Research map / Large language models / AI and society
Intelligent OPC Engineer Assistant for Semiconductor Manufacturing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an Intelligent OPC Engineer Assistant that combines RL-based OPC recipe search with a multimodal LLM agent pipeline (feature pool mining, zero-shot labeling, decision tree) to generate OPC recipes.
Problem
OPC recipe development (EPE measurement point and fragmentation rules) is manual, expert-driven, and time-consuming, and RL-based recipe search is too slow for full-chip use.
Approach
Use PPO to optimize EPE measurement and fragment points per pattern, then use GPT-4o to mine binary layout features from images, label each point zero-shot, build a decision tree on those features with RL-derived movement classes, and have the LLM write recipe rules from the tree, with iterative feature self-improvement.
Key results
- On ICCAD13, OPC+LLM reduces EPE N to 0.89x and EPE D to 0.81x of the pure OPC engine, with runtime unchanged at 4s.
- On NVDLA, OPC+LLM reduces EPE N by 8% and EPE D by 12% relative to the OPC engine, with runtime 6s.
- RL optimization reduces ICCAD13 EPE D by about 24% but takes about 3 hours per run.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply the LLM-generated recipe instead of the plain OPC recipe | OPC · ICCAD13 | PVBand: 53328 -> 51271 | Helped |
| Apply the LLM-generated recipe instead of the plain OPC recipe | OPC · ICCAD13 | EPE N: 119.70 -> 107.00 | Helped |
| Apply the LLM-generated recipe instead of the plain OPC recipe | OPC · ICCAD13 | EPE D: 693.10 -> 561.70 | Helped |
| Use the LLM-generated recipe rather than RL optimization | OPC · ICCAD13 | Runtime: 3hr -> 4s | Helped |
| Apply the LLM-generated recipe instead of the plain OPC recipe | OPC · NVDLA | PVBand: 170899 -> 161056 | Helped |
| Apply the LLM-generated recipe instead of the plain OPC recipe | OPC · NVDLA | EPE N: 159.45 -> 146.95 | Helped |
| Apply the LLM-generated recipe instead of the plain OPC recipe | OPC · NVDLA | EPE D: 869.25 -> 766.10 | Helped |
| Use the LLM-generated recipe rather than RL optimization | OPC · NVDLA | Runtime: 3hr -> 6s | Helped |
| Apply RL recipe optimization instead of the plain OPC recipe | OPC · ICCAD13 | EPE N: 119.70 -> 105.60 | Helped |
| Apply RL recipe optimization instead of the plain OPC recipe | OPC · ICCAD13 | EPE D: 693.10 -> 525.90 | Helped |
| Apply RL recipe optimization instead of the plain OPC recipe | OPC · NVDLA | EPE N: 159.45 -> 139.40 | Helped |
| Apply RL recipe optimization instead of the plain OPC recipe | OPC · NVDLA | EPE D: 869.25 -> 741.05 | Helped |
Surprising findings
- The LLM-generated rule-based recipe performs comparably to the RL optimization while adding no runtime over the base OPC engine.
Limitations
- RL is computationally intensive and impractical for full-chip application on its own.
- Decision tree average precision ranges 0.77 to 0.90, falling below 0.9 for complex patterns.
Experimental setup
Datasets: ICCAD 2013 CAD contest OPC benchmark NVDLA full-chip layout clips (NanGate 45nm)
Benchmarks: ICCAD13 NVDLA clips
Metrics: PVBand EPE N EPE D runtime
Baselines: pure OPC engine (open-source OPC engine) RL-optimized OPC (OPC+RL)
Builds on: proximal policy optimization GPT-4o decision tree
Closest papers by meaning
- See it to Place it: Evolving Macro Placements with Vision Language Models (NeurIPS 2026 · similarity 0.82)
- Retrieval-Guided Reinforcement Learning for Boolean Circuit Minimization (ICLR 2024 · similarity 0.81)
- A Hierarchical Adaptive Multi-Task Reinforcement Learning Framework for Multiplier Circuit Design (ICML 2024 · similarity 0.81)
- Towards Automated RISC-V Microarchitecture Design with Reinforcement Learning (AAAI 2024 · similarity 0.80)
- AUTOCIRCUIT-RL: Reinforcement Learning-Driven LLM for Automated Circuit Topology Generation (ICML 2025 · similarity 0.80)
- Reinforcement Learning within Tree Search for Fast Macro Placement (ICML 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.