A double-scale feature-based photolithography hotspot detection method and related device

Abstract

The application relates to the technical field of photolithography hot spots, and provides a photolithography hot spot detection method based on double-scale features and related equipment, which comprises the following steps: obtaining a target circuit layout; performing double-scale feature fusion enhancement and detection on the target circuit layout by using a teacher model to obtain a first photolithography hot spot detection result and a first mask of the target circuit board; performing feature extraction and detection on the target circuit layout by using a student model to obtain a second photolithography hot spot detection result and a second mask of the target circuit layout; training the student model based on the first photolithography hot spot detection result, the first mask, the second photolithography hot spot detection result and the second mask to obtain a trained student model; and performing photolithography hot spot detection on a to-be-detected circuit layout by using the trained student model to obtain a photolithography hot spot detection result of the to-be-detected circuit layout. The method can improve the generality and precision of photolithography hot spot detection.

Description

Technical Field

[0001] This application relates to the field of photolithography hotspot technology, and in particular to a photolithography hotspot detection method and related equipment based on dual-scale features. Background Technology

[0002] As integrated circuit feature sizes continue to shrink to the nanometer level, the photolithography step in semiconductor manufacturing processes becomes increasingly complex. Due to the diffraction effect of light, the geometry on the design layout is distorted when transferred to the wafer, resulting in "lithography hotspots." Hotspot areas are prone to manufacturing defects such as short circuits and open circuits, severely impacting chip yield. Therefore, accurately and efficiently detecting lithography hotspots from the design layout before chip fabrication has become a key technical challenge in the field of Electronic Design Automation (EDA).

[0003] The current mainstream lithography hotspot detection methods can be divided into three categories: (1) Lithography simulation method: high accuracy but huge computational load, difficult to handle full-chip-scale layout. (2) Pattern matching method: fast speed but unable to detect unknown or deformed hotspot patterns, poor generalization ability. (3) Machine learning / deep learning method: has become mainstream in recent years, but still has the following prominent problems in practical applications: insufficient fine-grained detail capture ability; lack of interaction between local and neighborhood features; serious overlap of feature distribution, for "hard-to-classify" (HTC) mode, the false alarm rate of some advanced models exceeds 83%; large model scale, difficult to deploy.

[0004] Therefore, current methods for detecting lithographic hotspots suffer from limitations in versatility and accuracy. Summary of the Invention

[0005] This application provides a method and related equipment for detecting lithographic hotspots based on dual-scale features, which can solve the problems of poor versatility and accuracy in lithographic hotspot detection.

[0006] Firstly, this application provides a lithographic hotspot detection method based on dual-scale features. This lithographic hotspot detection method: Obtain the target circuit layout; The teacher model is used to perform dual-scale feature fusion enhancement and detection on the target circuit layout to obtain the first lithographic hot spot detection result and the first mask of the target circuit board. The student model is used to extract and detect features of the target circuit layout, and the second lithographic hot spot detection result and the second mask of the target circuit layout are obtained. Based on the first lithographic hotspot detection results, the first mask, the second lithographic hotspot detection results, and the second mask, the student model is trained to obtain the trained student model. The trained student model is used to perform lithographic hot spot detection on the circuit layout to be tested, and the lithographic hot spot detection results of the circuit layout to be tested are obtained.

[0007] Optionally, the teacher model is used to perform dual-scale feature fusion enhancement and detection on the target circuit layout to obtain the first lithographic hotspot detection result and the first mask of the target circuit board, including: The teacher model is used to extract dual-scale features from the target circuit layout, resulting in large-scale and small-scale features of the target circuit layout. Large-scale features and small-scale features are concatenated to obtain fused features; Based on the fusion features, the first lithographic hot spot detection result of the target circuit layout is obtained; Based on the first lithographic hotspot detection results, the classification loss is calculated, and the first mask of the target circuit layout is calculated based on the classification loss.

[0008] Optionally, lithographic hotspot detection is performed based on fusion features to obtain the first lithographic hotspot detection result of the target circuit layout, including: Sparse contrast enhancement is applied to the fused features to obtain enhanced features; The enhanced features are classified and detected to obtain the first lithographic hotspot detection result of the target circuit layout.

[0009] Optionally, based on the detection results of the first lithography hotspot, the classification loss is calculated, including: Through the formula:

[0010] Calculate classification loss ; in, Indicates category weight, This indicates the degree of attention paid to difficult-to-classify samples. This indicates the detection results of the first lithography hotspot; The first mask of the target circuit layout is calculated based on classification loss, including: Through the formula:

[0011]

[0012] Calculate the normalized first mask ; in, Represents the gradient. Indicates the target circuit layout. It is a constant.

[0013] Optionally, the student model is used to extract and detect features of the target circuit layout to obtain the second lithographic hotspot detection result and the second mask of the target circuit layout, including: Using a student model, feature extraction is performed on the target circuit layout to obtain image features; The image features are classified and detected to obtain the second lithographic hotspot detection result of the target circuit layout; The second mask of the target circuit layout is obtained by calculating the mask based on the image features.

[0014] Optionally, based on the first lithographic hotspot detection results, the first mask, the second lithographic hotspot detection results, and all second masks, the student model is trained to obtain the trained student model, including: A classification loss function is constructed based on the first and second lithographic hotspot detection results. Construct a mask loss function based on the first mask and the second mask; Construct a comprehensive loss function based on the classification loss function and the masking loss function; The student model is trained using a comprehensive loss function to obtain the trained student model.

[0015] Optionally, the classification loss function is:

[0016] in, This represents the value of the classification loss function. Indicates the temperature compensation coefficient. Denotes KL divergence, Indicates adoption Detection results of the second photolithography hotspot after softening. Indicates adoption Detection results of the first lithographic hotspot after softening; The mask loss function is:

[0017] in, This represents the value of the mask loss function. Indicates the first mask The mask at the location, Indicating the second mask The mask value at that location.

[0018] Optionally, the comprehensive loss function is:

[0019] in, This represents the value of the comprehensive loss function. Indicates hard label loss. These represent the weight coefficients of the classification loss function. This represents the weighting coefficients of the mask loss function.

[0020] Secondly, this application provides a photolithographic hotspot detection device based on dual-scale features, comprising: The acquisition module is used to acquire the target circuit layout; The first detection module is used to perform dual-scale feature fusion enhancement and detection on the target circuit layout using the teacher model, and obtain the first lithographic hot spot detection result and the first mask of the target circuit board. The second detection module is used to extract and detect features of the target circuit layout using the student model, and obtain the second lithographic hot spot detection result and the second mask of the target circuit layout. The training module is used to train the student model based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result, and the second mask, so as to obtain the trained student model. The third detection module is used to perform lithographic hotspot detection on the circuit layout to be tested using the trained student model, and obtain the lithographic hotspot detection results of the circuit layout to be tested.

[0021] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned lithographic hotspot detection method based on dual-scale features.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for detecting lithographic hotspots based on dual-scale features.

[0023] The above-mentioned solution in this application has the following beneficial effects: In the embodiments of this application, a target circuit layout is obtained, and then a teacher model is used to perform dual-scale feature fusion enhancement and detection on the target circuit layout to obtain the first lithographic hotspot detection result and the first mask of the target circuit board. Then, a student model is used to extract and detect features of the target circuit layout to obtain the second lithographic hotspot detection result and the second mask of the target circuit layout. Then, based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result and the second mask, the student model is trained to obtain the trained student model. Finally, the trained student model is used to perform lithographic hotspot detection on the circuit layout to be detected to obtain the lithographic hotspot detection result of the circuit layout to be detected. Specifically, the lithographic hotspot detection results and output masks of the teacher model are used to train the student model. This allows the mask generated by the teacher model to guide the student model to focus on geometric regions, enabling the student model to achieve deep interaction with feature distribution and cross-scale information and effective transfer of spatial attention. This improves the lithographic hotspot detection performance of the student model. Using the trained student model for lithographic hotspot detection allows for direct deployment of the model without the need for model training during use, reducing the difficulty of model deployment and effectively improving the versatility and accuracy of lithographic hotspot detection.

[0024] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart of a lithographic hotspot detection method based on dual-scale features provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a teacher model provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a sliding cross attention module provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a lithographic hotspot detection device based on dual-scale features provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0027] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0028] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0029] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0030] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0031] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0032] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0033] To address the limitations of existing lithographic hotspot detection methods in terms of versatility and accuracy, this application provides a lithographic hotspot detection method based on dual-scale features. This method acquires a target circuit layout, then uses a teacher model to perform dual-scale feature fusion enhancement and detection on the target circuit layout, obtaining a first lithographic hotspot detection result and a first mask. Next, a student model is used to extract and detect features from the target circuit layout, obtaining a second lithographic hotspot detection result and a second mask. Based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result, and the second mask, the student model is trained to obtain a trained student model. Finally, the trained student model is used to perform lithographic hotspot detection on the circuit layout to be detected, obtaining the lithographic hotspot detection result of the circuit layout to be detected. Specifically, the lithographic hotspot detection results and output masks of the teacher model are used to train the student model. This allows the mask generated by the teacher model to guide the student model to focus on geometric regions, enabling the student model to achieve deep interaction with feature distribution and cross-scale information and effective transfer of spatial attention. This improves the lithographic hotspot detection performance of the student model. Using the trained student model for lithographic hotspot detection allows for direct deployment of the model without the need for model training during use, reducing the difficulty of model deployment and effectively improving the versatility and accuracy of lithographic hotspot detection.

[0034] The following is an illustrative example of the lithographic hotspot detection method based on dual-scale features provided in this application.

[0035] like Figure 1 As shown, the lithographic hotspot detection method based on dual-scale features provided in this application includes the following steps: Step 11: Obtain the target circuit layout.

[0036] The target circuit layout described above is a design layout for an integrated circuit.

[0037] In some embodiments of this application, the target circuit layout can be obtained using integrated circuit layout design software (such as KLayout).

[0038] Step 12: Use the teacher model to perform dual-scale feature fusion enhancement and detection on the target circuit layout to obtain the first lithographic hot spot detection result and the first mask of the target circuit board.

[0039] The aforementioned first lithographic hotspot detection result is used to describe whether the circuit area corresponding to each pixel in the target circuit layout, as predicted by the teacher model, is a lithographic hotspot, and can be a classification probability, etc. The aforementioned first mask includes the binary mask corresponding to each pixel in the target circuit layout, as predicted by the teacher model, with a value of 1 or 0. 1 indicates that the pixel has an impact on lithographic hotspot detection, and 0 indicates that the pixel has no impact on lithographic hotspot detection.

[0040] In some embodiments of this application, the steps described above for using a teacher model to perform dual-scale feature fusion enhancement and detection on the target circuit layout to obtain the first lithographic hotspot detection result and the first mask of the target circuit board include: The first step is to use the teacher model to extract dual-scale features from the target circuit layout, thereby obtaining the large-scale and small-scale features of the target circuit layout.

[0041] For example, the teacher model includes a feature extraction network for bi-branch feature extraction. This network includes a large-scale segmentation module (used to segment the target circuit layout into large-scale images using algorithms such as image segmentation), a large-scale SwinTransformer module (used to extract features from the input large-scale image, which is a Swin Transformer model), an upsampling module, a small-scale segmentation module (used to segment the target circuit layout into small-scale images using algorithms such as image segmentation, where the scale of the output image is smaller than that of the large-scale segmentation module), a small-scale SwinTransformer module (used to extract features from the input small-scale image, which is a Swin Transformer model), and two sliding cross-attention modules (where the cross-attention operation is based on a sliding window).

[0042] After the large-scale segmentation module segments the target circuit layout, the large-scale image is input into the large-scale SwinTransformer module for feature extraction. The feature extraction result is then upsampled by the upsampling module. Simultaneously, in another branch, the small-scale segmentation module segments the target circuit layout, and the small-scale image is input into the small-scale SwinTransformer module for feature extraction. Then, the feature extraction results of both the small-scale and large-scale branches are input into two sliding cross-attention modules. In each sliding cross-attention module, sliding cross-attention operations are performed on the input feature extraction results of the small-scale and large-scale branches, and the large-scale and small-scale features are output respectively.

[0043] The second step is to stitch together the large-scale features and the small-scale features to obtain the fused features.

[0044] For example, the teacher model also includes a splicing module (used to execute splicing functions such as concat). The input of this module is connected to the output of two sliding cross attention modules. Large-scale features and small-scale features are input into the splicing module for splicing to obtain fused features.

[0045] The third step is to perform lithographic hotspot detection based on the fusion features to obtain the first lithographic hotspot detection result of the target circuit layout.

[0046] Specifically, sparse contrast enhancement is performed on the fused features to obtain enhanced features; the enhanced features are then classified and detected to obtain the first lithographic hotspot detection result of the target circuit layout.

[0047] For example, the teacher model also includes a convolution module (which can be a convolution operation conv), a sparse contrast enhancement module, a classification head, and a mask calculation module connected in sequence. The input of the convolution module is connected to the input of the splicing module. The output of the classification head outputs the first lithographic hotspot detection result, and the output of the mask calculation module outputs the first mask. The sparse contrast enhancement module can use a graph neural network-based model to perform sparse contrast enhancement on the fused features to obtain enhanced features. The classification head can use fully connected layers to classify and detect the enhanced features to obtain the first lithographic hotspot detection result of the target circuit layout. The operation process of the mask calculation module is the process of calculating the first mask described later.

[0048] The process of sparse contrast enhancement is as follows: (1) Construction of Top-k sparse similarity graph: Fusion features Flattened , .

[0049] Calculate any two feature vectors in the fused features using the following formula. , Cosine similarity between :

[0050] in Represents the dot product. Let L2 norm be denoted. To reduce computational complexity and focus on the most relevant regions, only the k neighbors with the highest similarity for each feature vector (k=10) are retained, constructing a sparse adjacency matrix. :

[0051] (2) Utilize graph convolution to aggregate neighborhood optimization features: Based on the sparse adjacency matrix E, the edge weights are defined as similarity scaled by temperature:

[0052] Among them temperature parameters =0.5. Then, the neighbor features are aggregated through graph convolution to obtain the aggregated semantic context of each node. :

[0053] in Represents a node The neighborhood group, , Let be the degrees of nodes i and j. This aggregation operation allows each node to absorb semantically similar surrounding information.

[0054] (3) Enhanced characteristic nonlinear saturation contrast This application employs non-linear saturation updates, thus avoiding the training instability issues that may arise from linear updates.

[0055] in =0.1 is the update intensity hyperparameter. This formula moves features away from their neighborhood aggregation centers, with the step size and difference magnitude monotonically increasing but bounded above. At that time, the step size is approximately... The update amount approaches zero, maintaining compactness within the class; when the difference is large, the step size approaches... This maintains inter-class distances and avoids training instability caused by outliers. Unlike the linear updates of existing contrastive learning methods (such as ContraNorm), the update step size approaches [value missing] when the features differ significantly from their neighborhoods. This avoids the abnormal outlier disturbances and training instability that may be caused by linear updates, and can effectively alleviate the problems of dimensionality collapse and feature overlap in the feature space.

[0056] (4) Feature normalization and reshaping Update the feature vector Perform L2 normalization, then reshape back to the original spatial dimensions:

[0057] Output the final enhanced features For subsequent classification purposes.

[0058] The fourth step is to calculate the classification loss based on the first lithographic hotspot detection result, and then calculate the first mask of the target circuit layout based on the classification loss.

[0059] Specifically, through the formula:

[0060] Calculate classification loss .

[0061] in, Indicates category weight, This indicates the degree of attention paid to difficult-to-classify samples. This indicates the detection result of the first lithography hotspot.

[0062] Through the formula:

[0063]

[0064] Calculate the normalized first mask .

[0065] in, Represents the gradient. Indicates the target circuit layout. It is a constant.

[0066] It is understandable that the process of calculating the classification loss based on the first lithographic hotspot detection result and calculating the first mask of the target circuit layout based on the classification loss in this step is the same as the operation process of the mask calculation module in the teacher model.

[0067] In some embodiments of this application, the structure of the above-described teacher model is as follows: Figure 2 As shown, the input X ([B, 3, 1200, 1200]) is split into two paths, which enter the large-scale segmentation (patch size=4) and small-scale segmentation (patch size=2) respectively. The output of the large-scale segmentation is connected to the large-scale Swin Transformer module. Its output is upsampled and combined with the features of the small-scale Swin Transformer module (output [B, 128, 14, 14]) connected to the small-scale segmentation output. These features are then input into two sliding cross-attention modules. The outputs of the two sliding cross-attention modules are concatenated to form the output of the dual-scale feature fusion module, which is then input into a convolutional module Conv(1x1). The inputs are then sequentially fed into the graph sparse contrast enhancement module and the feedforward network to obtain the classification result. One output of the classification result is the output classification, and the other output is used to calculate the classification loss. Calculated through backpropagation After normalization and binarization, the output is a hotspot saliency mask with the shape [B, 1, 224, 224]. .

[0068] The small-scale branch employs a lightweight Swing Transformer with a patch size of 2, an embedding dimension of 16, a layer count of {1,1,2,1}, 4 attention heads, and an output resolution of 14×14, capturing sub-micron-level geometric details such as line width, line spacing, and corners. The large-scale branch also employs a lightweight Swing Transformer with a patch size of 4, an embedding dimension of 16, a layer count of {1,1,2,1}, 4 attention heads, and an output resolution of 7×7, capturing global context such as neighborhood layout and region density. Large-scale features are upsampled to 14×14 using transposed convolutions (4×4 kernel, 2 stride).

[0069] The sliding cross attention module structure is as follows Figure 3 As shown, the dual-scale feature ([B, 128, 14, 14]) is split into two paths, which are respectively fed into large-scale feature linearization and small-scale feature linearization. The output feature enters the window cross-attention module, and then... and The cross-attention is calculated, and the two results are fed into the offset window cross-attention module. After window sliding (window sliding scale = window size / / 2, sliding adaptation size) and linearization to generate new Q, K, and V, they are passed again. and Cross-attention is calculated, and then the feature is obtained by window sliding adaptation, residual connection and feedforward network. The output of the two feedforward networks is spliced ​​to obtain the feature [B, 256, 14, 14], which is then input into the graph-based sparse contrast enhancement module.

[0070] The calculation process is as follows: (1) First, large-scale features Small-scale features Divide into 7x7 windows.

[0071] (2) Further enhance the query of this branch by using the key / value of the other branch:

[0072]

[0073] in , , Depend on Obtained by linear transformation, , , Depend on Obtained by linear transformation, For the embedded dimension.

[0074] (3) Next, calculate the cross attention of the offset window. Offset the window by 3 pixels and then calculate the cross attention according to the above formula.

[0075] (4) Finally, the residuals are concatenated along the channel dimension to obtain the fused feature map. .

[0076] Step 13: Use the student model to extract and detect features of the target circuit layout to obtain the second lithographic hot spot detection result and the second mask of the target circuit layout.

[0077] The aforementioned second lithographic hotspot detection result is used to describe whether the circuit area corresponding to each pixel in the target circuit layout, as predicted by the student model, is a lithographic hotspot, and can be a classification probability, etc. The aforementioned second mask includes the binary mask corresponding to each pixel in the target circuit layout, as predicted by the student model.

[0078] In some embodiments of this application, the steps of using a student model to extract and detect features of the target circuit layout to obtain the second lithographic hotspot detection result and the second mask of the target circuit layout include: The first step is to use the student model to extract features from the target circuit layout to obtain image features.

[0079] For example, the student model includes a series of convolutional layers, a classification head, and a masking terminal. These convolutional layers can be used to extract features from the target circuit layout, obtaining image features. These image features are then input into the classification head and the masking terminal for computation. The classification head outputs the second lithographic hotspot detection result, and the masking terminal outputs the second mask. For instance, the input image undergoes the following four convolutions to obtain the feature tensor F: Conv1: 3×3 convolution, stride 2, channels 3→32, BatchNorm+ReLU, output 112×112×32.

[0080] Conv2: 3×3 convolution, stride 2, channels 32→64, BatchNorm+ReLU, output 56×56×64.

[0081] Conv3: 3×3 convolution, stride 2, channels 64→128, BatchNorm+ReLU, output 28×28×128.

[0082] Conv4: 3×3 convolution, stride 2, channels 128→256, BatchNorm+ReLU, output 14×14×256.

[0083] The second step is to classify and detect the image features to obtain the second lithographic hotspot detection results of the target circuit layout.

[0084] For example, the classification head in the student model can be a classification head composed of a fully connected layer and an activation function softmax. This classification head is used to classify and detect image features to obtain the second lithographic hotspot detection result of the target circuit layout.

[0085] The third step is to calculate the mask based on the image features to obtain the second mask of the target circuit layout.

[0086] For example, the mask in the student model can be composed of four consecutive transposed convolutions, dimensionality reduction convolutions, and activation functions. The image features are progressively upsampled using four transposed convolutions: the first level upsamples from 14×14 to 28×28 (transposed convolution kernel 4×4, stride 2), the second level upsamples to 56×56, the third level upsamples to 112×112, and the fourth level upsamples to 224×224. Each transposed convolution is followed by a normalization layer and a ReLU activation function, then dimensionality is reduced to 1 channel by a 1×1 convolution, activated by Sigmoid, and the output is the second mask.

[0087] It is understandable that the calculation process in step 13 above is the same as the calculation process of the student model.

[0088] Step 14: Based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result, and the second mask, train the student model to obtain the trained student model.

[0089] In some embodiments of this application, the steps of training the student model based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result, and the second mask to obtain the trained student model include: The first step is to construct a classification loss function based on the detection results of the first and second lithographic hotspots.

[0090] Specifically, the classification loss function is:

[0091] in, This represents the value of the classification loss function. Indicates the temperature compensation coefficient. Denotes KL divergence, Indicates adoption Detection results of the second photolithography hotspot after softening. Indicates adoption Results of the first photolithography hotspot detection after softening:

[0092]

[0093] in, The first lithographic hotspot detection result output by the teacher model. The second lithographic hotspot detection result is output for the student model.

[0094] The second step is to construct a mask loss function based on the first and second masks.

[0095] Specifically, the mask loss function is:

[0096] in, This represents the value of the mask loss function. Indicates the first mask The mask value at that location, Indicating the second mask The mask value at that location.

[0097] The third step is to construct a comprehensive loss function based on the classification loss function and the mask loss function.

[0098] Specifically, the comprehensive loss function is:

[0099] in, This represents the value of the comprehensive loss function. This represents the hard label loss (calculated using the focal loss function to ensure that the student model does not deviate from the real lithographic hotspots of the target circuit layout). The weight coefficients of the classification loss function (can be 0.5). This represents the weighting coefficient of the mask loss function (which can be 0.8).

[0100] The fourth step is to train the student model using the comprehensive loss function to obtain the trained student model.

[0101] For example, the Adam optimizer or similar tool can be used to train the student model using a comprehensive loss function, modulating the model parameters in the student model until the comprehensive loss function converges (e.g., the value of the comprehensive loss function is less than a preset value), thus obtaining the trained student model.

[0102] It is understood that when training a model, multiple samples are usually used for multiple training sessions. The above steps in this application only describe the processing of one sample. The processing of multiple samples is the same as the above steps.

[0103] Step 15: Use the trained student model to perform photolithographic hotspot detection on the circuit layout to be tested, and obtain the photolithographic hotspot detection results of the circuit layout to be tested.

[0104] The circuit layout to be tested described above is the design layout of an integrated circuit that requires photolithographic hotspot detection. The photolithographic hotspot detection results are used to describe whether the circuit area corresponding to each pixel in the circuit layout to be tested is a photolithographic hotspot, and can be classified as a probability, etc.

[0105] Specifically, the circuit layout to be detected is input into the trained student model, features are extracted from the circuit layout to obtain image features, and the image features are classified and detected to obtain the lithographic hotspot detection results of the circuit layout to be detected.

[0106] It is worth mentioning that by using the lithographic hotspot detection results and output masks of the teacher model to train the student model, the mask generated by the teacher model can guide the student model to focus on geometric regions. This enables the student model to achieve deep interaction with feature distribution and cross-scale information and effectively transfer spatial attention, thereby improving the lithographic hotspot detection performance of the student model. Using the trained student model for lithographic hotspot detection allows for direct deployment of the model without the need for model training during use, reducing the difficulty of model deployment and effectively improving the versatility and accuracy of lithographic hotspot detection.

[0107] Furthermore, this application has the following advantages: 1. Dual-scale feature fusion: By using parallel small-scale branches (capturing submicron-level geometric details such as line width, line spacing, and corners) and large-scale branches (capturing global context such as neighborhood layout and region density), the limitations of single-scale receptive fields are overcome, and the ability to perceive hotspot and key areas in complex maps is significantly improved.

[0108] 2. Sliding cross attention: By employing window cross attention and offset window cross attention, bidirectional information injection between local details and neighborhood context is achieved, enabling the model to dynamically adjust the attention weights for features at different scales, effectively enhancing the interaction between local and global features.

[0109] 3. Graph-based sparse contrast enhancement: By constructing a Top-k sparse similarity graph, graph convolutional neighborhood aggregation, and contrastive feature updates, the inter-class distance is actively increased and the intra-class distance is reduced without introducing an additional contrastive loss function. This effectively alleviates the overlap problem between hot and non-hot samples in the feature space and improves the clarity of the classification boundary.

[0110] 4. Hotspot Saliency Mask Distillation: The teacher model generates a binary saliency mask using gradient integration, accurately labeling the pixel regions (such as dense lines, corners, and line ends) that contribute most to hotspot classification in the map. The student model learns this mask through a mask prediction head, thus forcing the student to focus on the same key geometric regions as the teacher, inheriting the teacher's discriminative ability while maintaining lightweight design. Compared to methods using only classification distillation, this application further reduces the false positive rate through saliency mask distillation.

[0111] 5. Categorical Distillation: Using temperature-softened KL divergence loss, the student model mimics the output probability distribution of the teacher model, thereby learning the teacher model's detailed understanding of inter-category similarity and improving the generalization ability of the student model.

[0112] 6. Lightweight Deployment Capability: The student model uses a lightweight convolutional neural network (with approximately 1 / 4 the number of parameters of the teacher model), resulting in fast inference speed and suitability for deployment in actual production lines. After distillation training, only the classification branch of the student model needs to be retained; the mask prediction head is only used for training and does not increase inference overhead.

[0113] 7. Overall Results: Through the aforementioned multi-scale feature enhancement, feature distribution optimization, and double distillation strategy, this application effectively controls the false positive rate while maintaining a high recall rate. Compared with the method using only classification distillation, this application further guides students to focus on key spatial regions through saliency mask distillation, achieving better false positive rate control. At the same time, the number of model parameters is significantly reduced, and the inference speed is significantly improved, making it suitable for actual production line deployment.

[0114] The following is an exemplary description of the lithographic hotspot detection device based on dual-scale features provided in this application.

[0115] like Figure 4 As shown, this application provides a lithography hotspot detection device based on dual-scale features. The lithography hotspot detection device 400 based on dual-scale features includes: Acquisition module 401 is used to acquire the target circuit layout; The first detection module 402 is used to perform dual-scale feature fusion enhancement and detection on the target circuit layout using the teacher model, so as to obtain the first lithographic hot spot detection result and the first mask of the target circuit board. The second detection module 403 is used to extract and detect features of the target circuit layout using the student model, and obtain the second lithographic hot spot detection result and the second mask of the target circuit layout. Training module 404 is used to train the student model based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result, and the second mask to obtain the trained student model. The third detection module 405 is used to perform lithographic hot spot detection on the circuit layout to be tested using the trained student model, and obtain the lithographic hot spot detection result of the circuit layout to be tested.

[0116] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0117] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0118] like Figure 5 As shown, an embodiment of this application provides a terminal device, wherein the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 5 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0119] Specifically, when the processor D100 executes the computer program D102, it acquires the target circuit layout, then uses the teacher model to perform dual-scale feature fusion enhancement and detection on the target circuit layout to obtain the first lithographic hotspot detection result and the first mask of the target circuit board. Then, it uses the student model to perform feature extraction and detection on the target circuit layout to obtain the second lithographic hotspot detection result and the second mask of the target circuit layout. Then, based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result, and the second mask, it trains the student model to obtain the trained student model. Finally, it uses the trained student model to perform lithographic hotspot detection on the circuit layout to be detected to obtain the lithographic hotspot detection result of the circuit layout to be detected. Specifically, the lithographic hotspot detection results and output masks of the teacher model are used to train the student model. This allows the mask generated by the teacher model to guide the student model to focus on geometric regions, enabling the student model to achieve deep interaction with feature distribution and cross-scale information and effective transfer of spatial attention. This improves the lithographic hotspot detection performance of the student model. Using the trained student model for lithographic hotspot detection allows for direct deployment of the model without the need for model training during use, reducing the difficulty of model deployment and effectively improving the versatility and accuracy of lithographic hotspot detection.

[0120] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0121] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0122] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0123] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0124] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above-described embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a dual-scale feature-based lithography hotspot detection method apparatus / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0125] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0126] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0127] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A method for detecting lithographic hotspots based on dual-scale features, characterized in that, include: Obtain the target circuit layout; The target circuit layout is enhanced and detected by dual-scale feature fusion using a teacher model to obtain the first lithographic hot spot detection result and the first mask of the target circuit board. The target circuit layout is feature extracted and detected using a student model to obtain the second lithographic hot spot detection result and the second mask of the target circuit layout. Based on the first lithographic hotspot detection result, the first mask, the second lithographic hotspot detection result, and the second mask, the student model is trained to obtain the trained student model. The trained student model is used to perform lithographic hotspot detection on the circuit layout to be tested, and the lithographic hotspot detection result of the circuit layout to be tested is obtained.

2. The photolithographic hotspot detection method according to claim 1, characterized in that, The method of using a teacher model to perform dual-scale feature fusion enhancement and detection on the target circuit layout to obtain the first lithographic hotspot detection result and the first mask of the target circuit board includes: The target circuit layout is subjected to dual-scale feature extraction using a teacher model to obtain the large-scale and small-scale features of the target circuit layout. The large-scale features and the small-scale features are concatenated to obtain the fused features; Based on the fusion features, photolithography hotspot detection is performed to obtain the first photolithography hotspot detection result of the target circuit layout; Based on the first photolithography hotspot detection result, a classification loss is calculated, and a first mask of the target circuit layout is calculated based on the classification loss.

3. The photolithographic hotspot detection method according to claim 2, characterized in that, The step of performing photolithographic hotspot detection based on the fusion features to obtain the first photolithographic hotspot detection result of the target circuit layout includes: The fused features are then subjected to sparse contrast enhancement to obtain enhanced features; The enhanced features are classified and detected to obtain the first lithographic hotspot detection result of the target circuit layout.

4. The photolithographic hotspot detection method according to claim 3, characterized in that, The step of calculating the classification loss based on the first photolithography hotspot detection result includes: Through the formula: Calculate classification loss ; in, Indicates category weight, This indicates the degree of attention paid to difficult-to-classify samples. This indicates the detection results of the first lithography hotspot; The calculation of the first mask of the target circuit layout based on the classification loss includes: Through the formula: Calculate the normalized first mask ; in, Represents the gradient. Indicates the target circuit layout. It is a constant.

5. The photolithographic hotspot detection method according to claim 1, characterized in that, The step of using a student model to extract and detect features of the target circuit layout to obtain the second lithographic hotspot detection result and the second mask of the target circuit layout includes: Using the student model, feature extraction is performed on the target circuit layout to obtain image features; The image features are classified and detected to obtain the second lithographic hotspot detection result of the target circuit layout; The second mask of the target circuit layout is obtained by performing mask calculation based on the image features.

6. The photolithographic hotspot detection method according to claim 1, characterized in that, The student model is trained based on the first lithographic hotspot detection results, the first mask, the second lithographic hotspot detection results, and all second masks to obtain the trained student model, including: A classification loss function is constructed based on the first and second lithographic hotspot detection results. Construct a mask loss function based on the first mask and the second mask; Construct a comprehensive loss function based on the classification loss function and the masking loss function; The student model is trained using the comprehensive loss function to obtain the trained student model.

7. The photolithographic hotspot detection method according to claim 6, characterized in that, The classification loss function is: in, This represents the value of the classification loss function. Indicates the temperature compensation coefficient. Denotes KL divergence, Indicates adoption Detection results of the second photolithography hotspot after softening. Indicates adoption Detection results of the first lithographic hotspot after softening; The mask loss function is: in, This represents the value of the mask loss function. Indicates the first mask The mask at the location, Indicating the second mask The mask value at that location.

8. The photolithographic hotspot detection method according to claim 7, characterized in that, The comprehensive loss function is: in, This represents the value of the comprehensive loss function. Indicates hard label loss, The weight coefficients of the classification loss function are represented by their respective values. This represents the weighting coefficients of the mask loss function.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the lithographic hotspot detection method based on dual-scale features as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the lithographic hotspot detection method based on dual-scale features as described in any one of claims 1 to 8.

Images