural differences; Subsequently, the graph-level representation is input into a single layer. Network, and through Layer output prediction vector , , represent the probability that a sample is a hotspot or a non-hotspot, respectively.

[0014] During training, positive sample pairs are constructed. Compared with negative sample pairs ,in It belongs to the hot topic category. This category is not considered a hot topic; the training objective is to learn the loss through comparison. Optimization aims to bring similar samples closer together in the feature space and disperse samples from different classes: in Cosine similarity; Simultaneously, a weighted cross-entropy loss is designed: And combined with a weighted combination of contrastive learning loss: This is to fully consider the importance and proportion of positive and negative samples during the training process.

[0015] The specific implementation method of the hotspot re-detection method is as follows: The samples predicted as hotspots by the graph contrast learning method are extracted and, together with the real hotspot samples and obvious non-hotspot samples, are used to construct a refined training dataset. First, each candidate hotspot is converted into an image representation and input into a ResNetV2 network for feature extraction. Then, the extracted feature vectors are mapped to predicted probabilities through fully connected layers and softmax layers. and , respectively representing the first The probability that a sample belongs to a hotspot or a non-hotspot; The loss function is designed as follows: .

[0016] The technical effects to be achieved by the embodiments of the present invention are as follows: (1) A hierarchical geometry-topology collaborative representation method (Hotag) is proposed. By modeling complex polygonal units as graph nodes, the local geometric attributes and global topological interaction relationships are jointly encoded, maintaining the structural integrity of the layout and effectively improving the model's ability to recognize complex shapes and subtle differences. The structural expression capability is significantly improved, and geometric and topological features are fully integrated. In the first stage, this invention introduces a multi-level graph representation construction and feature encoding method, abstracting each polygon in the chip layout as a graph node. By constructing node attributes and relationships between nodes, multi-level geometry-topology information fusion is achieved.

[0017] (2) Based on graph representation, a contrastive learning mechanism is introduced to construct hot and non-hot sample pairs for feature constraints, significantly improving the model's discriminative ability and robustness in imbalanced scenarios. The contrastive learning mechanism enhances feature discriminativeness and improves the data imbalance problem. In the second stage, this invention designs a feature separation objective function based on contrastive learning. By maximizing the similarity between samples of the same class and minimizing the distance between samples of different classes, the model automatically forms a class clustering effect in the high-dimensional embedding space. Simultaneously, this invention introduces weight parameters into the traditional cross-entropy loss to dynamically balance the contributions of positive and negative samples in the loss function, improving the model's generalization performance.

[0018] (3) A refinement mechanism based on image re-detection is introduced in the third stage to re-evaluate hotspot and non-hotspot samples with similar structures in the first two stages, thereby effectively reducing the false alarm rate and enhancing the stability of the detection results. The introduction of the re-detection stage systematically suppresses false alarms and improves the reliability and practicality of detection. This invention innovatively designs a hotspot re-detection mechanism in the third stage, using a ResNetV2 backbone network and a pre-activated residual block structure to re-evaluate and reclassify candidate samples predicted as hotspots in the second stage. The introduction of this stage significantly reduces the false alarm rate without significantly increasing the computational overhead and further improves the robustness of the overall detection system.

[0019] (4) Multi-stage collaborative design achieves a high-precision and robust detection system. The three-stage detection framework proposed in this invention forms a complete closed loop from structural modeling to feature enhancement and then to false alarm correction. The three complement each other and progress step by step: the first stage ensures the complete expression of the map structure, the second stage optimizes the feature discrimination boundary, and the third stage corrects the classification error, thus achieving a dual improvement in model accuracy and stability. Attached Figure Description

[0020] The above and other objects and features of this disclosure will become clearer from the following description taken in conjunction with the accompanying drawings.

[0021] Figure 1 This is a schematic diagram illustrating the architecture of a geometric topology-aware lithographic hotspot detection method based on graph contrast learning according to an embodiment of the present disclosure; Figure 2 This is a schematic diagram illustrating photolithographic hotspots according to an embodiment of the present disclosure; Figure 3 This is a schematic diagram illustrating the construction of a topology diagram according to an embodiment of the present disclosure; Figure 4 This is a schematic diagram illustrating external topological features according to an embodiment of the present disclosure; Figure 5 This is a schematic diagram illustrating a graph embedding vector generation process according to an embodiment of the present disclosure. Detailed Implementation

[0022] The following detailed embodiments are provided to aid the reader in gaining a comprehensive understanding of the methods, apparatus, and / or systems described herein. However, various changes, modifications, and equivalents of the methods, apparatus, and / or systems described herein will become apparent upon understanding this disclosure. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but may be changed as will become clear upon understanding this disclosure, except for operations that must occur in a specific order. Furthermore, for clarity and conciseness, descriptions of features known in the art may be omitted.

[0023] The features described herein may be implemented in different forms and should not be construed as limited to the examples described herein. Rather, the examples described herein are provided only to illustrate some of the many feasible ways of implementing the methods, apparatus, and / or systems described herein, which will become clear upon understanding the disclosure of this application.

[0024] As used herein, the term “and / or” includes any one of the associated listed items and any combination of any two or more.

[0025] Although terms such as “first,” “second,” and “third” may be used herein to describe various components, assemblies, regions, layers, or parts, these components, assemblies, regions, layers, or parts should not be limited by these terms. Rather, these terms are used only to distinguish one component, assembly, region, layer, or part from another. Thus, without departing from the teaching of the examples described herein, the first component, first assembly, first region, first layer, or first part referred to as the first component, first assembly, first region, first layer, or first part may also be referred to as the second component, second assembly, second region, second layer, or second part.

[0026] In the specification, when an element (such as a layer, region, or substrate) is described as being "on" another element, "connected to," or "bonded to" another element, the element may be directly "on" another element, directly "connected to," or "bonded to" the other element, or one or more other elements may be present in between. Conversely, when an element is described as being "directly on" another element, "directly connected to," or "directly bonded to" another element, no other elements may be present in between.

[0027] The terminology used herein is for the purpose of describing various examples only and is not intended to limit disclosure. Unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. The terms “comprising,” “including,” and “having” indicate the presence of the described features, quantities, operations, components, elements, and / or combinations thereof, but do not preclude the presence or addition of one or more other features, quantities, operations, components, elements, and / or combinations thereof.

[0028] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains upon understanding this disclosure. Unless expressly defined herein, terms (such as those defined in a general dictionary) shall be interpreted as having a meaning consistent with their meaning in the context of the relevant field and in this disclosure, and shall not be interpreted in an idealized or overly formalistic manner.

[0029] Furthermore, in the description of the examples, detailed descriptions of well-known related structures or functions will be omitted when it is believed that such detailed descriptions would lead to a vague interpretation of this disclosure.

[0030] Figure 1 This is a schematic diagram illustrating a geometric topology-aware lithographic hotspot detection method based on graph contrast learning according to an embodiment of the present disclosure.

[0031] To achieve the aforementioned objectives, this invention proposes a novel photolithography hotspot detection framework, as follows: Figure 1 As shown, this method first constructs a hierarchical geometry-topology-aware graph representation using a novel multi-level graph representation construction and feature encoding method to capture multi-scale spatial dependencies and distinguish between hotspot and non-hotspot patterns. Then, graph contrastive learning is employed to optimize node embedding, enhancing feature separability. Finally, an image-based complex detection mechanism is introduced to refine the judgment of predicted hotspots, thereby systematically suppressing false alarms and improving detection robustness. Compared with existing technologies, this method can efficiently and accurately identify lithographic hotspots in chip layouts, balancing structural integrity, data balance, and false alarm control. This method fully integrates the advantages of graph structure modeling and image feature learning, balancing detection accuracy, model scalability, and dynamic adaptability.

[0032] The following sections define the problem of lithographic hotspot detection and describe the multi-level graph representation construction and feature encoding method stages, graph comparison learning stage, and hotspot re-detection stage included in this framework.

[0033] Hotspot detection problem definition The following defines the problem of lithographic hotspot detection and introduces relevant performance evaluation indicators.

[0034] Photolithography hotspot detection: Given a set of chip design layout samples containing both hotspots and non-hotspots, such as... Figure 2 As shown, the red boxes indicate the locations of lithographic hotspots. The goal of the hotspot detection model is to identify all real hotspots in the layout sample while maintaining detection accuracy and minimizing false alarms. This problem can be formalized as a binary classification task: For the input layout sample... ,Model Predict its category label The model needs to maximize detection accuracy while minimizing the false alarm rate, thereby balancing recognition performance and robustness. As shown in Figure 2, (a) and (b) are layout examples with and without hotspots, with hotspot areas marked by red circles, and (c) and (d) are the corresponding lithographic imaging simulation results.

[0035] accuracy Detection accuracy measures the proportion of real hotspots correctly identified by the model, and is defined as follows: Where TP (True Positive) represents the number of samples correctly identified as hotspots, and FN (False Negative) represents the number of real hotspots that are incorrectly identified as non-hotspots.

[0036] False Alarm Ratio: Where FP (False Positive) represents the number of non-hotspot samples that were incorrectly identified as hotspots, and TN (True Negative) represents the number of samples that were correctly identified as non-hotspots.

[0037] Multi-level graph representation construction and feature encoding methods After defining the problem of lithographic hotspot detection, this invention provides a novel graph representation method, Hotag, and its feature encoding and embedding generation method for lithographic hotspot detection, aiming to solve the problems of structural distortion, topological distortion, and insufficient feature information in the polygon decomposition stage of existing graph-based layout representation methods.

[0038] (I) Graph Structure Construction In this invention, the chip layout is represented as a graph structure. Where V is the set of nodes, each node corresponding to a polygon in the map; A is the set of node attributes, including the internal geometric information and external topological information of the nodes; and E is the set of edges, representing the adjacency relationships between nodes, such as... Figure 3 As shown, this applies only when the horizontal distance between the two polygons is less than a threshold. And the vertical distance is less than the threshold. Only when the threshold is reached is an edge established between the corresponding nodes. and The dimensions are determined based on the features of the photolithography process. Using this graph structure, a layout representation with attributes can be constructed, providing a foundation for subsequent feature extraction and embedding generation.

[0039] (II) Node Feature Extraction This invention further defines the internal geometric features and external topological features of nodes to describe the shape and spatial interaction of polygons. The internal geometric features are the characteristics of its shape and geometric structure, and their extracted features are as follows: Shape dimensions: the maximum width and length of the polygon; Shape complexity: Measured by the number of rectangles inside the polygon and the shape type (such as "rectangle", "E-shaped", "L-shaped", etc.).

[0040] External topological features describe the spatial relationships between the polygon and its neighboring polygons, such as... Figure 4 As shown, its extracted features are as follows: Shape interaction: Count the shape type and quantity of adjacent polygons; Distance interaction: Measure the horizontal, vertical, and minimum Euclidean distance between adjacent polygons; Diagonal relationship: When the vertices of two polygons overlap in the horizontal or vertical direction but do not share an edge, a diagonal relationship is defined.

[0041] The above features are used to fully describe the structural dependencies between polygons, providing a basis for the hotspot formation mechanism.

[0042] (III) Graph Embedding Generation Method This invention provides a graph embedding generation method based on a graph transformer and a graph convolutional network (GCN), the steps of which include: (1) Use the above two steps to represent the layout as a graph structure. ; (2) Use a graph converter to perform weighted aggregation of node neighbors through an attention mechanism to achieve node attribute encoding; (3) Use graph convolutional networks to aggregate the features of neighboring nodes and generate node embedding vectors; (4) Multi-layer graph converters are superimposed to achieve the fusion of local geometric dependencies and global topological interactions.

[0043] Where the node represents the first The layer update formula is: in, For attention aggregation function, This is a multilayer perceptron used to enhance the representation of node features. The overall graph embedding process is as follows: Figure 5 As shown.

[0044] Graph Comparison Learning Method In lithography hotspot detection, hotspot patterns typically constitute only a tiny fraction of the chip layout, while the number of non-hotspot samples far exceeds that of hotspot samples. This extreme imbalance in data distribution poses a significant challenge to traditional training methods based on cross-entropy loss. Specifically, cross-entropy loss overemphasizes majority class (non-hotspot) samples when calculating gradients, making it difficult for the model to fully learn the discriminative features of minority class hotspots during training. This leads to significant biases in model prediction: on the one hand, some true hotspots are not correctly identified, resulting in a decrease in recall; on the other hand, due to the large number of non-hotspot samples, the model is prone to misclassifying structurally similar non-hotspots as hotspots, increasing the false positive rate. To address this issue, this invention proposes a training objective based on contrastive learning, aiming to significantly enhance the feature separability of the model in imbalanced data environments, thereby improving overall detection performance. The specific process is as follows: Given a layout embedding representation as ,in For the number of nodes, The node feature dimension is used. A graph-level representation is obtained by performing max pooling on the nodes. This vector aggregates the geometric and topological information of the layout, effectively representing subtle structural differences. Subsequently, the graph-level representation is input into a single layer. Network, and through Layer output prediction vector , , represent the probability that a sample is a hotspot or a non-hotspot, respectively.

[0045] To enhance the ability to identify minority hotspots, this invention constructs positive sample pairs. Compared with negative sample pairs ,in It belongs to the hot topic category. This belongs to a non-hotspot category. The training objective is to learn the loss through comparison. Optimization aims to bring similar samples closer together in the feature space and disperse samples from different classes: in Let be the cosine similarity.

[0046] Meanwhile, to further alleviate the class imbalance problem, a weighted cross-entropy loss is designed: And combined with a weighted combination of contrastive learning loss: This combined loss fully considers the importance and proportion of positive and negative samples during training. It ensures the distinction between positive and negative samples while avoiding the loss of gradient information due to an insufficient number of minority class samples, thereby improving the stability and convergence of the model on imbalanced datasets.

[0047] Through this module's design, the model can effectively learn minority class hotspot features in highly imbalanced data environments, improving prediction accuracy and providing high-quality candidate hotspot samples for subsequent refinement stages. Furthermore, the Hottag graph representation method synergistically works with the contrastive learning training objective, enabling the model to fully utilize the geometric and topological information of polygons, enhancing sensitivity to subtle changes and complex structures, thereby ensuring the overall reliability and robustness of the detection. This method is suitable for full-chip hotspot detection scenarios with large-scale layouts, significantly improving detection efficiency and accuracy, and providing strong support for manufacturing process optimization.

[0048] Hotspot re-detection method In actual lithography hotspot detection, even if the model can effectively distinguish between hotspot and non-hotspot patterns through the comparative learning phase, a certain proportion of false alarms are still unavoidable due to the complex polygonal structure and various minute deformations in the layout. These false alarms mainly originate from non-hotspot regions with shapes highly similar to hotspots, whose structural features and topological relationships are difficult to completely distinguish from real hotspots in the initial detection stage. Without further processing, these false alarms not only increase the cost of subsequent process verification but may also lead to performance anomalies and yield reductions during chip manufacturing. Therefore, to improve the reliability of detection results, a dedicated hotspot refinement mechanism must be introduced based on the initial detection.

[0049] This invention introduces a hotspot re-detection module in the third stage. The core objective of this module is to re-evaluate samples with similar structures but different properties using the candidate hotspot set obtained in the preliminary stage, thereby significantly reducing the false positive rate. Specifically, the third stage extracts the samples predicted as hotspots in the second stage and constructs a refined training dataset together with the actual hotspot samples and obvious non-hotspot samples. Compared with the original map training set, this dataset has been carefully selected, and the class ratio is closer to balance, thus mitigating the gradient bias problem caused by data imbalance and enabling the model to focus on learning subtle structural differences during the refinement stage.

[0050] In the model design during the refinement stage, this invention adopts the ResNetV2 network as the backbone architecture. ResNetV2 improves the gradient propagation path through pre-activated residual blocks, enhancing the convergence speed of deep networks while reducing the gradient vanishing problem. The specific process includes: first, converting each candidate hotspot into an image representation, which is then input into the ResNetV2 network for feature extraction; subsequently, the extracted feature vectors are mapped to predicted probabilities through fully connected layers and softmax layers. and , respectively representing the first The probability that a sample belongs to a hotspot or a non-hotspot. The loss function is designed as follows: The hotspot refinement stage not only supplements the preliminary detection results but is also an indispensable core component of the entire three-stage hotspot detection framework (3S-HDF). Through a carefully constructed training dataset, the deep feature extraction capabilities of the ResNetV2 network, and a refined loss function design, this stage enables the model to possess high discriminative power and stability when handling complex layouts, subtle structural changes, and highly similar non-hotspot samples, providing reliable hotspot detection technology support for actual large-scale chip manufacturing.

[0051] While some embodiments of this disclosure have been shown and described, those skilled in the art will understand that modifications may be made to these embodiments without departing from the principles and spirit of this disclosure, which are defined by the claims and their equivalents.

Claims

1. A geometric topology-aware lithographic hotspot detection method based on graph contrastive learning, characterized in that, For the input layout sample ,Model Predict its category label Specifically, firstly, a novel multi-level graph representation construction and feature encoding method is used to construct a hierarchical geometry-topology-aware graph representation to capture multi-scale spatial dependencies and distinguish between hot and non-hot patterns; then, a graph contrastive learning method is used to optimize node embedding and enhance feature separability. Finally, an image-based complex detection mechanism is introduced to refine the determination of predicted hotspots.

2. The geometric topology-aware lithographic hotspot detection method based on graph contrast learning as described in claim 1, characterized in that, The specific implementation of the multi-level graph representation construction and feature encoding method is as follows: the graph structure construction provides the foundation for feature extraction and embedding generation; the node feature extraction and external topological features describe the structural dependencies between polygons; and the graph embedding generation method enhances the node feature expression capability.

3. The geometric topology-aware lithographic hotspot detection method based on graph contrast learning as described in claim 2, characterized in that, The specific method for constructing the graph structure is as follows: representing the chip layout as a graph structure. Where V is the set of nodes, each node corresponding to a polygon in the map; A is the set of node attributes, including the internal geometry and external topology of the nodes; and E is the set of edges, representing the adjacency relationships between nodes, where an edge is considered only if the horizontal distance between two polygons is less than a threshold. And the vertical distance is less than the threshold. When, establish edges between corresponding nodes, threshold. and Determined based on the feature dimensions of the photolithography process.

4. The geometric topology-aware lithographic hotspot detection method based on graph contrast learning as described in claim 2, characterized in that, The features extracted from the node features specifically include: Shape dimensions: the maximum width and length of the polygon; Shape complexity: Measured by the number of rectangles inside the polygon and the shape type; The external topological feature extraction features specifically include: Shape interaction: Count the shape type and quantity of adjacent polygons; Distance interaction: Measure the horizontal, vertical, and minimum Euclidean distance between adjacent polygons; Diagonal relationship: When the vertices of two polygons overlap in the horizontal or vertical direction but do not share an edge, a diagonal relationship is defined.

5. The geometric topology-aware lithographic hotspot detection method based on graph contrast learning as described in claim 2, characterized in that, The graph embedding generation method is specifically a graph embedding generation method based on graph transformers and graph convolutional networks, and its steps include: Representing the map as a graph structure ; A graph converter is used to perform weighted aggregation of node neighbors through an attention mechanism to achieve node attribute encoding; A graph convolutional network is used to aggregate the features of neighboring nodes to generate node embedding vectors; Multi-layer graph converters are overlaid to achieve the fusion of local geometric dependencies and global topological interactions; Where the node represents the first The layer update formula is: in, For attention aggregation function, It is a multilayer perceptron.

6. The geometric topology-aware lithographic hotspot detection method based on graph contrastive learning as described in claim 2, characterized in that, The specific method of the graph comparison learning method is as follows: Given a layout embedding representation as ,in For the number of nodes, For node feature dimensions; By performing max pooling on the nodes, a graph-level representation is obtained. This vector gathers the geometric and topological information of the layout, and can effectively express subtle structural differences; Subsequently, the graph-level representation is input into a single layer. Network, and through Layer output prediction vector , representing the probability that a sample is a hotspot or a non-hotspot, respectively; During training, positive sample pairs are constructed. Compared with negative sample pairs ,in It belongs to the hot topic category. It belongs to the non-hot topic category; The training objective is to learn loss through comparison. Optimization aims to bring similar samples closer together in the feature space and disperse samples from different classes: in Cosine similarity; Simultaneously, a weighted cross-entropy loss is designed: And combined with a weighted combination of contrastive learning loss: This is to fully consider the importance and proportion of positive and negative samples during the training process.

7. The geometric topology-aware lithographic hotspot detection method based on graph contrastive learning as described in claim 6, characterized in that, The specific implementation method of the hotspot re-detection method is as follows: The samples predicted as hotspots by the graph contrast learning method are extracted and, together with the real hotspot samples and obvious non-hotspot samples, are used to construct a refined training dataset. First, each candidate hotspot is converted into an image representation and input into a ResNetV2 network for feature extraction. Then, the extracted feature vectors are mapped to predicted probabilities through fully connected layers and softmax layers. and , respectively representing the first The probability that a sample belongs to a hotspot or a non-hotspot; The loss function is designed as follows: 。

Images