Layout lithography hotspot detection method and system based on multi-directional feature enhancement network

Abstract

The present invention provides a method and system for detecting hotspots in lithography layouts based on a multi-directional feature enhancement network, and relates to the technical field of lithography hotspot detection. The present invention constructs a multi-directional feature enhancement network for the detection task of lithography layouts, which includes: a feature extraction unit, a feature fusion unit, and a classification prediction unit; wherein the feature extraction unit includes: 4 feature extraction layers; each feature extraction layer includes: a head convolution layer, a multi-directional feature enhancement layer, a channel-priority convolution attention layer, and a residual connection layer. The present invention can improve the sensitivity and robustness of the network, and can also effectively reduce false alarms, improve the detection accuracy of lithography layouts, and achieve a balance between the detection accuracy and generalization performance of multiple types of defects in complex layout scenarios.

Description

Technical Field

[0001] The present invention relates to the field of lithography hotspot detection technology, and more specifically, to: 1. a layout lithography hotspot detection method based on a multi-directional feature enhancement network; 2. a layout lithography hotspot detection system based on a multi-directional feature enhancement network. Background Art

[0002] Lithography hotspots are specific areas of the lithography process where defects (such as shorts, opens, and pattern distortion) are likely to occur due to limitations in the design pattern or process conditions. These defects can easily lead to functional failures and reduced yield in semiconductor integrated circuit (IC) chips. Therefore, to ensure optimal final performance and yield of IC chips, lithography hotspot detection is essential prior to actual IC manufacturing.

[0003] Existing methods for detecting lithography hotspots primarily include: 1. Lithography simulation; 2. Combining pattern matching with machine learning. Method 1 is not adaptable to highly complex design layouts and is insufficient for comprehensive layout-wide inspection. While method 2 addresses the technical limitations of method 1, its specific design is complex and challenging. Furthermore, some existing solutions have demonstrated difficulty balancing detection accuracy and generalization performance for multiple defect types in complex layout scenarios. Summary of the Invention

[0004] Based on this, it is necessary for the present invention to address the problem that the existing method of detecting lithography hotspots by combining pattern matching with machine learning cannot balance the detection accuracy and generalization performance of multiple types of defects in complex layout scenarios, and provide a layout lithography hotspot detection method and system based on a multi-directional feature enhancement network.

[0005] The present invention is achieved by adopting the following technical solutions:

[0006] In a first aspect, the present invention discloses a layout lithography hotspot detection method based on a multi-directional feature enhancement network, comprising:

[0007] Step 1: Get the photolithography pattern to be tested P ;

[0008] Step 2: P Input the trained multi-directional feature enhancement network for processing to obtain the detection results Result .

[0009] Among them, the multi-directional feature enhancement network includes: feature extraction part, feature fusion part, and classification prediction part.

[0010] 1. Feature extraction unit includes: 4 feature extraction layers. m The output of the feature extraction layer is taken as them +1 input for feature extraction layer; m ∈[1,2,3]; the input of the first feature extraction layer is P ; The output of the fourth feature extraction layer is multi-directional features X .

[0011] The feature extraction layer includes: head convolution layer, multi-directional feature enhancement layer, channel-priority convolution attention layer, and residual connection layer.

[0012] In any feature extraction layer: the head convolution layer is used to perform 3×3 convolution processing on the input of the feature extraction layer, the multi-directional feature enhancement layer is used to extract and enhance multi-directional features on the output of the head convolution layer, the channel-priority convolution attention layer is used to perform spatial and channel attention calculations on the output of the multi-directional feature enhancement layer, and the residual connection layer is used to add the input of the feature extraction layer to the output of the channel-priority convolution attention layer to obtain the output of the feature extraction layer.

[0013] 2. Feature fusion unit is used to X Perform multi-scale fusion to obtain multi-scale features Y .

[0014] 3. The classification prediction unit is used to Y Perform classification prediction to obtain Result .

[0015] This layout lithography hotspot detection method based on a multi-directional feature enhancement network implements the method or process according to an embodiment of the present disclosure.

[0016] In a second aspect, the present invention discloses a layout lithography hotspot detection system based on a multi-directional feature enhancement network, which uses the layout lithography hotspot detection method based on a multi-directional feature enhancement network disclosed in the first aspect.

[0017] The layout lithography hotspot detection system based on multi-directional feature enhancement network includes: a layout acquisition module and a hotspot detection module.

[0018] The layout acquisition module is used to obtain the lithography layout to be tested P The hotspot detection module is used to P Input the trained multi-directional feature enhancement network for processing to obtain the detection results Result .

[0019] This layout lithography hotspot detection system based on a multi-directional feature enhancement network implements the method or process according to the embodiment of the present disclosure.

[0020] In a third aspect, the present invention discloses a computer program product, comprising a computer program. When executed by a processor, the computer program implements the steps of the layout lithography hotspot detection method based on a multi-directional feature enhancement network as disclosed in the first aspect.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The present invention constructs a multi-directional feature enhancement network for the detection task of lithography patterns, which can improve the sensitivity and robustness of the network, effectively reduce false alarms, improve the detection accuracy of lithography patterns, and achieve a balance between the detection accuracy and generalization performance of multiple types of defects in complex layout scenarios.

[0023] 2. The multi-directional feature enhancement network of the present invention not only uses convolution layers in different directions in the multi-directional feature enhancement layer to extract multi-scale features of the layout in the feature extraction part so that the network can focus on the directional features that are crucial to the target edge and shape in the lithography layout, but also combines the features extracted in different directions in the channel dimension to create a more comprehensive feature map; wherein the high-frequency directional feature input is multiplied in the attention processing layer to enhance the relevant features; a channel-priority convolution attention mechanism is added to each feature extraction layer to enhance the feature extraction ability of the model, so that the network can pay more attention to important areas, thereby making the network more refined; and then the feature fusion part is used to enhance the network's perception of the corresponding features in the context, so that the network can pay attention to the detailed directional features and overall features of the target at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 Flowchart of a layout lithography hotspot detection method based on a multi-directional feature enhancement network provided in Example 1 of the present invention;

[0026] Figure 2 for Figure 1 The network structure diagram of the feature extraction part;

[0027] Figure 3 for Figure 2 Network structure diagram of the multi-directional feature enhancement layer;

[0028] Figure 4 for Figure 3 Network structure diagram of the attention processing layer;

[0029] Figure 5 for Figure 3 The network structure diagram of the feature capture layer;

[0030] Figure 6 for Figure 1 The network structure diagram of the feature fusion unit;

[0031] Figure 7 for Figure 1 Network structure diagram of the classification prediction part. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component. When a component is considered to be "fixed to" another component, it may be directly fixed to the other component or there may be a central component.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used herein in the specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "or / and" as used herein includes any and all combinations of one or more of the associated listed items.

[0035] First of all, it should be noted that after a principle analysis of the existing lithography hotspot detection methods, it was found that the coordination between feature extraction and feature fusion is not good: traditional feature decomposition strategies are often limited to a single frequency band or simple directional division, and fail to effectively integrate high-frequency local details and low-frequency global layout information, resulting in insufficient detection sensitivity for complex process deviations.

[0036] Based on this conclusion, the present invention proposes a new architecture design for feature extraction and feature fusion, and provides a multi-directional feature enhancement network (abbreviated as Mdfe-Net).

[0037] Example 1

[0038] See Figure 1, which shows a flow chart of a layout lithography hotspot detection method based on a multi-directional feature enhancement network provided in this embodiment 1, which includes:

[0039] Step 1: Get the photolithography pattern to be tested P ;

[0040] Step 2: P Input the trained multi-directional feature enhancement network for processing to obtain the detection results Result .

[0041] One of the core aspects of this invention is the provision of a newly designed multi-directional feature enhancement network. It is important to note that this invention requires the use of a trained Mdfe-Net with optimized network parameters. The Mdfe-Net training process can be referenced to conventional neural network training methods and will not be further elaborated here.

[0042] like Figure 1 As shown in the figure, it also shows the architectural design of Mdfe-Net, including: feature extraction unit, feature fusion unit, and classification prediction unit.

[0043] Ⅰ. Feature extraction part aims to P Processed into multi-directional features X .

[0044] See Figure 2 , the feature extraction part includes: 4 feature extraction layers. Among them, the input of the first feature extraction layer is P ;No. m The output of the feature extraction layer is taken as the m +1 input for feature extraction layer; m ∈[1,2,3]; the output of the fourth feature extraction layer is multi-directional features X .

[0045] That is to say, the feature extraction unit can be regarded as performing four stages of feature extraction, which can be expressed as follows:

[0046] ;

[0047] Where Resnet-Block(.) represents the feature extraction layer.

[0048] like Figure 2 As shown in the figure, the feature extraction layer includes: head convolution layer, multi-directional feature enhancement layer, channel-priority convolution attention layer, and residual connection layer.

[0049] In general, in any feature extraction layer:

[0050] The head convolution layer is used as the input of the feature extraction layer (set as x) Perform 3×3 convolution processing;

[0051] The multi-directional feature enhancement layer is used to extract and enhance multi-directional features of the output of the head convolutional layer;

[0052] The channel-priority convolutional attention layer is used to perform spatial and channel attention calculations on the output of the multi-directional feature enhancement layer;

[0053] The residual connection layer is used to add the input of the feature extraction layer to the output of the channel-priority convolutional attention layer to obtain the output of the feature extraction layer (set as y ).

[0054] That is to say, for the first feature extraction layer - its x for P ; For the 4th feature extraction layer—— y for X .

[0055] Of course, the processing of the above feature extraction layer can also be expressed as:

[0056] ;

[0057] Where, CPCA (.) represents the channel-priority convolutional attention layer; MDFE (.) represents the multi-directional feature enhancement layer; Represents the head convolutional layer.

[0058] 101. It should be noted that the second core of the present invention lies in the structural design of the multi-directional feature enhancement layer:

[0059] Since hot spots in lithography patterns are often caused by abnormal changes in local structures, which often manifest as subtle local structures, corners, edges, and some irregular local deformations in the image, it is crucial to capture these local features, especially the accurate modeling and recognition of Manhattan polygonal structural features. The multi-directional feature enhancement layer aims to improve the model's ability to capture details in lithography patterns and accurately locate potential hot spots through sophisticated feature extraction and enhancement mechanisms. The design of the multi-directional feature enhancement layer combines multiple directional convolutions, attention processing, and multi-scale feature learning to achieve efficient positioning of hot spots in the complex structure of the lithography pattern.

[0060] See Figure 3 ,The multi-directional feature enhancement layer includes four first-class convolutional layers, three parent product layers, three second-class convolutional layers, four third-class convolutional layers, one feature capture layer, three attention processing layers, and one parent overlay layer.

[0061] In general, in any multi-directional feature enhancement layer:

[0062] Four first-class convolutional layers are used to take the input from the multi-directional feature enhancement layer Input The features of horizontal high frequency, vertical high frequency, diagonal high frequency and low frequency directions are extracted respectively.

[0063] It should be noted that although the convolution kernel specifications of the first-class convolution layer are all 2×2, the specific designs of the convolution kernels of the four first-class convolution layers are different: the first-class convolution layer is designed as: , which corresponds to the horizontal high-frequency direction; the second type of convolutional layer is designed as: , which corresponds to the vertical high-frequency direction; the third type of convolutional layer is designed as: , which corresponds to the diagonal high-frequency direction; the fourth first-class convolutional layer is designed as: , which corresponds to the low-frequency direction. In this way, convolutional layers in different directions capture different local structural features in the image: convolutional layers in the horizontal and vertical high-frequency directions focus on edge information and straight line structures in the image, which often determine the location of potential hotspots in the lithography pattern, especially changes in edges or long strip structures; convolutional layers in the diagonal high-frequency direction capture diagonal features in the image, which are crucial for the oblique line structures that may appear in the lithography pattern; convolutional layers in the low-frequency direction help extract the macroscopic structure of the image to focus on the overall layout, which is particularly important for understanding the global structure of the pattern, especially some pattern information with large-scale changes.

[0064] The first parent product layer is used to combine the horizontal high-frequency features output by the first class convolution layer with the weight coefficient W H Multiplication; the second parent product layer is used to combine the vertical high-frequency features output by the second first-class convolutional layer with the weight coefficient W V Multiplication; the third parent product layer is used to combine the diagonal high-frequency features output by the third first-class convolutional layer with the weight coefficient W D Multiply.

[0065] It should be noted that W H 、 W V 、 W D are all learnable parameters; the weighted operation of the three parent product layers enables the network to automatically learn the relative importance of different directional features during training, and then dynamically adjust the contribution of each directional feature, thereby highlighting the directional information that is most sensitive to potential hotspot locations.

[0066] No. m The second convolutional layer is used to mThe output of the parent product layer is convolved. It should be noted that the convolution kernel size of the second type of convolution layer is 3×3.

[0067] No. m The three types of convolution layers are used to perform convolution processing on the input of the multi-directional feature enhancement layer. It should be noted that the convolution kernel size of the three types of convolution layers is 1×1.

[0068] No. m The attention processing layer is used to m The output of the second-class convolutional layer is m The output of the three convolutional layers is used for feature enhancement. For the attention processing layer, please refer to the subsequent description and will not be expanded here.

[0069] The feature capture layer is used to extract multi-scale features from the low-frequency features output by the fourth first-class convolutional layer. For details about the feature capture layer, see the subsequent description and will not be expanded here.

[0070] The fourth three-class convolutional layer is used to convolve the output of the feature capture layer. It is important to note that the convolution operation here is to adjust the number of channels to ensure that the low-frequency directional features and other directional features are consistent in the channel dimension.

[0071] The parent overlay layer is used to add the outputs of the three attention processing layers and the output of the fourth three-class convolutional layer to obtain the output of the multi-directional feature enhancement layer. Output .

[0072] Of course, the processing of the above multi-directional feature enhancement layer can also be expressed as follows:

[0073] ;

[0074] Where, CAM (.) indicates the attention processing layer; Conv 1(.) indicates a type of convolutional layer; Conv 2(.) indicates the second type of convolutional layer; Conv 3(.) indicates three types of convolutional layers; Inception (.) indicates the feature capture layer.

[0075] In addition, see Figure 4 , for the attention processing layer - it includes: 1 maximum pooling layer, 1 average pooling layer, 1 sub-overlay layer, and 1 sub-product layer.

[0076] In any attention processing layer:

[0077] The maximum pooling layer is used as the input to the attention processing layer in 1 Perform maximum pooling processing;

[0078] The average pooling layer is used to input the attention processing layer in 1. Perform average pooling processing;

[0079] The sub-overlay layer is used to add the output of the maximum pooling layer and the output of the average pooling layer;

[0080] The sub-product layer is used to combine the output of the sub-overlay layer with the input of the attention processing layer. in 2 is multiplied to get the output of the attention processing layer out .

[0081] Of course, the processing of the above attention processing layer can also be expressed as:

[0082] ;

[0083] Where, Maxpool (.) indicates the maximum pooling layer; Avgpool (.) indicates an average pooling layer.

[0084] It should be noted that for any multi-directional feature enhancement layer m Attention processing layer: its input in 1 is the first layer of the multi-directional feature enhancement layer m The output and input of the second convolutional layer in 2 is the first layer of the multi-directional feature enhancement layer m The output of three convolutional layers.

[0085] In other words, the attention layer performs two pooling operations to obtain feature weights, which are then multiplied by the convolutional high-frequency feature input to enhance the relevant features. This allows the horizontal, vertical, and diagonal high-frequency features to be processed by the attention layer, effectively helping the model focus on key areas and further improving detection accuracy.

[0086] In addition, see Figure 5 , for the feature capture layer - it includes: 1 maximum pooling layer, 1 two-class convolution layer, 4 three-class convolution layers, 1 four-class convolution layer, and 1 splicing layer.

[0087] In any feature capture layer:

[0088] The maximum pooling layer is used to capture the input of the feature layer In Perform maximum pooling processing;

[0089] The first three-class convolutional layer is used to perform convolution on the output of the maximum pooling layer;

[0090] The second three-type convolutional layer is used to perform convolution processing on the input of the feature capture layer;

[0091] The four-class convolution layer is used to perform convolution processing on the output of the second three-class convolution layer. It should be noted that the convolution kernel size of the four-class convolution layer is 5×5.

[0092] The third three-type convolutional layer is used to perform convolution processing on the input of the feature capture layer;

[0093] The second type of convolutional layer is used to perform convolution processing on the output of the third type of convolutional layer;

[0094] The fourth three-type convolutional layer is used to perform convolution processing on the input of the feature capture layer;

[0095] The concatenation layer is used to concatenate the output of the first three-class convolution layer, the output of the fourth-class convolution layer, the output of the second-class convolution layer, and the output of the fourth three-class convolution layer to obtain the output of the feature capture layer. Out .

[0096] Of course, the processing of the above feature capture layer can also be expressed as:

[0097] ;

[0098] Where, Concat (.) indicates the splicing layer; Maxpool (.) indicates the maximum pooling layer; Conv 2(.) indicates the second type of convolutional layer; Conv 3(.) indicates three types of convolutional layers; Conv 4(.) indicates four types of convolutional layers.

[0099] Low-frequency directional features are not processed using weights like other directional features. Instead, multi-scale feature extraction is performed through a feature capture layer: the feature capture layer uses multiple convolution kernels of different sizes to capture structural information in the lithography pattern at multiple scales. Especially when processing images with significant scale changes, this multi-scale learning method can provide the model with more comprehensive global information.

[0100] Therefore, by extracting the horizontal high-frequency, vertical high-frequency, diagonal high-frequency and low-frequency features of the image respectively, the directional features in the image can be effectively captured and the model's perception of the image context can be improved. Among them. Horizontal, vertical and diagonal high frequencies help extract detail features, especially edge information, which is very important for the recognition of various straight line structures in the lithography layout. Low frequency can extract global information of the image, highlight the overall structure, and help the network understand the macro layout of the layout. The four directional features will be combined in the channel dimension to form a more comprehensive feature map, so that the network can not only focus on the directional features of the target (such as details of edges and corners), but also retain the overall structural information, thereby improving the model's recognition ability for different types of lithography layout structures. In addition, by defining learnable weights (W H 、 W V 、 W D ) to optimize the contribution of the three directional features - these weights enable the network to automatically adjust the impact of each channel on the final output during training, thereby ensuring that the network can accurately highlight the most important information in each direction during hotspot detection, further improving the accuracy and robustness of the model.

[0101] In summary, the feature extraction at each stage is equipped with a multi-directional feature enhancement layer, which extracts components in different directions by directional decomposing the feature map and weights them through learnable parameters. This enables the network to adaptively capture important features in each direction - in the lithography pattern, details in different directions may represent different types of defects. The weighted feature map is sent to the attention processing layer for further processing - it reduces the dimension through convolution and generates spatial attention, allowing the network to focus on key areas in the image, thereby enhancing sensitivity to important information. The channel-priority convolutional attention layer was then introduced, which provides the network with stronger feature representation capabilities through its spatial and channel attention mechanisms - in the detection of hotspots in the lithography pattern, details in certain areas may be crucial for the identification of defects, and channels represent different types of features. In the channel-priority convolutional attention layer, the spatial attention mechanism can help the network focus on important areas in the image, especially potential hotspots, while the channel attention mechanism enables the network to pay more attention to key feature information by weighting different feature channels. This alternating weighting of space and channels can effectively improve the network's perception of key areas and features, especially when facing complex or small defects, and can improve detection accuracy.

[0102] 102. The channel-priority convolutional attention layer aims to enhance the expressiveness of feature maps through spatial and channel attention mechanisms, thereby improving the performance in hotspot detection in lithography patterns.

[0103] In summary, the channel-priority convolutional attention layer begins by using the spatial attention mechanism to further enhance the features of spatially important regions. It then introduces the channel-priority attention mechanism to improve the perception of key features by weighting different channels. The weighting process of the spatial and channel-priority attention mechanisms alternates between optimizing the two dimensions of the feature map. The spatial attention mechanism focuses on specific regions of the image, while the channel-priority attention mechanism helps the network identify which feature channels are critical for the final hotspot detection task. By combining these two attention mechanisms, the channel-priority convolutional attention layer effectively improves the network's perception of key regions and features, thereby optimizing the network's feature representation.

[0104] Because the channel-priority convolutional attention layer is located at the end of each stage, the network can focus more on important areas, improving the recognition accuracy of complex lithography patterns and potential defects. This makes the entire network structure more refined and can provide higher accuracy and stronger robustness in practical applications.

[0105] II. Feature Fusion is aimed at X Perform multi-scale fusion to obtain multi-scale features Y .

[0106] See Figure 6 ,The feature fusion part includes: 2 middle convolution layers, 3 hole convolution layers, 1 global average pooling layer, and 1 splicing layer.

[0107] In general, in the feature fusion part:

[0108] The first middle convolutional layer is used to X Perform 1×1 convolution processing.

[0109] 3 dilated convolutional layers are used to X Correspondingly, features at three different scales are captured. It should be noted that while the three dilated convolutional layers use convolution kernels of similar specifications, their dilation ratios are different: the first dilated convolutional layer has a 3×3 kernel size and a dilation ratio of 6; the second has a 3×3 kernel size and a dilation ratio of 12; and the third has a 3×3 kernel size and a dilation ratio of 18. This allows for the perception of contextual information at different scales: dilated convolutional layers with smaller dilation ratios focus on local information, while those with larger dilation ratios capture broader contextual information.

[0110] The global average pooling layer is used to X Perform global average pooling processing;

[0111] The concatenation layer is used to concatenate the output of the first middle convolutional layer, the output of the three dilated convolutional layers, and the output of the global average pooling layer;

[0112] The second middle convolutional layer is used to perform 1×1 convolution on the output of the concatenated layer to obtain Y .

[0113] Of course, the processing of the above feature fusion part can also be expressed as follows:

[0114] ;

[0115] Where, Conv 1×1 (.) indicates the middle convolutional layer; Conv 3×3,6 (.) represents the first dilated convolutional layer; Conv 3×3,12 (.) represents the second dilated convolutional layer; Conv 3×3,18 (.) indicates the third dilated convolutional layer; Imagepooling(.) indicates the global average pooling layer.

[0116] The feature fusion unit, through a combination of dilated convolution and global average pooling, enables the network to effectively integrate multi-scale information. This not only improves the perception of complex patterns and large-scale defects, but also more accurately identifies hotspots at different scales, significantly enhancing hotspot detection in lithography patterns. The feature fusion unit is placed after the feature extraction unit and before the classification prediction unit, further optimizing the network's feature representation and enabling efficient learning and recognition in multi-scale contexts.

[0117] III. Classification prediction part is used for Y Perform classification prediction to obtain Result .

[0118] See Figure 7 The classification prediction part includes: 1 tail convolution layer, 1 batch normalization layer, 1 activation function layer, 1 global average pooling layer, 1 random inactivation layer, and 1 fully connected layer.

[0119] The tail convolution layer is used to Y Perform 3×3 convolution processing;

[0120] The batch normalization layer is used to perform batch normalization on the output of the tail convolutional layer;

[0121] The activation function layer is used to process the output of the batch normalization layer through the Relu activation function;

[0122] The global average pooling layer is used to perform global average pooling on the output of the activation function layer;

[0123] The random dropout layer is used to process the output of the global average pooling layer to prevent overfitting;

[0124] The fully connected layer is used to map the output of the random inactivation layer to obtain Result .

[0125] Need to explain, Result That is, whether the representation is P A lithography hotspot is detected in the P Classify to category 1 (label set to hotspot); if not, P Classify into category 2 (label set to nohotspot).

[0126] Of course, the processing of the above classification prediction part can also be expressed as follows:

[0127] ;

[0128] Where, Conv 3×3 (.) indicates the tail convolution layer; BN (.) represents the batch normalization layer; Relu (.) represents the activation function layer; Imagepooling (.) represents the global average pooling layer; dropout (.) indicates a random dropout layer; FC (.) indicates a fully connected layer.

[0129] Simulation Verification

[0130] To demonstrate the effectiveness and superiority of this method (referred to as Mdfe-Net), this Example 1 conducted a comparative experiment:

[0131] Based on two known benchmark datasets ICCAD 2012 and ICCAD 2019, we divided the training set and test set into three comparison scenarios, as shown in Table 1.

[0132] Table 1 Comparison scenarios

[0133]

[0134] In the table, Train HS represents the number of hotspots in the training set; Train NHS represents the number of non-hotspots in the training set; Test HS represents the number of hotspots in the test set; and Test NHS represents the number of non-hotspots in the test set.

[0135] For scenario 1, traditional machine learning methods (including: SPIE'15, ICCAD'16, TCAD'18) and deep learning methods (MsDenseNet'24) are introduced for comparison. The results are shown in Table 2.

[0136] Table 2 Comparison results 1

[0137]

[0138] Table 2 shows that Mdfe-Net performs exceptionally well on the ICCAD 2012 dataset: its accuracy reaches 99.6%, a significant improvement over other methods; its false alarm rate is 15.5%, significantly lower than other methods; and its F1 score reaches 0.72, significantly higher than other methods. This demonstrates that Mdfe-Net has an advantage in balancing precision and recall.

[0139] For scenarios 2 and 3, traditional machine learning methods (including DAC'19 and TCAD'18) were introduced for comparison. The results are shown in Table 3.

[0140] Table 3 Comparison results 2

[0141]

[0142] From Table 3 we can see that:

[0143] In scenario 2, Mdfe-Net achieved an accuracy of 85.2%, outperforming both TCAD'18 and DAC'19. Although Mdfe-Net had a false positive rate of 3.2%, its F1 score of 0.73 (higher than both TCAD'18 and DAC'19) demonstrated its overall competitive performance.

[0144] In scenario three, Mdfe-Net achieved an accuracy of 92.9%, significantly outperforming both TCAD'18 and DAC'19. Its false positive rate was 83.4%, significantly lower than those of TCAD'18 and DAC'19. Mdfe-Net achieved an F1 score of 0.67 (higher than both TCAD'18 and DAC'19), demonstrating a strong performance in terms of balanced performance.

[0145] In summary, Mdfe-Net has high sensitivity and robustness, high detection accuracy, and achieves a balance between detection accuracy and generalization performance for multiple types of defects in complex layout scenarios.

[0146] Example 2

[0147] This embodiment 2 provides a layout lithography hotspot detection system based on a multi-directional feature enhancement network, which uses the layout lithography hotspot detection method based on a multi-directional feature enhancement network provided in embodiment 1.

[0148] The layout lithography hotspot detection system based on multi-directional feature enhancement network includes: a layout acquisition module and a hotspot detection module.

[0149] The layout acquisition module is configured to: obtain the lithography layout to be tested P .

[0150] The hotspot detection module is configured to: P Input the trained multi-directional feature enhancement network for processing to obtain the detection results Result .

[0151] Since this system uses the layout lithography hotspot detection method based on the multi-directional feature enhancement network in Example 1, it also has the same effect and will not be repeated here.

[0152] Example 3

[0153] This embodiment 3 discloses a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of the layout lithography hotspot detection method based on a multi-directional feature enhancement network disclosed in embodiment 1 are implemented.

[0154] Computer devices can be either mobile or fixed terminals. Examples of the former include mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals). Examples of the latter include digital TVs and desktop computers.

[0155] This embodiment 3 also discloses a readable storage medium, which stores computer program instructions. When the computer program instructions are read and executed by a processor, the steps of the layout lithography hotspot detection method based on the multi-directional feature enhancement network disclosed in embodiment 1 are executed.

[0156] Among them, the readable storage medium may include, but is not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0157] This embodiment 3 further discloses a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the layout lithography hotspot detection method based on a multi-directional feature enhancement network disclosed in embodiment 1 are implemented.

[0158] It should be noted that the computer program for executing the above-mentioned operations can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The above-mentioned computer program can be executed entirely on the user's computer, partially on the user's computer, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN).

[0159] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A layout lithography hotspot detection method based on a multi-directional feature enhancement network, characterized in that: include: Step 1: Get the photolithography pattern to be tested P ; Step 2: P Input the trained multi-directional feature enhancement network for processing to obtain the detection results Result ; Among them, the multi-directional feature enhancement network includes: Feature extraction unit, which includes: 4 feature extraction layers; m The output of the feature extraction layer is taken as the m +1 input for feature extraction layer; m ∈[1,2,3]; the input of the first feature extraction layer is P ; The output of the fourth feature extraction layer is multi-directional features X ; The feature extraction layer includes: a head convolution layer, a multi-directional feature enhancement layer, a channel-priority convolution attention layer, and a residual connection layer; in any feature extraction layer: the head convolution layer is used to perform 3×3 convolution processing on the input of the feature extraction layer, the multi-directional feature enhancement layer is used to perform multi-directional feature extraction and enhancement on the output of the head convolution layer, the channel-priority convolution attention layer is used to perform spatial and channel attention calculations on the output of the multi-directional feature enhancement layer, and the residual connection layer is used to add the input of the feature extraction layer to the output of the channel-priority convolution attention layer to obtain the output of the feature extraction layer; the multi-directional feature enhancement layer includes: 4 first-class convolution layers, 3 parent product layers, 3 second-class convolution layers, 4 third-class convolution layers, 1 feature capture layer, 3 attention processing layers, and 1 parent overlay layer; 4 first-class convolution layers are used to obtain the output of the feature extraction layer from the input of the multi-directional feature enhancement layer. Input The features of horizontal high frequency, vertical high frequency, diagonal high frequency and low frequency directions are extracted respectively; Feature fusion unit, which is used to X Perform multi-scale fusion to obtain multi-scale features Y ;as well as Classification prediction unit, which is used to Y Perform classification prediction to obtain Result .

2. The layout lithography hotspot detection method based on a multi-directional feature enhancement network according to claim 1, characterized in that: In any multi-directional feature enhancement layer: the first parent product layer is used to combine the horizontal high-frequency features output by the first class convolution layer with the weight coefficient W H Multiplication; the second parent product layer is used to combine the vertical features output by the second first-class convolutional layer with the weight coefficient W V Multiplication; the third parent product layer is used to combine the diagonal high-frequency features output by the third first-class convolutional layer with the weight coefficient W D multiplication; m The second convolutional layer is used to m The output of the first parent product layer is convolved; m The three types of convolutional layers are used to perform convolution processing on the input of the multi-directional feature enhancement layer; m The attention processing layer is used to m The output of the second-class convolutional layer is m The output of the three types of convolutional layers is used for feature enhancement; The feature capture layer is used to perform multi-scale feature extraction on the low-frequency features output by the fourth one-class convolution layer; the fourth three-class convolution layer is used to perform convolution processing on the output of the feature capture layer; the parent overlay layer is used to add the output of the three attention processing layers and the output of the fourth three-class convolution layer to obtain the output of the multi-directional feature enhancement layer. Output .

3. The layout lithography hotspot detection method based on a multi-directional feature enhancement network according to claim 2, characterized in that: The attention processing layer includes: 1 maximum pooling layer, 1 average pooling layer, 1 sub-overlay layer, and 1 sub-product layer; In any attention processing layer: the maximum pooling layer is used for the input of the attention processing layer in 1 performs maximum pooling; the average pooling layer is used as the input to the attention processing layer in 1 performs average pooling processing; the sub-overlay layer is used to add the output of the maximum pooling layer and the output of the average pooling layer; the sub-product layer is used to add the output of the sub-overlay layer to the input of the attention processing layer in 2 is multiplied to get the output of the attention processing layer out ; Among them, for any multi-directional feature enhancement layer m Attention processing layer: its input in 1 is the first layer of the multi-directional feature enhancement layer m The output and input of the second convolutional layer in 2 is the first layer of the multi-directional feature enhancement layer m The output of three convolutional layers.

4. The layout lithography hotspot detection method based on a multi-directional feature enhancement network according to claim 2, characterized in that: The feature capture layer includes: 1 maximum pooling layer, 1 two-class convolution layer, 4 three-class convolution layers, 1 four-class convolution layer, and 1 splicing layer; In any feature capture layer: the maximum pooling layer is used for the input of the feature capture layer In Perform maximum pooling processing; the first three-category convolution layer is used to perform convolution processing on the output of the maximum pooling layer; the second three-category convolution layer is used to perform convolution processing on the input of the feature capture layer; the fourth-category convolution layer is used to perform convolution processing on the output of the second three-category convolution layer; the third three-category convolution layer is used to perform convolution processing on the input of the feature capture layer; the second-category convolution layer is used to perform convolution processing on the output of the third three-category convolution layer; the fourth three-category convolution layer is used to perform convolution processing on the input of the feature capture layer; the splicing layer is used to splice the output of the first three-category convolution layer, the output of the fourth-category convolution layer, the output of the second-category convolution layer, and the output of the fourth three-category convolution layer to obtain the output of the feature capture layer Out .

5. The layout lithography hotspot detection method based on a multi-directional feature enhancement network according to claim 4, characterized in that: The convolution kernel size of the first type of convolution layer is 2×2; the convolution kernel size of the second type of convolution layer is 3×3; the convolution kernel size of the third type of convolution layer is 1×1; and the convolution kernel size of the fourth type of convolution layer is 5×5.

6. The layout lithography hotspot detection method based on a multi-directional feature enhancement network according to claim 1, characterized in that: The feature fusion unit includes: 2 middle convolution layers, 3 hole convolution layers, 1 global average pooling layer, and 1 splicing layer; In the feature fusion part: the first middle convolutional layer is used to X Perform 1×1 convolution processing; 3 hole convolution layers are used to X The features of three scales are captured accordingly; the global average pooling layer is used to X Perform global average pooling processing; the splicing layer is used to splice the output of the first middle convolution layer, the output of the three void convolution layers, and the output of the global average pooling layer; the second middle convolution layer is used to perform 1×1 convolution processing on the output of the splicing layer to obtain Y .

7. The layout lithography hotspot detection method based on a multi-directional feature enhancement network according to claim 6, characterized in that: The convolution kernel size of the first atrous convolution layer is 3×3, and the dilation ratio is 6; the convolution kernel size of the second atrous convolution layer is 3×3, and the dilation ratio is 12; the convolution kernel size of the third atrous convolution layer is 3×3, and the dilation ratio is 18.

8. The layout lithography hotspot detection method based on a multi-directional feature enhancement network according to claim 1, characterized in that: The classification prediction part includes: 1 tail convolution layer, 1 batch normalization layer, 1 activation function layer, 1 global average pooling layer, 1 random inactivation layer, and 1 fully connected layer; The tail convolution layer is used to Y Perform 3×3 convolution processing; the batch normalization layer is used to perform batch normalization processing on the output of the tail convolution layer; the activation function layer is used to process the output of the batch normalization layer through the Relu activation function; the global average pooling layer is used to perform global average pooling processing on the output of the activation function layer; the random inactivation layer is used to process the output of the global average pooling layer to prevent overfitting; the fully connected layer is used to map the output of the random inactivation layer to obtain Result .

9. A layout lithography hotspot detection system based on a multi-directional feature enhancement network, characterized in that: It uses the layout lithography hotspot detection method based on a multi-directional feature enhancement network as described in any one of claims 1 to 8; The layout lithography hotspot detection system based on the multi-directional feature enhancement network includes: Layout acquisition module, which is used to obtain the lithography layout to be tested P ;as well as Hotspot detection module, which is used to P Input the trained multi-directional feature enhancement network for processing to obtain the detection results Result .

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the layout lithography hotspot detection method based on a multi-directional feature enhancement network are implemented as described in any one of claims 1 to 8.

Images