Calculation photoetching modeling and hot spot detection method and device based on self-supervised learning

Abstract

The invention relates to a computational lithography modeling and hot spot detection method and device based on self-supervised learning, and the method comprises the steps: constructing a neural network model which comprises a layout feature encoder, a geometric reconstruction branch and a lithography space image prediction branch, and enabling a differentiable optical simulation layer to be built in the lithography space image prediction branch so as to integrate physical prior; after an input layout is masked based on label-free mask plate graph data, the model is driven to execute pre-training of geometric reconstruction and photoetching space image prediction at the same time, an original unmasked layout is processed through the differentiable optical simulation layer to generate a pseudo-true value, and a supervision signal is provided for photoetching space image prediction; and on the basis of the pre-training model, performing fine tuning by using a small amount of labeled data to realize hotspot detection. The differentiable optical simulation layer is integrated into the self-supervised learning framework, so that the model can learn the optical law from massive label-free data, the dependence on labeled data is reduced, and the hotspot detection capability and accuracy of unknown graphs are improved.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and semiconductor manufacturing technology, specifically to a computational lithography modeling and hotspot detection method and apparatus based on self-supervised learning. Background Technology

[0002] With integrated circuit manufacturing processes advancing to 5-nanometer and more advanced technology nodes, photolithography, as a core component, faces severe challenges due to the optical diffraction limit. To ensure that the patterns on the wafer accurately replicate the design layout, the industry commonly employs resolution enhancement techniques, such as inversion lithography and light source-mask co-optimization techniques. The application of these techniques results in highly complex curved patterns on the mask, rather than traditional rectangles or polygons.

[0003] During the photolithography imaging process, such complex patterns are highly susceptible to defects in specific areas due to optical proximity effects and variations in the process window. These defects can manifest as bridging, open circuits, or narrowing of linewidths. These defect-prone areas are known as photolithography hotspots. Accurate and rapid detection of photolithography hotspots across the entire chip is crucial for ensuring and improving chip manufacturing yields at advanced process nodes. Traditional detection methods rely on physical lithography simulation. While this method can provide high-precision imaging prediction results, its computational process is extremely time-consuming and cannot meet the demands of rapid verification of the entire chip layout in large-scale production.

[0004] In recent years, deep learning-based methods have been introduced into hotspot detection tasks due to their efficient reasoning capabilities. However, existing deep learning methods mostly employ supervised learning paradigms, and their performance is highly dependent on large-scale, high-quality labeled data. In the field of photolithography, obtaining samples with hotspot labels is extremely costly, typically requiring expensive wafer exposure experiments or time-consuming precise physical simulations. This results in limited training dataset size and a severe class imbalance problem.

[0005] To utilize massive amounts of unlabeled lithography layout data, self-supervised learning offers a viable solution. However, directly applying general computer vision self-supervised learning frameworks to lithography layout data presents a representation mismatch problem. General self-supervised methods typically construct pre-training tasks based on geometric similarity, using data augmentation techniques such as rotation, cropping, or mask reconstruction to enable the model to learn geometric structural features. However, in the lithography process, the relationship between the geometry of the mask pattern and its imaging result on the wafer is a highly nonlinear optical transformation. Visually vastly different mask patterns may produce the same imaging effect due to optical equivalence; conversely, geometrically very similar patterns may result in significant differences in imaging quality due to only minor variations in sub-resolution auxiliary patterns. Existing self-supervised methods fail to capture this implicit lithography physical law. The features learned by their pre-trained models represent the geometry of the layout more than its printability, leading to insufficient generalization ability and a high false positive rate when faced with complex, unseen curved masks. Therefore, how to integrate the physical mechanism of photolithography into the self-supervised learning process so that the model can learn feature representations with optical perception capabilities on unlabeled data is a problem that urgently needs to be solved in the current technology field. Summary of the Invention

[0006] One aspect of this invention provides a computational lithography modeling and hotspot detection method based on self-supervised learning, which mainly solves the technical problems of existing deep learning hotspot detection methods relying on a large amount of expensive labeled data, and general self-supervised learning methods having poor model generalization ability and low detection accuracy due to their inability to capture the physical laws of lithography.

[0007] According to one aspect of the present invention, a computational lithography modeling and hotspot detection method based on self-supervised learning is provided, comprising the following steps: Step 1: Construct an unlabeled mask layout dataset, which includes multiple mask layout data fragments extracted and segmented from the chip design layout database; Step 2: Construct a self-supervised pre-trained neural network model based on physical perception. This model includes a layout feature encoder, a geometric reconstruction branch, and a lithographic spatial image prediction branch with an embedded differentiable optical simulation layer. Step 3, perform self-supervised pre-training: perform masking operation on the mask layout data segment and input it into the layout feature encoder to obtain latent features; use the geometric reconstruction branch to calculate the geometric reconstruction loss based on the latent features; use the lithographic spatial image prediction branch to predict the lithographic spatial image based on the latent features, and calculate the optical physical loss between it and the pseudo-true lithographic spatial image generated by the differentiable optical simulation layer based on the original mask layout; Step 4: Update the network parameters of the self-supervised pre-trained neural network model based on the weighted sum of the geometric reconstruction loss and the optical physical loss; Step 5: Fine-tune the self-supervised pre-trained map feature encoder using sample data with hotspot labels to obtain the hotspot detection model; Step 6: Apply the hotspot detection model to perform hotspot detection on the entire chip mask layout.

[0008] Furthermore, the layout feature encoder employs a backbone network based on an attention-based Transformer architecture.

[0009] Furthermore, the step of performing a masking operation on the mask pattern data segment in step three specifically includes: using the differentiable optical simulation layer to calculate the gradient of the lithographic spatial image intensity relative to each pixel of the mask pattern data segment, generating a sensitivity map, wherein the sensitivity map quantifies the degree of influence of the change of each pixel on the mask on the lithographic imaging result; and setting the probability distribution of the masking operation according to the sensitivity map, so that the area with the higher value on the sensitivity map has a higher probability of being masked.

[0010] Furthermore, the differentiable optical simulation layer is constructed based on a coherent system summation model and includes a set of optical convolution kernels. These optical convolution kernels are generated by decomposing the coherent system summation model according to the numerical aperture of the lithography machine, the exposure wavelength, and the shape of the light source.

[0011] Furthermore, after step four and before step five, a step of performing contrastive learning enhancement based on optical equivalence is included. This step includes: applying an optical equivalent perturbation to the mask layout data segment to generate a perturbed mask layout data segment, wherein the optical equivalent perturbation refers to adjusting the position or size of the sub-resolution auxiliary graphics in the mask layout data segment without violating design rule constraints; treating the original mask layout data segment and the perturbed mask layout data segment as a pair of positive samples, and inputting the pair of positive samples into the layout feature encoder to obtain a pair of feature vectors; calculating the contrast loss, and updating the network parameters of the layout feature encoder by reducing the distance between the pair of feature vectors of the pair of positive samples in the feature space, while increasing the distance between the feature representations of different mask layout data segments in the feature space. The range for adjusting the sub-resolution auxiliary graphics is preferably ±1nm to 5nm.

[0012] Further, step five specifically includes: removing the geometric reconstruction branch and the lithographic spatial image prediction branch, and connecting the classification module to the output of the layout feature encoder; retaining the pre-trained network parameters of the layout feature encoder as initial values; and using a dataset with hotspot labels, performing supervised training on the network parameters of the layout feature encoder and the classification module by calculating the classification cross-entropy loss.

[0013] Furthermore, step one also includes performing random rotation and flipping and other data augmentation processing on the segmented mask layout data fragments.

[0014] Furthermore, step six specifically includes: dividing the full-chip mask layout to be detected into multiple data segments of the same size using a sliding window strategy with overlapping regions; inputting each data segment into the hotspot detection model and outputting a hotspot probability value for each data segment; and integrating the prediction results of all data segments to generate a hotspot distribution map covering the entire chip.

[0015] According to another aspect of the present invention, a computational lithography modeling and hotspot detection apparatus based on self-supervised learning is also provided, comprising: The data construction module is used to construct an unlabeled mask layout dataset, which includes multiple mask layout data fragments extracted and segmented from a chip design layout database; The model building module is used to build a self-supervised pre-trained neural network model based on physical perception. The model includes a layout feature encoder, a geometric reconstruction branch, and a lithographic spatial image prediction branch with an embedded differentiable optical simulation layer. The pre-training module is used to perform masking operations on the mask layout data fragments and input them into the layout feature encoder to obtain latent features. It then uses the geometric reconstruction branch to calculate the geometric reconstruction loss based on the latent features, uses the lithographic spatial image prediction branch to predict the lithographic spatial image based on the latent features, and calculates the optical physical loss between the predicted lithographic spatial image and the pseudo-true lithographic spatial image generated by the differentiable optical simulation layer based on the original mask layout. Finally, it updates the network parameters of the neural network model based on the weighted sum of the geometric reconstruction loss and the optical physical loss. The fine-tuning module is used to fine-tune the self-supervised pre-trained map feature encoder using sample data with hotspot labels to obtain a hotspot detection model. The detection application module is used to perform hotspot detection on the entire chip mask layout using the hotspot detection model.

[0016] This invention addresses the mismatch between geometric and physical representations in general self-supervised methods by introducing a differentiable optical simulation layer into a self-supervised learning framework. This allows the neural network model to autonomously learn complex nonlinear mappings between mask patterns and wafer imaging from massive amounts of unlabeled images without any manual annotation. Since the core training process does not rely on expensive hotspot labels and only requires a very small number of labeled samples for fine-tuning, data acquisition costs are significantly reduced. More importantly, because the model learns underlying optical diffraction patterns rather than specific geometric modes, it exhibits good generalization ability even for complex curves not present in the training set, accurately predicting their imaging risks and thus helping to reduce the false positive and false negative rates of hotspot detection. Attached Figure Description

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of a computational lithography modeling and hotspot detection method based on self-supervised learning provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of a self-supervised pre-trained neural network model based on physical perception constructed in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram illustrating the principle of the self-supervised pre-training process in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram illustrating the application of model fine-tuning and hotspot detection in an embodiment of the present invention.

[0022] Figure 5 This is a graph showing the experimental results of hotspot detection accuracy comparison provided in an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the sensitivity map generation process of the physical guidance mask strategy in this embodiment of the invention.

[0024] Figure 7 This is a comparison diagram of the mask pattern and its photolithographic imaging effect on the wafer in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] The core idea of ​​this invention is to construct a self-supervised learning framework capable of simulating the physical process of photolithography. This is achieved by designing a pre-trained model that incorporates both geometric reconstruction and optical prediction tasks, and by utilizing a differentiable optical simulation layer embedded within the model as prior physical knowledge to generate supervisory signals. This allows the neural network to learn deep feature representations with optical perception capabilities from massive amounts of unlabeled mask pattern data without manual annotation. These feature representations more accurately reflect the printability of the mask pattern, thereby achieving higher accuracy and generalization ability in subsequent hotspot detection tasks.

[0027] Example 1 This embodiment provides a computational lithography modeling and hotspot detection method based on self-supervised learning, referring to... Figure 1 The method includes the following steps: Step S100: Construct a large-scale unlabeled mask layout dataset.

[0028] In this embodiment, this step is used to prepare sufficient and diverse training data for subsequent self-supervised pre-training. Specifically, mask layout data in GDSII or OASIS format is extracted from integrated circuit design databases of one or more advanced process nodes (e.g., 5nm technology nodes). Preferably, the key pattern layers with the most stringent requirements for lithographic imaging quality, such as polysilicon layers, metal interconnect layers, or contact hole layers, are selected. During extraction, not only the geometric information of the main pattern is included, but also the sub-resolution auxiliary patterns around it that play a crucial role in imaging are fully preserved.

[0029] like Figure 7 The diagram illustrates a comparison between the mask layout and the wafer imaging effect. The left side shows the mask layout optimized by inversion lithography, containing a main pattern with complex curves and multiple sub-resolution auxiliary patterns (SRAFs) around it; the right side shows the corresponding wafer image. Due to the effects of optical diffraction and optical proximity, the imaging pattern on the wafer exhibits physical phenomena such as corner rounding, line-end contraction, and linewidth variations. SRAFs themselves do not directly image; their function is to improve the imaging quality of the main pattern. The method of this invention accurately captures these physical transformation laws by establishing a highly nonlinear mapping model between the mask and the imaging.

[0030] Subsequently, to adapt to the input requirements of the neural network model, the continuous layout data is sliced ​​into image data segments of predetermined sizes. In this embodiment, the size of each data segment can be set to 1024x1024 pixels, and the pixel resolution can be set to 1 nanometer per pixel. This setting achieves a good balance between capturing graphic details and controlling computational overhead. Then, the data segments are preprocessed to convert them into a binary image format, i.e., the pixel values ​​of the areas covered by the mask pattern are set to 1, and the pixel values ​​of the background areas are set to 0. To further expand the dataset size and improve the model's generalization ability, data augmentation operations are performed on the generated binary images. These operations include, but are not limited to, random geometric transformations, such as rotations of 0, 90, 180, or 270 degrees, and horizontal or vertical flips. These transformations are physically equivalent in photolithography and help the model learn rotation-invariant features. After the above processing, a large-scale unlabeled training set is finally formed, denoted as Dunlabeled = {Mi} (i=1...N), where Mi represents the i-th preprocessed mask layout data segment, and N is the total number of data segments. For example, the number of N can be one million or more.

[0031] Step S200: Construct a self-supervised pre-trained neural network model based on physical perception.

[0032] Reference Figure 2 The neural network model constructed in this embodiment adopts an encoder-decoder architecture and is designed with a structure that includes a dual-stream decoding branch. Specifically, the model consists of three parts: a layout feature encoder 201, a geometric reconstruction branch 202, and a lithographic spatial image prediction branch 203.

[0033] The layout feature encoder 201, denoted as E, maps the input mask layout data fragment M to a high-dimensional latent feature space. Preferably, this encoder employs a Swin Transformer network structure. Compared to traditional convolutional neural networks, the Swin Transformer, through its shift window and multi-head self-attention mechanism, can more effectively capture long-range spatial dependencies and complex topological structures existing in the mask pattern, which is crucial for accurately simulating the optical proximity effect.

[0034] Geometric reconstruction branch 202, denoted as The decoder is relatively simple in structure, and can be composed of several upsampling layers and convolutional layers. It receives latent features from the output of the layout feature encoder 201, and its task is to reconstruct the original mask layout. Its output can be represented as... .

[0035] Photolithographic spatial image prediction branch 203, denoted as This is a key component for realizing the core idea of ​​this invention. This branch also receives the latent features output by the layout feature encoder 201, but its prediction target is not the mask layout itself, but rather the lithographic spatial image formed by the layout under predetermined lithography process parameters. The core of this branch is the differentiable optical simulation layer 204. This layer is built based on a physical model, and its calculation process follows the theory of partially coherent imaging. In this embodiment, the Coherent System Decomposition (SOCS) model is used to approximate this physical process. Specifically, firstly, based on the actual parameters of the lithography system, including the exposure wavelength... The numerical aperture (NA) of the objective lens (e.g., 193 nm), and the specific shape distribution of the light source are pre-calculated and a set of fixed optical convolution kernels are generated using the SOCS decomposition algorithm. and the corresponding eigenvalues These convolutional kernels physically represent different coherent components of a partially coherent light source. The calculation process for this layer can be represented by the following formula: Where K is the total number of optical convolution kernels obtained after the coherent system summation model decomposition, and M is the input mask pattern. For the k-th eigenvalue, For the k-th optical convolution kernel, This represents a two-dimensional convolution operation. This represents the square of the modulus used to calculate light intensity. During the self-supervised pre-training phase, this set of optical convolutional kernels... and its eigenvalues They are fixed and do not participate in gradient updates. They are embedded in the neural network as a kind of structured physical prior knowledge to guide the model in learning the laws of optical transformation.

[0036] Step S300: Perform self-supervised pre-training based on mask image modeling and optical consistency.

[0037] Reference Figure 3 This step uses the unlabeled dataset constructed in step S100 to train the model constructed in step S200. The core of the training process lies in simultaneously optimizing two objectives: the accuracy of geometric reconstruction and the consistency of optical prediction.

[0038] Specifically, for the input mask pattern M, a masking operation is first performed to generate Mmasked. To make the pre-training task more challenging and closer to the essence of photolithography physics, this embodiment preferably adopts a physically guided masking strategy. Specifically, using the existing differentiable optical simulation layer 204 in the model, the gradient of the photolithographic spatial image intensity Iproxy relative to each pixel of the input mask M is calculated, and the absolute value of the gradient is used as the sensitivity map, denoted as S = | Iproxy / M|. A higher value in the sensitivity map S indicates a greater impact of changes in mask pixels at the corresponding location on the final imaging result. For example... Figure 6 As shown, the generation process of the sensitivity map in the physical guidance masking strategy is demonstrated: First, the original mask layout M is obtained, and the sensitivity map S is calculated using a differentiable optical simulation layer, where the bright areas represent key features that have a significant impact on the imaging results (such as corners and line ends); then, a mask probability distribution is generated based on the sensitivity map, and the original layout is randomly sampled to generate the masked image Mmasked.

[0039] Subsequently, the masking probability is set based on the sensitivity map S, so that highly sensitive areas (such as the corners and ends of the main graphic, or the location of sub-resolution auxiliary graphics) have a higher probability of being occluded. This strategy forces the model to understand and utilize the optical proximity effect to infer information about the occluded areas, rather than simply performing pixel interpolation.

[0040] The masked image Input the layout feature encoder E to obtain latent features Then, the two tasks are executed in parallel: First, the geometric reconstruction task. Send to geometric reconstruction branch The reconstructed mask pattern is obtained. The reconstruction results were calculated. The geometric reconstruction loss is defined by the difference between the original mask pattern M and the mask pattern M in the occluded region. The mean squared error or binary cross-entropy loss function can be used.

[0041] Secondly, the task of optical consistency. Feed into the lithographic spatial image prediction branch The predicted lithographic spatial image is obtained. To supervise this prediction, this invention designs a pseudo-truth value generation mechanism: the original, unmasked, complete mask pattern M is directly input into the differentiable optical simulation layer for a forward calculation to obtain the proxy lithographic spatial image. .this It is used as a monitoring signal, i.e., the "pseudo-true value". Subsequently, the predicted lithographic spatial image is calculated. With pseudo-truth value The difference between them is defined as optical physical loss. For example, L1 or L2 loss can be used. This design makes the entire self-supervised learning process a closed loop, without the need for any external simulation tools or real wafer data.

[0042] Finally, the two loss functions are weighted and summed to obtain the total training loss function: .in, and The weight hyperparameter is used to balance the importance of the two tasks. It is calculated based on the total loss using the standard backpropagation algorithm and optimizers such as AdamW. To iteratively update the map feature encoder E and the geometric reconstruction branch And lithographic spatial image prediction branch The trainable parameters in.

[0043] Step S400: Optionally, perform contrastive learning enhancement based on optical equivalence.

[0044] To further enhance the model's robustness to minor changes in the photolithography process, this embodiment also introduces a contrastive learning mechanism. The core of this step is constructing optically equivalent positive sample pairs. Specifically, for a given mask pattern M, positive samples M' are generated by applying small perturbations that do not significantly alter the final imaging result. An effective perturbation method is to randomly fine-tune the position or size of the sub-resolution auxiliary pattern within the limits allowed by the design rules, for example, adjusting it within a range of ±1nm to 5nm. Since the sub-resolution auxiliary pattern itself is not directly imaged, its small changes are largely optically equivalent.

[0045] The original mask M and the perturbated mask M' are considered as a pair of positive samples. They are then passed through a layout feature encoder E to obtain the corresponding feature vectors z and z'. The encoder is then trained using a contrastive loss function (such as InfoNCE loss). The goal of this loss function is to minimize the distance between the positive sample pair (z, z') and the feature representations of any other sample (negative sample) in the feature space, while maximizing the distance in the feature space. In this way, the model learns a feature representation that is insensitive to optically irrelevant geometric details.

[0046] Step S500: Perform model fine-tuning and hotspot detection.

[0047] Reference Figure 4 After completing one or more of the above steps of self-supervised pre-training, the layout feature encoder E already possesses powerful feature extraction capabilities that incorporate lithographic physics knowledge. At this point, the model is fine-tuned using a dataset Dlabeled = {(Mj, yj)} (j=1...P) with hotspot labels, which is much smaller than the unlabeled dataset, to adapt to the downstream hotspot detection task.

[0048] Specifically, the parameters of the pre-trained map feature encoder E are retained, and its subsequent geometric reconstruction branch and lithographic spatial image prediction branch are removed. At the output of encoder E, a newly created, lightweight classification module 401 is connected. This classification module can consist of a global average pooling layer and one or more fully connected layers, and its final output is a probability value between 0 and 1, representing the likelihood that the input mask map data fragment is a hotspot. Then, a labeled dataset is used... The model is trained in a supervised manner by minimizing the cross-entropy loss function commonly used in classification tasks. During the fine-tuning phase, one can choose to update only the parameters of the classification module 401, or simultaneously fine-tune all or some of the parameters of the encoder E. Because pre-training provides excellent initial parameter values, the fine-tuning process typically converges with only a small number of labeled samples and a few training epochs, achieving high performance.

[0049] Step S600: Apply the trained model to perform full-chip hotspot detection.

[0050] The finally trained hotspot detection model is deployed in the production process. The full-chip mask layout to be detected is first divided into a series of standard-sized data segments using the same slicing and preprocessing methods as in step S100. To avoid boundary effects, a sliding window strategy with overlapping regions can be used during slicing. Then, these data segments are input one by one into the fine-tuned hotspot detection model, which calculates a hotspot probability for each segment. Finally, the prediction results of all segments are integrated and mapped back to the original full-chip coordinate system to generate a full-chip hotspot distribution map. In this map, high-probability areas are prominently marked, providing precise guidance for subsequent layout corrections or process adjustments.

[0051] To verify the effectiveness of the method described in this invention, this embodiment compares and analyzes the performance of the self-supervised learning method based on physical perception (hereinafter referred to as the method of this invention) with various existing technical solutions under predetermined experimental conditions.

[0052] The hardware environment used in the experiment included a computing server equipped with multiple high-performance graphics processors, and the experimental dataset used metal layer mask layouts of logic chips at a certain 5-nanometer process node. Specifically, the unlabeled mask layout dataset used in the pre-training phase contained 1 million mask fragments of 1024x1024 pixels, while the labeled dataset used in the fine-tuning and testing phases contained 2000 samples labeled with hotspot tags.

[0053] like Figure 5As shown in the figure, the performance results of four comparative research methods are illustrated: First, the traditional physical simulation method uses a high-precision physical imaging model to simulate and predict the layout, and determines hotspots by calculating the graphic features under a predetermined process window. Experimental results show that its detection accuracy is 92%. Although this method has high physical accuracy, the computational overhead of full-chip scanning is extremely high, making it difficult to meet the needs of rapid iteration. Second, the supervised deep learning method directly trains a deep neural network model from scratch on the aforementioned labeled dataset using a supervised learning paradigm. Experimental results show that its accuracy is 88%. Due to the limited size of the labeled samples, the model cannot fully cover the distribution characteristics of complex layout graphics, resulting in poor generalization ability when dealing with unseen graphics. Furthermore, the general self-supervised learning method uses a mask autoencoder architecture for pre-training on an unlabeled mask layout dataset. However, the pre-training task only includes geometric reconstruction and does not include the lithographic spatial image prediction branch and optical physical loss proposed in this invention. The experimental results show that the accuracy is only 78%, which proves that there is a mismatch between the feature representation based solely on geometric similarity and the lithographic physical process, and it cannot effectively extract the printability features of the mask. Finally, the method of this invention uses the dual-stream decoder architecture with an embedded differentiable optical simulation layer described in this invention for self-supervised pre-training, combined with optical equivalent contrastive learning enhancement, and finally fine-tuned using the same labeled dataset. The experimental results show that the accuracy reaches 95%. In summary, this invention introduces a physical perception mechanism and optical consistency constraints into a self-supervised learning framework, enabling the model to autonomously extract deep features containing optical diffraction patterns from massive amounts of unlabeled data. Compared with traditional supervised methods and general self-supervised learning methods, this invention achieves higher detection accuracy with a small amount of labeled data, effectively solves the representation mismatch problem, and significantly improves the model's accuracy and generalization ability in detecting hotspots on complex curve masks.

[0054] Example 2 This embodiment provides a computational lithography modeling and hotspot detection device based on self-supervised learning. The internal structure of this device can be implemented by computer program code. When the program code is executed by a processor, it implements the method described in Embodiment 1 above. This device can be a standalone server or a functional module integrated into electronic design automation (EDA) software. The device includes: The data construction module is used to perform the function described in step S100, namely, to obtain mask layout data from the chip design database and process it into a large-scale set of unlabeled data fragments suitable for neural network training.

[0055] The model building module is used to perform the function described in step S200, namely, to build a neural network model that includes a layout feature encoder, a geometric reconstruction branch, and a lithographic spatial image prediction branch with an embedded differentiable optical simulation layer.

[0056] The pre-training module performs the functions described in steps S300 and optionally S400. This module is responsible for self-supervised pre-training of the model on an unlabeled dataset through a dual task of mask image modeling and optical consistency, as well as contrastive learning based on optical equivalence, enabling the layout feature encoder to learn features related to lithography physics.

[0057] The fine-tuning module performs the function described in step S500. This module uses a small amount of labeled sample data, connects it to the classification module after the pre-trained map feature encoder, and performs supervised fine-tuning on the model to enable it to have accurate hotspot classification capabilities.

[0058] The detection application module is used to perform the function described in step S600. This module receives the full chip layout to be detected, calls the fine-tuned model to perform segment-by-segment inference, and finally generates a hotspot distribution map of the entire chip.

[0059] The specific implementation methods and internal working principles of the above modules have been described in detail in Example 1, and will not be repeated here.

[0060] Example 3 This embodiment illustrates the implementation process of the method of the present invention in a specific application scenario and demonstrates its necessity compared to existing technologies. This application scenario involves hotspot detection of key metal layers in a 3-nanometer technology node logic chip generated using advanced inverse lithography (ILT) technology.

[0061] In this scenario, the technical challenges are twofold. First, the mask patterns generated by the ILT algorithm exhibit highly complex freeform curves, whose geometric features differ significantly from the more regular patterns found in previous technology nodes (such as 7nm or 5nm). Therefore, when a supervised deep learning model trained directly on older node data is used for hotspot detection, the model's generalization ability is insufficient, resulting in high false negative and false positive rates. Second, since this 3nm process is a newly developed platform, a sufficient number of wafer-verified labeled hotspot samples have not yet been accumulated, making it impossible to build a labeled dataset of sufficient scale and diversity to support training a high-performance supervised model from scratch. While traditional full-chip physical simulation methods are accurate for verification, their computational time cost is extremely high for layouts containing complex curve patterns, failing to meet the requirements of rapid design iteration cycles.

[0062] To address the aforementioned problems, the technical solution provided by this invention can be applied. First, step S100 is executed to collect a large amount of unlabeled mask pattern data generated by the 3nm ILT tool. Although this data lacks hotspot labels, it accurately reflects the distribution characteristics of the mask pattern under the target process.

[0063] Next, steps S200 and S300 are executed. A physically-aware, self-supervised pre-trained model is constructed. The differentiable optical simulation layer is configured according to the core parameters of the 3nm lithography platform, for example, setting the exposure wavelength to 13.5nm (corresponding to extreme ultraviolet lithography), and pre-calculating a set of corresponding optical convolution kernels based on the parameters of the actual light source and projection optical system. Subsequently, self-supervised pre-training based on mask image modeling and optical consistency is performed on the collected large-scale unlabeled curve context layout data. During this process, the model is forced to simultaneously recover the complete geometry of locally occluded complex curve patterns and predict their lithographic spatial image under the target lithography system. This task forces the model to learn the intrinsic physical relationship between the geometric construction of the curve pattern and optical diffraction effects, rather than simply memorizing specific geometric patterns.

[0064] After pre-training, step S500 is executed for model fine-tuning. At this point, only a small labeled dataset (e.g., containing only a few hundred positive samples and a few thousand negative samples) needs to be obtained through precise physical simulation or inspection of small-batch pilot-production wafers. This small dataset is then used to fine-tune the pre-trained encoder, which already possesses a deep physical understanding. Because the model already has a strong feature extraction foundation, the fine-tuning process converges quickly and achieves high-precision hotspot classification capabilities.

[0065] Finally, step S600 is executed to apply the fine-tuned model to hotspot detection across the entire chip layout. This model can quickly process complex curves across the entire chip and accurately identify high-risk areas that were difficult for previous models to determine. This embodiment demonstrates that the method of the present invention, through physical perception-based self-supervised learning, effectively overcomes the challenges of insufficient labeled data and inadequate model generalization ability faced in the early stages of new process node development, providing an efficient and reliable technical path for hotspot detection in advanced computational lithography.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A computational lithography modeling and hotspot detection method based on self-supervised learning, characterized in that, Includes the following steps: Step 1: Construct an unlabeled mask layout dataset, which includes multiple mask layout data fragments extracted and segmented from the chip design layout database; Step 2: Construct a self-supervised pre-trained neural network model based on physical perception. This model includes a layout feature encoder, a geometric reconstruction branch, and a lithographic spatial image prediction branch with an embedded differentiable optical simulation layer. Step 3, perform self-supervised pre-training: perform masking operation on the mask layout data segment and input it into the layout feature encoder to obtain latent features; use the geometric reconstruction branch to calculate the geometric reconstruction loss based on the latent features; use the lithographic spatial image prediction branch to predict the lithographic spatial image based on the latent features, and calculate the optical physical loss between it and the pseudo-true lithographic spatial image generated by the differentiable optical simulation layer based on the original mask layout; Step 4: Update the network parameters of the self-supervised pre-trained neural network model based on the weighted sum of the geometric reconstruction loss and the optical physical loss; Step 5: Fine-tune the self-supervised pre-trained map feature encoder using sample data with hotspot labels to obtain the hotspot detection model; Step 6: Apply the hotspot detection model to perform hotspot detection on the entire chip mask layout.

2. The method according to claim 1, characterized in that, The layout feature encoder employs a backbone network based on an attention mechanism-based Transformer architecture.

3. The method according to claim 1, characterized in that, The step of performing a masking operation on the mask layout data segment in step three specifically includes: Using the differentiable optical simulation layer, the gradient of the intensity of the lithographic spatial image relative to each pixel of the mask pattern data segment is calculated to generate a sensitivity map. The sensitivity map quantifies the degree of influence of the change of each pixel on the mask on the lithographic imaging result. The probability distribution of the masking operation is set according to the sensitivity map, so that the higher the value of the area on the sensitivity map, the higher the probability of the area being masked.

4. The method according to claim 1, characterized in that, The differentiable optical simulation layer is constructed based on a coherent system summation model and contains a set of optical convolution kernels. These optical convolution kernels are generated by decomposing the coherent system summation model according to the numerical aperture of the lithography machine, the exposure wavelength, and the shape of the light source.

5. The method according to claim 1, characterized in that, After step four and before step five, a step of performing contrastive learning enhancement based on optical equivalence is also included, which includes: An optical equivalent perturbation is applied to the mask layout data segment to generate a perturbed mask layout data segment. The optical equivalent perturbation refers to adjusting the position or size of the sub-resolution auxiliary graphics in the mask layout data segment without violating design rule constraints. The original mask layout data segment and the perturbed mask layout data segment are regarded as a pair of positive samples, and the pair of positive samples are respectively input into the layout feature encoder to obtain a pair of feature vectors; The contrast loss is calculated by reducing the distance between the pair of feature vectors of the positive samples in the feature space, while increasing the distance between the feature representations of different mask layout data fragments in the feature space, thereby updating the network parameters of the layout feature encoder.

6. The method according to claim 5, characterized in that, The adjustment of the position or size of the sub-resolution auxiliary graphic specifically involves random fine-tuning within the range of ±1nm-5nm.

7. The method according to claim 1, characterized in that, The step of fine-tuning using sample data with hotspot tags in step five specifically includes: Remove the geometric reconstruction branch and the lithographic spatial image prediction branch, and connect the classification module to the output of the layout feature encoder; The network parameters of the pre-trained map feature encoder are retained as initial values; The classification module consists of at least one fully connected layer, used to output the probability that the input mask layout data segment is a hotspot; Using a dataset with hotspot labels, supervised training is performed on the network parameters of the map feature encoder and the classification module by calculating the classification cross-entropy loss.

8. The method according to claim 1, characterized in that, The step of constructing the unlabeled mask layout dataset in step one further includes performing data augmentation on the segmented mask layout data fragments, the data augmentation including random rotation and flipping operations.

9. The method according to claim 1, characterized in that, The step of applying the hotspot detection model to perform full-chip hotspot detection in step six specifically includes: The full-chip mask layout to be inspected is divided into multiple data segments of the same size using a sliding window strategy with overlapping areas; The data segments are input one by one into the hotspot detection model, and a hotspot probability value is output for each data segment. The prediction results of all data segments are integrated to generate a hotspot distribution map covering the entire chip.

10. A computational lithography modeling and hotspot detection device based on self-supervised learning, characterized in that, include: A data construction module is used to construct an unlabeled mask layout dataset, which includes multiple mask layout data fragments extracted and segmented from a chip design layout database; The model building module is used to build a self-supervised pre-trained neural network model based on physical perception. The model includes a layout feature encoder, a geometric reconstruction branch, and a lithographic spatial image prediction branch with an embedded differentiable optical simulation layer. A pre-training module is used to perform self-supervised pre-training: masking the mask layout data fragments and inputting them into the layout feature encoder to obtain latent features; calculating the geometric reconstruction loss based on the latent features using the geometric reconstruction branch; predicting the lithographic spatial image based on the latent features using the lithographic spatial image prediction branch, and calculating the optical-physical loss between the predicted lithographic spatial image and the pseudo-true lithographic spatial image generated by the differentiable optical simulation layer based on the original mask layout; and updating the parameters of the neural network model according to the weighted sum of the geometric reconstruction loss and the optical-physical loss. The fine-tuning module is used to fine-tune the self-supervised pre-trained map feature encoder using sample data with hotspot labels to obtain a hotspot detection model. The detection application module is used to perform hotspot detection on the entire chip mask layout using the hotspot detection model.

Images