Research map / Multi-agent systems and games / LLM agents
AutoMS: Multi-Agent Evolutionary Search for Cross-Physics Inverse Microstructure Design
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 AutoMS, a multi-agent LLM framework that wraps the MIND microstructure generator in an evolutionary search (SAES) with simulation-based local gradient approximation, reaching 83.8% success over 17 cross-physics inverse design tasks.
Problem
Inverse design of microstructures under coupled cross-physics objectives, where topology optimization is expensive and deep generative models produce physically invalid (hallucinated) designs.
Approach
Reformulates inverse microstructure design as LLM-driven evolutionary search: a hierarchical multi-agent orchestration layer (Manager, Parser, Generator, Simulator, Reporter) translates intent and coordinates a numerical SAES layer that estimates local gradients via weighted least squares from simulation history, updates MIND's conditioning parameters, and uses Pareto selection with adaptive weigh
Key results
- AutoMS achieves 83.8% success rate on 17 tasks vs 43.7% for NSGA-II, 66.7% for CMA-ES, 65.0% for BayesOpt and 53.3% for ReAct.
- AutoMS reports MRE of 0.0140 versus 0.4276 for NSGA-II and BPM of 94.2%.
- AutoMS reduces wall-clock time by 23.3% relative to ReAct (2180.5 s vs 2843.9 s average).
Surprising findings
- Removing SAES raises constraint satisfaction rate to 76.2%, higher than the full system's 59.6%, while lowering full success rate, which the authors attribute to a partial-satisfaction trap.
- The one-shot MIND baseline reaches 35.3% success, with 0% on several hard tasks, indicating the generator alone is insufficient.
Limitations
- Reliance on iterative trial-and-error simulation calls, which the authors note as a cost to reduce via explicit physical laws.
- Remaining failures on coupled tri-field and narrow feasible-region tasks.
- Evaluation uses a fixed benchmark designed by the authors and a single base generator (MIND).
Experimental setup
Datasets: MIND training dataset (over 130,000 pre-computed microstructures)
Benchmarks: 17-task cross-physics inverse microstructure design benchmark
Metrics: success rate constraint satisfaction rate mean relative error best property match quality score iterations wall-clock time
Baselines: NSGA-II ReAct BayesOpt CMA-ES Single-Agent + SAES CoT + SAES One-shot MIND
Builds on: MIND NSGA-II GraphRAG ReAct CMA-ES
Closest papers by meaning
- MACS: Multi-Agent Reinforcement Learning for Optimization of Crystal Structures (NeurIPS 2025 · similarity 0.85)
- Co-PiLOT: Constrained Physics-Informed Latent Optimization for Target-Driven Inverse Design (NeurIPS 2026 · similarity 0.85)
- LLM and Simulation as Bilevel Optimizers: A New Paradigm to Advance Physical Scientific Discovery (ICML 2024 · similarity 0.84)
- 3DID: Direct 3D Inverse Design for Aerodynamics with Physics-Aware Optimization (NeurIPS 2025 · similarity 0.84)
- LEIA: Learned Environment for Interactive Architected Materials (NeurIPS 2026 · similarity 0.84)
- Debate2Create: Robot Co-design via Multi-Agent LLM Debate (ICML 2026 · similarity 0.83)
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.