Integrated circuit photoetching hot spot detection method, storage medium and equipment

Abstract

The invention belongs to the field of integrated circuit design, and particularly relates to an integrated circuit photoetching hot spot detection method, a storage medium and equipment. The method comprises the following steps: constructing a layout classification model for identifying whether an input layout contains photoetching hot spots or not; the layout classification model sequentially comprises a front feature extraction module, a plurality of rear feature extraction modules, a feature fusion module and a full connection layer. The front feature extraction module is composed of common convolution, degrouping convolution, interactive convolution and an attention layer. The post-feature extraction module is composed of an enhanced interaction layer and an attention layer. Performing data enhancement on the hot-spot layout and sampling the non-hot-spot layout to form a balanced data set; and training the layout classification model by using the BCE loss introduced with label smoothing as a loss function. And inputting the slice image of the original layout into the trained layout classification model to realize photoetching hot spot detection. According to the invention, the problem that the detection precision and the false alarm rate are unbalanced in the existing scheme is solved.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit design, specifically relating to a method for detecting lithographic hotspots in integrated circuits, and the corresponding storage medium, computer program product, and layout hotspot detection equipment. Background Technology

[0002] Photolithography is the process of transferring patterns from an integrated circuit layout to a wafer. As integrated circuit manufacturing processes continue to shrink, the problems encountered in nanometer-level photolithography are becoming increasingly severe. Detecting photolithographic hotspots and avoiding defective patterns in the early stages of layout design remains a key technology for improving chip production yield. Therefore, engineers have proposed various manufacturable methods, such as design rule checking and resolution enhancement techniques, which can provide reliable solutions to photolithography problems. However, traditional photolithographic hotspot detection methods use corresponding optical models to simulate actual lighting conditions for optical simulation. While this method offers high accuracy and is beneficial for optical proximity correction, it is extremely time-consuming.

[0003] Current mainstream lithographic hotspot detection methods fall into three categories: optical simulation-based hotspot detection methods, template matching-based hotspot detection methods, and deep learning-based hotspot detection methods. While optical simulation-based methods offer high accuracy, they are extremely time-consuming, particularly for large-scale lithography layouts requiring global simulation. Template matching methods are accurate and fast, but rely heavily on existing hotspot databases and cannot detect previously unseen hotspot patterns. Deep learning-based hotspot detection methods learn key features that form hotspots on the lithography layout using appropriate models, enabling them to identify unseen hotspot patterns. However, limited by model design and database sample size, the classification accuracy achieved by this approach is currently below expectations, and false positives are possible. Summary of the Invention

[0004] To address the shortcomings of existing deep learning-based integrated circuit hotspot detection methods in terms of detection accuracy and false alarm rate, this invention provides a method for detecting lithographic hotspots in integrated circuits, along with corresponding storage media, computer program products, and layout hotspot detection equipment.

[0005] This invention is achieved using the following technical solution: A method for detecting photolithographic hotspots in integrated circuits, comprising the following steps: Construct a layout classification model to identify whether an input layout contains lithographic hotspots; the layout classification model includes a pre-feature extraction module, multiple post-feature extraction modules, a feature fusion module, and a fully connected layer.

[0006] The pre-feature extraction module consists of one 7×7 convolution, one ungrouping convolution, one interactive convolution, and one attention layer. The post-feature extraction module consists of one enhanced interactive layer and one attention layer. The ungrouping convolution is used to divide the input features into high-frequency and low-frequency channels according to a preset channel ratio; the feature information of the low-frequency channels is processed by max pooling and convolution to obtain low-frequency features. F L The high-frequency features are obtained by convolution processing of the high-frequency channel feature information. F H In interactive convolution, low-frequency features of the input... F L After convolution and upsampling processing compared with after convolution processing F H Feature concatenation is performed to obtain new high-frequency features. High-frequency features of the input F H After convolution and pooling processing and after convolution processing F L Feature concatenation is performed to obtain new low-frequency features. The enhanced interactive layer in the post-feature extraction module consists of multiple interactive convolutions with residual connections and different kernels. The attention layer calculates attention and weights the high-frequency and low-frequency features of the input to achieve feature enhancement; the feature fusion module fuses the high-frequency and low-frequency features of the input to obtain fused features; the fully connected layer generates classification results based on the fused features of the input.

[0007] The original dataset is constructed by acquiring labeled hot and non-hot maps. Data augmentation is performed on the hot maps, and non-hot maps are sampled to create a balanced dataset. A BCE loss with label smoothing is used as the loss function. The map classification model is trained using the balanced dataset, and model parameters that meet the performance metrics are retained.

[0008] The sliced ​​images of the original layout are input into the trained layout classification model to achieve lithographic hotspot detection.

[0009] As a further improvement of the present invention, in the pre-feature extraction module, the convolution kernels of the convolution modules used in the ungrouping convolution and interactive convolution are both 3×3; the stride of the pooling operation is 2.

[0010] And / or, the expression for the feature processing in interactive convolution is: ; In the above formula, This indicates a convolution operation with a 3×3 kernel; Indicates an upsampling operation; Indicates pooling operation; This indicates feature splicing.

[0011] As a further improvement of the present invention, the enhanced interaction layer consists of two 3×3 first interactive convolutions and one 1×1 second interactive convolution. The enhanced interaction layer contains two branches; the input features are processed sequentially by a second interactive convolution in one branch and by two first interactive convolutions in the other branch; the features output from the two branches are residually concatenated according to the high-frequency channel and the low-frequency channel, respectively, and then used as the output of the enhanced interaction layer.

[0012] As a further improvement of this invention, the attention layers in both the pre-feature extraction module and the post-feature extraction module employ a foreground attention mechanism, Outlook Attention, which incorporates relative spatial location information. Furthermore, the attention layers independently calculate and weight the attention for both high-frequency and low-frequency features input from the previous layer of the network; thereby partially focusing on the influence of features within local regions in the high-frequency features, and introducing more local and even global information into the low-frequency features.

[0013] As a further improvement of this invention, the feature fusion module consists of four multi-branch modules (SAMBB) incorporating spatial attention, one upsampling module (Upsample), one 3×3 convolutional module (Conv3), one 1×1 convolutional module (Conv1), and one channel attention module (CA). In the feature fusion module, the input high-frequency features... F H After being processed sequentially by three SAMBB modules and one Upsample module, the low-frequency characteristics are compared with those processed by one SAMBB module. F L The features are then stitched together by channel; the stitched features are then processed by the Conv3 module, CA module, and Conv1 module to obtain the final fused features.

[0014] As a further improvement of this invention, the SAMBB multi-branch module with spatial attention is composed of a spatial attention module SA and four 3×3 convolutional modules Conv3. In the SAMBB module, the input features are first processed by SA, and the resulting features enter two branches, which are then processed by three convolutional modules and one convolutional module, respectively. The features output from the two branches are concatenated to obtain the final output of SAMBB.

[0015] As a further improvement of this invention, during the training phase, the data augmentation method for hotspot maps includes image rotation and flipping; the augmented image obtained through data augmentation has the same classification label as the original image. Non-hotspot maps are sampled according to a preset ratio so that the data scale of hotspot and non-hotspot maps in the sample dataset of each round of training is similar.

[0016] As a further improvement to this invention, a loss function for label smoothing BCE loss is introduced. L bce as follows: ; In the above formula, y i Indicates the first i The classification label for each map, y=1 or 0; This indicates the category label after applying the smoothing strategy; The input labels are the output of the network model. y i The probability of; N Indicates the number of samples in the map.

[0017] As a further improvement of the present invention, during the testing phase, recall, false positive rate (FAR), and F1 score are selected as performance evaluation metrics for the network model.

[0018] The present invention also includes a storage medium storing a computer program. When executed by a processor, the computer program creates a layout classification model trained by the aforementioned integrated circuit lithography hotspot detection method, and is used to classify whether a slice image of the input original layout belongs to a lithography hotspot.

[0019] The present invention also includes a computer program product comprising a computer program that, when executed by a processor, creates a layout classification model trained by the aforementioned integrated circuit lithography hotspot detection method, and is used to classify whether a slice image of the input original layout belongs to a lithography hotspot.

[0020] The present invention also includes a layout hotspot detection device, which includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it creates a lithographic hotspot detection system; the lithographic hotspot detection system includes a data preprocessing module and a recognition module.

[0021] The data preprocessing module uses Python's pya and klayout libraries to convert the input raw map files in GDS or OAS format into images according to the rule that 1 dBu corresponds to 1 pixel. It then automatically segments the converted raw map into multiple slice images based on the marker layers of the map slices. These segmented slice images are then sequentially input into the recognition module.

[0022] The recognition module runs a layout classification model trained using the aforementioned integrated circuit lithography hotspot detection method. This layout classification model is used to classify whether the input slice image belongs to a lithography hotspot.

[0023] The technical solution provided by this invention has the following beneficial effects: This invention designs a targeted layout classification model by combining the image features of the lithography hotspots in integrated circuits. In the backbone network, this invention extracts different frequency features of the layout more fully through a feature frequency separation strategy. It also incorporates interactions between frequency domain features, residual connections of multiple convolutional kernels, and an attention mechanism based on relative spatial location to better compensate for the shortcomings of convolutional neural networks. This enables the full learning and extraction of multi-level features from lithography heatmaps. In the neck network, this invention enhances the model's receptive field by including a multi-branch fusion module combined with channel and spatial attention mechanisms. It fully fuses high-frequency details and low-frequency structural features extracted from high-frequency and low-frequency channels, enhancing the network model's feature representation capability and avoiding the loss of global information.

[0024] During the training phase, this invention addresses the imbalance between positive and negative samples by employing targeted random sampling and sample augmentation for different types of sample data. Furthermore, it utilizes a binary cross-entropy loss method incorporating label balancing measurements to improve the training process. Ultimately, this achieves high-precision hotspot detection while reducing false positives. On some open-source datasets, it even demonstrates a detection accuracy of up to 99.7% with no false positives, exhibiting excellent detection performance for previously unseen hotspot patterns and good generalization ability. Attached Figure Description

[0025] Figure 1 This is a flowchart of the steps of the integrated circuit photolithography hotspot detection method provided in Embodiment 1 of the present invention.

[0026] Figure 2 This is a network architecture diagram of the layout classification model constructed in Embodiment 1 of the present invention.

[0027] Figure 3 This is a schematic diagram illustrating the principle of ungrouping convolution in Embodiment 1 of the present invention.

[0028] Figure 4 This is a schematic diagram of the principle of interactive convolution in Embodiment 1 of the present invention.

[0029] Figure 5 This is a schematic diagram of the foreground attention mechanism in Embodiment 1 of the present invention.

[0030] Figure 6 This is a schematic diagram of the enhanced interaction layer in Embodiment 1 of the present invention.

[0031] Figure 7 This is a schematic diagram of the feature fusion module provided in Embodiment 1 of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] Example 1 In analyzing existing neural network-based schemes for identifying lithographic hotspots in layouts, this embodiment found that the main factors limiting the performance improvement of these schemes include the following: First, most existing solutions use convolutional neural networks (CNNs) for feature extraction, but CNNs themselves have certain limitations. While convolutional kernels implicitly learn the frequency features of the input data, this learning process typically favors low- to mid-frequency information. The crucial pooling operation actively suppresses high-frequency components, which precisely correspond to sharp edges, minute protrusions, or depressions in the layout—details crucial for hotspot formation. For example, the spacing of specific periodic structures in the layout (corresponding to specific spatial frequencies) easily induces standing wave effects. Isolated and dense patterns have drastically different spectral distributions, resulting in significantly different imaging behaviors; however, standard CNN architectures lack explicit mechanisms to model, analyze, and utilize the global or local frequency information of the input layout.

[0034] Secondly, convolutional neural network approaches use pooling layers or strided convolutions to perform spatial downsampling to expand the receptive field. However, multiple downsampling steps significantly reduce the spatial resolution of deep feature maps. For the tiny (nanoscale), finely detailed defects commonly found in lithographic hotspots, their precise geometric shape and edge information easily become blurred or even lost in deep feature maps, severely impacting the model's localization accuracy and ability to distinguish subtle differences.

[0035] Furthermore, standard CNN convolutional kernels themselves lack rotation invariance. Lithographic hotspots are typically strongly correlated with geometric structures in specific directions (such as 45-degree angled lines or line ends with specific spacing). The model may produce inconsistent responses to image segments with similar structures but different orientations after rotation, leading to unstable detection results. Some lithographic hotspots originate from interactions between spatially distant graphics. The shallow receptive fields of CNN layers are insufficient to effectively capture such long-distance spatial relationships.

[0036] Finally, during network model training, due to intellectual property protection reasons, the scale of real-world map data available is limited, and the sample data mainly consists of slice images that do not contain lithographic hotspots, while the number of slice images containing lithographic hotspots is extremely small. The data scale of different types of map samples is highly unbalanced. Using such sample data for training will lead to difficulties in network model convergence and insufficient generalization.

[0037] To address the aforementioned issues, this embodiment provides a method for detecting lithographic hotspots in integrated circuits. This method employs a novel decomposition convolution and interactive convolution in the feature extraction module, decomposing layout features into high-frequency and low-frequency components. Feature interaction is achieved across different frequency channels, and an attention mechanism is incorporated to introduce relative spatial location information, better compensating for the shortcomings of convolutional neural networks. This allows the network model to better focus on features that may cause lithographic defects due to different frequency components. In the feature fusion stage, this embodiment innovatively introduces a multi-branch fusion strategy that integrates spatial attention and channel attention, and utilizes multiple convolutional kernels to enhance the model's receptive field, avoiding the loss of global information. The improved network in this embodiment, combined with the newly designed training strategy, enables the network model to achieve higher layout hotspot detection accuracy, reduce false positive rates, and improve the generalization ability of the network model.

[0038] Specifically, such as Figure 1 As shown, the method for detecting photolithographic hotspots in integrated circuits provided in this embodiment includes the following steps: S1: Construct a layout classification model to identify whether an input layout contains lithographic hotspots. For example... Figure 2 As shown, the map classification model sequentially includes a pre-feature extraction module, multiple post-feature extraction modules, a feature fusion module, and a fully connected layer. To make the principles and performance of the map classification model constructed in this embodiment clearer, the following provides a detailed explanation of the three key functional modules: the pre-feature extraction module, the post-feature extraction module, and the feature fusion module. (a) Pre-feature extraction module The pre-feature extraction module provided in this embodiment consists of one 7×7 convolution (denoted as Conv), one ungrouping convolution (denoted as DC-Conv), one interactive convolution (denoted as Inter-Conv), and one attention layer. The first 7×7 convolution processes the input layout data through a convolutional module with a large kernel, resulting in a feature map with 64 channels. The obtained feature map can be represented using X∈R C×H×W Let C represent the number of channels of the feature, which is 64 in this embodiment. H and W represent the height and width of the feature map, respectively.

[0039] Next, in the ungrouped convolution input to the feature map extracted by the 7×7 convolution, the module structure of the ungrouped convolution in this embodiment is as follows: Figure 3 As shown, it consists of a max-pooling layer with a step size of 2. Maxpool and two 3×3 convolutional modules (denoted as in this embodiment) Conv3) Composition. The ungrouping convolution is used to divide the input features into high-frequency and low-frequency channels according to a preset channel ratio; in this embodiment, the ratio of the number of channels in the high-frequency and low-frequency channels is 7:1. The feature information divided into the low-frequency channels is then processed by max pooling and convolution to obtain low-frequency features. F L The feature information divided into high-frequency channels is processed by convolution to obtain high-frequency features. F H In this embodiment, feature downsampling of the low-frequency channel is performed using pooling operations, which reduces its spatial size by half, thereby focusing on low-frequency features; while the features of the high-frequency channel are not downsampled, but only conventional convolution is used, so high-frequency features can be focused on.

[0040] After decomposing the convolution, this embodiment employs a newly designed interactive convolution to convolve the inputs of two different frequency components separately to extract different types of features; and introduces cross-frequency domain connections between features of different frequencies to promote feature interaction, then uses nonlinear activation layers to enhance feature representation capabilities, and introduces residual connections for features of different frequencies to optimize training and avoid gradient vanishing or gradient exploding. Specifically, as follows... Figure 4 As shown, the interactive convolution contains four convolutional modules and one upsampling module. Upsample and 1 pooling module Maxpool In this embodiment, the convolutional modules in the interactive convolution all use a 3×3 kernel. Conv 3; while the step size for the upsampling and downsampling operations corresponding to the upsampling and pooling modules is 2.

[0041] In detail, such as Figure 4 As shown, in interactive convolution, the low-frequency features of the input... F L After convolution and upsampling processing compared with after convolution processing F H Feature concatenation is performed to obtain new high-frequency features. High-frequency features of the input F H After convolution and pooling processing and after convolution processing F L Feature concatenation is performed to obtain new low-frequency features. In detail, the feature processing in interactive convolution is expressed as follows: ; In the above formula, This indicates a convolution operation with a 3×3 kernel; Indicates an upsampling operation; Indicates pooling operation; This indicates feature splicing.

[0042] After cross-frequency domain feature interaction processing via interactive convolution, new high-frequency and low-frequency features that simultaneously contain feature information from both low-frequency and high-frequency channels can be obtained. In this embodiment, the foreground attention mechanism Outlook Attention, which incorporates relative spatial location information, is further employed as the attention layer in the pre-feature extraction module. Attention is independently calculated and weighted for the high-frequency and low-frequency features input from the previous layer of the network, thereby achieving feature enhancement.

[0043] Specifically, the foreground attention mechanism used in this embodiment is an improved self-attention mechanism, which can introduce relative spatial position information while avoiding the computationally expensive QKV dot product operation. Its working principle is as follows: Figure 5 As shown in the figure, Outlook Attention employs a sliding window mechanism, which enables fine-grained local encoding of location information. Specifically, for each spatial location in the feature map ( i , j The foreground attention mechanism generates attention weights within a local window (k*k) centered on this point, and then performs a weighted summation within the window. Unlike traditional self-attention computation, the foreground attention mechanism reshapes the generated local window, thus preserving, to some extent, key positional information in the visual task. This window-based reshaping method, which preserves key positional relationships, is particularly suitable for modeling regions containing local features such as inflection points and line edges that influence neighboring areas.

[0044] It is particularly important to emphasize that this embodiment uses OutlookAttention for feature enhancement in both high-frequency and low-frequency channels of the network. The high-frequency channel focuses on the feature influence in local areas, while the low-frequency channel, due to its smaller spatial size, introduces more local and even global information.

[0045] (ii) Post-feature extraction module In the backbone network of the map classification model provided in this embodiment, multiple post-feature extraction layers can be connected after the pre-feature extraction module. The number of post-feature extraction layers in this embodiment can be optimized based on various factors such as model performance and training difficulty in practical applications; in this embodiment, the preferred number of post-feature extraction layers is three. Each post-feature extraction module consists of one enhancement interaction layer and one attention layer. That is, several loops of enhancement interaction layers and attention layers are connected after the attention layer of the pre-feature extraction module.

[0046] In this embodiment, the enhanced interaction layer consists of multiple layers with different convolutional kernels, such as those with residual connections. Figure 4The interactive convolution shown enables more efficient interaction and fusion of extracted high-frequency and low-frequency features. Specifically, as... Figure 6 As shown, the enhanced interaction layer in this embodiment consists of three aforementioned interactive convolutions. Two of these interactive convolutions have 3×3 kernels, referred to as the first interactive convolution, denoted as Inter-Conv3. The third interactive convolution has a 1×1 kernel, referred to as the second interactive convolution, denoted as Inter-Conv1. The enhanced interaction layer contains two branches. The input features are processed sequentially through two second interactive convolutions in one branch and through two first interactive convolutions in the other branch. The features output from the two branches are residually concatenated according to their high-frequency and low-frequency channels, and then used as the output of the enhanced interaction layer. That is, the added interaction layer introduces residual connections between every two interactive convolutions, and these residual connections are implemented using interactive convolutions with smaller internal kernels. It is important to emphasize that the residual connections in the two branches refer to concatenating the feature maps of the high-frequency and low-frequency channels obtained from one branch to the corresponding channel components of the other branch.

[0047] The attention mechanism in the post-feature extraction module is the same foreground attention mechanism as that in the former extraction model. The function and principle of this module can be found in the previous description, and will not be repeated in this embodiment.

[0048] (III) Feature Fusion Module In the backbone network, the original layout, after being processed by a pre-feature extraction module and multiple post-feature extraction modules, yields high-frequency and low-frequency features that enable deep information interaction. To facilitate the fully connected layer's combination of extracted high-frequency details and low-frequency structures for classifying the layout release containing lithographic hotspots, this embodiment uses a feature fusion module to fuse the input high-frequency and low-frequency features to obtain fused features. Specifically, in this embodiment... To effectively fuse high-frequency details and low-frequency structure, this embodiment employs a feature fusion module that incorporates spatial attention, multi-branch modules, and channel attention. For example... Figure 7 As shown, the SAMBB multi-branch module with spatial attention introduced in this embodiment consists of a spatial attention module (SA) and four 3×3 convolutional modules (Conv3). In the SAMBB module, the input features are first processed by the SA, and the spatial attention mechanism guides the network to focus on potential hotspot key regions. The resulting features then enter the multi-branch module, which contains two branches, and are processed by three convolutional modules and one convolutional module in each branch, respectively. The features output from the two branches are concatenated to obtain the final output of SAMBB. The multi-branch module effectively enhances the ability to extract salient details from the feature map.

[0049] Specifically, the feature fusion module used in this embodiment consists of four multi-branch modules (SAMBB) incorporating spatial attention, one upsampling module (Upsample), one 3×3 convolutional module (Conv3), one 1×1 convolutional module (Conv1), and one channel attention module (CA). In the feature fusion module, the input high-frequency features... F H The features are processed sequentially through three SAMBB modules to progressively enhance detailed features, followed by an upsampling module to match the spatial dimension of the low-frequency features. The resulting feature information of the same size is then compared with the low-frequency features processed by one SAMBB module. F L Channel-wise concatenation is performed; the concatenated features are then processed by the Conv3, CA, and Conv1 modules to obtain the final fused features. The first 3×3 convolution is used for initial smoothing, the CA module performs slight feature recalibration based on a general attention mechanism, and the 1×1 convolutional layer enables fine-tuning of the fused features.

[0050] Finally, the fully connected layer generates a classification result based on the input fusion features. Specifically, the fully connected layer is used to map the fusion features extracted, enhanced, and fused from the backbone network and neck network onto a specific probability distribution using a classification function, and thereby determine whether the current map slice image is a hotspot image or a non-hotspot image.

[0051] S2: Obtain labeled hot and non-hot maps to form the original dataset; perform data augmentation on the hot maps and sample the non-hot maps to form a balanced dataset. Use BCE loss with label smoothing as the loss function; train the map classification model using the balanced dataset; retain model parameters that meet the performance metrics.

[0052] Considering that the number of lithographic landmasses without hotspots is far greater than the number of lithographic landmasses containing hotspots in the available landmass slices during the training phase, this embodiment makes special adjustments to the sample data in the training set during the training phase to avoid the network model performance being affected by the imbalance of positive and negative samples. Specifically, on the one hand, this embodiment performs data augmentation on the relatively small number of hotspot landmasses (referring to landmass slice images containing lithographic hotspots) to expand the sample. Data augmentation methods that can be used include rotating the original image at a specified angle and mirroring the original image. The augmented image obtained through data augmentation has the same classification label as the original image. On the other hand, this embodiment also samples non-hotspot landmasses (referring to landmass slice images without lithographic hotspots) according to a preset ratio to ensure that the data scale of hotspot and non-hotspot landmasses in the sample dataset of each training round is similar.

[0053] In practical applications, a 10% random sampling rate can be chosen to ensure that the training sample data remains balanced at each stage. Considering that random sampling may lead to insufficient learning of sample features, this embodiment can also increase the number of training epochs. For example, setting the number of training epochs to 1500 ensures that all samples can be applied to the actual training of the network model, thereby ensuring the efficient utilization of limited sample data.

[0054] To address the severe imbalance between hot and non-hot maps in the map dataset, this embodiment introduces a label smoothing strategy in addition to the binary cross-entropy (BCE) loss used in the training phase to improve the training process. Specifically, the loss function of the label-smoothed BCE loss is... L bce as follows: ; In the above formula, y i Indicates the first i The classification label for each layout is y=1 or 0; y=1 indicates that the layout contains lithographic hotspots, and y=0 indicates that the layout does not contain lithographic hotspots. This indicates the category label after applying the smoothing strategy; The input labels are the output of the network model. y i The probability of; N Indicates the number of samples in the map.

[0055] After training is completed, during the testing phase, this embodiment can select recall, false positive rate (FAR), and F1 score as performance evaluation metrics for the network model.

[0056] S3: Input the sliced ​​images of the original layout into the trained layout classification model to achieve lithographic hotspot detection.

[0057] In practical applications, the original layout design file is usually first converted into image data, and then the large-scale original layout is segmented according to function according to preset rules, resulting in multiple slice images of the original layout. The network model in this embodiment, after training and achieving the required performance, can identify whether each slice image contains lithographic hotspots.

[0058] Example 2 The method for detecting lithographic hotspots in integrated circuits provided in Example 1 is essentially a data processing method. In order to apply this method in practice, this example further provides a storage medium, a computer program product, and a layout hotspot detection device that can implement this method.

[0059] Specifically, the storage medium provided in this embodiment stores a computer program. When the computer program is executed by the processor, it creates a layout classification model trained by the integrated circuit lithography hotspot detection method as described in Embodiment 1, and is used to classify whether a slice image of the input original layout belongs to a lithography hotspot.

[0060] The computer program product provided in this embodiment includes a computer program that, when executed by a processor, creates a layout classification model trained by the integrated circuit lithography hotspot detection method as described in Embodiment 1, and is used to classify whether a slice image of the input original layout belongs to a lithography hotspot.

[0061] The layout hotspot detection device provided in this embodiment includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it creates a lithography hotspot detection system. The lithography hotspot detection system includes two parts: a data preprocessing module and a recognition module. The data preprocessing module uses Python's pya and klayout libraries to convert the input raw layout file in GDS or OAS format into an image according to the rule that 1 dBu corresponds to 1 pixel. It then automatically segments the converted raw layout into multiple slice images based on the marker layer of the layout slices. The segmented slice images are sequentially input into the recognition module. The recognition module runs a layout classification model trained using the integrated circuit lithography hotspot detection method of Embodiment 1. The layout classification model is used to classify whether the input slice image belongs to a lithography hotspot.

[0062] The hotspot detection device provided in this embodiment is essentially a computer device. In practical applications, it samples independent computer devices, such as laptops, tablets, desktop computers, or large and medium-sized computer devices that can execute computer programs, such as rack servers, blade servers, tower servers, or cabinet servers (including independent servers or server clusters composed of multiple servers).

[0063] The computer device in this embodiment includes, but is not limited to, a memory and a processor that can be interconnected via a system bus. In this embodiment, the memory (i.e., the readable storage medium) includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory can be an internal storage unit of the computer device, such as the hard disk or RAM of the computer device. In other embodiments, the memory can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Of course, the memory can also include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. Furthermore, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0064] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of a computer device.

[0065] Simulation test To verify the performance of the integrated circuit lithography hotspot detection method provided by this invention, technicians simulated and trained the relevant scheme and tested the performance of the trained network model.

[0066] I. Dataset and Control Group This experiment tested the performance of the proposed solution on two datasets, ICCAD2012 and ICCAD2019-1. To more intuitively reflect the performance level of the proposed solution, several existing hotspot detection schemes were also selected as control groups for comparative testing. The sample size of each dataset mentioned above is shown in the table below: Table 1: Sample Size of the Sample Dataset

[0067] The control group included ASPDAC*19, TCAD*22, TODAES*22, and GLSVISLI*24. ASPDAC*19 alleviated the problem of insufficient training data through semi-supervised learning, and combined it with a self-paced multi-task learning method to allow the model to learn easy samples first and then gradually learn difficult samples to improve the classifier's discriminative ability. TCAD*22 designed an end-to-end hotspot detector using a fully convolutional network combined with a transfer learning scheme, which greatly reduced the amount of data required and enabled fast detection across the entire chip. TODAES*22 used Bayesian optimized neural network architecture search to dynamically design the hotspot detection model, avoiding manual parameter tuning and architecture design, thus improving design efficiency. GLSVISLI*24, based on EfficientNet, adjusted the number of convolutions and aspect ratio during the process to better preserve fine-grained geometric details, and combined cross-layer connection feature fusion to reduce false positives and improve F1 score.

[0068] II. Experimental Parameter Settings In this experiment, to avoid excessive random fluctuations and ensure that the initial settings of training parameters are completely consistent for each experiment, a fixed random seed was used during the network initialization phase.

[0069] The optimization algorithm used during the training phase was ADAMW. The batch size of the training model was 128. The number of training epochs was set to 1500 in the experiment. The initial learning rate was set to 5e-4, and the learning rate was adjusted by the cosine annealing algorithm with an annealing period of 30 epochs.

[0070] III. Experimental Results and Analysis In this experiment, recall, false alarm rate (FAR), and F1 score were selected as performance evaluation metrics for the network models in this invention and the control group. Recall is the proportion of detected hotspot samples out of the total number of hotspot samples; it can assess the detection accuracy of lithographic hotspots output by each network model. FAR represents the proportion of falsely detected non-hotspot samples out of all non-hotspot samples. The F1 score is the harmonic mean of precision and recall, which can balance the evaluation of the model's detection accuracy and false alarms.

[0071] The performance of the network models of this invention and the control group on various evaluation metrics in the experiment is shown in the table below: Table 2: Network model performance of the present invention and the control group

[0072] Analysis of the superscript data reveals that the proposed solution achieves excellent performance in both Recall and FAR metrics. Furthermore, the F1 score of this invention is the best on both datasets, demonstrating that the invention achieves a balance between accuracy and false alarm rate, thus fulfilling the intended design goals.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting photolithographic hot spots in integrated circuits, characterized in that, It includes: A layout classification model is constructed to identify whether an input layout contains lithographic hotspots; the layout classification model includes a pre-feature extraction module, multiple post-feature extraction modules, a feature fusion module, and a fully connected layer in sequence; The pre-feature extraction module consists of one 7×7 convolution, one ungrouping convolution, one interactive convolution, and one attention layer; the post-feature extraction module consists of one enhanced interactive layer and one attention layer. The ungrouping convolution is used to divide the input features into high-frequency channels and low-frequency channels according to a preset channel ratio; the feature information of the low-frequency channels is processed by max pooling and convolution to obtain low-frequency features. F L The high-frequency features are obtained by convolution processing of the high-frequency channel feature information. F H In interactive convolution, the input... F L After convolution and upsampling processing compared with after convolution processing F H Feature splicing yields new high-frequency features ; Input F H After convolution and pooling processing and after convolution processing F L New low-frequency features are obtained by feature splicing. The enhanced interaction layer consists of multiple interactive convolutions with different kernels that introduce residual connections; the attention layer calculates attention for the high-frequency and low-frequency features of the input and weights them to achieve feature enhancement; the feature fusion module fuses the high-frequency and low-frequency features of the input to obtain fused features; the fully connected layer generates classification results based on the fused features of the input. Obtain labeled hot and non-hot maps to form the original dataset; perform data augmentation on the hot maps and sample the non-hot maps to form a balanced dataset; We adopted the BCE loss with label smoothing as the loss function; and used a balanced dataset to train the map classification model; retaining the model parameters that meet the performance metrics. The sliced ​​images of the original layout are input into the trained layout classification model to achieve lithographic hotspot detection.

2. The method for detecting photolithographic hot spots in integrated circuits as described in claim 1, characterized in that, In the aforementioned pre-feature extraction module, the convolution kernels of the ungrouping convolution and interactive convolution modules are both 3×3; the stride of the pooling operation is 2. And / or, the expression for the feature processing in the interactive convolution is: ； In the above formula, This indicates a convolution operation with a 3×3 kernel; Indicates an upsampling operation; Indicates pooling operation; This indicates feature splicing.

3. The method for detecting photolithographic hot spots in integrated circuits as described in claim 2, characterized in that: The enhanced interaction layer consists of two first interactive convolutions with a kernel size of 3×3 and one second interactive convolution with a kernel size of 1×1; the enhanced interaction layer contains two branches; The input features are processed by a second interactive convolution in one branch and by two first interactive convolutions in another branch. The features output from the two branches are then residually connected according to the high-frequency and low-frequency channels, and used as the output of the enhanced interactive layer.

4. The method for detecting photolithographic hot spots in integrated circuits as described in claim 1, characterized in that: Both the pre-feature extraction module and the post-feature extraction module employ the Outlook Attention mechanism, which incorporates relative spatial location information. Furthermore, each attention layer independently calculates and weights the high-frequency and low-frequency features input from the previous layer of the network. This allows for partial focus on the influence of features within local regions in the high-frequency features, while introducing more local and even global information in the low-frequency features.

5. The method for detecting photolithographic hot spots in integrated circuits as described in claim 1, characterized in that: The feature fusion module consists of four multi-branch modules (SAMBB) incorporating spatial attention, one upsampling module (Upsample), one 3×3 convolutional module (Conv3), one 1×1 convolutional module (Conv1), and one channel attention module (CA). In the feature fusion module, the input high-frequency features... F H After being processed sequentially by three SAMBB modules and one Upsample module, the low-frequency characteristics are compared with those processed by one SAMBB module. F L The features are then stitched together by channel; the stitched features are then processed by the Conv3 module, CA module, and Conv1 module to obtain the final fused features.

6. The method for detecting photolithographic hot spots in integrated circuits as described in claim 5, characterized in that: The SAMBB (Spatial Attention Multi-Branch Module) consists of a spatial attention module (SA) and four 3×3 convolutional modules. In the SAMBB module, the input features are first processed by SA, and the resulting features enter two branches, which are then processed by three convolutional modules and one convolutional module, respectively. The features output from the two branches are concatenated to form the final output of SAMBB.

7. The method for detecting photolithographic hot spots in integrated circuits as described in claim 1, characterized in that: During the training phase, data augmentation methods for hotspot maps include image rotation and flipping; the augmented images obtained through data augmentation have the same classification labels as the original images. Non-hotspot maps are sampled according to a preset ratio so that the data size of hotspot maps and non-hotspot maps in the sample dataset of each round of training is similar. And / or, introduce a loss function with label smoothing BCE loss. L bce as follows: ； In the above formula, y i Indicates the first i The classification label for each map, y=1 or 0; This indicates the category label after applying the smoothing strategy; The input labels are the output of the network model. y i The probability of; N Indicates the number of samples in the map; And / or, during the testing phase, recall, false positive rate (FAR), and F1 score are selected as performance evaluation metrics for the network model.

8. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it creates a layout classification model trained by the integrated circuit lithography hotspot detection method as described in any one of claims 1-7, and uses it to classify whether the slice image of the input original layout belongs to a lithography hotspot.

9. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it creates a layout classification model trained by the integrated circuit lithography hotspot detection method as described in any one of claims 1-7, and uses it to classify whether the slice image of the input original layout belongs to a lithography hotspot.

10. A map hotspot detection device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, it creates a lithography hotspot detection system; the lithography hotspot detection system includes a data preprocessing module and a recognition module; The data preprocessing module uses Python's pya and klayout libraries to convert the input raw map file in GDS or OAS format into an image according to the rule that 1 dBu corresponds to 1 pixel; and automatically segments the converted raw map into multiple slice images according to the marker layer of the map slice; the segmented slice images are then input into the recognition module in sequence; The identification module runs a layout classification model trained by the integrated circuit lithography hotspot detection method as described in any one of claims 1-7. The layout classification model is used to classify whether the input slice image belongs to a lithography hotspot.

Images