A Map Hotspot Detection Method Based on Multi-Feature Representation Learning and Its Storage Medium

Abstract

This invention belongs to the field of integrated circuit design, specifically relating to a layout hotspot detection method and its storage medium based on multi-feature representation learning. The method first constructs a novel layout hotspot detection model consisting of a feature reconstruction network, a feature fusion network, and a classifier. The feature reconstruction network uses a U-shaped network to extract reconstructed features at different scales. A multi-feature modeling interaction module is introduced into the encoder to extract and model hotspot information including visual and geometric features. Feature enhancement is achieved through inter-layer cross-coding and inter-layer guided decoding modules. Then, a BCE loss with label smoothing is used as the loss function, and the layout hotspot detection model is trained in two stages using imbalanced and balanced datasets. Finally, the model is applied to identify hotspots in the input layout. This invention solves the problem of insufficient extraction and representation of layout features in existing solutions, leading to insufficient accuracy and efficiency in hotspot identification.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit design, specifically relating to a layout hotspot detection method based on multi-feature representation learning, and its corresponding storage medium and detection equipment for photolithographic hotspots in the layout. Background Technology

[0002] As semiconductor process nodes continue to advance, the minimum feature size of integrated circuits is much smaller than the wavelength of the photolithography light source. Due to the diffraction effect of light, the exposure system cannot ideally image the mask pattern. The resulting optical proximity effect (OPE) easily leads to layout distortion, causing pinch-offs, bridging, and edge placement errors (EPE) hotspots on the wafer surface. After physical verification and layout design are completed, hotspot detection is required. To address this phenomenon, the industry typically uses optical proximity correction (OPC) technology to compensate for the imaging error caused by OPE by modifying the mask layout, making the actual exposed pattern closer to the target pattern. However, OPC technology still cannot completely solve the exposure distortion problem caused by complex masks. The photolithography pattern after OPC mask optimization may still have hotspot patterns. In order to improve the tape-out success rate or product yield, hotspot detection must be performed before actual chip manufacturing.

[0003] Existing map hotspot detection (LHD) methods can be categorized into two types based on the data type of the input: graph-based methods and image-based methods. Graph-based methods achieve highly efficient hotspot detection through graph construction and encoding, as well as graph operations. Although graph-based methods are generally faster than image-based methods, they all suffer from unacceptable false alarms. This is because their detection performance relies on a high-quality graph construction and encoding process, which can easily lead to incomplete capture of map information, thus degrading the performance of graph-based models.

[0004] In recent years, image-based methods have mainly followed three research routes: lithography simulation-based methods, pattern matching-based methods, and machine learning-based methods. Lithography simulation-based methods simulate real defect areas by mathematically modeling the actual lithography process, thus achieving high-precision hotspot detection and are frequently used solutions in mainstream EDA tools. However, lithography simulation comes at the cost of significant computational resources and time. Pattern matching-based methods build feature libraries by collecting known hotspot patterns and analyzing their features, then matching any incoming pattern against these features. While fast, pattern matching-based methods often fail to identify hotspot patterns not present in the feature library, leading to high false alarm rates. Machine learning-based methods build machine learning models by learning from existing training data. Due to the model's generalization ability, it can detect previously unseen hotspots. This method primarily relies on handcrafted features and has poor robustness across different technology scenarios. With the development of deep learning, many methods based on convolutional neural networks (CNNs) have emerged, achieving significant progress in detection performance. While existing CNN-based hotspot detection methods have achieved good results, their lack of learning of multi-feature representations of the layout is a bottleneck hindering further performance improvement. Therefore, current solutions struggle to identify complex hotspot patterns and avoid them in subsequent tape-out processes, necessitating the development of more efficient layout hotspot detection algorithms. Summary of the Invention

[0005] To address the problem that existing solutions are insufficient in extracting and representing layout features, leading to inadequate accuracy and efficiency in hotspot identification, this invention provides a layout hotspot detection method based on multi-feature representation learning, along with its corresponding storage medium and a detection device for lithographic hotspots in the layout.

[0006] This invention is achieved using the following technical solution:

[0007] A map hotspot detection method based on multi-feature representation learning, comprising:

[0008] A layout hotspot detection model is constructed to identify whether an input layout contains lithographic hotspots. This model comprises a feature reconstruction network, a feature fusion network, and a ResNet-based classifier. The feature reconstruction network extracts reconstruction features of different scales from the input layout; the feature fusion network adaptively fuses these reconstruction features across layers to obtain a fused feature; and the classifier generates a classification result based on the fused feature.

[0009] The feature reconstruction network employs a U-shaped network consisting of a four-layer encoder and a three-layer decoder, with reconstructed features D1-D4 decreasing in scale according to the input layout. The encoder layer 1 includes a ResBlock and a multi-feature modeling interaction module (MFMIM) connected sequentially; layers 2-4 include a ResBlock, an inter-layer cross-coding module (ICEM), and an MFMIM connected sequentially. In the encoder, the outputs of the ResBlock and MFMIM of the previous layer serve as the inputs to the ICEM and ResBlock of the next layer, respectively. Each layer of the decoder includes one inter-layer guided decoding module (IGDM) and two ResBlocks; the output of the ResBlock of the next layer in the decoder serves as the input to the IGDM and ResBlock of the previous layer; layers 1-3 of the encoder are skipped to layers 1-3 of the decoder; the output of the encoder layer 4 is denoted as D4 and serves as the input to the decoder layer 3; the outputs of layers 1-3 of the decoder are denoted as D1-D3.

[0010] The original dataset is constructed by obtaining hot and non-hot maps containing label information; BCE loss with label smoothing is used as the loss function. L BCE The hotspot detection model is trained in two stages. The first stage uses the original dataset for training; the second stage randomly selects non-hotspot maps with the same number as hotspot maps to form a balanced dataset for training; model parameters that meet the performance metrics after training are retained.

[0011] The sliced ​​images of the original layout are input into the trained layout hotspot detection model to achieve lithographic hotspot detection.

[0012] The present invention also includes a storage medium storing a computer program, which, when executed by a processor, creates a layout hotspot detection model trained by the aforementioned layout hotspot detection method based on multi-feature representation learning, and is used to classify whether the input original layout belongs to a lithographic hotspot.

[0013] The present invention also includes a computer program product comprising a computer program that, when executed by a processor, creates a layout hotspot detection model trained by the aforementioned layout hotspot detection method based on multi-feature representation learning, and is used to classify whether the input original layout belongs to lithographic hotspots.

[0014] The present invention also includes a detection device for lithographic hotspots in a layout, comprising 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 layout hotspot detection model trained by the aforementioned layout hotspot detection method based on multi-feature representation learning, thereby enabling the identification of lithographic hotspots in the input layout.

[0015] The technical solution provided by this invention has the following beneficial effects:

[0016] This invention designs a novel map hotspot detection model, which utilizes a feature reconstruction network and a feature fusion network to achieve deep extraction and comprehensive interactive learning of visual and geometric features contained in the map. This enables the network model to acquire reconstructed features rich in information expression and robust features with multi-scale dynamic fusion. The feature reconstruction network integrates the reconstruction of map features into map hotspot detection through multi-feature representation learning, while the feature fusion network recalibrates and aggregates feature maps from different levels. This progressive information enhancement strengthens the network's feature representation capabilities. Consequently, the model can acquire richer hotspot information, efficiently combine and fully utilize the hotspots themselves and surrounding information, and improve the model's ability to identify hotspots contained in the map.

[0017] In the training phase of the network model, this embodiment proposes a two-stage training strategy. The strategy utilizes an imbalanced dataset in the first stage and a balanced dataset in the second stage. The training process employs labeled smooth BCE loss, and the learning rate is dynamically adjusted as the training progresses. This training strategy effectively ensures the stability of the network learning process. By learning the overall data characteristics first and then learning the characteristics of specific data, it effectively mitigates the network's excessive focus caused by a large number of non-hotspot data. Attached Figure Description

[0018] Figure 1 This is a flowchart of the steps of the map hotspot detection method based on multi-feature representation learning provided in Embodiment 1 of the present invention.

[0019] Figure 2 This is a network architecture diagram of the hotspot detection model constructed in Embodiment 1 of the present invention.

[0020] Figure 3 This is a schematic diagram of the multi-feature modeling interaction module provided in Embodiment 1 of the present invention.

[0021] Figure 4 This is a schematic diagram of the hypergraph computing layer provided in Embodiment 1 of the present invention.

[0022] Figure 5 This is a schematic diagram of the multi-feature interaction network provided in Embodiment 1 of the present invention.

[0023] Figure 6 This is a schematic diagram of the interlayer cross-coding module provided in Embodiment 1 of the present invention.

[0024] Figure 7 This is a schematic diagram of the detail-preserving downsampling module provided in Embodiment 1 of the present invention.

[0025] Figure 8 This is a schematic diagram of the interlayer guidance decoding module provided in Embodiment 1 of the present invention.

[0026] Figure 9 This is a schematic diagram of the feature fusion network provided in Embodiment 1 of the present invention.

[0027] Figure 10 This is a heatmap showing the partial hotspot identification results of the present invention's solution in a test experiment. Detailed Implementation

[0028] 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.

[0029] Example 1

[0030] In this embodiment, our researchers discovered in their study of hotspot detection on map layouts that, although existing CNN-based LHD methods have achieved good results, they can only learn the visual features of hotspots and lack attention to topological structure, i.e., geometric features. This is a bottleneck hindering the improvement of model performance. In fact, map layout images inherently lack rich color information, and CNNs mainly rely on the attention maps of hotspots and their surrounding information to infer hotspots. Furthermore, since there is almost no contrast and SCR (signal-to-noise ratio) between hotspots and non-hotspots in map layout images, extracting geometric features from the map to enhance the feature differences between them should improve the accuracy of hotspot detection.

[0031] Building upon this, this embodiment views hotspot detection as a multi-feature representation learning process. Hotspot features can be broadly categorized into visual features and geometric features. Geometric features represent the geometric topology of hotspots, effectively expressing the complex graph topology relationships within the image. Visual features can further include local and global contextual features of hotspots. Local features reflect the spatial information of the hotspot itself, while global contextual features typically have a larger receptive field, capturing more global contextual information and reflecting the spatial information surrounding the hotspot. Comprehensively extracting these various feature types from the image and efficiently modeling them will yield richer hotspot information and contribute to improving the accuracy and efficiency of hotspot detection.

[0032] Based on the above technical approach, this embodiment designs a novel map hotspot detection model. This model extracts visual and geometric features from a map image, organizes them appropriately to form a new fused feature, and finally classifies whether the input map contains hotspots based on this fused feature. Specifically, the map hotspot detection model designed in this embodiment includes three key components: a feature reconstruction network (Rcc Net), a feature fusion network (ACFF Net), and a ResNet-based classifier. The feature reconstruction network extracts reconstruction features at different scales from the input map. It employs a U-shaped network based on a residual structure to enrich the diversity of the extracted map features. The U-shaped network introduces a multi-feature modeling interaction module (MFMIM) at the encoder end to organize the extracted visual and geometric features at different levels through interactive learning, thereby enhancing the network's representational ability. The feature fusion network adaptively fuses the various reconstruction features across layers to obtain the fused feature. This process aggregates weighted features, avoiding direct connections between each feature map and achieving the integration of multi-scale reconstruction features. The classifier uses ResNet (such as the classic ResNet18) as the final feature extraction and decision network, thereby generating classification results based on whether the map contains lithographic hotspots according to the fused features.

[0033] Specifically, such as Figure 1 As shown, the map hotspot detection method based on multi-feature representation learning provided in this embodiment includes the following steps:

[0034] S1: Construct a layout hotspot detection model to identify whether the input layout contains lithographic hotspots; such as Figure 2 As shown, the hotspot detection model consists of a feature reconstruction network, a feature fusion network, and a classifier connected in sequence.

[0035] Among them, combined Figure 2 In this embodiment, the feature reconstruction network adopts a U-shaped network consisting of a four-layer encoder and a three-layer decoder. Under this structure, the feature reconstruction network can extract four reconstructed features D1~D4 at different scales, which contain various feature information such as visual features and geometric features, based on the input layout.

[0036] Specifically, in the encoder of the feature reconstruction network provided in this embodiment, the first layer of the encoder includes a ResBlock and a Multi-Feature Modeling Interaction Module (MFMIM) connected in sequence. Each ResBlock includes a 3×3 convolutional layer (Conv3), a batch normalization layer (BN), and a RuLU activation layer connected in sequence; the input of Conv3 and the output of BN are residually concatenated. The second to fourth layers of the encoder include a ResBlock, an Inter-Layer Cross-Encoding Module (ICEM), and an MFMIM connected in sequence. In the encoder, the outputs of the ResBlock and MFMIM of the previous layer serve as the inputs of the ICEM and ResBlock of the next layer, respectively.

[0037] At the decoder end of the feature reconstruction network, each layer of the decoder includes one Interlayer Guided Decoding Module (IGDM) and two ResBlocks connected sequentially. The output of the ResBlock of the next layer in the decoder serves as the input to the IGDM and ResBlock of the previous layer; layers 1-3 of the encoder are skipped to layers 1-3 of the decoder. Furthermore, the output of layer 4 of the encoder, denoted as D4, serves as the input to layer 3 of the decoder; the outputs of layers 1-3 of the decoder are denoted as D1-D3.

[0038] In the hotspot detection model provided in this embodiment, the feature reconstruction network and the feature fusion network are the two core components. Within the feature reconstruction network, the multi-feature modeling interaction module, the inter-layer cross-coding module, and the inter-layer guided decoding module constitute key components. To make the performance and advantages of the hotspot detection model provided in this embodiment clearer, the structure and principles of the multi-feature modeling interaction module, the inter-layer cross-coding module, the inter-layer guided decoding module, and the feature fusion network are described in detail below:

[0039] (I) Multi-feature modeling interaction module

[0040] This embodiment embeds an MFMIM in each layer of the encoder of the feature reconstruction network to achieve multi-feature modeling and interactive learning, thereby improving global context awareness on a larger scale and helping to reduce false positives about hotspots. Figure 3 As shown, the multi-feature modeling interaction module includes a multi-feature modeling network (MFM) and a multi-feature interaction network (MFI) connected in sequence. The MFM network consists of two 1×1 convolutional layers (Conv1), one visual feature modeling module (VFM), and one geometric feature modeling module (GFM). Furthermore, the VFM includes four 3×3 convolutional layers (Conv3), and the GFM includes four hypergraph computation layers (HCM).

[0041] In the multi-feature modeling interaction module, the input features are convolved by the first Conv1 and then used as the input of the first layer Conv3 in VFM; the output of Conv3 in the same layer is used as the input of HCM; the output of HCM in the previous layer is used as the input of the next layer Conv3; the output of the first Conv1 is processed by the Sigmoid activation function and then multiplied by the output elements of each HCM layer and then concatenated by channel. The resulting features are then processed by the second Conv1 and MFI in sequence to obtain the output features of the multi-feature modeling interaction module.

[0042] In the MFM network, considering the limited information content of the landscape image, which can easily lead to feature redundancy, a 1×1 convolution is first used to compress the channel dimension of the feature map at each stage from C to C / 4. Channel compression helps to compact the features. Then, a 3×3 convolution with a sequential structure is used layer by layer to increase the receptive field, thereby capturing a larger visual range and enhancing the ability to model the global context. However, the continuous convolution of the sequential structure can easily cause severe dilution of the information in the original feature map. Therefore, this embodiment captures spatial features from multiple different receptive fields, which also provides a more complete visual feature representation. Finally, the MFM network integrates features from different receptive fields to construct more balanced local features and global context features. To further enhance the efficient transmission of balanced features in continuous convolution, MFM performs a sigmoid operation on the compressed features to obtain self-attention, and then multiplies it element-wise with the output features of different receptive fields, making the attention skip-connection injection method direct and efficient. Finally, after concatenating the features of multi-receptive field spatial information according to the channel dimension, a 1×1 convolution is used to fuse the output.

[0043] The Visual Feature Modeling (VFM) module and the Geometric Feature Modeling (GFM) module constitute the core of the MFM network. In the Geometric Feature Modeling (GFM) module, such as... Figure 4 As shown, the data processing of the Hypergraph Computational Layer (HCM) includes two stages: hypergraph construction and hypergraph convolution. In the hypergraph construction stage, the vertex set consists of deconstructed visual features, and the hyperedge set consists of edges between the central feature point and all feature points within a specified distance threshold. In the hypergraph convolution stage, high-order learning is performed on the feature points using spatial domain hypergraph convolution with residuals.

[0044] The expression for the data processing procedure in the hypergraph computation layer is as follows:

[0045] ;

[0046] In the above formula, X This represents the input features of the hypergraph computation layer; D v and D e These represent the diagonal matrices of the vertices and hyperedges, respectively. HThis represents the correlation matrix of the constructed hypergraph; It is a trainable parameter; HCM ( X , H ) represents the output characteristics of the hypergraph computation layer.

[0047] Based on the visual features of the map image extracted by the MFM network and the potential high-order global correlations, this embodiment further uses a multi-feature interaction network (MFI) to interactively learn the extracted multiple features, thereby enhancing the fine-grained information flow and representational synergy of the multiple features. Figure 5 As shown, the Multi-Feature Interaction Network (MFI) includes a channel attention module (SE) and a spatial attention module (SA) connected sequentially. These are used to learn the importance of each channel and each spatial location, respectively. This embodiment employs a sequential channel-to-spatial order, which maximizes the effectiveness of the attention mechanism.

[0048] Combination Figure 5 As can be seen, in the channel attention module, the input features are first subjected to global average pooling. The resulting features are then processed by two fully connected layers (FC) and a sigmoid activation function, and then element-wise multiplied with the input features of the channel attention module. The resulting features are then matrix-sumped with the input features of the channel attention module to obtain the attention-weighted channel features. In the spatial attention module, the input features are subjected to global average pooling in both the horizontal and vertical directions to obtain the lowest frequency components. Then, the two sets of lowest frequency components in different directions are concatenated by channel. The concatenated features are processed by one-dimensional convolution, and then nonlinearly mapped by an SE mechanism composed of two one-dimensional convolutions to generate more compact and complete low-frequency features in two different directions. The two low-frequency features are processed by a sigmoid activation function and then fused by matrix multiplication. The resulting fused features are then element-wise multiplied with the input features of the spatial attention module. The resulting features are then matrix-sumped with the input features of the spatial attention module to obtain the attention-weighted spatial features.

[0049] (ii) Interlayer cross-coding module

[0050] In different layers of the encoder, feature information can be extracted from the layout at different scales, thereby obtaining low-level and high-level features at different scales. Low-level features often rely on the receptive field information in high-level features, while also providing detailed information for high-level features. Therefore, this embodiment designs an inter-layer cross-coding module ICEM with an integrated cross-attention mechanism to effectively utilize the high-level and low-level features input between different coding layers. The information advantage brought by this dual-branch path helps to eliminate noise while injecting large receptive field information and improving coding efficiency.

[0051] Specifically, such as Figure 6As shown, the interlayer cross-coding module ICEM in this embodiment is used to process the high-level features output by two ResBlocks in adjacent layers of the encoder. F high and low-level features F low Feature fusion is performed. Specifically, ICEM consists of one detail-preserving downsampling module (DPD), two attention heads (Attention Head1 and Attention Head2) with shared weights, and three 1×1 convolutional modules (Conv1). The schematic diagram of the DPD module is shown below. Figure 7 As shown, it can extract feature points at intervals through slicing operations to form four sets of downsampled feature maps. Then, these maps are concatenated with the schematic map after max pooling by channel, and then processed by a 1×1 convolutional module. The attention heads Attention Head1 and Attention Head2 are composed of a 3×3 convolutional module Conv3, a 1×1 convolutional module Conv1, and a Sigmoid activation layer connected in sequence.

[0052] In ICEM, F high As input to Attention Head1, the output of Attention Head1 is compared with the output processed by DPD. F low After element-wise multiplication, the result is processed by the first Conv1. The output of the first Conv1 serves as the input to AttentionHead2; the output of AttentionHead2 is then multiplied by... F high After element-wise multiplication, the data is processed by a second Conv1. The features output from the first and second Conv1 are then added together as matrices, and processed by a third Conv1 to obtain the fused features output by ICEM.

[0053] In the dual-branch fusion structure of ICEM, the low-level feature branch preserves details through DPD. F low Linear downsampling. DPD primarily extracts feature points at intervals through slicing operations, forming four sets of downsampled feature maps. This operation converts spatial information into channel information, achieving downsampling without information loss. To compensate for potential shortcomings of slicing operations, it retains the traditional max-pooling downsampling path. The semantic information of high-level feature branches is used to eliminate noise. This includes using... F high After removing background noise from the semantic information in the DPD output features, the features are refined through 1×1 convolution, and then guided... F highAdd hotspot details. Finally, fuse the two complementary features obtained through a 1×1 convolution and output the result.

[0054] (III) Interlayer Guidance Decoding Module

[0055] At the decoder end of the feature reconstruction network, the detailed information provided in the low-level features is needed to restore the feature information of different scales extracted by the encoder to the low-dimensional space layer by layer. For this task, this embodiment designs a special inter-layer guided decoding module (IGDM), which is used to process the high-level features input from adjacent layers in the decoder. E high and low-level features E low Feature fusion is then performed. Global information at different levels is then used to generate complementary attention that selectively emphasizes features, helping to narrow the semantic gap in shallow features.

[0056] like Figure 8 As shown, IGDM consists of two channel attention modules (SE) and a sigmoid activation function. E high and E low In each of the two branches, attention extraction and feature weighting are performed via the SE module, and then combined with... E low Element-wise multiplication is performed to obtain the fused features output by IGDM. In this embodiment, IGDM captures two sets of attention through a dual-branch path-based SE mechanism, and then the attention is summed and averaged before being used to weight the low-level features.

[0057] (iv) Feature Fusion Network

[0058] The feature fusion network in this embodiment can dynamically adjust the receptive field based on reconstructed features at different scales, providing the network model with a wider range of dynamic representations and significantly enhancing the network's ability to learn multi-feature representations. Specifically, to obtain cross-scale reconstructed features at different levels, the feature reconstruction network uses each layer of the decoder as output, resulting in four feature maps denoted as D1~D4. To achieve adaptive cross-layer fusion of D1~D4, the feature fusion network designed in this embodiment first applies 1×1 convolutions and upsampling to D2~D4 progressively to adjust them to the same size as D1. Then, matrix addition is used to combine these features to generate a global feature descriptor to selectively emphasize the importance of features at different scales. This descriptor produces a more compact feature representation through compression. Then, four sets of parallel re-extraction are used to obtain feature descriptors at different scales, which can recalibrate the feature maps from different streams. Next, Softmax is applied to the descriptors to generate attention activation. The adaptively recalibrated multi-scale features are then element-wise multiplied to integrate information from different levels.

[0059] Specifically, such as Figure 9 As shown, the feature fusion network consists of three 1×1 convolutional layers (Conv1), three bilinear interpolation layers (Up), one global max pooling layer (GAP), five fully connected layers (FC), and a softmax layer. The global max pooling layer (GAP) and the fully connected layers (FC) constitute a channel attention mechanism. The data processing of the feature fusion network includes:

[0060] 1×1 convolution and bilinear interpolation are performed on D1~D4 to obtain intermediate features of the same scale. ;Will The global features are obtained by adding the matrices; global average pooling and fully connected layers are then applied to the global features to obtain the global feature descriptors. s ; through four fully connected layers of different scales respectively s The data is processed and then subjected to Softmax to extract feature descriptors at different scales. s 1~ s 4. [The sentence is incomplete and requires more context to be translated accurately.] s 1~ s 4. As attention, feature weights are applied to D1~D4 respectively; the resulting feature matrices are summed to obtain the fused feature output by the feature fusion network. D out .

[0061] S2: Obtain the hot and non-hot maps containing label information to form the original dataset; use BCE loss with label smoothing as the loss function. L BCEThe hotspot detection model is trained in two stages. The first stage uses the original dataset for training; the second stage randomly selects non-hotspot maps with the same number as hotspot maps to form a balanced dataset for training; model parameters that meet the performance metrics after training are retained.

[0062] In the training phase of the hotspot detection model designed in this embodiment, considering the extremely unbalanced proportion of images containing hotspots and those without, and the high similarity of the graphics, the neural network is prone to overfitting to non-hotspot images, which carries the risk of getting trapped in local optima. Inspired by the inherent dynamic scale vision capability of the human visual system—focusing on smaller details from larger ones—this embodiment proposes a two-stage training strategy to adapt to this task.

[0063] In the first stage, all non-hotspot images are used for training to observe the overall data distribution. In the second stage, a portion of non-hotspot images, roughly the same size as the hotspots, are randomly selected for training to balance the classes and help the network learn the differences between them. Furthermore, in the early stages of training, this embodiment uses an constant learning rate to ensure the model explores a sufficiently broad solution space, helping it better learn the overall characteristics of the data and providing excellent inspiration for later learning. In the later stages of training, this embodiment uses a periodically decreasing cosine learning rate to give the model the opportunity to take larger steps, facilitating escape from local optima. The training strategy employed in this embodiment fundamentally solves the class imbalance problem and allows for rapid fluctuations to help the model escape local optima, significantly improving the performance of the hotspot detection network.

[0064] Furthermore, this embodiment employs BCE loss with label smoothing during the training phase. L BCE To evaluate the difference between the classification results and the true labels. Specifically, L BCE The expression is as follows:

[0065] ;

[0066] In the above formula, y, and These represent the classification result, the true hard label, and the true soft label, respectively. smooth Represents the smoothing factor; in this embodiment smooth The value is 0.2.

[0067] S3: Input the sliced ​​image of the original layout into the trained layout hotspot detection model to achieve lithographic hotspot detection.

[0068] In the practical application of this embodiment, based on the trained layout hotspot detection model, technicians can segment the original large-scale integrated circuit layout into multiple slice images of specified sizes and input them sequentially into the layout hotspot detection model. The model can identify whether each slice image contains lithographic hotspots and output the corresponding classification results. In this way, it is possible to determine whether lithographic hotspots exist in the corresponding regions of each slice image in the original layout, thereby achieving high-precision and efficient detection of all lithographic hotspots contained in the entire integrated circuit layout.

[0069] Example 2

[0070] To better apply the solution in Embodiment 1, this embodiment further provides corresponding storage media, computer program products, and detection equipment for photolithographic hot spots in the layout.

[0071] The storage medium contains a computer program. When the computer program is executed by the processor, it creates a layout hotspot detection model trained by the layout hotspot detection method based on multi-feature representation learning as shown in Example 1, and uses it to classify whether the input original layout belongs to a lithographic hotspot.

[0072] The computer program product includes a computer program that, when executed by a processor, creates a layout hotspot detection model trained by the layout hotspot detection method based on multi-feature representation learning as shown in Example 1, and is used to classify whether the input original layout belongs to lithographic hotspots.

[0073] The device for detecting lithographic hotspots in this layout is essentially a computing 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 layout hotspot detection model trained by the layout hotspot detection method based on multi-feature representation learning as shown in Example 1, thereby enabling the identification of lithographic hotspots in the input layout.

[0074] In practical applications, the computer equipment can be a standalone computer device, such as a laptop, tablet, desktop computer, or a rack server, blade server, tower server, or cabinet server (including standalone servers or server clusters composed of multiple servers) capable of executing computer programs.

[0075] 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.

[0076] 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.

[0077] Simulation test

[0078] To verify the performance of the map hotspot detection method based on multi-feature representation learning provided by this invention, technicians simulated and trained the relevant scheme and tested the performance of the trained network model.

[0079] I. Overview of Experimental Methods

[0080] This experiment tested the performance of the proposed solution (denoted as MFRLNet) on two datasets, ICCAD12 and ICCAD19. To more intuitively reflect the performance level of the proposed solution, the existing hotspot detection solution SOTA was also selected as a control group for comparison. The relevant data set information is shown in Table 1.

[0081] Table 1: Relevant information about the sample dataset

[0082]

[0083] This experiment used AdamW as the optimizer with a learning rate of 1e-3. The learning rate scheduler was a CombinedLR, which combines StepLR and CosineAnnealingLR, with a periodic cosine descent period of 30. The training process lasted a total of 900 epochs with a batch size of 96. For fairness, all experiments were conducted on the same single RTX 3090 Ti GPU server. During the experiments, accuracy (Acc), false positive rate (Fa), and F1 score were used as performance metrics for different schemes.

[0084] II. Performance Comparison

[0085] The performance of the LHD protocol in the experimental and control groups on the ICCAD12 and ICCAD19 datasets is shown in Table 2:

[0086] Table 2: Performance of the present invention and the control group

[0087]

[0088] Analysis of the data in the table above reveals that, compared to state-of-the-art (SOTA) methods on the ICCAD12 and ICCAD19 datasets, the MFRLNet designed in this invention achieves the best performance on most evaluation metrics and demonstrates the most outstanding overall performance. For example, the method of this invention achieves 99.6% Acc and 3.7% Fa on ICCAD12, with a comprehensive F1 score of 0.908, far superior to other methods. In ICCAD19-1, the Acc of the TCAD'25 and DATE'23 methods is 12.8% and 13.1% higher than that of the method of this invention, respectively, but the Fa of this invention is only about one-quarter of theirs. The F1 score shows that the comprehensive performance of the method of this invention is 0.163 and 0.167 higher than theirs, indicating that the algorithm of this invention achieves an excellent balance between avoiding hotspot detection errors and detecting as many as possible. It is important to note that false alarms or missed detections in hotspot detection tasks can have serious adverse effects on chip manufacturing. Compared to the TCAD'25 and TCAD'22 methods on ICCAD19-2, the method of this invention sacrifices a small portion of Acc for a significant improvement in Fa, and achieves optimal performance in F1, which is well worth the effort. Overall, the method of this invention demonstrates excellent performance while balancing various metrics. It can also be seen that other methods exhibit lower Acc because the lack of visual feature modeling makes it impossible to identify hotspot graphics, while the lack of geometric feature modeling makes it difficult to distinguish similar hotspots from non-hotspot graphics, resulting in higher Fa. Furthermore, because the ICCAD19-2 validation set includes a large number of unseen hotspots from the training set, the MFRLNet of this invention outperforms ICCAD19-1 on ICCAD19-2. This is also the reason why Acc and Fa are higher in ICCAD19-2, as this increases the difficulty of distinguishing between hotspots and non-hotspots.

[0089] III. Explainability

[0090] This experiment visualized heatmaps of some layout images using the ICCAD19-1 dataset. The results are as follows: Figure 10 As shown.

[0091] Heatmaps can provide a reliable interpretation of hotspot predictions. The red areas in a heatmap represent key regions used to predict hotspots. In correctly predicted heatmaps, these red areas largely coincide with the actual hotspot areas, indicating that the hotspot predictions can be accurately and reasonably explained. In incorrectly predicted heatmaps, these red areas largely differ from the actual hotspot areas, indicating that the model has not considered important regions for decision-making.

[0092] Analysis of the data in the figure shows that the present invention can effectively detect areas where hotspots may occur in the layout image, clearly demonstrating the rationality of the overall solution.

[0093] IV. Ablation Experiment

[0094] To verify the contribution of each component in the MFRLNet layout hotspot detection model designed in this invention to the overall performance of the scheme, an ablation study was conducted on the ICCAD19-1 dataset.

[0095] The ablation results of the overall architecture are shown in Table 3:

[0096] Table 3: Ablation Experiment Results of the MFRLNet Model

[0097]

[0098] Analysis of the data in the table above reveals that RecNet improves the performance of the classifier, and the use of ACFFNet can achieve even better results, which demonstrates the rationality of the overall architecture of this invention.

[0099] Furthermore, this invention continued the ablation study on the internal components of RecNet. The corresponding ablation experimental results are shown in Table 4:

[0100] Table 4: Ablation Experiment Results of RecNet

[0101]

[0102] Analysis of the superscript data reveals that adding each component sequentially significantly improves the model, verifying that the internal components designed in RecNet in this invention can work effectively together.

[0103] 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 map hotspot detection method based on multi-feature representation learning, characterized in that, It includes: A landscape hotspot detection model is constructed, consisting of a feature reconstruction network, a feature fusion network, and a ResNet-based classifier. The feature reconstruction network uses a U-shaped network consisting of a four-layer encoder and a three-layer decoder, and the output features D1 to D4 are reconstructed in descending order of scale. The first layer of the encoder includes a ResBlock and a multi-feature modeling interaction module MFMIM connected in sequence. The second to fourth layers include a ResBlock, an inter-layer cross-coding module ICEM, and an MFMIM connected in sequence. The outputs of the ResBlock and MFMIM of the previous layer in the encoder are used as the inputs of the ICEM and ResBlock of the next layer, respectively. Each layer of the decoder includes one inter-layer guided decoding module IGDM and two ResBlocks. The output of the ResBlock of the next layer in the decoder is used as the input of the IGDM and ResBlock of the previous layer. The first to third layers of the encoder are skipped connections with the first to third layers of the decoder. The output of the fourth layer of the encoder is denoted as D4 and used as the input of the third layer of the decoder. The outputs of layers 1-3 of the decoder are denoted as D1-D3; the feature fusion network is used to adaptively fuse the reconstructed features across layers to obtain the fused features; The classifier is used to generate classification results based on the fused features; The Multi-Feature Modeling Interaction Module (MFMIM) comprises a Multi-Feature Modeling Network (MFM) and a Multi-Feature Interaction Network (MFI) connected sequentially. The MFM network consists of two 1×1 convolutional layers (Conv1), a Visual Feature Modeling Module (VFM), and a Geometric Feature Modeling Module (GFM). The VFM includes four 3×3 convolutional layers (Conv3), and the GFM includes four Hypergraph Computation Layers (HCM). The data processing of the HCM includes hypergraph construction and hypergraph convolution. In the hypergraph construction stage, the vertex set is composed of deconstructed visual features, and the hyperedge set is composed of edges between all feature points within a specified distance threshold from the center feature point. In the hypergraph convolution stage, high-order learning is performed on the feature points using spatial domain hypergraph convolution with residuals. The Multi-Feature Interaction Network (MFI) consists of a channel attention module (SE) and a spatial attention module (SA) connected sequentially. In the channel attention module, the input features are first subjected to global average pooling. The resulting features are then processed by two fully connected layers (FC) and a sigmoid activation function. After that, they are multiplied element-wise with the input features of the channel attention module. The resulting features are then added to the input features of the channel attention module to obtain the attention-weighted channel features. In the spatial attention module, the input features are first obtained by global average pooling in the horizontal and vertical directions to obtain the lowest frequency components. Then, the two sets of lowest frequency components in different directions are concatenated by channel. After the concatenated features are processed by one-dimensional convolution, they are non-linearly mapped by the SE mechanism composed of two one-dimensional convolutions to generate more compact and complete low frequency features in two different directions. Two low-frequency features are processed by the Sigmoid activation function and then fused by matrix multiplication. The resulting fused features are then multiplied element-wise with the input features of the spatial attention module. The resulting features are then added to the input features of the spatial attention module to obtain the attention-weighted spatial features. The interlayer cross-coding module ICEM is used to process high-level features of the outputs of two ResBlocks in adjacent layers of the encoder. F high and low-level features F low Perform feature fusion; A BCE loss with label smoothing is adopted. First, the hot spot detection model is trained using the original imbalanced dataset consisting of hot spots and non-hot spots. Then, it is trained using the recombined balanced dataset. The model parameters that meet the performance index after the two-stage training are retained and used to perform lithographic hot spot detection on slice images of the original map.

2. The map hotspot detection method based on multi-feature representation learning as described in claim 1, characterized in that: In the multi-feature modeling interaction module, the input features are convolved by the first Conv1 and then used as the input of the first layer Conv3 in VFM; the output of Conv3 in the same layer is used as the input of HCM; the output of HCM in the previous layer is used as the input of the next layer Conv3. The output of the first Conv1 is processed by the Sigmoid activation function, multiplied by the output elements of each HCM layer, and then concatenated by channel. The resulting features are then processed by the second Conv1 and MFI in sequence to obtain the output features of the multi-feature modeling interaction module.

3. The map hotspot detection method based on multi-feature representation learning as described in claim 2, characterized in that: The expression for the data processing procedure of the hypergraph computing layer is as follows: ； In the above formula, X This represents the input features of the hypergraph computation layer; D v and D e These represent the diagonal matrices of the vertices and hyperedges, respectively. H This represents the correlation matrix of the constructed hypergraph; It is a trainable parameter; This represents the output characteristics of the hypergraph computation layer.

4. The map hotspot detection method based on multi-feature representation learning as described in claim 1, characterized in that: ICEM consists of one detail-preserving downsampling module (DPD), two shared weight attention heads (Attention Head1 and Attention Head2), and three 1×1 convolutional modules (Conv1). F high As input to Attention Head1, the output of Attention Head1 is compared with the output of DPD. F low After element-wise multiplication, the result is processed by the first Conv1; the output of the first Conv1 serves as the input to Attention Head 2; the output of Attention Head 2 is then multiplied by... F high After element-wise multiplication, the feature outputs from the first and second Conv1 processes are added together as a matrix, and then processed by the third Conv1 to obtain the fused feature output by ICEM.

5. The map hotspot detection method based on multi-feature representation learning as described in claim 1, characterized in that: The Interlayer Guided Decoding (IGDM) module is used to process high-level features of adjacent layer inputs in the decoder. E high and low-level features E low Perform feature fusion; IGDM consists of a two-channel attention module (SE) and a sigmoid activation function; E high and E low In each of the two branches, attention extraction and feature weighting are performed via the SE module, and then combined with... E low Element-wise multiplication is performed; thus, the fusion characteristics of the IGDM output are obtained.

6. The map hotspot detection method based on multi-feature representation learning as described in claim 1, characterized in that: Feature fusion The network data processing process includes: 1×1 convolution and bilinear interpolation are performed on D1~D4 to obtain intermediate features of the same scale. ;Will The global features are obtained by adding the matrices together. Global feature descriptors are obtained by performing global average pooling and fully connected layer processing on the global features. s ; through four fully connected layers of different scales and the Softmax activation function from s Feature descriptors of different scales were extracted from the data. s 1~ s 4; will s 1~ s 4. As attention, the features of D1 to D4 are weighted respectively, and the resulting feature matrices are summed to obtain the fused features output by the feature fusion network. D out .

7. The map hotspot detection method based on multi-feature representation learning as described in claim 1, characterized in that: BCE loss with label smoothing L BCE The expression is as follows: ； In the above formula, y, and These represent the classification result, the true hard label, and the true soft label, respectively. smooth This represents the smoothing factor.

8. The map hotspot detection method based on multi-feature representation learning as described in claim 1, characterized in that: In the early stages of training, an constant learning rate is used; in the later stages of training, a periodically decreasing cosine learning rate is used.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it creates a layout hotspot detection model trained by the layout hotspot detection method based on multi-feature representation learning as described in any one of claims 1-8, and uses it to identify whether the input original layout belongs to a lithographic hotspot.

10. A detection device for photolithographic hotspots in a layout, 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 layout hotspot detection model trained by the layout hotspot detection method based on multi-feature representation learning as described in any one of claims 1-8, thereby enabling the identification of lithographic hotspots in the input layout.

Images