A Subresolution Auxiliary Feature Generation Method Based on Large Inference Model

Abstract

This invention discloses a sub-resolution auxiliary feature generation method based on a large inference model. It first generates a seed set from the input target layout, then expands it to obtain a candidate SRAF set, ultimately outputting high-quality SRAF results. By explicitly generating a chain to learn the intrinsic logic of the SRAF, it improves the ability to comply with constraints and the robustness of generation. The output full-process trajectory can trace any final SRAF back to its corresponding seed and candidate, facilitating the location of violations and parameter tuning. Furthermore, seed de-clustering simplifies and reduces sequence length and inference overhead. Combined with reinforcement learning imaging rewards, it further reduces EPE and PVB under compliance conditions. This invention can be used to enhance mask resolution for different layout layers, such as metal layers and via layers, improving imaging quality and expanding the lithography process window.

Description

Technical Field

[0001] This invention relates to the fields of computer-aided design and computational lithography for integrated circuits, and particularly to a sub-resolution auxiliary feature generation method based on a large inference model. Background Technology

[0002] As integrated circuit process nodes continue to shrink, the printability and process margin of target patterns in mask layouts are constantly decreasing. Subresolution auxiliary features (SRAF), as a resolution enhancement technique, improve the aerial image and enhance resistance to process variations by inserting tiny auxiliary structures around the target pattern that do not image on the wafer, thereby reducing edge offset and suppressing mismatch caused by process variations.

[0003] Existing SRAF generation methods mainly include rule-based methods, model-based methods, and machine learning-based methods. Rule-based methods rely on engineering experience and rule tables, generating SRAFs through geometric construction and rule verification. For example, Chinese invention patent application CN116243553A discloses a method for generating curve SRAFs, verifying MRC, and manufacturing masks. However, these methods have limited accuracy and flexibility. Model-based methods typically combine lithography simulation and rule checking, offering higher accuracy but consuming significant computational resources and having long runtimes. Machine learning-based methods can improve generation speed, as exemplified by Chinese invention patent application CN120652731A, which discloses a lithography mask method and apparatus. However, existing end-to-end prediction schemes often only learn the "input-result" mapping, ignoring the inherent "seed-candidate-selection" logic in the SRAF generation process. This results in insufficient robustness to process / rule constraints, and a lack of traceable process information when spacing or size violations occur, making it difficult to locate failure modes and perform engineering debugging.

[0004] Therefore, there is a need for an automatic SRAF generation method that can explicitly model the SRAF generation process, output interpretable intermediate decisions, and simultaneously take into account rule constraints and image quality, so as to improve the usability and debuggability of industrial applications. Summary of the Invention

[0005] The problem this invention aims to solve is to provide a sub-resolution auxiliary feature generation method based on a large inference model. This method encodes the SRAF workflow into a structured generation chain and aligns it through supervised fine-tuning and reinforcement learning, enabling the model to infer in stages and output explicit generation trajectories. This optimizes imaging quality and improves generation accuracy, robustness, and interpretability while satisfying mask rule checking (MRC).

[0006] This invention adopts the following technical solution: a sub-resolution auxiliary feature generation method based on a large inference model, comprising a training phase and an inference phase, the specific steps of which are as follows:

[0007] Training phase:

[0008] Step 1: Obtain the target graphic layout data and photolithography process conditions, and convert the layout polygons into a serialized representation as model input;

[0009] Step 2, Simplify the seed: Generate an SRAF seed set based on the aerial image intensity gradient within the insertable region, and simplify the seed by declustering to obtain a simplified seed set;

[0010] Step 3, Candidate Regions: Starting with the simplified seed, generate strip-shaped candidate SRAF regions along the tangential / normal directions of the target boundary, and calculate the weights and rankings for the candidate regions;

[0011] Step 4, Final SRAF: Perform mask rule checks on the candidate regions sequentially to obtain the candidate set that satisfies the constraints, and determine the final SRAF set based on the candidate set;

[0012] Step 5: Encapsulate the entire process of simplifying the seed-candidate region-final SRAF into a structured generation chain and output it in a fixed label format;

[0013] Step 6: Based on the generated chain data, perform supervised fine-tuning on the basic large language model to enable it to learn the generation logic and output format of staged reasoning;

[0014] Step 7: During the training phase, a reinforcement learning method based on group relative policy optimization is used to perform alignment training on the supervised fine-tuned large language model. A reward function is constructed using format, process consistency, rule constraints, and imaging quality to improve generation performance.

[0015] Reasoning stage:

[0016] During the inference phase, the target graphic layout and prompt information are input, and the complete generation chain and the final SRAF polygon set are output for subsequent OPC and lithography simulation.

[0017] As a preferred embodiment, in step 5, the structured generation chain is encapsulated and output with a fixed tag format, as follows:

[0018] The simplified seed, candidate region, and final SRAF are each placed in a predefined label, represented as: <seeds> …< / seeds> " <candidates> …< / candidates> " <srafs> …< / srafs> ";in, <seeds> …< / seeds> Indicates the range of SRAF seeds. <candidates> …< / candidates> Indicates the range of SRAF candidates. <srafs> …< / srafs>This indicates the range of the final generated SRAF.

[0019] Each stage of the polygon is represented by a sequence of vertex coordinates, and the decision trajectory of the entire process is explicitly preserved in the output.

[0020] As a preferred embodiment, in step 6, the supervised fine-tuning involves constructing prompt words that include SRAF generation logic and rules. , map sequence With tips As input, a chain will be generated. and the final SRAF Used as a supervisory signal for fine-tuning training.

[0021] The goal of oversight and fine-tuning is to maximize the following objectives:

[0022] ;

[0023] in, This represents the model's predicted probability for word segmentation. Indicates model parameters, These are prompt words that contain the SRAF generation logic and rules. Represents a sequence of map layouts. Indicates the SRAF generation chain, This indicates the final generated SRAF.

[0024] As a preferred embodiment, in step 7, reinforcement learning alignment is performed, sampling multiple generative chains from the same map sample. Group them together and use the GRPO update strategy;

[0025] The reward function includes at least: format reward, used to detect correct label structure; process reward, used to determine whether the final SRAF comes from a candidate or a seed; rule reward, used to determine that the MRC is free of violations; and imaging reward, used to perform lithography simulation scoring based on EPE and PVB indicators.

[0026] Furthermore, the purpose of the reinforcement learning is to maximize the following objective:

[0027] ;

[0028] in, and For hyperparameters, Represents the reference model. This represents the ratio of the probability predictions for word segmentation made by the old and new models. Indicates the current model, Denotes KL divergence, This indicates the relative strength of each SRAF result within its group. This indicates the number of results generated in each group. This indicates the training objective of reinforcement learning. This represents the clipping function.

[0029] Among them, the ratio of the probability predictions for word segmentation by the old and new models. The calculation is as follows:

[0030] ;

[0031] in, Indicates the generated first One SRAF generation chain, ; Indicates a time step. This represents the model before the update.

[0032] Furthermore, a set of reward values ​​is assigned to the output results within each group. ,pass This reflects the relative strength of each answer within its group, resulting in a standardized score that guides the model to select the most competitive output.

[0033] ;

[0034] in, Indicates the generated first The reward for each SRAF generating chain The function represents averaging the rewards within a group. The function represents the standard deviation of the group's rewards.

[0035] The present invention also provides: an electronic device, comprising:

[0036] One or more processors;

[0037] A storage device on which one or more programs are stored;

[0038] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the sub-resolution auxiliary feature generation methods described above.

[0039] The present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in any of the above-described sub-resolution auxiliary feature generation methods.

[0040] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0041] 1. The method of this invention can be used to enhance the mask resolution of different printing plates such as metal layers and via layers. By explicitly generating chains to learn the internal logic of SRAF, it improves the ability to comply with constraints and the robustness of generation. The output full-process trajectory can trace any final SRAF back to its corresponding seed and candidate, which is convenient for locating the cause of violation and adjusting parameters. By simplifying seed declustering, the sequence length is reduced and the inference overhead is reduced. Combined with the imaging reward of reinforcement learning, EPE and PVB are further reduced and the process window is improved under the premise of compliance.

[0042] 2. The generation method of this invention uses a combination strategy of "process coding + supervised fine-tuning + enhanced alignment" to transform SRAF generation from a black-box prediction that only outputs results into an interpretable, traceable, and constraint-aware staged reasoning process, which is suitable for high-quality automatic generation of SRAFs at different levels of the map. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the model supervision and fine-tuning process of the present invention;

[0044] Figure 2 This is a schematic diagram of the reinforcement learning process of the model of the present invention;

[0045] Figure 3 This is an illustration of the reinforcement learning reward system of the present invention;

[0046] Figure 4 The diagram shows the experimental results of the present invention on the metal layer and the via layer. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0048] In one embodiment of the present invention, the SRAF generation method based on the large inference model of the present invention, combined with manufacturability and imaging quality evaluation, achieves adaptive optimization. The training process of the model is as follows: Figure 1 and Figure 2 As shown, the monitoring and fine-tuning process is as follows: Figure 1 As shown, the reinforcement learning process is as follows: Figure 2 As shown. Specifically, it includes the following steps:

[0049] Step 1: Obtain the target graphic layout data and photolithography process conditions, and convert the GDS layout target polygon into a serialized representation. , , as model input;

[0050] Step 2: Generate an SRAF seed set based on the aerial image intensity gradient within the insertable region, and simplify the seeds by declustering to obtain a simplified seed set;

[0051] Step 3: Starting with the simplified seed, generate strip-shaped candidate SRAF regions along the tangential / normal directions of the target boundary, and calculate the weights and sorting of the candidate regions;

[0052] Step 4: Perform mask rule checks on the candidate regions sequentially to obtain the candidate set that meets the constraints, and determine the final SRAF set based on the candidate set;

[0053] Step 5: Use a word segmenter to perform word segmentation and encoding on each stage to obtain the reference result of the model. , …indicates the word segmentation result;

[0054] Step 6: Serialize the GDS layout , The design prompts and system prompts, together, are segmented and encoded by the word segmenter and then input into the basic large language model for forward propagation. The model selects the word segment with the highest probability as the output. , …;

[0055] Step 7: Compare the reference results with the model output to calculate the loss, and use backpropagation to update the model parameters. After multiple parameter iterations, a supervised fine-tuned large language model can be obtained.

[0056] Step 8: For the large language model after supervised fine-tuning, a reinforcement learning strategy based on Group Relative Policy Optimization (GRPO) is used for further training: the model samples a set of output results for each input, resulting in SRAF generation chain 1, SRAF generation chain 2, etc. The result of each sample is calculated according to the reward model, and after group normalization, the intra-group advantage of each result can be obtained, and the model is updated accordingly.

[0057] Specifically, steps 2 to 4 can be implemented using EDA software, but the resulting SRAF seed is quite complex, so the seed needs to be further simplified in step 2.

[0058] In this embodiment, blue noise sampling is used to disperse the sticky seeds into independent unit squares. This greatly reduces the training cost of the model while maintaining the lithographic semantic information, and at the same time maintains the data balance of the seed, candidate and final SRAF stages.

[0059] ;

[0060] in, This represents the i-th blue noise sampling point. Represents a set of blue noise. Represents the original set of seeds. This represents the simplified seed set.

[0061] Specifically, the implementation process of steps 5 to 7 includes:

[0062] 1) Generation chain and output format: The simplified seed, candidate region, and final SRAF are placed in predefined labels, such as " <seeds> …< / seeds> " <candidates> …< / candidates> " <srafs> …< / srafs> ";in, <seeds> …< / seeds> Indicates the range of SRAF seeds. <candidates> …< / candidates> Indicates the range of SRAF candidates. <srafs> …< / srafs> This indicates the range of the final generated SRAF. The polygons at each stage are represented by a sequence of vertex coordinates, thus explicitly preserving the decision trajectory of the entire process in the output, facilitating engineering debugging.

[0063] 2) Supervised Fine-tuning: Construct a cue word P containing SRAF generation logic and rules, taking the layout sequence X and the cue word P as input, and generating the chain... and the final SRAF This serves as a supervisory signal for fine-tuning training, enabling the model to learn not only the output results but also stage-by-stage reasoning and intermediate decisions.

[0064] In this embodiment, the objective of supervised fine-tuning is to maximize the following objectives:

[0065] ;

[0066] in, This represents the model's predicted probability for word segmentation. Indicates model parameters, These are prompt words that contain the SRAF generation logic and rules. Represents a sequence of map layouts. Indicates the SRAF generation chain, This indicates the final generated SRAF.

[0067] 3) Reinforcement learning alignment: such as Figure 3As shown, multiple generation chains are sampled from the same GDS map sample to form a group, and the GRPO update strategy is adopted based on the output sequence. The reward model includes: format reward (to check the format and confirm that the label structure is correct), process reward (to check the process, and the final SRAF comes from the candidate and the candidate comes from the seed), rule reward (to check the mask rule and confirm that the MRC is free of violations), and lithography reward (based on lithography simulation scoring based on indicators such as EPE and PVB, so as to optimize the imaging quality while ensuring compliance).

[0068] Specifically, in Figure 3 The document provides four scenarios to illustrate the scoring of the reward model in this embodiment: Figure 3 (a) indicates that the output result has a formatting error, in which case a reward of -5 points will be given; Figure 3 (b) indicates that there is a process error in the output result, in which case a reward of -3 points will be given; Figure 3 (c) indicates that there is an error in the output result, in which case a reward of -3 points is given; Figure 3 (d) in the text indicates that the output does not contain any of the above errors, and a reward is given based on its lithography result.

[0069] Furthermore, the fabrication process on the metal layer and via layer in this embodiment is as follows: Figure 4 As shown: a seed set is first generated for the input target pattern (via or metal layer), and then expanded to obtain a candidate SRAF set, finally outputting a high-quality SRAF result; this process makes the model output traceable, easy to debug, and can remain stable under different pattern density conditions.

[0070] In particular, blue noise sampling is used to disperse the cohesive seeds into independent unit squares, which greatly reduces the training cost of the model while maintaining the lithographic semantic information, and at the same time maintains the data balance of the seed, candidate and final SRAF stages.

[0071] Furthermore, the method of the present invention was tested on the ICCAD2013 and ICCAD 2020 datasets to verify the performance of the present invention in the metal layer and via layer.

[0072] The experimental framework of this invention was implemented using PyTorch, with training and inference performed on a single NVIDIA H800 GPU. Training data for the via layer was sourced from ICCAD 2020, and training data for the metal layer was sourced from a publicly available metal layer dataset. Furthermore, over 100,000 SRAF generation chains were generated using EDA software according to predetermined rules for training. Comparisons with current state-of-the-art results on the metal layer and via layer are shown in Tables 1 and 2 below.

[0073] Among them, SL-SRAF is a method for inserting SRAF on a two-dimensional mesh plane using machine learning; SODL-SRAF is a method for inserting SRAF based on supervised dictionary learning; RL-SRAF is a method for inserting SRAF based on reinforcement learning; CTM-SRAF is a method for inserting SRAF based on a continuous transmission mask; LLM-SRAF is a method for inserting SRAF based on a large language model; Rule-based SRAF is a rule-based SRAF insertion method; and Ridge-based SRAF is a method for inserting SRAF based on the contour lines of a continuous transmission mask.

[0074] Table 1: Comparison with current state-of-the-art results on through-hole layers

[0075]

[0076] Table 2: Comparison with the current best results on the metal layer

[0077]

[0078] Experimental results show that, compared with existing end-to-end SRAF prediction methods, the method of this invention achieves state-of-the-art performance on both the metal layer and the via layer. For testing, the via layer uses an eight-group dense contact array and a ten-group sparse contact array benchmark consistent with existing work to evaluate the PVB (process variation band) area and EPE (edge ​​placement error) length; compared with state-of-the-art methods, the average PVB area is reduced by approximately 5.4% and the EPE by approximately 9.6%. The metal layer was evaluated on a 32 nm metal layer layout provided by the ICCAD 2013 CAD competition. Compared with state-of-the-art results, the average PVB area and the number of EPEs are both reduced by approximately 6.1%, demonstrating its adaptability to complex metal layer patterns.

[0079] In summary, this invention improves the ability to comply with constraints and the robustness of generation by explicitly generating chains to learn the internal logic of SRAFs; the output full-process trajectory can trace any final SRAF back to its corresponding seed and candidate, which is convenient for locating the cause of violation and adjusting parameters; the seed declustering simplifies and reduces sequence length and inference overhead; and the imaging reward of reinforcement learning further reduces EPE and PVB and improves the process window under the premise of compliance.

[0080] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A sub-resolution auxiliary feature generation method based on a large inference model, characterized in that, Includes the following steps: Training phase: Step 1: Obtain the target graphic layout data and photolithography process conditions, and convert the layout polygons into a serialized representation as model input; Step 2, Simplify the seed: Generate an SRAF seed set based on the aerial image intensity gradient within the insertable region, and simplify the seed by declustering to obtain a simplified seed set; Step 3, Candidate Regions: Starting with the simplified seed, generate strip-shaped candidate SRAF regions along the tangential or normal direction of the target boundary, and calculate the weights and sorting of the candidate regions; Step 4, Final SRAF: Perform mask rule checks on the candidate regions sequentially to obtain the candidate set that satisfies the constraints, and determine the final SRAF set based on the candidate set; Step 5: Encapsulate the entire process of simplifying the seed-candidate region-final SRAF into a structured generation chain and output it in a fixed label format; Step 6: Perform supervised fine-tuning on the basic large language model based on the generative chain data, so that it learns the generation logic and output format of staged reasoning; The aforementioned monitoring and fine-tuning constructs prompt words that include SRAF generation logic and rules. , map sequence With prompt words As input, SRAF generation chain and the final generated SRAF Used as a supervisory signal for fine-tuning training. The goal of oversight and fine-tuning is to maximize the following objectives: ； in, This represents the model's predicted probability for word segmentation. Indicates model parameters, Indicates the SRAF generation chain, This represents the final generated SRAF; Step 7: Use a reinforcement learning method based on group relative policy optimization to perform alignment training on the supervised fine-tuned large language model, and construct a reward function using format, process consistency, rule constraints and imaging quality to improve generation performance; Reinforcement learning alignment is performed by sampling multiple generative chains from the same map sample. Form a group and use the GRPO update strategy; The reward function includes at least: format reward, used to detect correct label structure; process reward, used to determine whether the final SRAF comes from a candidate or a seed; rule reward, used to determine that the MRC is free of violations; and imaging reward, used to perform lithography simulation scoring based on EPE and PVB indicators. Reasoning stage: Input the target graphic layout and prompt information, output the complete generation chain and provide the final SRAF set for subsequent OPC and lithography simulation.

2. The sub-resolution auxiliary feature generation method according to claim 1, characterized in that, In step 5, the structured generation chain is encapsulated and output with a fixed tag format, as follows: The simplified seed, candidate region, and final SRAF are each placed in a predefined label, represented as: <seeds> …< / seeds> "" <candidates> …< / candidates> "and" <srafs> …< / srafs> ";in, <seeds> …< / seeds> Indicates the range of SRAF seeds. <candidates> …< / candidates> Indicates the range of SRAF candidates. <srafs> …< / srafs> Indicates the range of the final generated SRAF; Each stage of the polygon is represented by a sequence of vertex coordinates, and the decision trajectory of the entire process is explicitly preserved in the output.

3. The sub-resolution auxiliary feature generation method according to claim 1, characterized in that, The purpose of reinforcement learning is to maximize the following objectives: ； in, and For hyperparameters, Represents the reference model. This represents the ratio of the probability predictions for word segmentation made by the old and new models: Indicates the current model, Denotes KL divergence, This indicates the relative strength of each SRAF result within its group. This indicates the number of results generated for each group. This indicates the training objective of reinforcement learning. This represents the clipping function.

4. The sub-resolution auxiliary feature generation method according to claim 3, characterized in that, The ratio of the probability predictions for word segmentation by the old and new models The calculation is as follows: ； in, Indicates the generated first One SRAF generation chain, ; Indicates a time step. This represents the model before the update.

5. The sub-resolution auxiliary feature generation method according to claim 4, characterized in that, Assign a set of reward values ​​to the output of each group. ,pass Reflecting the relative strength of each SRAF result within its group, a standardized score is generated to guide the model in selecting the most competitive output: ； in, Indicates the generated first The reward for each SRAF generating chain, The function represents averaging the rewards within a group. The function represents the standard deviation of the group's rewards.

6. The sub-resolution auxiliary feature generation method according to claim 1, characterized in that, The simplified seed is obtained by using blue noise sampling to disperse the clump-together seeds into independent unit squares, as follows: ； in, Indicates the first One blue noise sampling point, Represents a set of blue noise. Represents the original set of seeds. This represents the simplified seed set.

7. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the sub-resolution auxiliary feature generation method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the sub-resolution auxiliary feature generation method according to any one of claims 1 to 6.

Images