Deep learning lithography hotspot detection method and system combined with physical model

Abstract

The application provides a photolithography hotspot detection method and system combining a physical model, relates to the technical field of photolithography processes, and comprises the following steps: acquiring design layout and mask layout data; calculating light intensity gradient and surface stress distribution by using a double physical field model to form a physical sensitive feature map; inputting the feature map into a physical constraint neural network for reconstruction and mapping; constructing a state transition matrix based on a compensation state sequence, calculating a transition cost to determine an optimal compensation path, and generating layout compensation data. The application realizes high-precision hotspot detection and effective layout compensation.

Description

Technical Field

[0001] This invention relates to photolithography technology, and more particularly to a deep learning-based photolithography hotspot detection method and system that incorporates a physical model. Background Technology

[0002] With the continuous development of semiconductor manufacturing processes, the feature size of integrated circuit design has reached the nanometer level, and traditional photolithography processes are facing physical limits. At these limits, the distribution of surface stress generated during optical proximity effect and chemical mechanical polishing becomes extremely complex. These physical phenomena make photolithographic patterns prone to hot spots.

[0003] Traditional hotspot detection methods primarily rely on optical model simulations or purely data-driven machine learning approaches. Optical model-based hotspot detection methods require complete lithography imaging simulations; while the physical mechanisms are well-defined, their computational complexity is high, resulting in low efficiency for large-scale integrated circuit layout processing. Purely data-driven deep learning methods, while offering higher processing efficiency, suffer from poor interpretability due to the lack of physical constraints and limited generalization ability when facing new processes or layout design patterns.

[0004] Traditional hotspot detection methods typically consider optical proximity effects or chemical mechanical polishing effects separately, lacking a comprehensive analysis of multi-physics coupling effects, leading to insufficient detection accuracy under complex process conditions. While improving detection efficiency, these methods often rely on black-box models, making it difficult to provide effective physical explanations and specific suggestions for layout optimization. Existing hotspot compensation techniques mostly employ single-physics correction methods, lacking consideration of the interaction between light intensity gradient and surface stress, resulting in unstable compensation effects, especially when process conditions change, making it difficult to maintain the compensation effect. Summary of the Invention

[0005] This invention provides a deep learning-based lithography hotspot detection method and system that incorporates a physical model, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a deep learning-based lithographic hotspot detection method incorporating a physical model, comprising:

[0007] The design layout data and mask layout data are obtained. The light intensity gradient and surface stress distribution are calculated using a dual-physics field model of optical proximity effect and chemical mechanical polishing. The physical sensitive feature map is obtained through tensor coupling calculation.

[0008] The physically sensitive feature map is input into the physical constraint neural network, which includes a feature reconstruction layer and a feature mapping layer. The feature reconstruction layer reconstructs the physically sensitive feature map based on the physical correlation between light intensity gradient and surface stress. The feature mapping layer maps the reconstructed feature map to the hotspot detection space and outputs hotspot detection data containing hotspot location information and physical feature information.

[0009] Hotspot detection data are mapped to light intensity compensation space and stress compensation space respectively to obtain compensation state sequence. The physical differences of continuous compensation states in the compensation state sequence are calculated to construct state transition matrix.

[0010] The transition cost of the compensated state is calculated based on the state transition matrix. The optimal compensation path is determined based on the transition cost. The compensation parameters are extracted from the optimal compensation path and applied to the mask layout data to generate layout compensation data.

[0011] The design layout data and mask layout data are acquired. A dual-physics model combining optical proximity effect and chemical mechanical polishing is used to calculate the light intensity gradient and surface stress distribution. Physically sensitive feature maps are obtained through tensor coupling calculations, including:

[0012] Obtain design layout data and mask layout data, perform edge detection on the design layout data, and obtain edge contour data;

[0013] A dual-physics model of optical proximity effect and chemical mechanical polishing is constructed based on edge contour data. The initial light intensity gradient distribution and the initial surface stress distribution are calculated using the dual-physics model.

[0014] The initial light intensity gradient distribution is used as a constraint to correct the initial surface stress distribution. The corrected surface stress distribution is then fed back to update the initial light intensity gradient distribution. The final light intensity gradient distribution and the final surface stress distribution are obtained through alternating iterations.

[0015] The final light intensity gradient distribution and the final surface stress distribution are converted into light intensity gradient tensor and surface stress tensor, respectively. The physical correlation coefficient is obtained by tensor coupling calculation. The light intensity gradient tensor and surface stress tensor are weighted and combined according to the physical correlation coefficient to generate a physically sensitive feature map.

[0016] The feature reconstruction layer reconstructs the physically sensitive feature map based on the physical correlation between light intensity gradient and surface stress. The feature mapping layer maps the reconstructed feature map to the hotspot detection space, outputting hotspot detection data containing hotspot location information and physical feature information, including:

[0017] The distribution information of light intensity gradient and surface stress is extracted from the physical sensitivity feature map, and the physical correlation between the light intensity gradient and surface stress is analyzed to obtain the physical field interaction intensity.

[0018] The feature constraint function is constructed based on the physical field interaction intensity. The feature reconstruction layer uses the feature constraint function to reconstruct the physical sensitive feature map and generate reconstructed features.

[0019] The reconstructed features are decomposed into multi-scale features, and the feature weights of the multi-scale features are calculated based on the physical field interaction intensity. The multi-scale features are then weighted and combined based on the feature weights to generate the reconstructed feature map.

[0020] The feature mapping layer constructs a feature mapping function based on the physical field interaction intensity, inputs the reconstructed feature map into the feature mapping function, and the feature mapping function performs feature enhancement on the reconstructed feature map based on the physical field interaction intensity to generate an enhanced feature map.

[0021] Hotspot location information and physical feature information are extracted from the enhanced feature map, and the hotspot location information and physical feature information are combined to generate hotspot detection data.

[0022] The construction and training of the physical constraint neural network includes:

[0023] Obtain the physical sensitivity feature map and the corresponding labeled hotspot location information, and construct a training sample set;

[0024] A physically constrained neural network structure is constructed, including a feature reconstruction layer and a feature mapping layer. The feature reconstruction layer adopts a parallel convolutional structure, and the feature mapping layer adopts an attention enhancement unit.

[0025] Light intensity gradient and surface stress information are extracted from the physical sensitivity feature map. Based on the light intensity gradient and surface stress information, a feature constraint function is constructed. The feature constraint function is combined with the labeled hotspot location information to construct a physical constraint loss function.

[0026] The training sample set is input into the physical constraint neural network, the physical constraint loss function value is calculated, and the network parameters are optimized based on the physical constraint loss function value to obtain the optimized network parameters.

[0027] The verification data is input into the physical constraint neural network configured with optimized network parameters. When the error between the output hotspot detection result and the labeled hotspot location information in the verification data is less than a preset error threshold, the optimized network parameters are configured in the physical constraint neural network for the generation of hotspot detection data.

[0028] Hotspot detection data are mapped to light intensity compensation space and stress compensation space respectively to obtain a compensation state sequence. The physical differences between consecutive compensation states in the compensation state sequence are calculated to construct a state transition matrix, including:

[0029] Extract light intensity and stress features from hotspot detection data and calculate feature correlation.

[0030] A physical constraint function is constructed based on the feature correlation. The physical constraint function is then used to decompose the light intensity feature and stress feature into constraints, resulting in light intensity constraint features and stress constraint features.

[0031] A light intensity compensation space is constructed based on the light intensity constraint features, and a stress compensation space is constructed based on the stress constraint features. The hotspot detection data is mapped to the light intensity compensation space and the stress compensation space respectively to obtain a compensation state sequence.

[0032] Calculate the light intensity difference and stress difference between two temporally adjacent compensation states in the compensation state sequence, and construct a state transition matrix based on the light intensity difference and stress difference.

[0033] The transition cost of the compensated state is calculated based on the state transition matrix, and the optimal compensation path is determined based on the transition cost, including:

[0034] Extract the energy and gradient features of adjacent compensated states in the state transition matrix, and construct physical constraint functions based on the correlation between the energy and gradient features;

[0035] The energy constraint value corresponding to the energy feature and the gradient constraint value corresponding to the gradient feature are calculated based on the physical constraint function. The energy constraint value and the gradient constraint value are then weighted and combined to obtain the transition cost between adjacent compensation states.

[0036] A path optimization objective function is constructed based on the physical constraint function and the transfer cost. A dynamic programming equation is established based on the path optimization objective function, and the optimal compensation path is obtained by iteratively solving the dynamic programming equation.

[0037] Extracting compensation parameters from the optimal compensation path and applying these parameters to the mask layout data to generate layout compensation data includes:

[0038] Extract the compensation state sequence from the optimal compensation path, and extract the initial compensation parameters from the compensation state sequence;

[0039] The compensation state sequence is divided into multiple local compensation regions in the spatial dimension, and the spatial compensation features are obtained by performing feature correlation analysis on adjacent local compensation regions. In the temporal dimension, it is divided into multiple compensation stages, and the temporal compensation features are obtained by performing evolution analysis on adjacent compensation stages.

[0040] Based on the spatial compensation characteristics and the temporal compensation characteristics, spatial constraints and temporal constraints of the compensation parameters are established, and the spatial constraints and temporal constraints are combined to construct the compensation parameter constraint function.

[0041] The initial compensation parameters are optimized using the compensation parameter constraint function to obtain the optimized compensation parameters, and the local structural features of the mask layout data are analyzed to determine the compensation area.

[0042] The compensation area is determined based on the local structural features of the mask layout data, and the optimized compensation parameters are applied to the compensation area to generate layout compensation data.

[0043] A second aspect of the present invention provides a deep learning-based lithography hotspot detection system incorporating a physical model, comprising:

[0044] The first unit is used to acquire design layout data and mask layout data, calculate the light intensity gradient and surface stress distribution using the dual-physics field model of optical proximity effect and chemical mechanical polishing, and obtain the physical sensitive feature map through tensor coupling calculation.

[0045] The second unit is used to input the physically sensitive feature map into the physical constraint neural network. The physical constraint neural network includes a feature reconstruction layer and a feature mapping layer. The feature reconstruction layer reconstructs the physically sensitive feature map based on the physical correlation between light intensity gradient and surface stress. The feature mapping layer maps the reconstructed feature map to the hotspot detection space and outputs hotspot detection data containing hotspot location information and physical feature information.

[0046] The third unit is used to map the hotspot detection data to the light intensity compensation space and the stress compensation space respectively to obtain the compensation state sequence, and to calculate the physical differences of the continuous compensation states in the compensation state sequence to construct the state transition matrix.

[0047] The fourth unit is used to calculate the transition cost of the compensation state based on the state transition matrix, determine the optimal compensation path based on the transition cost, extract compensation parameters from the optimal compensation path, and apply the compensation parameters to the mask layout data to generate layout compensation data.

[0048] A third aspect of the present invention provides an electronic device, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] In this embodiment, the present invention combines a deep learning-based lithography hotspot detection method with a physical model. Through the deep fusion of physical sensitive feature extraction and neural network mapping, high-precision detection and localization of lithography hotspots are achieved, improving the accuracy and interpretability of hotspot detection. The optical proximity effect and the dual physical fields of chemical mechanical polishing are coupled and analyzed to construct a physically constrained neural network architecture, which effectively captures the physical essence of hotspot formation, avoids the blind spots of purely data-driven methods, and reduces the false detection rate and false negative rate of hotspots. Based on the strategy of optimizing the compensation path according to the state transition matrix, the entire process from hotspot detection to layout compensation is automated, improving the efficiency and quality of hotspot compensation in the chip manufacturing process, and effectively improving chip yield and performance consistency. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the deep learning-based lithography hotspot detection method combined with a physical model, as described in an embodiment of the present invention.

[0055] Figure 2 This is a flowchart illustrating the compensation parameter optimization and layout data generation process in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention, as shown below. Figure 1 As shown, the deep learning-based lithography hotspot detection method combining a physical model includes:

[0059] The design layout data and mask layout data are obtained. The light intensity gradient and surface stress distribution are calculated using a dual-physics field model of optical proximity effect and chemical mechanical polishing. The physical sensitive feature map is obtained through tensor coupling calculation.

[0060] The physically sensitive feature map is input into the physical constraint neural network, which includes a feature reconstruction layer and a feature mapping layer. The feature reconstruction layer reconstructs the physically sensitive feature map based on the physical correlation between light intensity gradient and surface stress. The feature mapping layer maps the reconstructed feature map to the hotspot detection space and outputs hotspot detection data containing hotspot location information and physical feature information.

[0061] Hotspot detection data are mapped to light intensity compensation space and stress compensation space respectively to obtain compensation state sequence. The physical differences of continuous compensation states in the compensation state sequence are calculated to construct state transition matrix.

[0062] The transition cost of the compensated state is calculated based on the state transition matrix. The optimal compensation path is determined based on the transition cost. The compensation parameters are extracted from the optimal compensation path and applied to the mask layout data to generate layout compensation data.

[0063] In one optional implementation, design layout data and mask layout data are acquired, and the light intensity gradient and surface stress distribution are calculated using a dual-physics model of optical proximity effect and chemical mechanical polishing. The physically sensitive feature map is then obtained through tensor coupling calculations, including:

[0064] Obtain design layout data and mask layout data, perform edge detection on the design layout data, and obtain edge contour data;

[0065] A dual-physics model of optical proximity effect and chemical mechanical polishing is constructed based on edge contour data. The initial light intensity gradient distribution and the initial surface stress distribution are calculated using the dual-physics model.

[0066] The initial light intensity gradient distribution is used as a constraint to correct the initial surface stress distribution. The corrected surface stress distribution is then fed back to update the initial light intensity gradient distribution. The final light intensity gradient distribution and the final surface stress distribution are obtained through alternating iterations.

[0067] The final light intensity gradient distribution and the final surface stress distribution are converted into light intensity gradient tensor and surface stress tensor, respectively. The physical correlation coefficient is obtained by tensor coupling calculation. The light intensity gradient tensor and surface stress tensor are weighted and combined according to the physical correlation coefficient to generate a physically sensitive feature map.

[0068] In a preferred embodiment, integrated circuit design layout data and corresponding mask layout data are first acquired. The design layout data is typically stored in GDSII format and contains information on multiple layers, such as polysilicon layers and metal layers. The mask layout data is the actual manufacturing data after optical proximity correction. The Canny edge detection algorithm is used to process the design layout data to extract edge contour information. First, a Gaussian filter is used to smooth the layout data, with a kernel size of 5×5 pixels and a σ value of 1.4. The image gradient magnitude and direction are calculated, and non-maximum suppression is performed. Complete edge contour data is obtained through a dual-threshold algorithm and connectivity analysis. For a typical 90nm process node design layout, the high threshold is set to 30% of the maximum image gradient value, and the low threshold is set to 40% of the high threshold, thus effectively extracting the edge contours.

[0069] Based on the acquired edge contour data, a two-physics model for optical proximity effect and chemical mechanical polishing (CMP) was constructed. The optical proximity effect model employs Hopkins optical imaging theory, combining the mask pattern with partially coherent illumination conditions of the light source, and considering the aberrations of the optical system. The light source was set as ring illumination with an internal σ value of 0.7 and an external σ value of 0.9, a numerical aperture (NA) of 0.85, and a working wavelength of 193 nm. In the CMP model, a hybrid approach combining density-based and step-based models was adopted, with the density window size set to 100 μm × 100 μm and the step-based influence radius set to 50 μm. The material removal rate parameter in the model was set to 200 nm / min, the polishing pad elastic modulus to 500 MPa, and the pressure distribution uniformity to 85%, based on actual process conditions.

[0070] The initial light intensity gradient distribution and initial surface stress distribution were calculated using a constructed two-physics model. For a typical polysilicon gate structure, the initial light intensity gradient was calculated using a spatial frequency domain method, considering the Fourier transform of the mask layout and the optical transfer function, with a sampling point spacing of 5 nm. The initial surface stress distribution was calculated using the finite element method, with a mesh size of 10 nm, considering a material hardness parameter of 9 GPa and a Poisson's ratio of 0.17. At the edges of the layout, the light intensity gradient typically exhibits a higher gradient value, while in uniform regions, the gradient value tends to zero; the surface stress shows significant changes in regions with large variations in the pattern density of the layout, with stress values ​​reaching up to 120 MPa at the edges of a typical gate structure.

[0071] To improve computational accuracy, the initial light intensity gradient distribution was used as a constraint to correct the initial surface stress distribution. In regions where the light intensity gradient exceeded a preset threshold (set as 20% of the maximum gradient value), the stress distribution was locally refined, with the mesh size reduced to 5 nm, and material parameters adjusted to reflect the influence of photochemical reactions on material properties during exposure. The corrected surface stress distribution was then used to update the light intensity gradient distribution, considering the effect of stress-induced refractive index changes on light propagation. Ten iterations were performed, with the convergence criterion for each iteration being that the relative rate of change between the stress field and the light intensity gradient field was less than 1%.

[0072] After iteration, the final light intensity gradient distribution and surface stress distribution are obtained and converted into light intensity gradient tensor and surface stress tensor forms, respectively. The light intensity gradient tensor contains components in the x and y directions, while the surface stress tensor contains components such as normal stress and shear stress. The physical correlation coefficient between these two types of tensors is calculated using a normalized cross-correlation method with a sliding window size of 50 nm × 50 nm and a step size of 10 nm. For regions with significant light intensity gradients and drastic stress changes, the correlation coefficient is typically higher than 0.75; while for regions with gentle light intensity gradients or stress changes, the correlation coefficient is typically lower than 0.3.

[0073] Based on the calculated physical correlation coefficient, the light intensity gradient tensor and the surface stress tensor are weighted and combined. In regions with high correlation coefficients, the light intensity gradient weight is set to 0.6 and the stress tensor weight is set to 0.4; in regions with low correlation coefficients, the light intensity gradient weight is set to 0.8 and the stress tensor weight is set to 0.2. This weighted combination generates a physical sensitivity feature map, which visually reflects areas in the layout with manufacturing defect risks. For a design region containing multiple gate structures, the feature map clearly shows high-risk areas at gate gaps and gate corners, with risk values ​​approximately 2.5 times higher than the average area. These areas are more prone to defects such as short circuits or open circuits during subsequent manufacturing processes.

[0074] By establishing a dual-physics model of optical proximity effect and chemical mechanical polishing, the co-calculation and dynamic optimization of light intensity gradient and surface stress distribution were achieved. Tensor coupling was employed for correlation analysis of the physical fields, enabling a more accurate characterization of physical interactions during the manufacturing process. Iterative optimization and weighted combination were used to generate physically sensitive feature maps, improving the accuracy of feature extraction and better reflecting process defects and potential risks. This approach overcomes the limitations of traditional single-physics models in accurately describing complex processes, providing a more reliable theoretical basis and technical support for defect prediction and process optimization in chip manufacturing.

[0075] In one optional implementation, the feature reconstruction layer reconstructs the physically sensitive feature map based on the physical correlation between light intensity gradient and surface stress, and the feature mapping layer maps the reconstructed feature map to the hotspot detection space, outputting hotspot detection data containing hotspot location information and physical feature information, including:

[0076] The distribution information of light intensity gradient and surface stress is extracted from the physical sensitivity feature map, and the physical correlation between the light intensity gradient and surface stress is analyzed to obtain the physical field interaction intensity.

[0077] The feature constraint function is constructed based on the physical field interaction intensity. The feature reconstruction layer uses the feature constraint function to reconstruct the physical sensitive feature map and generate reconstructed features.

[0078] The reconstructed features are decomposed into multi-scale features, and the feature weights of the multi-scale features are calculated based on the physical field interaction intensity. The multi-scale features are then weighted and combined based on the feature weights to generate the reconstructed feature map.

[0079] The feature mapping layer constructs a feature mapping function based on the physical field interaction intensity, inputs the reconstructed feature map into the feature mapping function, and the feature mapping function performs feature enhancement on the reconstructed feature map based on the physical field interaction intensity to generate an enhanced feature map.

[0080] Hotspot location information and physical feature information are extracted from the enhanced feature map, and the hotspot location information and physical feature information are combined to generate hotspot detection data.

[0081] During feature reconstruction, the distribution information of light intensity gradient and surface stress is extracted from the physically sensitive feature map. A sliding window technique is used to scan the physically sensitive feature map, with the window size set to 7×7 pixels. For each window region, the light intensity gradient vector G is calculated. i and surface stress tensor S i The light intensity gradient is obtained by the difference in light intensity values ​​between adjacent pixels, while surface stress is estimated based on the light intensity distribution pattern. For example, when a light intensity value of 100 cd / m² is detected... 2 Mutation to 250 cd / m 2 The region corresponds to a light intensity gradient of 150 cd / m². 2 / pixel, at this point the estimated surface stress is approximately 75MPa.

[0082] The analysis of the physical field interaction intensity is achieved by constructing a correlation mapping between light intensity gradient and surface stress. A 512×512 correlation matrix is ​​established, where each element represents the correspondence between a specific light intensity gradient value and a surface stress value. For example, in wafer inspection, when the light intensity gradient reaches 180 cd / m², the correlation is established. 2When the surface stress is 90 MPa and the physical field interaction intensity reaches a peak of 0.85, the physical field interaction intensity reaches a peak of 0.85.

[0083] Based on the obtained physical field interaction intensity, a feature constraint function is constructed. This function adopts a piecewise structure design: when the physical field interaction intensity is in the range of 0 to 0.3, the function value increases linearly; when the interaction intensity is in the range of 0.3 to 0.7, the function value increases quadratically; and when the interaction intensity exceeds 0.7, the function value increases exponentially. This design can effectively amplify the feature representation of high interaction intensity regions. The feature reconstruction layer uses this constraint function to transform the physical sensitive feature map and generate reconstructed features. For a typical 28nm process wafer image, the constraint function maps the original feature value range from [0, 1] to [0, 2.5], significantly enhancing the features of regions with high physical field interaction intensity.

[0084] The multi-scale decomposition of reconstructed features is achieved through wavelet transform. The reconstructed features are decomposed into four scale levels, corresponding to microscopic defects (5nm-20nm), small structural defects (20nm-50nm), medium structural defects (50nm-100nm), and large structural defects (above 100nm) in the wafer image, respectively. For each scale level, a corresponding feature weight is calculated based on the physical field interaction strength. When an abnormal surface stress region is detected and the physical field interaction strength is 0.82, the weight for the microscopic defect scale is 0.4, the weight for the small structural defect scale is 0.3, the weight for the medium structural defect scale is 0.2, and the weight for the large structural defect scale is 0.1. Based on these weights, the multi-scale features are weighted and combined to generate the final reconstructed feature map.

[0085] The feature mapping layer constructs a feature mapping function based on the physical field interaction intensity. This function is designed as an adaptive enhancement structure, containing three key parameters: a base enhancement coefficient α, an interaction intensity sensitivity factor β, and a nonlinear adjustment factor γ. α is set to 1.2, β to 0.8, and γ to 1.5. The enhancement effect of the mapping function varies with different physical field interaction intensities. For example, when the interaction intensity is 0.3, the feature enhancement is approximately 15%; when the interaction intensity is 0.6, the enhancement is approximately 35%; and when the interaction intensity reaches 0.9, the enhancement can reach 80%. After inputting the reconstructed feature map into this mapping function, an enhanced feature map is generated. The contrast of hotspot regions in this enhanced feature map is significantly improved, making subsequent detection more accurate.

[0086] Hotspot location information and physical feature information are extracted from the enhanced feature map using an adaptive thresholding method. A base threshold is set at 1.5 times the average value of the feature map, while dynamically adjusted based on the physical field interaction intensity. For regions with an interaction intensity greater than 0.7, the threshold is reduced by approximately 20% to improve the detection rate; for regions with an interaction intensity less than 0.3, the threshold is increased by approximately 10% to reduce false detections.

[0087] For each detected hotspot, the system extracts information such as its center coordinates, area size, shape features, physical field interaction intensity value, and estimated surface stress value. This information is combined into a structured hotspot detection data package, which includes the hotspot ID, location coordinates (x, y), area size (unit: pixels), shape feature description (circularity, aspect ratio, etc.), interaction intensity value, and corresponding estimated surface stress value (MPa). This data package can be directly used for subsequent defect classification, severity assessment, and yield analysis.

[0088] By analyzing the physical correlation between light intensity gradient and surface stress, feature constraint functions and feature mapping functions were constructed, enabling multi-scale decomposition and weighted reconstruction of features. Feature enhancement based on physical field interaction intensity allows the reconstructed feature map to better capture key physical information from the process. Organically combining hotspot location information with physical feature information improves the accuracy and interpretability of hotspot detection. This feature reconstruction and mapping method based on physical correlation maintains the integrity of the original physical features while enhancing their expressive power, providing more reliable data support for hotspot detection and effectively improving the credibility of the detection results.

[0089] In one alternative implementation, the construction and training of the physically constrained neural network includes:

[0090] Obtain the physical sensitivity feature map and the corresponding labeled hotspot location information, and construct a training sample set;

[0091] A physically constrained neural network structure is constructed, including a feature reconstruction layer and a feature mapping layer. The feature reconstruction layer adopts a parallel convolutional structure, and the feature mapping layer adopts an attention enhancement unit.

[0092] Light intensity gradient and surface stress information are extracted from the physical sensitivity feature map. Based on the light intensity gradient and surface stress information, a feature constraint function is constructed. The feature constraint function is combined with the labeled hotspot location information to construct a physical constraint loss function.

[0093] The training sample set is input into the physical constraint neural network, the physical constraint loss function value is calculated, and the network parameters are optimized based on the physical constraint loss function value to obtain the optimized network parameters.

[0094] The verification data is input into the physical constraint neural network configured with optimized network parameters. When the error between the output hotspot detection result and the labeled hotspot location information in the verification data is less than a preset error threshold, the optimized network parameters are configured in the physical constraint neural network for the generation of hotspot detection data.

[0095] In constructing a physically constrained neural network, it is necessary to obtain physically sensitive feature maps and corresponding labeled hotspot location information to build a training sample set. The physically sensitive feature map can be a surface topography image, containing information such as light intensity distribution and surface deformation. Labeled hotspot location information is usually represented by coordinates (x, y), identifying the precise location of the actual hotspot. The training sample set consists of multiple pairs of such feature maps and labeled locations; for example, it can contain 1000 pairs of sample data, with 800 pairs used for training and 200 pairs used for validation.

[0096] The physically constrained neural network architecture includes a feature reconstruction layer and a feature mapping layer. The feature reconstruction layer employs a parallel convolutional structure with three parallel convolutional branches: the first branch uses a 3×3 kernel, the second branch uses a 5×5 kernel, and the third branch uses a 7×7 kernel. Each branch contains 32 kernels with a stride of 1 and SAME padding to maintain the feature map size. The feature maps extracted from each branch are concatenated along the channel dimension and then fused using a 1×1 convolutional layer to output a 64-channel feature map. The function of the feature reconstruction layer is to extract multi-scale features from different receptive fields, enhancing the network's ability to perceive multi-scale hotspot features.

[0097] The feature mapping layer employs an attention enhancement unit, which performs channel attention and spatial attention calculations on the input feature map. During channel attention calculation, global average pooling and global max pooling operations are performed on the input feature map to obtain two channel descriptors. These descriptors are then processed by fully connected layers with shared weights and merged, with the channel weights obtained using the sigmoid function. During spatial attention calculation, maximum and average values ​​are applied to the feature map along the channel dimension to obtain two spatial feature maps. These maps are concatenated and then processed using a 7×7 convolution and the sigmoid function to obtain spatial weights. Features are obtained by multiplying the input features first by the channel weights and then by the spatial weights. This approach adaptively adjusts the importance of different channels and spatial locations in the feature map, enhancing the perception of hotspot regions.

[0098] During training, light intensity gradient and surface stress information are extracted from the physically sensitive feature map. The light intensity gradient is calculated using the Sobel operator. Sobel convolution kernels are applied to the input feature map in both the horizontal and vertical directions, and then the gradient magnitude is calculated. For a 100×100 pixel feature map, a gradient magnitude map of the same size can be obtained, with values ​​ranging from 0 to 255. Surface stress information can be extracted from the feature map through a pre-established stress model or obtained from additional stress measurement data. It is typically represented as the principal components of the stress tensor, with values ​​ranging from 0 to 1000 MPa.

[0099] A feature constraint function is constructed based on light intensity gradient and surface stress information. This function is designed as a weighted sum of gradient and stress constraint terms. The gradient constraint term calculates the consistency between the predicted hotspot location and the high-gradient region, while the stress constraint term calculates the consistency between the predicted hotspot location and the high-stress region. The cross-entropy can be calculated separately for the predicted hotspot probability map, the normalized gradient map, and the stress map, and then summed using weighted methods. The gradient constraint weight can be set to 0.3, and the stress constraint weight can be set to 0.7.

[0100] The physical constraint loss function is composed of a feature constraint function and a standard classification loss function. The standard classification loss function uses cross-entropy loss to calculate the difference between the predicted hotspot probability map and the actual labeled hotspot locations. The final loss function is a weighted sum of the standard classification loss and the feature constraint function, with weight coefficients set to 0.6 and 0.4, respectively.

[0101] During training, the training sample set is input into the physical constraint neural network, the network output is calculated via forward propagation, and then the physical constraint loss function is calculated. The Adam optimizer is used for backpropagation, with an initial learning rate of 0.001. The learning rate decays to 0.1 times its original value every 50 epochs, and training lasts for a total of 200 epochs. Each mini-batch contains 32 samples. During training, the training loss and validation loss are recorded. Training is terminated early when the validation loss does not decrease for 10 consecutive epochs.

[0102] After training, validation data is used to evaluate model performance. The validation data is input into a physically constrained neural network configured with optimized network parameters, and the error between the output hotspot detection results and the labeled hotspot location information in the validation data is calculated. The error is calculated using the mean Euclidean distance, which is the average distance between the predicted hotspot location and the actual hotspot location. When this error is less than a preset error threshold (e.g., 2.5 pixels), the model performance is considered satisfactory, and the optimized network parameters are configured in the physically constrained neural network for generating hotspot detection data.

[0103] A physically constrained neural network, comprising feature reconstruction and feature mapping layers, was constructed, and parallel convolutional structures and attention enhancement units were introduced to improve the network's feature extraction and representation capabilities. Feature constraint functions and physically constrained loss functions, built based on light intensity gradient and surface stress information, better integrate physical laws into the network training process, improving the model's generalization performance. Validation data was used for model evaluation and parameter optimization, ensuring the accuracy and reliability of hotspot detection results. This deep learning method, incorporating physical constraints, guarantees the physical rationality of the detection results while fully leveraging the advantages of deep learning.

[0104] In one optional implementation, hotspot detection data is mapped to both a light intensity compensation space and a stress compensation space to obtain a compensation state sequence. The physical differences between consecutive compensation states in the compensation state sequence are then calculated to construct a state transition matrix, including:

[0105] Extract light intensity and stress features from hotspot detection data and calculate feature correlation.

[0106] A physical constraint function is constructed based on the feature correlation. The physical constraint function is then used to decompose the light intensity feature and stress feature into constraints, resulting in light intensity constraint features and stress constraint features.

[0107] A light intensity compensation space is constructed based on the light intensity constraint features, and a stress compensation space is constructed based on the stress constraint features. The hotspot detection data is mapped to the light intensity compensation space and the stress compensation space respectively to obtain a compensation state sequence.

[0108] Calculate the light intensity difference and stress difference between two temporally adjacent compensation states in the compensation state sequence, and construct a state transition matrix based on the light intensity difference and stress difference.

[0109] Light intensity and stress features are extracted from hotspot detection data, and their correlation is calculated to construct a physical constraint function. When processing the hotspot detection data, feature extraction algorithms are used to analyze the lithographic images, extracting light intensity and stress feature vectors. The light intensity feature vector includes parameters such as light intensity distribution, gradient change, and edge sharpness, while the stress feature vector includes parameters such as stress distribution, stress concentration, and stress gradient. After extraction, the Pearson correlation coefficient method is used to calculate the correlation between the light intensity and stress features. The correlation value ranges from -1 to 1; a correlation closer to 1 indicates a stronger positive correlation, a correlation closer to -1 indicates a stronger negative correlation, and a correlation close to 0 indicates virtually no correlation between the two features.

[0110] Based on the calculated feature correlations, a physical constraint function is constructed. This constraint function uses a weighted summation form, assigning higher weights to feature combinations with high correlation and lower weights to feature combinations with low correlation. Assuming a 10-dimensional light intensity feature vector and an 8-dimensional stress feature vector are extracted, the calculated correlation matrix is ​​10×8. By setting a threshold of 0.65, feature combinations with correlations greater than the threshold are selected to construct the physical constraint function. This constraint function expresses the physical relationship between light intensity and stress and can be used for subsequent feature decomposition and optimization.

[0111] Constraint decomposition is performed on light intensity and stress features using a constructed physical constraint function. The constraint decomposition process employs an iterative optimization algorithm. In each iteration, the features are adjusted according to the physical constraint function to ensure they meet the physical constraints. For light intensity features, dimensions with little correlation to stress features are filtered out using the physical constraint function, retaining key dimensions to form light intensity constrained features. Similarly, a similar operation is performed on stress features to obtain stress constrained features. The original 10-dimensional light intensity features may be reduced to 7-dimensional light intensity constrained features, and the 8-dimensional stress features may be reduced to 5-dimensional stress constrained features. This reduces data dimensionality while preserving important physical correlations.

[0112] A light intensity compensation space is constructed based on the obtained light intensity constraint features, and a stress compensation space is constructed based on the stress constraint features. The construction of the compensation spaces employs a nonlinear mapping method to map the constraint features into a high-dimensional space, allowing for a better representation of the relationships between features. The construction of the light intensity compensation space uses kernel function techniques to map the light intensity constraint features into a space that reflects the light intensity distribution pattern. The construction of the stress compensation space also uses kernel function techniques, but the parameter settings differ from those of the light intensity compensation space to reflect the specific characteristics of the stress distribution. Radial basis functions are used as kernel functions, and the kernel parameters are adaptively adjusted according to the data distribution characteristics.

[0113] Hotspot detection data are mapped to both the light intensity compensation space and the stress compensation space to obtain a compensation state sequence. The mapping process uses the previously constructed nonlinear mapping function to map each sample point in the original hotspot detection data to the two compensation spaces, forming compensation state points. These compensation state points are arranged in chronological order to form a compensation state sequence. Data collected at different moments during the photolithography process, after mapping, forms a trajectory in the compensation space, reflecting the evolution of the hotspot state during the photolithography process.

[0114] The light intensity difference and stress difference values ​​between two temporally adjacent compensated states in the compensated state sequence are calculated. For the state sequence in the light intensity compensation space, the Euclidean distance between two adjacent state points is calculated as the light intensity difference value; similarly, for the state sequence in the stress compensation space, the Euclidean distance is calculated as the stress difference value. For example, processing data from 100 time points yields 99 light intensity difference values ​​and 99 stress difference values, which reflect the intensity of hotspot states changing over time.

[0115] A state transition matrix is ​​constructed based on the differences in light intensity and stress. The state transition matrix describes the probability of a hotspot transitioning from one state to another; the rows and columns of the matrix correspond to different states, and the matrix elements represent the transition probabilities between corresponding states. During construction, the differences in light intensity and stress are normalized, and then a comprehensive difference value is calculated by combining the two values. The state transition matrix is ​​constructed based on this comprehensive difference value. The dimension of the state transition matrix depends on the number of defined states; hotspot states can be categorized into four types: normal, slightly abnormal, moderately abnormal, and severely abnormal, corresponding to a 4×4 state transition matrix.

[0116] By extracting light intensity and stress characteristics and analyzing their correlation, a physical constraint function was constructed, enabling effective processing of hotspot detection data. Combining a physical model with data-driven approaches fully utilizes the physical relationship between light intensity and stress during photolithography, improving the accuracy and reliability of hotspot detection. By constructing a compensation space and state transition matrix, a precise description of the hotspot state evolution process was achieved, providing strong support for the prediction and control of hotspots during photolithography.

[0117] In one optional implementation, calculating the transition cost of the compensation state based on the state transition matrix and determining the optimal compensation path based on the transition cost includes:

[0118] Extract the energy and gradient features of adjacent compensated states in the state transition matrix, and construct physical constraint functions based on the correlation between the energy and gradient features;

[0119] The energy constraint value corresponding to the energy feature and the gradient constraint value corresponding to the gradient feature are calculated based on the physical constraint function. The energy constraint value and the gradient constraint value are then weighted and combined to obtain the transition cost between adjacent compensation states.

[0120] A path optimization objective function is constructed based on the physical constraint function and the transfer cost. A dynamic programming equation is established based on the path optimization objective function, and the optimal compensation path is obtained by iteratively solving the dynamic programming equation.

[0121] First, when extracting the energy and gradient features of adjacent compensated states in the state transition matrix, each pair of adjacent compensated states is processed. The energy feature can be represented as the power consumption value of the compensated state, such as the electrical energy consumption of a mechanical system in a specific state, with units of W. The gradient feature can be represented as the rate of change of the compensated state, such as the change in position, angle, or velocity per unit time. For a specific robotic arm, assuming that when transitioning from compensated state A to compensated state B, the measured energy feature value is 15.7 W, and the gradient feature value is 2.3 degrees / s.

[0122] When constructing physical constraint functions based on the correlation between energy and gradient features, the inherent physical relationship between these two features is considered. A physical constraint function can be represented as a weighted combination of the two features, where the weights reflect the importance of different features in a specific application scenario. For example, in precision machining, gradient features may be more important and therefore given a higher weight; while in energy-constrained applications, energy features may be more critical.

[0123] The physical constraint function can be constructed using a piecewise linear function. When the energy characteristic value is below a certain threshold (e.g., 10W), the energy constraint value can be calculated linearly; when the energy characteristic value is above this threshold, an exponential relationship is used to reflect the penalty mechanism for high-energy-consumption states. Similarly, the constraint value of the gradient characteristic can also be calculated using a piecewise function. For example, when the gradient value is in the range of 1.0-3.0 degrees / s, a linear calculation is used; outside this range, a nonlinear penalty is applied.

[0124] In calculating the energy constraint and gradient constraint values ​​based on the physical constraint function, the measured energy and gradient characteristics are substituted into the physical constraint function for calculation. Taking the transition from state A to state B as an example, the energy characteristic value of 15.7W exceeds the threshold of 10W, therefore the energy constraint value is calculated as 10 + (15.7 - 10). 2 =35.49; the gradient eigenvalue of 2.3 degrees / s is within the standard range, and the gradient constraint value is calculated as 2.3 × 1.5 = 3.45.

[0125] When calculating the transition cost by weighting the energy constraint and gradient constraint values, appropriate weights are set according to the application scenario. Assuming that in the current application, the weight of the energy constraint is 0.4 and the weight of the gradient constraint is 0.6, then the transition cost from state A to state B is 35.49 × 0.4 + 3.45 × 0.6 = 16.27. This transition cost represents the total cost required to transition from state A to state B; a smaller value indicates a better transition.

[0126] When constructing a path optimization objective function based on physical constraint functions and transition costs, the goal is to find a series of intermediate states from the initial compensated state to the target compensated state, minimizing the total cost of the entire transition path. The path optimization objective function can be defined as the sum of the transition costs between all adjacent states on the path, and may also include additional constraints such as path smoothness requirements or time limits.

[0127] When establishing dynamic programming equations based on the path optimization objective function, the entire optimization problem is decomposed into a series of subproblems. Each node in the state space is defined as a compensation state, and the minimum cumulative cost to reach that node is defined. Then, the minimum cost path from the initial state to any other state can be determined through a recursive relationship. For each possible compensation state, the cumulative cost from all possible predecessor states is calculated, and the predecessor state with the minimum cumulative cost is selected as the optimal predecessor.

[0128] In the iterative solution of the dynamic programming equation, starting from the initial compensation state, the minimum cost to reach each possible compensation state and its corresponding predecessor state are calculated step by step. Assuming the initial state is state 1 and the target state is state 100, the cost of all possible paths from state 1 to state 2, state 1 to state 3, and so on, is calculated sequentially, and the optimal path information is recorded. By backtracking these optimal predecessor states, an optimal compensation path from the initial state to the target state can ultimately be obtained.

[0129] By analyzing the correlation between energy and gradient characteristics, a physical constraint function is constructed, enabling accurate calculation of the transition cost between adjacent compensation states. A weighted combination of energy and gradient constraints comprehensively considers the physical constraints during the compensation process. Based on the physical constraint function and the transition cost, a path optimization objective function is constructed, and a dynamic programming method is used to solve for the optimal compensation path, ensuring that the compensation strategy satisfies both physical constraints and optimality. This physically constraint-based path optimization method improves the rationality and efficiency of the compensation strategy, providing a reliable decision-making basis for process compensation.

[0130] like Figure 2 As shown, Figure 2 This is a flowchart illustrating the compensation parameter optimization and layout data generation process in an embodiment of the present invention.

[0131] In one optional implementation, extracting compensation parameters from the optimal compensation path and applying these parameters to the mask layout data to generate layout compensation data includes:

[0132] Extract the compensation state sequence from the optimal compensation path, and extract the initial compensation parameters from the compensation state sequence;

[0133] The compensation state sequence is divided into multiple local compensation regions in the spatial dimension, and the spatial compensation features are obtained by performing feature correlation analysis on adjacent local compensation regions. In the temporal dimension, it is divided into multiple compensation stages, and the temporal compensation features are obtained by performing evolution analysis on adjacent compensation stages.

[0134] Based on the spatial compensation characteristics and the temporal compensation characteristics, spatial constraints and temporal constraints of the compensation parameters are established, and the spatial constraints and temporal constraints are combined to construct the compensation parameter constraint function.

[0135] The initial compensation parameters are optimized using the compensation parameter constraint function to obtain the optimized compensation parameters, and the local structural features of the mask layout data are analyzed to determine the compensation area.

[0136] The compensation area is determined based on the local structural features of the mask layout data, and the optimized compensation parameters are applied to the compensation area to generate layout compensation data.

[0137] First, the optimal compensation path is analyzed to extract the compensation state sequence. The compensation state sequence refers to the changes in compensation values ​​at different times or spatial locations, including information such as compensation direction and compensation magnitude. For example, when analyzing an optimal compensation path, a series of compensation state points can be obtained, such as a compensation value of +2nm displacement in the X direction and -1nm displacement in the Y direction at path point (10, 20); and a compensation value of +3nm displacement in the X direction and -1.5nm displacement in the Y direction at path point (15, 25), etc. By collecting information from all these compensation state points, a complete compensation state sequence is formed, serving as the basis for the initial compensation parameters.

[0138] To more accurately analyze the distribution patterns of compensation parameters, the compensation state sequence needs to be divided into multiple local compensation regions in the spatial dimension. This division can be based on factors such as geometric similarity, functional similarity, or process sensitivity. For example, a layout area can be divided into multiple local compensation regions using a 5μm × 5μm grid. For adjacent local compensation regions, feature correlation analysis is performed to identify the similarities and differences between regions. Indicators such as the gradient of compensation parameter changes and correlation coefficients between adjacent regions can be calculated to determine the spatial compensation characteristics. For example, the analysis might show that in densely lined regions, the compensation parameters in adjacent regions gradually increase, with a growth rate of approximately 0.5 nm / μm; while in sparse regions, this growth rate is approximately 0.2 nm / μm.

[0139] The compensation state sequence is divided into multiple compensation stages along a temporal dimension. These stages can be based on key nodes in the process flow; for example, the compensation process can be divided into three stages: pre-lithography, during lithography, and post-lithography. Evolutionary analysis is performed on adjacent compensation stages to observe how the compensation parameters change over time. For example, the analysis might show that from the pre-lithography to the during lithography stage, the compensation parameter in the X direction increases by an average of 1.2 nm, while the compensation parameter in the Y direction decreases by an average of 0.8 nm. This temporal characteristic helps in understanding the dynamic evolution of compensation requirements.

[0140] Based on the spatial and temporal compensation characteristics obtained from the above analysis, spatial and temporal constraints are established for the compensation parameters. Spatial constraints ensure that the changes in compensation parameters in adjacent regions meet the requirements of smoothness and consistency, avoiding abrupt changes. For example, it can be stipulated that the difference in compensation parameters between two adjacent 5μm×5μm regions should not exceed 0.5nm. Temporal constraints ensure that the changes in compensation parameters at different stages conform to the requirements of the process flow, maintaining a reasonable transition. For example, it can be stipulated that the rate of change of compensation parameters from one stage to the next should not exceed 30%. These spatial and temporal constraints are combined to construct a compensation parameter constraint function. This constraint function can be expressed as a restriction condition on the compensation parameters, used in the subsequent optimization process.

[0141] The initial compensation parameters are optimized using the constructed compensation parameter constraint function. The optimization process can employ an iterative optimization method, adjusting the compensation parameters according to the constraints in each iteration until the convergence condition is met. For example, by continuously adjusting the compensation parameter values ​​to satisfy both spatial constraints (smooth transition between adjacent regions) and temporal constraints (reasonable evolution at different stages), an optimized set of compensation parameters is finally obtained. The initial compensation parameters in a certain region might be +2.7 nm in the X direction and -1.8 nm in the Y direction; after optimization, they are adjusted to +2.5 nm in the X direction and -1.6 nm in the Y direction to better meet the overall constraint requirements.

[0142] After optimizing the compensation parameters, it is necessary to analyze the local structural features of the mask layout data to determine the compensation application area. Different layout structures have different sensitivities to compensation, requiring targeted application. For example, fine-line structures, especially lines with a width of less than 100nm, require more precise compensation; while for large-area filled regions, the compensation accuracy requirement is relatively lower. By analyzing features such as line width, spacing, and corner density in the layout, key areas sensitive to process deviations are identified as the focus of compensation.

[0143] The optimized compensation parameters are applied to the defined compensation areas to generate layout compensation data. Based on the characteristics of each region and its corresponding optimal compensation parameters, the original mask layout data is locally adjusted. For example, for a critical line with a linewidth of 65nm, its position and shape on the mask are adjusted according to the optimized compensation parameters of +2.5nm in the X direction and -1.6nm in the Y direction; while for non-critical areas, smaller compensation values ​​may be applied or no compensation may be performed. Through this refined compensation application, the final generated layout compensation data can effectively reduce process deviations during manufacturing and improve chip manufacturing accuracy.

[0144] By performing multi-scale analysis of the compensation state sequence in both spatial and temporal dimensions, spatial and temporal compensation features are extracted, enabling comprehensive constraint optimization of compensation parameters. The compensation application area is determined by combining local structural features of the mask layout data, making the compensation more targeted. Establishing spatial and temporal constraints on the compensation parameters ensures the continuity and consistency of the compensation. This compensation parameter optimization method based on multi-dimensional feature analysis guarantees both the local accuracy of the compensation effect and maintains global coordination, significantly improving the accuracy and reliability of layout compensation.

[0145] A second aspect of this invention provides a deep learning-based lithography hotspot detection system incorporating a physical model, the system comprising:

[0146] The first unit is used to acquire design layout data and mask layout data, calculate the light intensity gradient and surface stress distribution using the dual-physics field model of optical proximity effect and chemical mechanical polishing, and obtain the physical sensitive feature map through tensor coupling calculation.

[0147] The second unit is used to input the physically sensitive feature map into the physical constraint neural network. The physical constraint neural network includes a feature reconstruction layer and a feature mapping layer. The feature reconstruction layer reconstructs the physically sensitive feature map based on the physical correlation between light intensity gradient and surface stress. The feature mapping layer maps the reconstructed feature map to the hotspot detection space and outputs hotspot detection data containing hotspot location information and physical feature information.

[0148] The third unit is used to map the hotspot detection data to the light intensity compensation space and the stress compensation space respectively to obtain the compensation state sequence, and to calculate the physical differences of the continuous compensation states in the compensation state sequence to construct the state transition matrix.

[0149] The fourth unit is used to calculate the transition cost of the compensation state based on the state transition matrix, determine the optimal compensation path based on the transition cost, extract compensation parameters from the optimal compensation path, and apply the compensation parameters to the mask layout data to generate layout compensation data.

[0150] A third aspect of the present invention provides an electronic device, comprising:

[0151] processor;

[0152] Memory used to store processor-executable instructions;

[0153] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0154] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0155] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting hotspots in photolithography using deep learning combined with a physical model, characterized in that, The method comprises the following steps: acquiring design layout data and mask layout data, performing edge detection on the design layout data to obtain edge profile data; constructing a double-physical-field model of optical proximity effect and chemical mechanical polishing according to the edge profile data, and calculating an initial light intensity gradient distribution and an initial surface stress distribution by using the double-physical-field model; modifying the initial surface stress distribution by taking the initial light intensity gradient distribution as a constraint condition, feeding back the modified surface stress distribution to update the initial light intensity gradient distribution, and obtaining a final light intensity gradient distribution and a final surface stress distribution through alternating iteration; converting the final light intensity gradient distribution and the final surface stress distribution into a light intensity gradient tensor and a surface stress tensor, calculating a physical correlation coefficient through tensor coupling, weighting and combining the light intensity gradient tensor and the surface stress tensor according to the physical correlation coefficient, and generating a physical sensitive feature map; inputting the physical sensitive feature map into a physical constraint neural network, wherein the physical constraint neural network comprises a feature reconstruction layer and a feature mapping layer, the feature reconstruction layer reconstructs the physical sensitive feature map based on the physical correlation between the light intensity gradient and the surface stress, and the feature mapping layer maps the reconstructed feature map to a hot spot detection space and outputs hot spot detection data containing hot spot position information and physical feature information; mapping the hot spot detection data to a light intensity compensation space and a stress compensation space to obtain a compensation state sequence, constructing a state transition matrix by calculating the physical difference of continuous compensation states in the compensation state sequence; calculating the transition cost of the compensation state based on the state transition matrix, determining the optimal compensation path according to the transition cost, extracting the compensation parameter from the optimal compensation path, and applying the compensation parameter to the mask layout data to generate layout compensation data.

2. The method of claim 1, wherein, The feature reconstruction layer reconstructs the physical sensitive feature map based on the physical correlation between the light intensity gradient and the surface stress, and the feature mapping layer maps the reconstructed feature map to a hot spot detection space and outputs hot spot detection data containing hot spot position information and physical feature information, which comprises: extracting the distribution information of the light intensity gradient and the surface stress from the physical sensitive feature map, analyzing the physical correlation between the light intensity gradient and the surface stress, and obtaining the physical field interaction intensity; constructing a feature constraint function based on the physical field interaction intensity, reconstructing the physical sensitive feature map by the feature constraint function in the feature reconstruction layer, and generating a reconstructed feature; decomposing the reconstructed feature into multi-scale features, calculating the feature weight of the multi-scale features according to the physical field interaction intensity, weighting and combining the multi-scale features based on the feature weight, and generating the reconstructed feature map; the feature mapping layer constructs a feature mapping function according to the physical field interaction intensity, inputs the reconstructed feature map into the feature mapping function, the feature mapping function performs feature enhancement on the reconstructed feature map based on the physical field interaction intensity, and generates an enhanced feature map; extracting the hot spot position information and the physical feature information from the enhanced feature map, and combining the hot spot position information and the physical feature information to generate the hot spot detection data.

3. The method of claim 1, wherein, The construction and training of the physical constraint neural network comprises: acquiring the physical sensitive feature map and the corresponding labeled hot spot position information, and constructing a training sample set; The physical constraint neural network structure is constructed, including a feature reconstruction layer and a feature mapping layer, the feature reconstruction layer adopts a parallel convolution structure, and the feature mapping layer adopts an attention enhanced unit; The light intensity gradient and surface stress information are extracted from the physical sensitive feature map, the feature constraint function is constructed based on the light intensity gradient and surface stress information, and the physical constraint loss function is constructed by combining the feature constraint function and the labeled hotspot position information; The training sample set is input into the physical constraint neural network, the physical constraint loss function value is calculated, the network parameters are optimized based on the physical constraint loss function value, and the optimized network parameters are obtained; The verification data is input into the physical constraint neural network configured with the optimized network parameters, when the error between the output hotspot detection result and the labeled hotspot position information in the verification data is less than a preset error threshold, the optimized network parameters are configured in the physical constraint neural network for hotspot detection data generation.

4. The method of claim 1, wherein, The hotspot detection data are respectively mapped to a light intensity compensation space and a stress compensation space to obtain a compensation state sequence, and the physical difference between consecutive compensation states in the compensation state sequence is calculated to construct a state transition matrix, including: The light intensity features and stress features in the hotspot detection data are extracted, and the feature correlation degree is calculated; The physical constraint function is constructed according to the feature correlation degree, and the light intensity constraint features and stress constraint features are obtained by using the physical constraint function to respectively constrain and decompose the light intensity features and stress features; The light intensity compensation space is constructed based on the light intensity constraint features, the stress compensation space is constructed based on the stress constraint features, and the hotspot detection data are respectively mapped to the light intensity compensation space and the stress compensation space to obtain a compensation state sequence; The light intensity difference value and the stress difference value between two compensation states in the compensation state sequence are calculated, and the state transition matrix is constructed according to the light intensity difference value and the stress difference value.

5. The method of claim 1, wherein, The transition cost of the compensation state is calculated based on the state transition matrix, and the optimal compensation path is determined according to the transition cost, including: The energy features and gradient features of adjacent compensation states in the state transition matrix are extracted, and the physical constraint function is constructed based on the correlation between the energy features and the gradient features; The energy constraint value corresponding to the energy features and the gradient constraint value corresponding to the gradient features are calculated based on the physical constraint function, and the transition cost between adjacent compensation states is obtained by weighted combination of the energy constraint value and the gradient constraint value; The path optimization objective function is constructed according to the physical constraint function and the transition cost, the dynamic programming equation is established based on the path optimization objective function, and the optimal compensation path is obtained by iteratively solving the dynamic programming equation.

6. The method of claim 1, wherein, The compensation parameters are extracted from the optimal compensation path, and the compensation parameters are applied to the mask layout data to generate layout compensation data, including: The compensation state sequence in the optimal compensation path is extracted, and the initial compensation parameters are extracted from the compensation state sequence; The compensation state sequence is divided into a plurality of local compensation regions in the spatial dimension, the spatial compensation features are obtained by feature correlation analysis on adjacent local compensation regions, and the time series compensation features are obtained by evolution analysis on adjacent compensation stages in the time series dimension. According to the space compensation feature and the time sequence compensation feature, a space constraint and a time sequence constraint of a compensation parameter are established, and the space constraint and the time sequence constraint are combined to construct a compensation parameter constraint function; An initial compensation parameter is optimized by using the compensation parameter constraint function to obtain an optimized compensation parameter, and a local structure feature of mask layout data is analyzed to determine a compensation action area; According to the local structure feature of the mask layout data, the optimized compensation parameter is applied to the compensation action area to generate layout compensation data.

7. A deep learning lithography hotspot detection system incorporating a physical model for implementing the method of any of the preceding claims 1-6, characterized in that, Comprise: A first unit is configured to obtain design layout data and mask layout data, calculate an optical intensity gradient and a surface stress distribution by using a double-physical-field model of optical proximity effect and chemical mechanical polishing, and obtain a physical sensitive feature map by tensor coupling calculation; A second unit is configured to input the physical sensitive feature map into a physical constraint neural network, the physical constraint neural network comprises a feature reconstruction layer and a feature mapping layer, the feature reconstruction layer reconstructs the physical sensitive feature map based on the physical correlation between the optical intensity gradient and the surface stress, and the feature mapping layer maps the reconstructed feature map to a hotspot detection space to output hotspot detection data containing hotspot position information and physical feature information; A third unit is configured to map the hotspot detection data to an optical intensity compensation space and a stress compensation space respectively to obtain a compensation state sequence, calculate a physical difference of continuous compensation states in the compensation state sequence to construct a state transition matrix; A fourth unit is configured to calculate a transition cost of the compensation state based on the state transition matrix, determine an optimal compensation path according to the transition cost, extract a compensation parameter from the optimal compensation path, and apply the compensation parameter to the mask layout data to generate layout compensation data.

8. An electronic device, comprising: Comprise: A processor; A memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 6.

Images