CN-122021945-A - Sub-resolution auxiliary feature generation method based on large reasoning model

CN122021945ACN 122021945 ACN122021945 ACN 122021945ACN-122021945-A

Abstract

The invention discloses a sub-resolution auxiliary feature generation method based on a large reasoning model, and generating a seed set for the input target layout, expanding to obtain a candidate SRAF set, and finally outputting a high-quality SRAF result. The inherent logic of SRAF is learned through an explicit generation chain, the following capability of constraint and the generation robustness are improved, any final SRAF can be traced to corresponding seeds and candidates through the output overall process track, the violation reasons and parameter adjustment are conveniently positioned, the sequence length is simplified and the reasoning cost is reduced through seed cluster removal, and EPE and PVB are further reduced on the premise of compliance by combining the imaging rewards of reinforcement learning. The method can be used for enhancing the mask resolution of different layout layers such as a metal layer, a through hole layer and the like, improving the imaging quality and expanding the photoetching process window.

Inventors

Zou Ningmu
LI JIALE
CHEN SILIN
ZHANG YUFENG
XU ZHILI
WANG YIQING
Zang Yuqian
DU YUAN

Assignees

南京大学

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. The sub-resolution auxiliary feature generation method based on the large reasoning model is characterized by comprising the following steps of: Training phase: step 1, obtaining target graphic layout data and photoetching process conditions, and converting layout polygons into serialized representations as model input; Step 2, simplifying seeds, namely generating SRAF seed sets in an insertable area according to the aerial image intensity gradient, and carrying out cluster removal and simplification on the seeds to obtain simplified seed sets; Step 3, candidate areas are generated by taking simplified seeds as starting points, and a bar-shaped candidate SRAF area is generated along the tangential direction or the normal direction of the target boundary, and weights and sequences are calculated for the candidate areas; Step 4, final SRAF, namely sequentially executing mask rule check on the candidate areas to obtain candidate sets meeting constraint, and determining a final SRAF set based on the candidate sets; step 5, packaging the whole process of simplifying the seed-candidate region-final SRAF into a structured generation chain, and outputting the structured generation chain in a fixed tag format; Step 6, performing supervision fine tuning on the basic large language model based on the generation chain data to enable the basic large language model to learn the generation logic and output format of the phased reasoning; Step 7, performing alignment training on the supervised and fine-tuned large language model by adopting a reinforcement learning method based on group relative strategy optimization, and constructing a reward function by using format, process consistency, rule constraint and imaging quality, so as to improve the generation performance; Reasoning: and inputting a target graphic layout and prompt information, outputting a complete generation chain, and giving out a final SRAF set for subsequent OPC and lithography simulation.
2. The sub-resolution assist feature generating method according to claim 1, wherein in step 5, the structured generation chain is encapsulated and outputted in a fixed tag format, the method comprising: Placing the reduced seed, the candidate region, and the final SRAF in predefined labels, respectively, expressed as "< seeds >," < candidates >, "and" < SRAFs >, "wherein < seeds >, < candidates > indicates the range of SRAF seeds, < SRAFs >, < candidates > indicates the range of SRAF candidates, < SRAFs > indicates the range of final SRAF generated. And expressing the polygons of each stage by adopting a vertex coordinate sequence, and explicitly retaining the whole process decision track in output.
3. The sub-resolution assist feature generation method according to claim 1, wherein in step 6, the supervisory fine tuning constructs a hint word comprising SRAF generation logic and rules Sequencing the layout Prompt word As input, SRAF generation chain Finally generated SRAF As a supervisory signal for the fine tuning training, The goal of the supervised fine tuning is to maximize the following goals: ; Wherein, the Representing the predicted probability of the model to word segmentation, The parameters of the model are represented by the parameters, Representing the SRAF generation chain, Representing the final generated SRAF.
4. The sub-resolution assist feature generating method as recited in claim 1, wherein in said step 7, reinforcement learning alignment is performed to sample a plurality of generation chains for a same layout sample A group is formed, and GRPO updating strategies are adopted; The reward function at least comprises format rewards used for detecting the correct label structure, process rewards used for determining that the final SRAF is derived from the candidate and the candidate is derived from the seed, rule rewards used for determining that MRC is free of violations and imaging rewards, and photoetching simulation scoring based on EPE and PVB indexes is carried out.
5. The sub-resolution assist feature generating method as recited in claim 4, wherein the purpose of the reinforcement learning is to maximize the following objectives: ; Wherein, the And Is the parameter of the ultrasonic wave to be used as the ultrasonic wave, The reference model is represented by a reference model, The ratio of the probability predictions of new and old models for segmentation: Representing the current model of the model, Represents the degree of divergence of KL, Representing the relative dominance of each SRAF result within the group to which it belongs, Representing the number of results generated for each group, Representing the training objectives of reinforcement learning, Representing a clipping function.
6. The method of claim 5, wherein the ratio of the new model to the probability prediction of the segmented word The calculation is as follows: ; Wherein, the Representation generation of the first The number of SRAFs is a chain of generation, ; The time step is indicated as such, Representing the model before updating.
7. The method of claim 6, wherein a set of prize values is assigned to each output outcome within a group By means of Reflecting the relative advantages of each SRAF result in the belonging group to obtain a standardized score, and guiding the model to select the output result with the highest competitive power: ; Wherein, the Representation generation of the first The individual SRAFs generate a chain of rewards, The function represents an averaging of the rewards within the group, The function represents the standard deviation for the rewards in the group.
8. The sub-resolution assist feature generation method according to claim 1, wherein the simplified seed is obtained by dispersing the stuck seeds into unit squares independent of each other by blue noise sampling, by: ; Wherein, the Represent the first A number of blue noise sampling points are used, Representing the set of blue noise, Representing the original set of seeds and, Representing the reduced seed set.
9. An electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon; the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the sub-resolution assist feature generation method of any of claims 1 to 8.
10. A computer-readable storage medium, having stored thereon a computer program which, when executed by a processor, implements the steps of the sub-resolution assist feature generation method according to any of claims 1 to 8.

Description

Sub-resolution auxiliary feature generation method based on large reasoning model Technical Field The invention relates to the technical field of integrated circuit computer aided design and computational lithography, in particular to a sub-resolution auxiliary feature generation method based on a large reasoning model. Background As integrated circuit process nodes continue to shrink, the printability and process margin of target patterns in the mask layout continue to decrease. Sub-resolution assist features (SRAFs) are one of the resolution enhancement techniques that improve aerial images and increase process fluctuation resistance by inserting tiny assist features around the target pattern that do not image on the wafer, thereby reducing edge offset and suppressing mismatch from process variations. Existing SRAF generation methods mainly include rule-based methods, model-based methods, and machine learning-based methods. The method based on rules relies on engineering experience and rule table, and adopts geometric construction and rule verification to generate SRAF, for example, chinese patent application with publication No. CN116243553A discloses a method for generating curve SRAF, a method for verifying MRC and a method for manufacturing mask, which have limited precision and flexibility. The model-based method generally combines lithography simulation and rule inspection, and has higher precision, but large calculation resource consumption and long running time. The method based on machine learning can improve the generation speed, for example, the Chinese patent application with publication number of CN120652731A discloses a photoetching mask method and a photoetching mask device. However, the existing end-to-end prediction scheme only learns the "input-result" mapping, ignores the inherent logic of "seed-candidate-screening" in the SRAF generation process, causes insufficient robustness to process/rule constraints, and lacks traceable process information when pitch or size violations occur, and is difficult to locate failure modes and perform engineering debugging. Therefore, there is a need for an automatic SRAF generation method that can explicitly model the SRAF generation process, output interpretable intermediate decisions, and compromise both rule constraints and imaging quality, to improve industrial application availability and scalability. Disclosure of Invention The invention aims to solve the problem of providing a sub-resolution auxiliary feature generation method based on a large reasoning model, which encodes SRAF workflow into a structured generation chain, and enables the model to reasoning and output an explicit generation track according to stages by supervising fine adjustment and reinforcement learning alignment, so that the imaging quality is optimized on the premise of meeting Mask Rule Checking (MRC), and the generation precision, robustness and interpretability are improved. The invention adopts the following technical scheme that the sub-resolution auxiliary characteristic generation method based on the large reasoning model comprises a training stage and a reasoning stage, and comprises the following specific steps: Training phase: step 1, obtaining target graphic layout data and photoetching process conditions, and converting layout polygons into serialized representations as model input; Step 2, simplifying seeds, namely generating SRAF seed sets in an insertable area according to the aerial image intensity gradient, and carrying out cluster removal and simplification on the seeds to obtain simplified seed sets; Step 3, candidate areas are generated by taking simplified seeds as starting points along tangential/normal directions of target boundaries, and weights and sequences are calculated for the candidate areas; Step 4, final SRAF, namely sequentially executing mask rule check on the candidate areas to obtain candidate sets meeting constraint, and determining a final SRAF set based on the candidate sets; step 5, packaging the whole process of simplifying the seed-candidate region-final SRAF into a structured generation chain, and outputting the structured generation chain in a fixed tag format; step 6, performing supervision fine tuning on the basic large language model based on the generation chain data to enable the basic large language model to learn a generation logic and an output format of the phased reasoning; step 7, in the training stage, a reinforcement learning method based on group relative strategy optimization is adopted to conduct alignment training on the large language model subjected to supervision and fine adjustment, and a reward function is constructed by using the format, process consistency, rule constraint and imaging quality, so that the generation performance is improved; Reasoning: the reasoning stage inputs the target graphic layout and prompt information, outputs a complete generation chain and gives out a final SRAF polygon set for sub