- The paper introduces a novel two-stage framework that combines reinforcement learning with large language models to optimize OPC recipes.
- It reduces edge placement error significantly by automating recipe development through RL exploration and decision tree summarization.
- The approach minimizes reliance on manual tuning and enhances scalability, marking a substantial step toward AI-driven semiconductor manufacturing.
Intelligent OPC Engineer Assistant for Semiconductor Manufacturing
The paper "Intelligent OPC Engineer Assistant for Semiconductor Manufacturing" explores the application of AI/LLM methodologies to optimize optical proximity correction (OPC) in semiconductor manufacturing. The authors, Guojin Chen, Haoyu Yang, Bei Yu, and Haoxing Ren, propose a novel two-stage framework that integrates reinforcement learning (RL) and LLMs to automate and enhance the efficiency of OPC recipe development. This research addresses a critical aspect of semiconductor manufacturing, where accurate pattern transfer onto wafers is essential for high-yield production of advanced integrated circuits.
Introduction
OPC is a computational technique used to counteract lithographic distortions caused by diffraction and interference during the photolithography process. Traditional OPC methods, which involve heuristic-based adjustments by experienced engineers, are time-consuming and less adaptable to new technology nodes. The authors aim to streamline this process by leveraging AI, specifically RL and LLMs, to automate the development of OPC recipes, thereby reducing the manual effort and expertise required.
Methodology
The methodology presented in the paper consists of two primary stages:
- Exploration with Reinforcement Learning:
- The first stage employs an RL-based framework to optimize OPC recipes. The RL agent explores the parameter space to identify optimal solutions for adjusting EPE measurement points and fragmentation points based on pattern characteristics. Proximal Policy Optimization (PPO) is utilized to maximize the expected cumulative reward, which is derived from the OPC loss function comprising Euclidean distance, edge placement error (EPE), and process variation band (PVB).
- Recipe Summarization with LLMs:
- In the second stage, LLMs are used to summarize the RL-generated OPC recipes. This involves transforming the RL outputs into a structured format, generating a feature pool using multi-modal large models (MLLMs), and annotating each point with relevant features. A decision tree is constructed from these annotated features and used to generate the final OPC recipes. This approach ensures that the insights gained from the RL optimization are translated into practical and reusable recipe rules.
Experimental Results
The authors evaluated their framework using datasets from two distinct processes: the ICCAD 2013 contest and the NVIDIA Deep Learning Accelerator (NVDLA). The results demonstrated that:
- For the ICCAD 2013 dataset:
- The RL optimization reduced the EPE count by 12% and the EPE distance by 24% compared to baseline OPC.
- The LLM-generated recipe showed comparable performance to RL with a reduction in EPE count by 11% and EPE distance by 19%, without significant runtime overhead.
- For the NVDLA dataset:
- The RL optimization reduced the EPE count by 13% and the EPE distance by 15%.
- The LLM-generated recipe achieved a reduction in EPE count by 8% and EPE distance by 12%, maintaining the same runtime as the baseline OPC engine.
Implications and Future Work
The integration of RL and LLMs for OPC recipe development has significant implications for both practical and theoretical advancements in semiconductor manufacturing:
- Practical Implications:
- The proposed framework reduces the dependency on human expertise and accelerates the OPC recipe development process. This is particularly beneficial for advanced technology nodes where the complexity of patterns and the number of layers increase.
- The automation of recipe generation through LLMs ensures scalability and adaptability to different OPC engines and process variations, potentially leading to broader adoption in the industry.
- Theoretical Implications:
- The use of RL and LLMs in OPC represents a novel application of AI in EDA (Electronic Design Automation) workflows. This approach could inspire further research into the integration of AI methodologies in other stages of semiconductor design and manufacturing processes.
- The decision tree framework and feature extraction methodologies developed in this paper could be applied to other optimization problems in computational lithography and beyond.
Looking ahead, future developments in AI could enhance the capabilities of this framework. For instance, advancements in multi-modal large models could improve feature extraction and annotation accuracy, while more sophisticated RL techniques could further optimize the exploration of the parameter space. Additionally, the framework's adaptability to various OPC engines and processes could be tested on a wider range of semiconductor manufacturing technologies, potentially leading to further refinements and broader applicability.
In conclusion, the Intelligent OPC Engineer Assistant represents a significant step forward in the automation and optimization of OPC recipe development. By effectively combining RL and LLMs, the authors have demonstrated a promising approach to addressing the challenges of modern semiconductor manufacturing, paving the way for more efficient and precise chip production.