Deep learning method, device and equipment for chip photoetching pattern optimization

Abstract

The invention relates to the technical field of chip design optimization, and discloses a deep learning method, a deep learning device and deep learning equipment for chip photoetching pattern optimization. Geometric topological features of a photoetching pattern are extracted through a dynamic sparse convolutional network, optimized feature coding of a photoetching process is carried out based on adaptability analysis, and a deep learning result is obtained. According to the method, a pattern is divided into a plurality of sub-regions which can be processed in parallel by utilizing a PPO reinforcement learning algorithm, a preset strategy library is configured for each sub-region, multi-target collaborative processing simulation is carried out, sub-region boundaries and processing strategies are optimized, and finally an efficient collaborative processing strategy is generated. The adaptability and efficiency of the chip manufacturing process are enhanced, and the problem that in the prior art, the performance bottleneck exists when the photoetching effect is optimized by depending on manual adjustment is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of chip design optimization, and in particular to a deep learning method, device and equipment for chip lithography pattern optimization. Background Art

[0002] With the continuous advancement of semiconductor technology, the size of chips is getting smaller and smaller, and the manufacturing process is becoming more and more complex. In the chip manufacturing process, the photolithography process is one of the most critical steps, which directly affects the performance, function and production cost of the chip. The optimization of the photolithography pattern is an important link to ensure the accuracy and efficiency of the photolithography process, especially at the deep micro-nano scale. The design and optimization of the photolithography pattern is crucial to achieving high-performance chips.

[0003] Traditional lithography pattern optimization methods usually rely on a large number of empirical rules and manual adjustments. Although these methods can optimize the lithography effect to a certain extent, they cannot fully cope with the ever-changing process requirements and complex lithography patterns. In addition, with the increasing complexity of chip design, traditional lithography optimization methods have become difficult to cope with the growing computing and optimization needs, especially in the miniaturized lithography process. The complexity and mutual influence of lithography patterns have caused traditional methods to face performance bottlenecks and insufficient precision. Summary of the Invention

[0004] The purpose of the present invention is to provide a deep learning method, device and equipment for chip lithography pattern optimization, aiming to solve the performance bottleneck problem of relying on manual adjustment to optimize the lithography effect in the existing technology.

[0005] The present invention is implemented as follows: In a first aspect, the present invention provides a deep learning method for chip lithography pattern optimization, comprising: Obtaining original design information of the lithographic pattern, and extracting local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern; Performing a photolithography process adaptability analysis on the geometric topology feature information, and performing feature encoding on the geometric topology feature information based on a result of the adaptability analysis to generate a photolithography process adaptability feature matrix of the photolithography pattern; Dividing the lithographic pattern into a plurality of sub-regions that can be processed in parallel according to the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configuring a corresponding preset strategy library for each of the sub-regions; According to the division status of each sub-region and the configured preset strategy library, a multi-target collaborative processing simulation is performed on the lithography pattern to adjust the boundary division form of the sub-region and confirm the regional processing strategy, so as to finally generate the collaborative processing strategy of the lithography pattern.

[0006] In a second aspect, the present invention provides a deep learning device for chip lithography pattern optimization, which is used to implement the deep learning method for chip lithography pattern optimization described in any one of the first aspects, comprising: A topology analysis module is used to obtain the original design information of the lithographic pattern and extract local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern; an adaptability analysis module, configured to perform adaptability analysis of the geometric topology feature information with respect to the lithography process, and to perform feature encoding on the geometric topology feature information based on the adaptability analysis result, so as to generate a lithography process adaptability feature matrix of the lithography pattern; A region division module is used to divide the lithographic pattern into a plurality of sub-regions that can be processed in parallel based on the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configure a corresponding preset strategy library for each sub-region; The collaborative simulation module is used to perform multi-target collaborative processing simulation on the lithography pattern according to the division status of each sub-region and the configured preset strategy library, so as to adjust the boundary division form of the sub-region and confirm the regional processing strategy, so as to finally generate the collaborative processing strategy of the lithography pattern.

[0007] In a third aspect, the present invention provides a deep learning device for chip lithography graphic optimization, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements a deep learning method for chip lithography graphic optimization as described in any one of the first aspects.

[0008] The present invention provides a deep learning method for chip lithography pattern optimization, which has the following beneficial effects: The present invention extracts the geometric topological features of the lithography pattern through a dynamic sparse convolutional network, performs optimized feature encoding of the lithography process based on adaptability analysis, utilizes the PPO reinforcement learning algorithm to divide the pattern into multiple sub-regions that can be processed in parallel, and configures a preset strategy library for each sub-region to perform multi-objective collaborative processing simulation, optimize the sub-region boundaries and processing strategies, and ultimately generate an efficient collaborative processing strategy. This method improves the automation level and accuracy of lithography pattern optimization, enhances the adaptability and efficiency of the chip manufacturing process, and solves the performance bottleneck problem of relying on manual adjustment to optimize the lithography effect in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 This is a schematic diagram of the steps of a deep learning method for chip lithography pattern optimization provided by an embodiment of the present invention; Figure 2 It is a structural schematic diagram of a deep learning device for chip lithography pattern optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0011] The implementation of the present invention is described in detail below with reference to specific embodiments.

[0012] Reference Figure 1 、 Figure 2 As shown, a preferred embodiment of the present invention is provided.

[0013] In a first aspect, the present invention provides a deep learning method for chip lithography pattern optimization, comprising: S1: obtaining original design information of the lithographic pattern, and extracting local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern; S2: performing adaptability analysis of the lithography process on the geometric topology feature information, and performing feature encoding on the geometric topology feature information based on the adaptability analysis result to generate a lithography process adaptability feature matrix of the lithography pattern; S3: dividing the lithographic pattern into a plurality of sub-regions that can be processed in parallel according to the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configuring a corresponding preset strategy library for each sub-region; S4: Based on the division status of each sub-region and the configured preset strategy library, a multi-target collaborative processing simulation is performed on the lithography pattern to adjust the boundary division form of the sub-region and confirm the regional processing strategy, so as to finally generate the collaborative processing strategy of the lithography pattern.

[0014] Specifically, in step S1 of the embodiment provided by the present invention, the original design information of the photolithography pattern is obtained, which usually includes the graphic design data of the chip, such as the circuit layout, the spatial distribution of graphic elements (such as lines, holes, etc.) and their size, shape and other information. These design information are generally stored in vector format or grid format. The original design information is usually a complex two-dimensional graphic and needs to be properly converted and processed. A common practice is to convert these graphic data into layered vector graphics. The purpose of this step is to represent the graphics as a discrete set of points, so that each point corresponds to a key position or feature in the photolithography pattern.

[0015] More specifically, after the graphic is converted into a discrete point set, an adjacency matrix needs to be constructed. The adjacency matrix is a matrix structure used to describe the relationship between points in the graphic. It is used to indicate whether the points are adjacent and the strength of the adjacency relationship. Through the adjacency matrix, the graphic can be further analyzed and key areas can be identified. By analyzing the adjacency matrix, key areas with special significance in the graphic can be marked. Key areas may involve important structures in the design or parts that may have a greater impact on the lithography process.

[0016] More specifically, a pre-trained dynamic sparse convolutional network (DSC-Net) model is used to extract graphic features from these labeled discrete point sets. The dynamic sparse convolutional network is a convolutional neural network designed specifically for processing sparse data (such as photolithography graphics). Through this model, local features of the graphics, such as curvature, density, shape complexity, etc., can be extracted.

[0017] More specifically, the dynamic sparse convolutional network processes the local features of graphic data through adaptive sparse coding and flexibly models sparse data. The convolution kernel configuration in the network (such as 3x3, 5x5, etc.) will be adjusted according to the results of adaptive sparse coding to ensure the optimal feature extraction effect. The network will dynamically adjust its sparse pattern parameters to adapt to different types of graphic features and improve the accuracy of graphic feature extraction. After processing by the dynamic sparse convolutional network, the geometric topological feature information of the lithographic graphic is finally obtained. This information can describe the geometric structure, topological connection relationship (such as the connection between different design elements), shape and size of the lithographic graphic.

[0018] It can be understood that the dynamic sparse convolutional network model can effectively extract high-dimensional geometric topological features from the design information of the lithography graphics. Through sparse coding and convolution kernel adjustment, the network can efficiently capture local changes and important features in the graphics. Due to the use of dynamic sparse convolutional networks, this method can process sparse data, especially for those lithography graphics with irregular, complex shapes and details, which can improve the accuracy and efficiency of feature extraction.

[0019] More specifically, by conducting in-depth analysis of the geometric topological features of the lithographic patterns, potential design flaws or inappropriate structural layouts can be identified, thereby reducing errors during the lithographic process and improving the quality and manufacturability of the final product. The dynamic sparse convolutional network can flexibly adjust the convolution kernel and sparsity pattern according to different lithographic pattern designs and processing requirements, ensuring adaptability to various lithographic designs. This adaptability helps achieve better optimization results in different lithographic processes. The obtained geometric topological feature information provides important input data for subsequent optimization steps such as lithographic process adaptability analysis, region division, and strategy configuration. These features will be used to further optimize the lithographic pattern processing process, thereby achieving more efficient and accurate chip manufacturing.

[0020] Specifically, in step S2 of the embodiment provided by the present invention, geometric topological feature information is extracted and prepared. The geometric topological feature information of the lithographic pattern obtained from the preceding module (such as a dynamic sparse convolutional network) includes: pattern curvature, edge sharpness, area density (pattern density), node connection method, topological relationship, closed loop structure, sharp angles, gap distance, etc., which serve as input for subsequent adaptive analysis.

[0021] More specifically, typical process conditions are extracted from the existing photolithography process knowledge base: light source type (such as 193nm ArF, EUV, etc.), numerical aperture (NA), illumination method (conventional illumination, polarized illumination, free-form illumination), exposure agent, photoresist type, process fluctuation parameters (such as focal length offset, dose disturbance), and the Monte Carlo method can be used to perform multiple rounds of perturbation sampling on the above parameters to simulate process fluctuations.

[0022] More specifically, based on simulation or deep model-based process adaptability assessment, optical simulation software (such as LithoSim, Calibre, etc.) is used to simulate the imaging results of graphics under different process conditions, extract indicators such as graphic collapse, bridging, fracture, and line width change, or use pre-trained graphic-process mapping models (such as CNN+RNN or Transformer) to predict the key manufacturing indicators of graphics under different processes (CD error, OPC convergence, offset risk, etc.), and output the "fitness score" of each local graphic area under different processes. More specifically, the adaptability of each region (or graphic element) under different process conditions is quantified and characterized. For example, curvature greater than a certain threshold + high-density areas have poor adaptability (high probability of collapse) when NA=0.93; thin lines + small spacing have a high risk of offset under polarized illumination; and square-corner structures have drastic changes in line width when the focus shifts. Ultimately, process-sensitive feature labels are formed for each local area of the graphic. The above adaptability labels are vectorized and converted into a format readable by machine learning to generate a lithography process adaptation feature matrix.

[0023] It is understandable that this analysis process establishes a clear connection between "graphic geometric topology" and "process manufacturing risk", which is the key to achieving adaptive optimization of lithography processes. It can identify high-risk areas in the graphics that are sensitive to certain types of processes (for example: extremely small spacing, bending angles, fine lines), thereby supporting targeted design corrections (such as OPC or rule adjustments). Traditional process simulation is very time-consuming, but through deep models + feature matrices, the compatibility of graphic design and process parameters can be quickly and large-scale evaluated, significantly reducing the iterative cost from design to manufacturing. The feature matrix can be used as an input state space for reinforcement learning algorithms (such as PPO, Nash-Q) to make training decisions in subsequent area division and processing strategy adjustments, thereby achieving regional-level adaptive optimization of lithography parameters.

[0024] Specifically, in step S3 of the embodiment provided by the present invention, the following information of each lithography graphic area is input as a state vector: geometric topological features (curvature, density, connectivity, etc.), lithography process adaptation features (adaptability score matrix), adjacency relationships between graphics or spatial layout encoding (such as convolution features, position vectors), and each "state" represents the local process and geometric state of a candidate graphic area.

[0025] More specifically, the action is defined as the division decision or merging strategy assignment for the graphic area: whether to divide the area into the current sub-area, whether to end the division of the current sub-area, start a new sub-area, and which process strategy to select for the current sub-area? More specifically, a reward function is designed. The reward function should measure the effect of the partitioning strategy on process parallelism and manufacturability optimization, which usually includes: Parallelism reward: The more balanced and independent the sub-regions are, the higher the reward; Process consistency reward: The closer the adaptation characteristics within each sub-region are, the higher the reward; Strategy matching reward: Whether the configured strategy is suitable for the risk characteristics of the region; Penalty: If the region is divided too finely or the strategy does not match, a penalty will be given. More specifically, PPO is a policy gradient algorithm with the advantages of good stability and easy tuning. The policy network inputs the state and outputs the action probability distribution. The value network inputs the state and outputs the value estimate of the current state. The PPO agent gradually makes divisions and policy allocation decisions on the graphic area based on the current policy network. The environment calculates the reward based on the division results and uses the PPO Clip loss function to perform gradient updates on the policy network to maintain the stability of the policy update. At the same time, the value network is updated to improve the accuracy of the reward estimate. The strategy is gradually optimized through multiple episodes, so that the division results tend to have a distribution with strong parallelism and high policy matching.

[0026] More specifically, the sub-region division and strategy configuration results are output, and the final result is the sub-region division plan (for example, the graphics are divided into regions A, B, C, D…), the process feature summary information of each sub-region (density, complexity, etc.), and the strategy configured for each sub-region.

[0027] It is understandable that by learning the optimal division method through PPO, the parallel processing potential of graphics is significantly improved, the manufacturing cycle and total cost are reduced, and the graphics within each sub-area have similar process adaptability, which is conducive to unifying exposure, focal length, and light source strategies, reducing error accumulation within the region, and reinforcing learning to automatically match the best strategy for areas of different complexity without the need for manual experience rules, greatly improving the degree of automation. The PPO model can be migrated to different lithography process platforms (such as EUV, ArF, etc.) by simply adjusting the state dimension and strategy set, and is highly flexible. The division and strategy configuration results are used as input to further optimize the scheduling, energy consumption control, thermal field regulation and other process links globally, thereby improving the entire line capacity and yield.

[0028] Specifically, in step S4 of the embodiment provided by the present invention, the sub-region division and the process adaptability characteristics of each sub-region from the previous PPO reinforcement learning are used, and the preset strategy library is the strategy configuration information for each sub-region, including photoresist type, exposure dose, scanning speed, light source type, etc.

[0029] More specifically, the multi-objective optimization goals include lithography yield: ensuring that the graphic area can still maintain sufficient dimensional tolerance under the influence of process; shortest processing time: reducing the processing time of each sub-area and optimizing overall production capacity; resource balance: balancing resource allocation among multiple sub-areas to avoid excessive local process load; conflict minimization: avoiding process interference between sub-areas (such as focal length offset, exposure overlap, etc.).

[0030] More specifically, lithography simulation tools (such as Calibre, Sentaurus Lithography, etc.) are used to perform preliminary simulations on each sub-area to determine the imaging effect of each area under a given strategy and simulate the impact of various process parameters on the lithography pattern, including focal length offset, uneven exposure dose, light source wavelength and other factors.

[0031] More specifically, taking into account the complexity of the sub-regions (such as areas with dense fine lines and areas with large curvature), dynamic adjustments are made to avoid excessive cross-interference at the boundaries of the regions and ensure the stability of the lithography process within the regions. The lithography error model is used to evaluate the processing results of the boundary regions, and the positions of the sub-region boundaries are adjusted to avoid the generation of redundant overlapping areas or affected areas. The regional processing strategy is confirmed, and the effectiveness of the current configuration strategy is judged through simulation. For example, whether the exposure dose is reasonable and whether the scanning speed is suitable for the morphology and geometric complexity of the current region. A multi-objective evaluation is performed on the processing strategy of each sub-region, including a comprehensive evaluation of multiple objectives such as processing time, imaging error, and process adaptability.

[0032] More specifically, a multi-objective optimization algorithm (such as NSGA-II, Pareto optimization, reinforcement learning, etc.) is used to adjust the division boundaries and strategy configuration of each sub-region to achieve the optimal global goal. After each simulation run, the quality and mutual relationship of each sub-region are analyzed, and the boundary division and processing strategy of each region are updated based on multi-objective evaluation indicators.

[0033] More specifically, based on the simulation results, the boundaries of the sub-regions are automatically adjusted, the relative positions of the boundaries are optimized, process conflicts are reduced, and the coordination between regions is improved. Different lithography strategies are configured according to the lithography characteristics of different sub-regions (such as high-density areas, low-density areas, corners, etc.) to ensure the overall lithography quality and production capacity.

[0034] More specifically, the final photolithography pattern collaborative processing strategy is generated, which includes: the division boundary of each sub-area, the optimal photolithography strategy selected for each sub-area (including exposure dose, scanning speed, light source wavelength, photoresist type, etc.), and the optimized values of process parameters (such as focal length, contrast, etc.) More specifically, the final lithography pattern and its coordinated processing strategy are input into the simulation tool for secondary verification to ensure that the process compatibility and yield of each sub-area under the new configuration meet the expectations. Based on the verification results, the boundary division and strategy configuration are further refined, and the cycle is iterated until all optimization goals are met.

[0035] It can be understood that through multi-objective collaborative simulation, it is possible to achieve a reasonable division of the lithography pattern, ensure that the processing of each sub-area can be seamlessly coordinated with other areas, and optimize the processing strategy of each area, making the lithography process more efficient. By dynamically adjusting the area division and processing strategy, process interference and overlap between areas are avoided, defects caused by process mismatch in the lithography process are reduced, and the overall lithography yield is improved. The division of sub-areas is optimized so that the lithography process in each area can be carried out independently and in parallel, greatly shortening the overall processing time and increasing production capacity through appropriate resource allocation. Through multi-objective optimization, the process load balance of each sub-area in the lithography process can be ensured, avoiding resource waste or quality problems caused by excessively high or low process requirements in some areas. The system can respond to parameter changes in the lithography process (such as light source wavelength, exposure dose, etc.) in real time and automatically adjust the sub-area division and strategy accordingly. It has a certain degree of process adaptability and flexibility. The generated collaborative processing strategy can be seamlessly integrated into the production process, providing precise guidance for subsequent process scheduling and control, and reducing the complexity of manual intervention.

[0036] The present invention provides a deep learning method for chip lithography pattern optimization, which has the following beneficial effects: The present invention extracts the geometric topological features of the lithography pattern through a dynamic sparse convolutional network, performs optimized feature encoding of the lithography process based on adaptability analysis, utilizes the PPO reinforcement learning algorithm to divide the pattern into multiple sub-regions that can be processed in parallel, and configures a preset strategy library for each sub-region to perform multi-objective collaborative processing simulation, optimize the sub-region boundaries and processing strategies, and ultimately generate an efficient collaborative processing strategy. This method improves the automation level and accuracy of lithography pattern optimization, enhances the adaptability and efficiency of the chip manufacturing process, and solves the performance bottleneck problem of relying on manual adjustment to optimize the lithography effect in the existing technology.

[0037] Preferably, the steps of obtaining original design information of the lithographic pattern and extracting local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern include: S11: acquiring original design information of the lithography pattern, and performing layered vector graphics conversion on the original design information to generate a discrete point set corresponding to the original design information; S12: constructing an adjacency matrix based on the discrete point set, and performing key area marking on the discrete point set by parsing the adjacency matrix; S13: Perform preliminary image curvature and density calculations on the discrete point sets labeled in key areas using a pre-trained dynamic sparse convolutional network model to generate corresponding adaptive sparse codes. S14: According to the adaptive sparse coding, the dynamic sparse convolutional network model is configured with a 3X3 convolution kernel or a 5X5 convolution kernel, and the sparse pattern parameters of the dynamic sparse convolutional network model are synchronously and dynamically adjusted to extract graphic features of the original design information, and obtain the geometric topological connection matrix of the lithography pattern as the geometric topological feature information of the lithography pattern.

[0038] Specifically, the original design information generally comes from the design file of the lithography graphics, which usually exists in the form of GDSII format, OASIS format, etc. The design file is imported and the data is parsed to extract the basic information of the graphics (such as graphic outline, hierarchy, path, etc.).

[0039] More specifically, each layer in the lithography pattern is decomposed and represented as a vector graphic. Geometric processing technology is used to convert each layer of design graphics from a vector graphic (such as polygons, line segments, etc.) into a discrete point set. This can be achieved by dividing the graphics into uniform grid points or extracting key points based on the edges / contours of the graphics. The generated discrete point set contains the precise position of the design graphics in space for subsequent processing.

[0040] More specifically, an adjacency matrix is constructed and region annotation is performed. Based on the position of the discrete point set, an adjacency matrix is constructed, which records the adjacency relationship between points. If two points are within a certain distance threshold, they are considered to be adjacent points. According to the adjacency matrix, the key areas in the point set are annotated. These areas may be high-density areas in the graphics, complex structural areas, or parts that have an important impact on the lithography process. This step is performed by analyzing the adjacency matrix, such as calculating the density, curvature and other features of the graphics, thereby identifying the structural characteristics of the graphics.

[0041] More specifically, a pre-trained dynamic sparse convolutional network (DSCNN) model is used to process a set of discrete points labeled in key areas, calculate the local curvature of each point in the lithographic pattern (for example, the degree of curvature between the point and its adjacent points), and calculate the pattern density of each area, for example, how the density of points changes within a specific radius.

[0042] More specifically, based on the curvature and density information of the image, an adaptive sparse code is generated through a dynamic sparse convolutional network. Sparse coding extracts important features from the image and compresses unimportant features, reducing data redundancy while retaining important information of the image.

[0043] More specifically, based on the generated adaptive sparse coding, the convolution kernel configuration of the convolution network is dynamically adjusted, and the convolution kernel of appropriate size (such as 3x3 or 5x5) is dynamically selected to extract features at different scales. The selection of the convolution kernel is directly related to the curvature and density characteristics of the target area. According to the structural characteristics of the graphics (such as high-density areas, curved areas, etc.), the sparse pattern parameters of the convolution network are dynamically adjusted. Sparse convolution reduces the computational cost by only calculating the important parts of the image. The adjustment of the sparse pattern parameters can make the network more efficient while increasing the attention to key areas.

[0044] More specifically, the original design information is processed through convolution operations and sparse coding, and the geometric topological features of the lithographic pattern are finally extracted. This feature includes the geometric structure of the lithographic pattern (such as connection relationship, curvature, density, etc.) and topological information (such as point adjacency relationship, area division, etc.). Based on the extracted geometric topological information, a geometric topological connection matrix is generated. This matrix records the spatial relationship and connection pattern between each element in the lithographic pattern. Each row in the matrix represents the connection relationship between a point or area and other points or areas, where 1 indicates that there is a connection and 0 indicates that there is no connection.

[0045] It is understandable that the dynamic sparse convolutional network model can extract the local geometric topological features of the lithographic pattern, including important information such as curvature and density. This method makes the geometric feature extraction of the lithographic pattern more efficient and avoids unnecessary computational redundancy in traditional methods. The introduction of sparse convolution mode greatly reduces the amount of calculation, especially when processing large-scale lithographic patterns. It can quickly extract key information, thereby accelerating the processing and analysis of lithographic patterns. Adaptive sparse coding and convolution kernel configuration can dynamically adjust parameters for different lithographic pattern features, thereby more accurately capturing local features in the pattern. This makes the model highly adaptable to a variety of graphic topologies.

[0046] More specifically, by generating a geometric topological connectivity matrix, the connectivity between regions within a pattern can be effectively represented. This is crucial for subsequent process optimization of lithographic patterns, such as region partitioning and process scheduling. This method supports multi-level and hierarchical analysis of patterns and is applicable to a wide range of complex lithographic designs. Model parameters can be dynamically adjusted based on the complexity of the lithographic pattern, enabling more efficient feature extraction.

[0047] Preferably, the step of performing adaptability analysis of the lithography process on the geometric topological feature information includes: S21: Retrieving from a photolithography process database several photolithography processes consisting of exposure wavelength, numerical aperture, photoresist type, and development condition elements; S22: simulating a lithography pattern production process of various lithography processes for the geometric topology feature information using a deep learning model pre-trained using lithography process data, so as to generate simulated production information of the various lithography processes relative to the lithography pattern; S23: Perform multi-dimensional adaptability index evaluation on simulated production information of various photolithography processes to obtain fidelity characteristics consisting of pattern edge roughness and critical dimension uniformity, manufacturability characteristics consisting of minimum resolvable line width, exposure latitude, and depth of focus, and potential defect characteristics consisting of short circuit risk and photoresist residue probability; S24: Analyze the graphic impact of process fluctuations on the photolithography process using the Monte Carlo method to generate sensitivity characteristics of the photolithography process. The fidelity characteristics, manufacturability characteristics, potential defect characteristics, and sensitivity characteristics of the photolithography process collectively serve as adaptability characteristics of the photolithography process.

[0048] Specifically, several different photolithography processes are retrieved from the photolithography process database. These processes include the following elements: exposure wavelength: different exposure wavelengths have different effects on the resolution and contrast of the graphics, numerical aperture (NA): affects the resolution and spot shape of the image, photoresist type: the chemical properties and photosensitivity of different photoresists have a significant impact on the imaging accuracy and photolithography performance of the graphics, development conditions: factors such as the time, temperature, and chemicals of the development process will affect the final quality of the photolithography graphics.

[0049] More specifically, a large-scale photolithography process database containing the above-mentioned process elements is selected to ensure the diversity and comprehensiveness of the process. Through data preprocessing and screening, relevant process data is extracted for subsequent analysis and simulation.

[0050] More specifically, a deep learning model simulates the pattern production process, trained based on historical data from photolithography processes. This model learns the impact of different photolithography processes on patterns, including the effects of exposure, development, depth of focus, and other factors on pattern quality. The trained model combines input geometric topological feature information with various photolithography process elements to simulate the production process of photolithography patterns under different processes, generating simulated production information for the pattern. This simulation includes the pattern's exposure and development processes, light scattering and absorption during exposure, the photoresist's reaction, and changes in pattern characteristics after development. Each photolithography process generates a set of simulated production information, reflecting the production results of the photolithography pattern under that process. This information will be used for subsequent adaptability assessments.

[0051] More specifically, a multi-dimensional adaptability index evaluation is performed on the simulated production information to obtain various features required for the adaptability analysis of the lithography process: Fidelity features include graphic edge roughness: reflecting whether the edge of the graphic is smooth and whether there is undesirable edge roughness, which affects the accuracy and readability of the finished graphic; critical dimension uniformity: evaluating the uniformity of critical dimensions (such as line width, spacing, etc.) in the lithography graphic to ensure process controllability and avoid functional failure due to uneven dimensions.

[0052] More specifically, manufacturability characteristics include minimum resolvable line width: reflecting the minimum graphic size that the photolithography process can clearly distinguish, which is a key indicator of the resolution of the photolithography process; exposure tolerance: refers to the tolerance range of exposure. A higher exposure tolerance means that the process is less sensitive to changes in exposure and can tolerate a certain degree of deviation; depth of focus: a larger depth of focus means that the photolithography process has a higher tolerance for changes in focal length, thereby improving precision during the manufacturing process.

[0053] More specifically, potential defect characteristics include short circuit risk: This indicates the possibility of short circuits in photolithography patterns, which is typically related to factors such as pattern spacing and line width. Short circuits can cause circuit failure. There is also the probability of photoresist residue: During the photolithography process, the photoresist may not be fully developed, resulting in residual material in the pattern, which can affect the integrity and quality of the pattern.

[0054] More specifically, the Monte Carlo method is used to simulate possible fluctuations in the process through multiple simulations and the introduction of random variables (such as exposure wavelength, depth of focus, development conditions, etc.). Each simulation will generate different production results. By analyzing the changes between these results, the sensitivity characteristics of the process to the photolithography pattern can be obtained.

[0055] More specifically, by statistically analyzing Monte Carlo simulation results, we calculate the sensitivity of the lithography process to pattern characteristics (such as edge roughness and line width). These characteristics indicate the extent to which different process variations affect pattern quality. By evaluating the sensitivity of different process parameters, we identify the process factors most susceptible to fluctuations, helping to optimize the lithography process.

[0056] More specifically, the fidelity characteristics, manufacturability characteristics, potential defect characteristics and sensitivity characteristics are integrated into the adaptability characteristics of the lithography process. These characteristics can comprehensively reflect the process's adaptability to graphic quality under different environments and conditions. Taking these characteristics into consideration, an adaptability score or ranking is provided for each lithography process, thereby helping to select the lithography process that is most suitable for a specific design graphic.

[0057] It is understandable that through the adaptability analysis of the lithography process, the impact of different lithography processes on the design graphics can be accurately evaluated, helping engineers to select the most suitable process and ensure high-quality production of the design graphics. Through multi-dimensional adaptability index evaluation, each parameter of the lithography process can be deeply analyzed, the configuration of process parameters can be optimized, and the production effect and manufacturability of the graphics can be improved. The evaluation of potential defect characteristics such as short circuit risks and photoresist residues can help to discover potential production problems, adjust the process in advance, and avoid quality problems in large-scale production.

[0058] More specifically, process fluctuation analysis using the Monte Carlo method can identify the sensitivity characteristics of the process, help engineers better respond to changes in actual production, and perform fault-tolerant design, thereby improving the stability and reliability of the lithography process. Simulating the lithography pattern production process through deep learning models can accelerate the optimization process of the lithography process and reduce the time and cost required by traditional experimental methods.

[0059] Preferably, the step of encoding the geometric topological feature information based on the result of the adaptability analysis to generate a lithography process adaptation feature matrix of the lithography pattern includes: S25: Binding various process parameters in the photolithography process with corresponding adaptability characteristics to form a process-adaptability mapping table; S26: using the geometric topological feature information as a basic map, and performing process-adaptability mapping factor analysis on each specific location of the basic map according to each lithography process in the mapping table, so as to obtain a set of process-adaptability mapping factors corresponding to the basic map for each lithography process; wherein the set of process-adaptability mapping factors includes each process-adaptability mapping factor of a lithography process with respect to the basic map, and the process-adaptability mapping factor is used to describe an intrinsic mapping relationship between the lithography pattern and the lithography parameters and adaptability of the lithography process; S27: performing information parsing and feature encoding under interactive verification on the process-adaptability mapping factor sets of various lithography processes to generate a lithography process adaptation feature matrix composed of a set of lithography process mapping feature vectors at specific locations of the basic map.

[0060] Specifically, each photolithography process has a set of process parameters (such as exposure wavelength, numerical aperture, photoresist type, etc.), and the previous adaptability analysis will generate a set of adaptability characteristics (such as fidelity, manufacturability, potential defects, etc.). The parameters of each photolithography process are bound to the corresponding adaptability characteristics to form a process-adaptability mapping table. This table records the mapping relationship between each parameter of each process and the corresponding adaptability characteristics, for example, exposure wavelength and minimum resolvable line width, photoresist type and pattern edge roughness, etc. The mapping table can be designed as a multi-dimensional structure, which contains the specific values of each process parameter and the adaptability evaluation results of the process on the corresponding geometric figure.

[0061] More specifically, the base atlas refers to a collection of geometric topological feature information, which includes all key parameters of the design graphics (such as line width, spacing, curvature, etc.). The base atlas is the original geometric feature of the lithography graphics and is used as the basis for subsequent adaptability mapping. For each lithography process, based on the process parameters in the mapping table and combined with the geometric feature information in the base atlas, the process-adaptability mapping factors of each position are parsed in turn. The mapping factors are used to describe the mapping relationship between the lithography process and the geometric features in the base atlas, including the impact of each process parameter on the graphics and adaptability evaluation. For example, the process mapping factor at a certain position can describe the graphics performance under specific exposure wavelength and focal depth conditions. In this process, different parameters of the lithography process (such as exposure latitude, focal depth, etc.) are combined with geometric features (such as line width, curvature, etc.) to form a process-adaptability mapping factor set. For each lithography process, a corresponding process-adaptability mapping factor set is generated, which contains the mapping factors of the lithography process at all positions.

[0062] More specifically, the process-adaptability mapping factor sets for different lithography processes are cross-validated to ensure that the adaptability characteristics of different processes are consistent and reflect the actual process performance. This verification process involves verifying the accuracy of the mapping factors through simulation or experimental data to ensure that the process adaptability assessment results truly reflect the actual performance in the production process. Feature encoding is then performed on the cross-validated process-adaptability mapping factor set. The purpose of encoding is to convert each process mapping factor into a numerical or vector form to facilitate subsequent analysis and calculation. Feature encoding can use different encoding methods, such as numerical encoding and vector encoding, depending on the analysis requirements.

[0063] More specifically, the process-adaptability mapping factors for each position are encoded to form a set of lithography process mapping feature vectors. Each vector represents the lithography process adaptability characteristic for the corresponding position in the base map. The vector contains information about the impact of the lithography process on that position, including lithography parameters, adaptability assessment, and other information. By organizing and arranging the feature vectors for all positions, a lithography process adaptability feature matrix is ultimately formed. Each row of this matrix corresponds to the lithography process adaptability characteristic of a specific position, and each column represents a specific process parameter or adaptability characteristic. This matrix can be used for subsequent analysis, optimization, machine learning modeling, and other applications.

[0064] It is understandable that by encoding and matrixing the adaptability characteristics of the lithography process, complex lithography process information can be converted into a systematic numerical representation, which is convenient for quantitative analysis and visualization. This method can help process engineers more clearly understand the impact of the lithography process on different graphics. By analyzing the process-adaptability mapping factor, the intrinsic mapping relationship between the lithography process parameters and the graphics can be revealed, which helps to optimize the parameter settings of the lithography process and reduce errors in the production of lithography graphics.

[0065] More specifically, the lithography process adaptation feature matrix provides a lithography process adaptability evaluation framework based on specific graphics. Through matrix analysis, the most critical influencing factors in the process can be identified, and a basis can be provided for process optimization, thereby improving the accuracy and efficiency of process optimization. By using the lithography process adaptation feature matrix, combined with machine learning and data analysis methods, the level of automated optimization of the process can be further improved. The feature matrix can be used as input to train the machine learning model, predict the impact of different processes on future production, and support intelligent process improvement. By customizing the process adaptation feature matrix for each geometric topological feature, it can quickly adapt to different graphic design requirements, promote the rapid customization and precise debugging of different lithography processes, and improve production efficiency.

[0066] Preferably, the steps of dividing the lithographic pattern into a plurality of sub-regions that can be processed in parallel according to the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configuring a corresponding preset strategy library for each sub-region include: S31: performing an initial round of region division processing on the geometric topological feature information according to the photolithography process adaptation feature matrix based on the PPO reinforcement learning algorithm, so as to divide a plurality of photolithography process regions based on the geometric topological feature information; S32: Evaluating the region division result according to a preset region division reward function to supervise the PPO reinforcement learning algorithm to perform region division optimization in a loop, thereby dividing the lithographic pattern into a plurality of sub-regions that can be processed in parallel; S33: Simulating the lithography effects of each sub-region under different strategies according to the lithography process adaptation feature matrix by generating an adversarial network model, so as to configure a corresponding preset strategy library for each sub-region.

[0067] Specifically, the PPO algorithm is a commonly used reinforcement learning algorithm used to optimize decision-making strategies, and is particularly suitable for complex and continuous environments. In this application, the PPO reinforcement learning algorithm is used to guide the process of lithography pattern area division. Based on the lithography process adaptation feature matrix (containing the adaptability analysis results of the lithography process to geometric topological features) and geometric topological feature information (such as line width, spacing, curvature, etc.), the PPO algorithm performs area division through several initial rounds. During this process, the PPO algorithm learns how to divide the entire lithography pattern into multiple sub-areas based on the adaptability characteristics of the lithography process to optimize the lithography effect and reduce errors in the process.

[0068] More specifically, through the initial strategy of the PPO algorithm, the graphic will be divided into several independent regions, each of which represents a specific process adaptability requirement in the lithography process adaptability feature matrix. These sub-regions may include different graphic structures or parameter regions.

[0069] More specifically, a region partitioning reward function is designed to evaluate the effects of different region partitioning results. The reward function is calculated based on the adaptability features in the lithography process adaptability matrix. The goal is to maximize the match between the lithography effect and the adaptability features. The reward function can consider multiple factors, such as: lithography accuracy (for example, minimum line width, pattern fidelity, etc.), adaptability (such as exposure tolerance, pattern defects, etc.), and production efficiency (for example, the parallel processing capability of each sub-region after pattern partitioning). More specifically, the PPO algorithm interacts with the feedback of the reward function and adjusts the partitioning strategy in each round of iteration. The optimization goal is to make the area division more reasonable, improve process adaptability and lithography effect. Through this cyclic process, the PPO algorithm will gradually improve the partitioning strategy to ensure that the final area division can support the optimal lithography process. After multiple rounds of optimization, the PPO reinforcement learning algorithm will eventually divide the lithography pattern into several sub-areas that can be processed in parallel. The division of each sub-area will take into account specific process parameters and adaptability characteristics, thereby maximizing the lithography process effect of the entire pattern.

[0070] More specifically, a generative adversarial network (GAN) is a deep learning model that can simulate and generate new data. In this application, GAN is used to simulate the lithography effects under different strategies. GAN consists of two parts: a generator (generating lithography effects) and a discriminator (evaluating the quality of lithography effects). Through training, the generator can generate lithography patterns close to reality, while the discriminator can help optimize the generator's effect.

[0071] More specifically, for each divided sub-region, a GAN model is used to simulate and analyze the lithography effects under different lithography process strategies. These simulations will provide feedback for subsequent process adjustments. The generated lithography effect simulations include key indicators such as pattern fidelity, defect rate, and exposure latitude, helping to verify the lithography process adaptability of each sub-region.

[0072] More specifically, based on the lithography effect simulated by GAN, a preset strategy library is constructed for each sub-area. This strategy library contains the optimal processing strategy for each sub-area under different lithography process parameters, such as exposure time, photoresist selection, exposure wavelength, depth of focus and other parameters. For each sub-area, the preset strategy library will provide a strategy selection for each possible process condition, thereby ensuring that the area can obtain the best process adaptability and lithography effect under different lithography conditions. The preset strategy library is not static, it will be dynamically updated according to the new lithography effect simulation and process data. In the actual production process, if new lithography requirements or changes in conditions arise, the strategy library can be continuously optimized and adjusted. The divided sub-areas and their corresponding preset strategy libraries support parallel processing of lithography graphics. Through parallel processing, different sub-areas can be optimized at the same time to improve the efficiency of lithography production.

[0073] It can be understood that through the combination of PPO reinforcement learning algorithm and GAN simulation, the lithography pattern area can be accurately divided, and the optimal process parameters can be selected according to the adaptability characteristics of different areas, thereby significantly improving the accuracy and adaptability of the lithography process. The lithography effect of each sub-area can be guaranteed through simulation and optimization of the strategy library. After dividing the lithography pattern into multiple sub-areas, parallel processing of multiple sub-areas can be achieved. This parallel processing improves production efficiency and shortens the production cycle. The independence of each sub-area ensures that the adjustments of different processes will not interfere with each other.

[0074] More specifically, by simulating the lithography effects under different strategies through generative adversarial networks (GANs), the performance of the lithography process under different conditions can be predicted in advance, and the strategy can be adjusted according to the simulation results, thereby effectively avoiding errors or defects that may occur in actual production. The PPO reinforcement learning algorithm continuously optimizes the area division strategy and can automatically adjust the process division plan according to actual needs and adaptability characteristics, reduce manual intervention, and realize intelligent lithography process optimization.

[0075] More specifically, as process conditions change, the dynamic updating of the strategy library ensures that the lithography process remains efficient and precise. Even when new pattern designs or lithography requirements change, the system can quickly adjust strategies to ensure production stability. This method can flexibly divide the lithography pattern according to its geometric characteristics and process adaptability requirements, providing customized process strategies. This makes production more flexible, meets the needs of different pattern types, and improves the overall efficiency of the production line.

[0076] Preferably, the preset reward function includes a sub-region internal adaptability consistency reward item and a sub-region boundary process effect interference penalty item.

[0077] Specifically, the geometric topological features within each sub-region need to be highly consistent with the adaptability of the lithography process. The adaptability consistency within the sub-region means that the lithography effect within the region will not have significant differences due to changes in process parameters (such as exposure time, photoresist type, exposure intensity, etc.). In order to quantify the adaptability consistency of the sub-region, the consistency can be measured by calculating the difference between the lithography effect of each point in the sub-region and the expected effect (for example, line width error, defect rate, etc.).

[0078] More specifically, the calculation method of the adaptability consistency bonus item usually involves evaluating the adaptability characteristics and process effects within each sub-region. The adaptability model of the lithography process (such as exposure tolerance, graphic fidelity, etc.) can be used to score each part within the region, and finally the adaptability consistency score of the entire sub-region is obtained.

[0079] More specifically, for each sub-region, the value of the reward item is proportional to the consistency of the process effect in the region, that is, if the adaptability difference within the sub-region is small, the reward is higher; if the difference is large, the reward is lower.

[0080] More specifically, during the training process of the PPO algorithm, the goal is to make the adaptability consistency of each sub-region as high as possible by adjusting the region division. Therefore, by adding the adaptability consistency reward item, the PPO algorithm can be guided to optimize the region division strategy so that the internal process adaptability of each sub-region is as uniform as possible.

[0081] More specifically, the sub-region division of the lithographic pattern involves the boundaries between different regions. The regions on the boundaries may interfere with each other. Especially in the lithographic process, the existence of boundaries may lead to inconsistent lithographic effects or defects. For example, if the exposure time, photoresist thickness, exposure intensity and other parameters between adjacent regions are not coordinated well, it may lead to obvious differences in the lithographic effects at the boundaries.

[0082] More specifically, the calculation of the boundary process interference penalty is achieved by evaluating the difference in lithography effects between adjacent sub-areas. In particular, the difference in process parameters near the boundary may have a serious impact on the lithography results. The penalty is calculated based on the degree of process interference between adjacent sub-areas. The greater the interference, the higher the value of the penalty. In actual calculations, the measurement of boundary interference can be based on the following factors: difference in exposure intensity, difference in photoresist type or thickness, difference in exposure time, difference in depth of focus or lithography accuracy. For each boundary, the PPO algorithm scores according to these interference factors and penalizes the reward function according to the degree of interference.

[0083] More specifically, during the PPO training process, by penalizing interference on the boundaries, the PPO algorithm will adjust the region division strategy to minimize the process differences between adjacent regions. This can make the lithography effects between adjacent regions more coordinated and avoid lithography defects or reduced accuracy caused by boundary interference.

[0084] Preferably, the steps of performing multi-objective collaborative processing simulation on the lithographic pattern according to the division status of each sub-region and the configured preset strategy library, adjusting the boundary division form of the sub-region and confirming the regional processing strategy, and finally generating the collaborative processing strategy of the lithographic pattern include: S41: constructing an optical proximity effect model and a photoresist flow model for each sub-region of the photoresist pattern according to each sub-region divided by the photoresist pattern and a corresponding preset strategy library, so as to simulate the distortion risk and photoresist residue risk of the edge portion of each sub-region under various preset strategies; S42: The edge of each sub-region is used as an intelligent analysis object to construct a game framework with distortion risk and photoresist residue risk as penalty evaluation mechanisms; S43: Iteratively solving the Pareto optimal boundary for the game understanding through a multi-objective optimization algorithm to adjust the boundary division form of the edge part of each sub-region, and deploying a corresponding regional processing strategy for each sub-region; S44: Summarizing the sub-regions whose boundary division forms have been adjusted and the corresponding regional processing strategies to generate a collaborative processing strategy for the lithography pattern.

[0085] Specifically, based on the overall layout of the photolithography pattern, it is first divided into multiple sub-areas. The division of each sub-area takes into account the photolithography process requirements, geometric characteristics, process parameters (such as exposure intensity, exposure time, photoresist thickness, etc.) and actual photolithography pattern characteristics. According to different photolithography requirements and goals, a preset strategy library is configured. Each strategy contains specific processing parameters and adjustment strategies under different photolithography conditions, such as exposure time, photoresist type and thickness, depth of focus, etc.

[0086] More specifically, during the lithography process, adjacent graphics may influence each other, resulting in distortion or distortion of the lithography effect. The optical proximity effect model can predict the mutual influence between different areas of the lithography graphics, especially the changes in the lithography effect in the boundary areas. For each sub-area, by constructing an optical proximity effect model, the impact of the area and the adjacent areas is evaluated, especially the graphic distortion that may occur at the edge of the sub-area.

[0087] More specifically, the photoresist flow model is used to simulate the flow behavior of photoresist during the exposure process, especially the distribution of photoresist at the edge. The flow of photoresist is affected by many factors, including exposure intensity, viscosity of photoresist, geometric characteristics of the pattern, etc. By establishing this model, it is possible to evaluate the residual photoresist at the edges of each sub-area and predict whether there is a risk of photoresist residue under different processing strategies.

[0088] More specifically, based on the optical proximity effect and photoresist flow model, the distortion risk and photoresist residue risk of the edge of each sub-area are simulated. The distortion risk reflects the difference between the graphic morphology and the expected graphic, and the photoresist residue risk refers to the situation where the photoresist cannot be completely removed due to uneven flow or insufficient exposure. A detailed risk assessment is performed on the edge of each sub-area to obtain the probability distribution of edge distortion and photoresist residue. Based on these evaluation results, a corresponding process adjustment strategy can be set for each sub-area.

[0089] More specifically, the edge parts of each sub-region are taken as the objects of intelligent analysis, and a game framework is constructed. In this framework, the boundary adjustment form of the sub-region and the selection of processing strategy will be affected by the penalty evaluation mechanism of distortion risk and photoresist residue risk. The core idea of this game framework is that the interaction between each sub-region and the boundary processing method are similar to the multi-party competition or cooperation in game theory. Each region hopes to minimize its risk and improve the quality of the entire photolithography pattern by adjusting its boundaries and strategies.

[0090] More specifically, a multi-objective optimization algorithm is used to iteratively solve the game. In each round of iteration, the strategy of each sub-region is updated through learning to gradually approach the Pareto optimal solution. The Pareto optimal solution means that the risk of a sub-region cannot be further reduced without increasing the risk of other sub-regions. The algorithm gradually optimizes the boundary division and processing strategy selection by alternately adjusting the boundary form and processing strategy of each sub-region, and finally generates an optimal regional processing plan.

[0091] More specifically, in the process of solving the game framework, the multi-objective optimization algorithm will output an optimal boundary division form, that is, the adjustment of the edge of each sub-region to ensure that the distortion risk and photoresist residue risk are minimized to the greatest extent. According to the optimal boundary division form, the corresponding regional processing strategy is deployed, including exposure parameters, photoresist thickness, exposure time, etc. The processing strategy of each sub-region will be fine-tuned according to its specific optical proximity effect and photoresist flow behavior to achieve high-quality results in the lithography process. Finally, by summarizing the boundary division and processing strategies of each sub-region, a complete collaborative processing strategy for lithography graphics is generated. This strategy will guide the actual processing process of the lithography graphics to achieve the optimal processing effect and quality.

[0092] It can be understood that through the optical proximity effect model, photoresist flow model and game framework, the entire photolithography pattern processing process is adjusted on the basis of multi-objective optimization. The boundary division and processing strategy of each sub-area can be optimized under multiple objectives (such as minimizing distortion, minimizing photoresist residue, etc.), thereby achieving higher photolithography pattern accuracy and quality. The application of multi-objective optimization algorithm makes the adjustment of boundary division and processing strategy intelligent. The algorithm can automatically adjust the boundary form and processing strategy of each sub-area under different process conditions, avoid human errors and improper adjustments, and improve the adaptability and efficiency of the photolithography process.

[0093] More specifically, by performing multi-objective collaborative processing simulation and optimization on the lithography patterns, the collaborative processing strategy finally generated can significantly improve the quality of the entire lithography pattern, reduce distortion and photoresist residue, and not only improve the pattern accuracy, but also reduce the defect rate and improve the production yield. The collaborative processing strategy of this method can be dynamically adjusted according to different lithography patterns and process requirements, thereby enhancing the flexibility and adaptability of the production process. Whether it is small-batch customized production or large-scale manufacturing, it can provide efficient production solutions. This method is not only suitable for specific lithography patterns, but also has certain versatility and scalability. By optimizing the strategy library and game framework, it can easily adapt to different types of lithography patterns and process requirements, and has broad application prospects.

[0094] Reference Figure 2 As shown, in a second aspect, the present invention provides a deep learning device for chip lithography pattern optimization, which is used to implement the deep learning method for chip lithography pattern optimization described in any one of the first aspects, comprising: A topology analysis module is used to obtain the original design information of the lithographic pattern and extract local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern; an adaptability analysis module, configured to perform adaptability analysis of the geometric topology feature information with respect to the lithography process, and to perform feature encoding on the geometric topology feature information based on the adaptability analysis result, so as to generate a lithography process adaptability feature matrix of the lithography pattern; A region division module is used to divide the lithographic pattern into a plurality of sub-regions that can be processed in parallel based on the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configure a corresponding preset strategy library for each sub-region; The collaborative simulation module is used to perform multi-target collaborative processing simulation on the lithography pattern according to the division status of each sub-region and the configured preset strategy library, so as to adjust the boundary division form of the sub-region and confirm the regional processing strategy, so as to finally generate the collaborative processing strategy of the lithography pattern.

[0095] In this embodiment, for the specific implementation of each module in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0096] In a third aspect, the present invention provides a deep learning device for chip lithography graphic optimization, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements a deep learning method for chip lithography graphic optimization as described in any one of the first aspects.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning method for chip lithography pattern optimization, characterized in that: include: Obtaining original design information of the lithographic pattern, and extracting local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern; Performing a photolithography process adaptability analysis on the geometric topology feature information, and performing feature encoding on the geometric topology feature information based on a result of the adaptability analysis to generate a photolithography process adaptability feature matrix of the photolithography pattern; Dividing the lithographic pattern into a plurality of sub-regions that can be processed in parallel according to the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configuring a corresponding preset strategy library for each of the sub-regions; According to the division status of each sub-region and the configured preset strategy library, a multi-target collaborative processing simulation is performed on the lithography pattern to adjust the boundary division form of the sub-region and confirm the regional processing strategy, so as to finally generate the collaborative processing strategy of the lithography pattern.

2. The deep learning method for chip lithography pattern optimization according to claim 1, characterized in that: The steps of obtaining original design information of the lithographic pattern and extracting local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern include: Obtaining original design information of a lithographic pattern, and performing layered vector graphics conversion on the original design information to generate a discrete point set corresponding to the original design information; Constructing an adjacency matrix based on the discrete point set, and performing key area marking on the discrete point set by parsing the adjacency matrix; A pre-trained dynamic sparse convolutional network model is used to perform preliminary image curvature and density calculations on discrete point sets labeled in key areas, generating corresponding adaptive sparse codes. The dynamic sparse convolutional network model is configured with a 3X3 convolution kernel or a 5X5 convolution kernel according to the adaptive sparse coding, and the sparse pattern parameters of the dynamic sparse convolutional network model are synchronously and dynamically adjusted to extract graphic features of the original design information and obtain the geometric topological connection matrix of the lithography pattern as the geometric topological feature information of the lithography pattern.

3. The deep learning method for chip lithography pattern optimization according to claim 1, characterized in that: The step of performing adaptability analysis of the lithography process on the geometric topological feature information includes: Retrieving from a photolithography process database several photolithography processes consisting of exposure wavelength, numerical aperture, photoresist type, and development condition elements; Simulating a lithographic pattern production process of various lithographic processes for the geometric topological feature information using a deep learning model pre-trained using lithographic process data to generate simulated production information of the various lithographic processes relative to the lithographic pattern; Conduct multi-dimensional adaptability evaluation of simulated production information for various photolithography processes to obtain fidelity characteristics consisting of pattern edge roughness and critical dimension uniformity, manufacturability characteristics consisting of minimum resolvable line width, exposure latitude, and depth of focus, and potential defect characteristics consisting of short circuit risk and photoresist residue probability. The Monte Carlo method is used to analyze the graphic impact caused by process fluctuations in the lithography process to generate sensitivity characteristics of the lithography process. The fidelity characteristics, manufacturability characteristics, potential defect characteristics and sensitivity characteristics of the lithography process are collectively used as adaptability characteristics of the lithography process.

4. The deep learning method for chip lithography pattern optimization according to claim 1, characterized in that: The step of encoding the geometric topological feature information based on the result of the adaptability analysis to generate a lithography process adaptation feature matrix of the lithography pattern includes: Binding various process parameters in the photolithography process with corresponding adaptability characteristics to form a process-adaptability mapping table; The geometric topological feature information is used as a basic map, and according to each lithography process in the mapping table, the basic map is sequentially analyzed for process-adaptability mapping factors at specific locations, so as to obtain a set of process-adaptability mapping factors corresponding to the basic map for each lithography process; wherein the set of process-adaptability mapping factors includes each process-adaptability mapping factor of a lithography process with respect to the basic map, and the process-adaptability mapping factor is used to describe the intrinsic mapping relationship between the lithography parameters and the adaptability of the lithography pattern and the lithography process; Information parsing and feature encoding are performed on the process-adaptability mapping factor sets of various lithography processes under interactive verification to generate a lithography process adaptation feature matrix composed of a set of lithography process mapping feature vectors at specific locations of the basic map.

5. The deep learning method for chip lithography pattern optimization according to claim 1, characterized in that: The steps of dividing the lithographic pattern into a plurality of sub-regions that can be processed in parallel according to the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configuring a corresponding preset strategy library for each sub-region include: Performing an initial round of region division processing on the geometric topological feature information according to the photolithography process adaptation feature matrix based on the PPO reinforcement learning algorithm, so as to divide a plurality of photolithography process regions based on the geometric topological feature information; Evaluating the region division result according to a preset region division reward function to supervise the PPO reinforcement learning algorithm to cyclically optimize the region division, thereby dividing the lithographic pattern into a plurality of sub-regions that can be processed in parallel; The lithography effect of each sub-region under different strategies is simulated by generating an adversarial network model according to the lithography process adaptation feature matrix, so as to configure a corresponding preset strategy library for each sub-region.

6. The deep learning method for chip lithography pattern optimization according to claim 5, characterized in that: The preset reward function includes a sub-region internal adaptability consistency reward item and a sub-region boundary process effect interference penalty item.

7. The deep learning method for chip lithography pattern optimization according to claim 1, characterized in that: The steps of performing multi-objective collaborative processing simulation on the lithographic pattern according to the division status of each sub-region and the configured preset strategy library to adjust the boundary division form of the sub-region and confirm the regional processing strategy to finally generate the collaborative processing strategy of the lithographic pattern include: Based on the sub-regions divided by the photolithographic pattern and the corresponding preset strategy library, an optical proximity effect model and a photoresist flow model are constructed for each sub-region of the photolithographic pattern to simulate the distortion risk and photoresist residue risk of the edge portion of each sub-region under various preset strategies; The edge of each sub-region is used as an intelligent analysis object to build a game framework with distortion risk and photoresist residue risk as the penalty evaluation mechanism; The game framework is iteratively solved for the Pareto optimal boundary through a multi-objective optimization algorithm to adjust the boundary division form of the edge part of each sub-region, and a corresponding regional processing strategy is deployed for each sub-region; The sub-regions adjusted in terms of boundary division and the corresponding regional processing strategies are summarized to generate a collaborative processing strategy for the lithography pattern.

8. A deep learning device for chip lithography pattern optimization, characterized in that: A deep learning method for implementing the chip lithography pattern optimization according to any one of claims 1 to 7, comprising: A topology analysis module is used to obtain the original design information of the lithographic pattern and extract local pattern features from the original design information using a pre-trained dynamic sparse convolutional network model to obtain geometric topological feature information of the lithographic pattern; an adaptability analysis module, configured to perform adaptability analysis of the geometric topology feature information with respect to the lithography process, and to perform feature encoding on the geometric topology feature information based on the adaptability analysis result, so as to generate a lithography process adaptability feature matrix of the lithography pattern; A region division module is used to divide the lithographic pattern into a plurality of sub-regions that can be processed in parallel based on the geometric topological feature information and the lithographic process adaptation feature matrix based on the PPO reinforcement learning algorithm, and configure a corresponding preset strategy library for each sub-region; The collaborative simulation module is used to perform multi-target collaborative processing simulation on the lithography pattern according to the division status of each sub-region and the configured preset strategy library, so as to adjust the boundary division form of the sub-region and confirm the regional processing strategy, so as to finally generate the collaborative processing strategy of the lithography pattern.

9. A deep learning device for chip lithography pattern optimization, comprising a memory and a processor, wherein the memory stores a deep learning program for chip lithography pattern optimization that can be run on the processor, characterized in that: When the processor executes the deep learning program for chip lithography pattern optimization, it implements the deep learning method for chip lithography pattern optimization described in any one of claims 1 to 7.

Images