Photoetching hot spot detection method and device based on dynamic subgraph feature fusion

Abstract

The invention discloses a photoetching hot spot detection method and device based on dynamic subgraph feature fusion, and relates to the technical field of integrated circuits, and the method comprises the following steps: orthogonally cutting a polygon in a to-be-detected photoetching layout slice sample into rectangles, taking the central point of each rectangle as a node, and constructing to-be-detected graph structure data in combination with embedded features associated with each node; determining the radius of a sub-graph according to the photoacid diffusion length and the local graph density associated with the to-be-measured photoetching layout slice sample; and through the trained dynamic sub-graph aggregation graph kernel convolutional network model, according to the sub-graph radius, constructing a sub-graph of each node by using the to-be-detected graph structure data, extracting features based on each sub-graph to carry out hot spot detection, and outputting a photoetching hot spot detection result. Based on the scheme, the radiuses of the sub-graphs are dynamically divided by adopting photoetching process parameters to reduce empirical static hypothesis errors, visual decomposition and complex structures of the analysis graphs are carried out in a sub-graph feature extraction and fusion mode to enhance the interpretability of hot spot detection, and the reliability of hot spot detection is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of integrated circuits, and in particular to a photolithography hotspot detection method and device based on dynamic subgraph feature fusion. BACKGROUND

[0002] Photolithography is a key technology in the manufacture of ultra-large-scale integrated circuits (ICs), and its process is to transfer the design patterns on a mask to a silicon wafer through exposure. Due to the influence of optical diffraction, some patterns on the mask will have imaging differences with the target topological structure, resulting in bridging, pinching or missing areas, etc. photolithography hotspots. In order to avoid the occurrence of hotspot defects, hotspot detection is introduced in the layout design stage to modify the layout to ensure the production cycle and yield.

[0003] Traditional hotspot detection methods mainly include pattern matching and machine learning. The pattern matching-based method matches the hotspot area by establishing a hotspot library for comparison, which relies on the hotspot library and thus has insufficient generalization ability. The machine learning-based method completes detection by learning the implicit relationship between the layout features designed by humans and their attributes, but it is difficult to capture nanoscale complex layouts due to the implementation method of the artificially designed features. With the rapid development of deep learning, deep learning-based methods have shown processing advantages compared to traditional methods. For example, the hotspot detection method based on convolutional neural networks (CNN) has shown significant detection accuracy and processing speed advantages due to the pattern recognition ability of the convolutional layer automatic feature extraction. However, as integrated circuit manufacturing develops towards ultra-large scale, and the input segment of the layout for hotspot detection usually has very high image resolution, the CNN-based method shows poor computational efficiency. At the same time, the graph neural network (GNN) method based on vector graph feature extraction can significantly improve the hotspot detection efficiency, but it has the limitations of high false positive rate (non-hotspot samples are falsely reported as hotspots) and weak interpretability, making it difficult to clearly understand the influence of the structural features of the input graph on the output, and its results are questioned whether it can reasonably explain the hotspot determination for a specific circuit layout. Therefore, it is necessary to improve the detection reliability of photolithography hotspots. SUMMARY

[0004] The present application provides a photolithography hotspot detection method and device based on dynamic subgraph feature fusion, which solves the technical problem of low detection reliability of existing photolithography hotspot detection methods.

[0005] The first aspect of the present application provides a photolithography hotspot detection method based on dynamic subgraph feature fusion, comprising: orthogonally cutting the polygons in the to-be-tested photolithography layout slice sample into rectangles, taking the center points of each rectangle as nodes, and constructing to-be-tested graph structure data in combination with the embedding features associated with each node; determine a subgraph radius according to a photoacid diffusion length and a local pattern density associated with the to-be-tested photolithography layout slice sample; construct subgraphs of each node using the to-be-tested graph structure data according to the subgraph radius through the trained dynamic subgraph aggregated graph kernel convolutional network model, and perform hotspot detection based on features extracted from each subgraph to output a photolithography hotspot detection result.

[0006] Further, the to-be-tested photolithography layout slice sample is orthogonally cut into rectangles, and the center point of each rectangle is taken as a node, and to-be-tested graph structure data is constructed by combining embedded features associated with each node, including: cut a plurality of polygons in the to-be-tested photolithography layout slice sample into rectangles along horizontal and vertical directions; take the center point of each rectangle as a node, and construct a quadrant region centered on each node; for any node, nodes in each quadrant region that are closest to the node and do not belong to the same polygon are taken as adjacent nodes to establish edges; determine node embedded features of each node and edge embedded features of each edge; construct to-be-tested graph structure data of the to-be-tested photolithography layout slice sample using each node, each edge, each node embedded feature, and each edge embedded feature.

[0007] Further, the subgraph radius is determined according to the photoacid diffusion length and the local pattern density associated with the to-be-tested photolithography layout slice sample, including: determine a local pattern density by performing ratio operation on a pattern block area in the to-be-tested photolithography layout slice sample and a total area of the sample; perform weighted summation operation on the local pattern density and the photoacid diffusion length associated with the to-be-tested photolithography layout slice sample to output a subgraph radius.

[0008] Further, the trained dynamic subgraph aggregated graph kernel convolutional network model includes an input layer, a graph kernel convolutional layer, and a multilayer perceptron; and the to-be-tested graph structure data is constructed using the to-be-tested graph structure data according to the subgraph radius through the trained dynamic subgraph aggregated graph kernel convolutional network model, and features are extracted based on each subgraph to perform hotspot detection and output a photolithography hotspot detection result, including: the to-be-tested graph structure data input into the trained dynamic subgraph aggregated graph kernel convolutional network model is received by the input layer and transmitted to the graph kernel convolutional layer; in the cascaded multilayer graph kernel convolutional layer, for any layer of graph kernel convolutional layer, subgraphs of each node are constructed using the to-be-tested graph structure data according to the subgraph radius, and corresponding feature maps are determined by performing feature extraction on each subgraph; The features are aggregated, and the aggregated features are input into a multi-layer perception machine for linear transformation and softmax operation, and a photolithography hotspot detection result is output.

[0009] Further, the embedding features include node embedding features and edge embedding features. The node embedding features include node coordinates, a length of a rectangle in which the node is located, and a width of the rectangle in which the node is located. The edge embedding features include an edge length.

[0010] Further, the determination process of the trained dynamic subgraph aggregated graph kernel convolutional network model includes: A training graph structure dataset and a test graph structure dataset are constructed based on a photolithography plate picture segment sample set; The training graph structure dataset is input into a dynamic subgraph aggregated graph kernel convolutional network model to be trained for hotspot detection, a cross-entropy loss function is used to calculate a loss function value, and model parameters are iteratively optimized based on the loss function value to determine an intermediate dynamic subgraph aggregated graph kernel convolutional network model; The test graph structure dataset is input into the intermediate dynamic subgraph aggregated graph kernel convolutional network model to determine an evaluation index; If the evaluation index meets an index threshold, the intermediate dynamic subgraph aggregated graph kernel convolutional network model is determined as the trained dynamic subgraph aggregated graph kernel convolutional network model.

[0011] The second aspect of the present application provides a photolithography hotspot detection device based on dynamic subgraph feature fusion, which includes: A graph data construction module is configured to orthogonally cut polygons in a photolithography plate slice sample into rectangles, use center points of the rectangles as nodes, and construct a to-be-tested graph structure data by combining embedding features associated with the nodes. A radius dynamic division module is configured to determine a subgraph radius based on a photoacid diffusion length and a local pattern density associated with the to-be-tested photolithography plate slice sample. A hotspot detection module is configured to construct subgraphs of the nodes by using the to-be-tested graph structure data according to the subgraph radius through a trained dynamic subgraph aggregated graph kernel convolutional network model, extract features based on the subgraphs, and output a photolithography hotspot detection result.

[0012] The third aspect of the present application provides a computer device including a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the photolithography hotspot detection method based on dynamic subgraph feature fusion.

[0013] The fourth aspect of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed to realize the photolithography hotspot detection method based on dynamic subgraph feature fusion according to any one of the above.

[0014] The fifth aspect of the present application provides a computer program product, which comprises computer programs / instructions, and the computer programs / instructions are executed by a processor to realize the photolithography hotspot detection method based on dynamic subgraph feature fusion according to any one of the above.

[0015] From the above technical solutions, the present application has the following advantages: The above scheme of the present application provides a photolithography hotspot detection method based on dynamic subgraph feature fusion, which comprises: cutting polygons in a to-be-detected photolithography layout slice sample into rectangles, taking the center points of the rectangles as nodes, and constructing to-be-detected graph structure data in combination with embedded features associated with the nodes; determining a subgraph radius according to a photoacid diffusion length and a local pattern density associated with the to-be-detected photolithography layout slice sample; constructing subgraphs of the nodes by using the to-be-detected graph structure data according to the subgraph radius through a trained dynamic subgraph aggregation graph kernel convolution network model, and performing hotspot detection based on features extracted from the subgraphs, and outputting a photolithography hotspot detection result. Based on the above scheme, the subgraph radius is dynamically divided based on physical driving of photolithography process parameters, which reduces the error of an empirical static assumption, the complex structure of a graph is visualized and analyzed through a subgraph feature extraction and fusion mode, the hotspot detection explainability is enhanced, and thus the hotspot detection reliability is improved as a whole. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0017] Figure 1 A step flowchart of a photolithography hotspot detection method based on dynamic subgraph feature fusion provided for the first embodiment of the present application is shown in the figure. Figure 2 A schematic diagram of graph structure data provided for the first embodiment of the present application is shown in the figure. Figure 3 A processing process schematic diagram of a dynamic subgraph aggregation graph kernel convolution network model provided for the first embodiment of the present application is shown in the figure. Figure 4 A processing process schematic diagram of a graph kernel convolution layer provided for the first embodiment of the present application is shown in the figure. Figure 5A hotspot / non-hotspot sample example schematic diagram provided for the first embodiment of the present application; Figure 6 A basic principle schematic diagram of a classical CNN, a classical GNN and a subgraph kernel convolution method provided for the first embodiment of the present application. Figure 7 A structural block diagram of a photolithography hotspot detection device based on dynamic subgraph feature fusion provided for the second embodiment of the present application. DETAILED DESCRIPTION

[0018] The embodiment of the present application provides a photolithography hotspot detection method and device based on dynamic subgraph feature fusion, and is used for solving the technical problem that the detection reliability of an existing photolithography hotspot detection method is low.

[0019] In order to make the application purpose, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the following described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0020] Term explanation: Graph: graph, a data structure.

[0021] CNN: convolutional neural network.

[0022] GNN: graph neural network.

[0023] KerGNN: graph kernel neural network model.

[0024] GAT: graph attention network, a variant of GNN.

[0025] GIN: graph isomorphism network, a variant of GNN.

[0026] DSGK: dynamic subgraph aggregation graph kernel convolution network model.

[0027] ODL: optical acid diffusion length.

[0028] Accuracy: an industry-recognized hotspot detection performance indicator, representing the proportion of correctly predicted samples in all samples.

[0029] Recall rate: an industry-recognized hotspot detection performance indicator, representing the ratio between correctly predicted positive samples and all positive samples (in this embodiment, the positive samples are hotspot samples, and the negative samples are non-hotspot samples).

[0030] False Positive Number: an industry-recognized hotspot detection performance indicator, representing the number of non-hotspot samples that are falsely reported as hotspot samples.

[0031] False Positive Rate: (False Positive Number / Non-Hotspot Sample Number) * 100%.

[0032] See Figure 1 The embodiment one of the present application provides a photolithography hotspot detection method based on dynamic subgraph feature fusion, comprising: Step 101: cutting the polygons in the to-be-tested photolithography layout slice sample into rectangles, taking the center points of the rectangles as nodes, and constructing to-be-tested graph structure data by combining the embedded features associated with the nodes.

[0033] To-be-tested, representing that the current is being detected.

[0034] Photolithography layout refers to a pattern file used for photolithography, containing complex polygons that map physical structures such as transistors and wires. Photolithography layout slice sample refers to a sample graph obtained by dividing a local area from a complete photolithography layout.

[0035] Embedded features refer to features representing the attributes of an object.

[0036] Graph structure data refers to a data structure composed of a set of nodes and a set of edges connecting the nodes.

[0037] In one specific embodiment of the present embodiment, step 101 comprises the following sub-steps: Cutting a plurality of polygons in the to-be-tested photolithography layout slice sample into rectangles along the horizontal and vertical directions; Taking the center points of the rectangles as nodes and constructing quadrant regions centered on each node; For any node, respectively establishing edges between the nodes in each quadrant region that are closest to the node and do not belong to the same polygon; Determining the node embedded features of each node and the edge embedded features of each edge; Using each node, each edge, each node embedded feature, and each edge embedded feature, constructing the to-be-tested graph structure data of the to-be-tested photolithography layout slice sample.

[0038] In one specific embodiment of the present embodiment, the embedded features include node embedded features and edge embedded features; The node embedded features include node coordinates, the length of the rectangle where the node is located, and the width of the rectangle where the node is located; The edge embedded features include edge length.

[0039] It should be noted that this embodiment uses the KerGNN model as a basis to set up a dynamic subgraph aggregation graph kernel convolution network model (DSGK-Dynamic Subgraph Aggregation Graph Kernel Convolution Network) for hotspot detection. For this purpose, the lithography pattern slice sample to be tested needs to be abstracted into a data form represented by a graph to obtain the graph structure data to be tested. Graph data consists of two parts: a node feature matrix and an edge index. The node matrix is ​​a two-dimensional tensor, where each row corresponds to a node and each column corresponds to its feature. The edge index is a two-dimensional tensor containing the start node and the end node, used to represent the connection relationship in the graph. like Figure 2 As shown, when constructing the specific graph structure data, in the sample slice of the photomask to be tested, each polygon is rectangularly cut along the horizontal and vertical directions to obtain multiple independent rectangles. It can be understood that the horizontal and vertical directions are aligned with the extension directions of the polygon's sides to facilitate cutting into rectangles; the center point of the rectangle is designated as a node. By forming edges through the pairwise connections between related nodes, the photolithography pattern slice sample under test is transformed into a corresponding undirected graph. , For a set of nodes, Let be the set of edges, containing Each node and Edge; where each node The embedding is described by three attributes: node embedding includes node coordinates. Length of the rectangle containing the node Width of the rectangle containing the node ,like Figure 2 As shown, For the first The coordinates of each node. index of a node , For the first The length of the rectangle containing each node For the first The width of the rectangle containing each node; each side Connecting the center points of adjacent rectangles, the edge embedding is represented by an attribute, namely the edge length, which indicates the shortest distance between the center points of the two rectangles, such as... Figure 2 As shown, For the first The node and the first Edges of each node, index of a node , the edge length of the edge between the first node and the second node; the edge length of the edge between the first node and the second node; the edge length of the edge between the first node and the second node; In one specific implementation, when constructing the edge of any node , the embodiment adopts a quadrant neighborhood filtering strategy, first establishes coordinate axes centered on the node to divide 4 quadrant regions, and takes the nearest node of the adjacent rectangle in each quadrant region as the corresponding adjacent node, but if the nearest node is located in the same polygon, the quadrant region is not connected, for example, as shown in Figure 2 , the 4 quadrant nearest nodes of the node are all located in other polygons, so all are connected to establish edges, the first quadrant of the node has no adjacent node, the nearest node of the second quadrant of the node is the same polygon, so it is not connected, and the third quadrant and the fourth quadrant of the node both establish edges with the corresponding nodes. It can be understood that if two nodes are connected to form an edge, they are adjacent nodes.

[0040] Step 102, determining the subgraph radius according to the photoacid diffusion length and the local pattern density associated with the to-be-tested photolithography layout slice sample.

[0041] Photoacid diffusion length refers to the diffusion distance of the photoacid generated after the photoresist is exposed and baked.

[0042] Local pattern density refers to the distribution density of physical structures such as transistors and wires (i.e., polygons) in the photolithography layout slice sample.

[0043] Subgraph radius refers to the radius of a specific subgraph in a graph. It can be understood as the range size of a local subgraph centered on a node, including all nodes and edges within a certain range around the node.

[0044] It should be noted that the subgraph division of the traditional graph kernel neural network usually adopts a fixed hop neighbor range or a static radius based manner, which ignores the physical constraints specific to photolithography. Since the hot spot is caused by the optical proximity effect, the subgraph extraction range should take into account the physical property of the photoacid diffusion length. In order to avoid missing detection in high-density areas of the photolithography pattern due to too small subgraphs, or including too many irrelevant patterns in low-density areas due to too large subgraphs, the local pattern density is set to measure the pattern distribution density. Based on the above considerations, the photoacid diffusion length and the local pattern density are used to dynamically divide the subgraph range based on the physical law of photolithography imaging. The model structure no longer depends on empirical setting, which can reduce the error of empirical assumption and avoid the limitation of traditional CNN and GNN methods which only rely on pattern texture feature extraction, thereby improving the physical interpretability of the model and reducing the false positive rate.

[0045] In one specific embodiment of the present embodiment, step 102 includes the following sub-steps: The local pattern density is determined by performing a ratio operation on the pattern block area in the to-be-tested photolithography pattern slice sample and the total sample area. The subgraph radius is output by performing a weighted summation operation based on the photoacid diffusion length associated with the to-be-tested photolithography pattern slice sample and the local pattern density.

[0046] The pattern block area refers to the area occupied by the patterns (i.e., polygons) of physical structures such as transistors and wires.

[0047] The total sample area refers to the area of the entire photolithography pattern slice sample, which includes the pattern block area and the blank area in the photolithography pattern slice sample.

[0048] It should be noted that the dynamic radius subgraph extraction proposed in the present embodiment calculates the subgraph range based on the photoacid diffusion length, considers the photolithography process conditions such as exposure dose and wavelength, ensures that the neighborhood covers the area actually affected by light scattering, and sets a density factor according to the area ratio of the photolithography pattern. In specific implementation, the calculation process includes:

[0049]

[0050]

[0051] In the formula, is the subgraph radius, is the photoacid diffusion length (typical 28nm process node is 20-30nm, 7nm / 5nm process node is 10-15nm), is a diffusion length weight (taking an empirical value of 1.5-2.0), is a local pattern density, is a density compensation coefficient (taking an empirical value of 0.1-0.3), is a photoacid diffusion coefficient, is a post-exposure bake time, is a pattern block area, is a total sample area, and the local pattern density represents a proportion of the pattern block area to the sample area.

[0052] Taking an IC CAD 2012 lithography layout slice as an example, the process node is 28 nm, and the mainstream DUV (193 nm wavelength) photoacid diffusion length is about 25 nm, taking an empirical value of 1.8, taking an empirical value of 0.2, the layout is a hotspot layout and is in a high-density area, and the calculation is 0.6, and the subgraph radius of the lithography layout is: , and then the subgraph is extracted with the node of each rectangle as the center and with a radius .

[0053] Step 103, constructing subgraphs of each node by using the to-be-tested graph structure data according to the subgraph radius through the trained dynamic subgraph aggregation graph kernel convolution network model, and performing hotspot detection based on feature extraction of each subgraph, and outputting a lithography hotspot detection result.

[0054] It should be noted that, after the dynamic subgraph aggregation graph kernel convolution network model is built, the model training is carried out to determine the trained dynamic subgraph aggregation graph kernel convolution network model, based on the graph structure detection principle of the KerGNN model, the subgraph is constructed based on the to-be-tested graph structure data according to the dynamically determined subgraph radius, the feature extraction and aggregation classification are performed to realize the hotspot detection, and thus the lithography hotspot detection result of the hotspots and non-hotspots in the graph is obtained.

[0055] In one specific embodiment of this embodiment, the trained dynamic subgraph aggregation graph kernel convolution network model includes an input layer, a graph kernel convolution layer, and a multilayer perceptron; and step 103 includes the following substeps: The to-be-tested graph structure data of the trained dynamic subgraph aggregation graph kernel convolution network model is input, received by the input layer, and transmitted to the graph kernel convolution layer; In the cascaded multilayer graph kernel convolution layer, for any layer of the graph kernel convolution layer, the subgraph of each node is constructed by using the to-be-tested graph structure data according to the subgraph radius, and the corresponding feature mapping is determined by performing feature extraction on each subgraph; After aggregating the feature maps, the data is input into a multilayer perceptron for linear transformation, followed by softmax operation, and the output is the lithographic hotspot detection result.

[0056] It should be noted that the dynamic subgraph aggregation graph kernel convolutional network model includes an input layer, graph kernel convolutional layers, and a multilayer perceptron (MLP); for example, it consists of one input layer, three graph kernel convolutional layers, and one two-layer multilayer perceptron, such as... Figure 3 and Figure 4 As shown: First, the structure data of the test pattern is constructed from slice samples of the photolithography pattern to be tested. It is sent to the input layer; Subsequently, in a single graph kernel convolutional layer, for each node... Center and radius The nodes and edges within the graph constitute a subgraph ,composition Sub-image Similar to the convolutional filters in CNNs, graph kernel convolutional layers use trainable hidden graphs. Perform subgraph feature extraction. For the first A graph convolution filter, The index of the graph convolution filter (which can be represented as the graph filter) The corresponding mapping dimensions, in specific implementation It can take the value 64). The number of convolutional filters in the graph, each Each contains a trainable adjacency matrix. and node attributes Based on subgraph Through graph kernel function Get each node exist Feature mapping in 3D space ,Right now ; will the first Feature maps extracted by layer graph kernel convolutional layers Passed to the next graph kernel convolutional layer, This serves as the hierarchical index for the graph kernel convolutional layers, continuing until feature extraction from three graph kernel convolutional layers is complete; Understandably, graph kernel functions A random walk kernel function can be used, which determines the similarity between two graphs by calculating the inner product between them. The process of the random walk kernel function includes:

[0057] In the formula, The first image is shown. The second image is shown. For the step count index, The total number of steps taken. For weighted sequences, Represents a series of weight values, It is an inner product graph. Let be the adjacency matrix of the inner product graph. index of a node , index of a node When calculating of When raised to the power of 1, its 1st power is 10. , The value of the nth element represents the nth element in the graph. and the There is a length between the two nodes. The number of public walking paths; After processing through three layers of graph kernel convolutional layers, for each layer, summation pooling is used to map the 64 features of each node. Summing yields the full-image features of each layer. To fully utilize the information from each layer of the model, the full-image features from all layers are then stitched together to obtain the total full-image features. The input is then fed into a multilayer perceptron, where it undergoes a linear transformation and is mapped to a two-dimensional output vector. This is represented by unique hotspot / non-hotspot labels, which are then normalized using softmax to obtain the lithographic hotspot detection results.

[0058] Understandably, by introducing a kernel function perspective through this subgraph feature aggregation method based on graph kernel convolution, the convolution method of standard CNN is extended to the graph domain. Each subgraph uses a hidden graph to extract local structural information, thereby constructing a multi-layer structure with better graph classification performance. Therefore, by visualizing the topology of the trained graph convolution filter, the local structural features of the input graph are further revealed. Compared with the traditional message aggregation method of GNN, this design significantly improves the interpretability and transparency of hotspot detection.

[0059] In one specific implementation of this embodiment, the process of determining the trained dynamic subgraph aggregation graph kernel convolutional network model includes: A training graph structure dataset and a test graph structure dataset were constructed based on a sample set of photolithography image segments. The training graph structure dataset is input into the dynamic subgraph aggregation graph kernel convolutional network model to be trained for hotspot detection. The cross-entropy loss function is used to calculate the loss function value, and the model parameters are iteratively optimized based on the loss function value to determine the intermediate dynamic subgraph aggregation graph kernel convolutional network model. The intermediate dynamic subgraph aggregated graph kernel convolutional network model is inputted into the test graph structure dataset to determine the evaluation index. If the evaluation index meets the index threshold, the intermediate dynamic subgraph aggregated graph kernel convolutional network model is taken as the trained dynamic subgraph aggregated graph kernel convolutional network model.

[0060] It should be noted that the evaluation index can include accuracy, recall rate, false positive number, or / and recall rate, and the specific index threshold can be set according to the corresponding evaluation index; Exemplarily, based on the ICCAD2012 competition benchmark dataset, the sample enhancement technology is used to solve the imbalance problem of positive and negative samples, for example, each hot spot sample is rotated by four angles (0 degrees, 90 degrees, 180 degrees, and 270 degrees), and each rotated instance is mirrored, so that each hot spot instance is expanded to eight copies. At the same time, the number of negative samples is reduced by randomly sampling non-hot spot data. Finally, the processed dataset contains eight times expanded hot spot data and part of non-hot spot data, and the proportion of hot spot and non-hot spot samples is balanced, so as to obtain a photomask picture segment sample set, which includes multiple hot spot photomask picture segment samples and non-hot spot photomask picture segment samples as shown in the following table. Figure 5 The photomask picture segment sample set is divided into a photomask picture segment sample training set and a photomask picture segment sample test set as shown in the following table. Table 1 ICCAD2012 original dataset and processed dataset statistics

[0061] According to step 101, the training graph structure dataset of the photomask picture segment sample training set and the test graph structure dataset of the photomask picture segment sample test set are constructed based on the PyTorch Geometric (PyG) library. Each graph (Graph) is marked as a label hot spot / non-hot spot, represented by a two-dimensional vector one-hot encoding. Each training sample and test sample is dynamically calculated according to step 102. The subgraph radius is calculated accordingly. The corresponding parameters are adjusted according to the sample process size and the type of photolithography machine. Model training and testing are performed on a hardware platform equipped with an Intel Xeon(R) Silver 4210R 2.4 GHz CPU, a GTX 4090 graphics card, and 64 GB of memory. The PyTorch deep learning framework is built. Each time, a graph structure data is inputted into the model for processing. Since the total number of center points of each sample is inconsistent, the number of input layer nodes is set to the maximum number of center points in all samples (in the ICCAD2012 dataset, the maximum number of center points is 75). For graphs with less than 75 nodes, the excess node attribute vector value is 0. In the model training stage, the training graph structure data of the training graph structure data set is input into the dynamic subgraph aggregated graph kernel convolutional neural network model, the cross entropy loss of the one-hot code of the hot spot / non-hot spot classification label and the model prediction probability is used as the cost function, and the adjacency matrix and node attribute of each layer of hidden graph are iteratively updated by back propagation, to obtain the trained model. For example, the initial learning rate of the model is set to 0.001, and 128 samples are trained in each batch; In the model evaluation stage, the test graph structure data of the test graph structure data set is input into the trained model for forward calculation, and the class with the maximum prediction probability of the model is compared with the classification label to evaluate the trained model. If the evaluation meets the expectation, the model is directly used as the trained model. If the expectation is not met, the model training is continued. In one verification of the embodiment, the forward calculation is performed, and the class with the maximum prediction probability of the model is compared with the classification label. When the training reaches 150 rounds, the model training effect is optimal, the hot spot detection accuracy reaches 98.8%, and the false positive rate is only 6%.

[0062] In order to better illustrate the technical effects of the embodiment, the DSGK model proposed in the embodiment is compared with the classic CNN method and the advanced GNN method in recent years. The core principle process of each method can be referred to Figure 6 As shown in the figure, the experimental data set comes from ICCAD2012, and the size of each detection version picture segment sample is 3.6*3.6um. Table 2 shows the experimental result comparison data: Table 2 Comparison of experimental results of the embodiment and other hot spot detection methods on ICCAD2012

[0063] In the CNN method, ResNet18 (a classic CNN model), TCAD’19, DAC’19 and TCAD’22 are used. Among them, the scheme of TCAD’19 proposes to convert the image into a super-resolution image with a self-defined channel number by using a feature tensor, to avoid spatial feature loss and improve the hot spot detection accuracy. The scheme of DAC’19 uses a binary neural network (BNN) for classification. The scheme of TCAD’22 applies attention modules and deep metric learning to hot spot and non-hot spot feature extraction. In the GNN method, GAT (graph attention model), GIN (graph isomorphism network model), DATE’22 and the DSGK of the embodiment are used. Among them, DATE’22 divides the layout pattern into rectangles, constructs a graph structure, and applies a GNN model for hot spot detection to greatly improve the detection efficiency. According to the results shown in Table 2, it can be seen that: 1) Compared with the CNN method, the model size of this embodiment (0.332M) is reduced by 1-3 orders of magnitude compared with the CNN model, and the model volume is compressed from MB to KB. The detection speed is still 25 times faster than the fastest BNN model of DAC'19. The detection accuracy is only 0.4% lower than DAC'19. The hotspot detection accuracy is comparable, but the number of false positives is the lowest, which is 243 fewer than TCAD'22, which has the fewest false positives. 2) Compared with other GNN methods, it has the lowest number of false positives; its accuracy is 0.4% higher than DATE'22, and significantly higher than GAT and GIN. Its inference time is 0.8s slower than GIN, but faster than both DATE'22 and GAT. In chip manufacturing, detection speed is a critical metric, and high-efficiency inference speed helps shorten the chip manufacturing cycle. Compared with other GNN methods, this technology has the highest accuracy and the fewest false alarms, indicating that it can effectively reduce misjudgments and lower subsequent processing costs in practical applications. Its inference speed is comparable to other GNN methods. Overall, it achieves a good balance between accuracy and efficiency. In summary, this embodiment provides an interpretable and highly efficient detection paradigm for improving the yield of advanced chip processes, and is particularly suitable for the increasingly severe optical complexity challenges in processes below 5nm.

[0064] In this embodiment of the invention, the physical drive of photolithography process parameters is used to dynamically divide the sub-graph radius, which reduces the error of empirical static assumptions. The complex structure of the graph is visualized and analyzed by sub-graph feature extraction and fusion, which enhances the interpretability of hotspot detection and thus improves the overall reliability of hotspot detection.

[0065] Please see Figure 7 The second embodiment of the present invention provides a lithographic hotspot detection device based on dynamic subgraph feature fusion, comprising: The graph data construction module 701 is used to orthogonally cut the polygons in the lithography pattern slice sample to be tested into rectangles, and use the center point of each rectangle as a node to construct the graph structure data to be tested by combining the embedded features associated with each node. The radius dynamic division module 702 is used to determine the sub-map radius based on the photoacid diffusion length and local pattern density associated with the photolithography pattern slice sample under test; The hotspot detection module 703 is used to construct subgraphs for each node using the structure data of the graph under test according to the subgraph radius through a trained dynamic subgraph aggregation graph kernel convolutional network model, and to perform hotspot detection based on the features extracted from each subgraph, and output the lithographic hotspot detection results.

[0066] In one specific embodiment of this example, the graph data construction module 701 is specifically used for: Cutting a plurality of polygons in a photolithography layout slice sample to be tested into rectangles along horizontal and vertical directions; Adopting the center points of the respective rectangles as nodes, and constructing a quadrant region with each node as the center; For any node, respectively establishing edges with the nodes in the respective quadrant regions that are closest to the node and do not belong to the same polygon; Determining node embedding features of the nodes and edge embedding features of the edges; Adopting the nodes, the edges, the node embedding features, and the edge embedding features to construct a to-be-tested graph structure data of the photolithography layout slice sample to be tested.

[0067] In one specific implementation of the embodiment, the radius dynamic division module 702 is specifically configured to: Adopting the ratio operation of the area of the graphic block in the photolithography layout slice sample to be tested and the total area of the sample to determine the local graphic density; Based on the weighted summation operation of the local graphic density and the photoacid diffusion length associated with the photolithography layout slice sample to be tested, the subgraph radius is output.

[0068] In one specific implementation of the embodiment, the trained dynamic subgraph aggregated graph kernel convolutional network model includes an input layer, a graph kernel convolutional layer, and a multilayer perceptron; the hotspot detection module 703 is specifically configured to: For the input to-be-tested graph structure data of the trained dynamic subgraph aggregated graph kernel convolutional network model, the input layer receives and transmits the to-be-tested graph structure data to the graph kernel convolutional layer; In the cascaded multilayer graph kernel convolutional layer, for any layer of the graph kernel convolutional layer, the subgraph of each node is constructed according to the subgraph radius using the to-be-tested graph structure data, and the feature mapping corresponding to each subgraph is determined by performing feature extraction on each subgraph; After aggregating the respective feature mappings, performing linear transformation on the aggregated feature mappings using the multilayer perceptron, and performing softmax operation, a photolithography hotspot detection result is output.

[0069] In one specific implementation of the embodiment, the embedding features include node embedding features and edge embedding features; The node embedding features include node coordinates, the length of the rectangle in which the node is located, and the width of the rectangle in which the node is located; The edge embedding features include edge length.

[0070] In one specific implementation of the embodiment, the model training module is further configured to: Based on the photolithography layout picture segment sample set, a training graph structure data set and a test graph structure data set are constructed; The training graph structure dataset is input into the dynamic subgraph aggregation graph kernel convolutional network model to be trained for hotspot detection, a cross-entropy loss function is used to calculate a loss function value, and the model parameters are iteratively optimized based on the loss function value to determine an intermediate dynamic subgraph aggregation graph kernel convolutional network model; The test graph structure dataset is input into the intermediate dynamic subgraph aggregation graph kernel convolutional network model to determine an evaluation index; If the evaluation index meets an index threshold, the intermediate dynamic subgraph aggregation graph kernel convolutional network model is taken as the trained dynamic subgraph aggregation graph kernel convolutional network model.

[0071] Embodiment three of the present application also provides a computer device including a memory and a processor, and the memory stores a computer program; the computer program is executed by the processor to make the processor execute the steps of the lithography hotspot detection method based on dynamic subgraph feature fusion according to any one of the above embodiments.

[0072] Embodiment four of the present application also provides a computer readable storage medium, which stores a computer program / instruction, and the computer program / instruction is executed by a processor to implement the steps of the lithography hotspot detection method based on dynamic subgraph feature fusion according to any one of the above embodiments.

[0073] Embodiment five of the present application also provides a computer program product, which includes a computer program / instruction, and the computer program / instruction is executed by a processor to implement the steps of the lithography hotspot detection method based on dynamic subgraph feature fusion according to any one of the above embodiments.

[0074] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described device and module can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0075] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed ones can be indirect coupling or communication connection through some interfaces, devices or modules, and can be electrical, mechanical or other forms.

[0076] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, i.e., may be located in one place, or may be distributed to multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0077] In addition, each functional module in each embodiment of the application can be integrated into a processing module, or each module can exist physically alone, or two or more modules can be integrated into one module. The integrated module can be realized in the form of hardware or in the form of a software functional module.

[0078] The integrated module, if realized in the form of a software functional module and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the application, essentially or the part that contributes to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0079] The above-described embodiments are only used to illustrate the technical solutions of the application, rather than limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to part of the technical features; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A photolithography hotspot detection method based on dynamic subgraph feature fusion, characterized in that, The method comprises the steps of: cutting polygons in a photolithography pattern slice sample into rectangles orthogonally, taking the center points of each rectangle as a node, and constructing a test graph structure data in combination with embedded features associated with each node; determining a subgraph radius according to the photoacid diffusion length and the local pattern density associated with the test photolithography pattern slice sample; constructing a subgraph of each node using the test graph structure data according to the subgraph radius through a trained dynamic subgraph aggregated graph kernel convolutional network model, and performing hotspot detection based on each subgraph to output a photolithography hotspot detection result.

2. The method for lithography hotspot detection based on dynamic subgraph feature fusion according to claim 1, characterized in that, The method of cutting polygons in a photolithography pattern slice sample into rectangles orthogonally, taking the center points of each rectangle as a node, and constructing a test graph structure data in combination with embedded features associated with each node comprises: cutting a plurality of polygons in a test photolithography pattern slice sample into rectangles along the horizontal and vertical directions; taking the center points of each rectangle as a node and constructing a quadrant region centered on each node; for any node, establishing edges with the nodes in each quadrant region that are closest to the node and do not belong to the same polygon; determining node embedding features of each node and edge embedding features of each edge; constructing a test graph structure data of the test photolithography pattern slice sample using each node, each edge, each node embedding feature, and each edge embedding feature.

3. The method of claim 1, wherein, The method of determining a subgraph radius according to the photoacid diffusion length and the local pattern density associated with the test photolithography pattern slice sample comprises: performing ratio operation on the area of a pattern block in the test photolithography pattern slice sample and the total area of the sample to determine the local pattern density; performing weighted summation operation on the local pattern density and the photoacid diffusion length associated with the test photolithography pattern slice sample to output a subgraph radius.

4. The method of claim 1, wherein, The trained dynamic subgraph aggregated graph kernel convolutional network model comprises an input layer, a graph kernel convolutional layer, and a multilayer perceptron; and the method of constructing a subgraph of each node using the test graph structure data according to the subgraph radius through the trained dynamic subgraph aggregated graph kernel convolutional network model, and performing hotspot detection based on each subgraph to output a photolithography hotspot detection result comprises: receiving and transmitting the test graph structure data input into the trained dynamic subgraph aggregated graph kernel convolutional network model through the input layer to the graph kernel convolutional layer; in the cascaded multilayer graph kernel convolutional layer, for any layer of graph kernel convolutional layer, constructing a subgraph of each node using the test graph structure data according to the subgraph radius, and performing feature extraction on each subgraph to determine the corresponding feature mapping; inputting the aggregated feature mappings into the multilayer perceptron after linear transformation and performing softmax operation to output a photolithography hotspot detection result.

5. The method of claim 1, wherein, The embedded features comprise node embedding features and edge embedding features; The node embedding features comprise node coordinates, the length of the rectangle where the node is located, and the width of the rectangle where the node is located; The edge embedding features comprise edge length.

6. The method of claim 1, wherein, The determination process of the trained dynamic subgraph aggregated graph kernel convolutional network model comprises: constructing a training graph structure dataset and a test graph structure dataset based on a photoetching plate picture segment sample set; inputting the training graph structure dataset into a dynamic subgraph aggregation graph kernel convolution network model to be trained to perform hotspot detection, calculating a loss function value by using a cross-entropy loss function, and iteratively optimizing model parameters based on the loss function value to determine an intermediate dynamic subgraph aggregation graph kernel convolution network model; inputting the test graph structure dataset into the intermediate dynamic subgraph aggregation graph kernel convolution network model to determine an evaluation index; if the evaluation index meets an index threshold, regarding the intermediate dynamic subgraph aggregation graph kernel convolution network model as a trained dynamic subgraph aggregation graph kernel convolution network model. 7.A device for detecting a hot spot in photolithography based on dynamic subgraph feature fusion, characterized in that, comprising: a graph data construction module configured to orthogonally cut polygons in a photoetching plate graph slice sample into rectangles, to take center points of the rectangles as nodes, and to construct a to-be-tested graph structure dataset by combining embedded features associated with the nodes; a radius dynamic division module configured to determine a subgraph radius according to a photoacid diffusion length and a local pattern density associated with the to-be-tested photoetching plate graph slice sample; a hotspot detection module configured to construct subgraphs of the nodes by using the to-be-tested graph structure dataset according to the subgraph radius through the trained dynamic subgraph aggregation graph kernel convolution network model, to perform hotspot detection based on features extracted from the subgraphs, and to output a photoetching hotspot detection result.

8. A computer device, comprising: comprising a memory and a processor, the memory storing a computer program, and the computer program being executed by the processor to cause the processor to perform steps of the photoetching hotspot detection method based on dynamic subgraph feature fusion according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement steps of the photoetching hotspot detection method based on dynamic subgraph feature fusion according to any one of claims 1-6.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement steps of the photoetching hotspot detection method based on dynamic subgraph feature fusion according to any one of claims 1-6.

Images