AutoMS: Multi-Agent Evolutionary Search for Cross-Physics Inverse Microstructure Design¶

🕒 Published (v1): 2026-03-28 08:51 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

AutoMS is a multi-agent neuro-symbolic framework that reformulates inverse microstructure design—finding geometries with specific cross-physics properties—as an LLM-driven evolutionary search. It pairs a hierarchical orchestration layer (LLMs for semantic planning and role decomposition) with a numerical Simulation-Aware Evolutionary Search (SAES) engine that uses local gradient approximation over black-box simulation feedback. On 17 cross-physics benchmark tasks it achieves 83.8% success rate, nearly doubling NSGA-II and outperforming all agentic and numerical baselines.

Problem¶

Designing microstructures that simultaneously satisfy coupled cross-physics objectives (e.g., thermal conductivity + mechanical stiffness + electrical conductivity) is a non-differentiable, high-dimensional optimization problem. Topology optimization stalls in non-convex landscapes; deep generative models ("one-shot" mappings) suffer from physical hallucination—producing geometrically plausible but physically invalid structures—because they optimize statistical likelihood rather than physical validity. Naive single-agent LLM approaches fail due to open-loop reasoning with no simulation grounding.

Method¶

AutoMS adopts a two-layer hierarchical architecture:

Orchestration Layer — A Manager agent coordinates a pipeline: a Parser agent (backed by GraphRAG) resolves natural-language intent into quantifiable targets in the mechanical conditioning space $\mathcal{X}$; a Generator agent calls the pre-trained MIND model $G_{\text{MIND}}: \mathcal{X} \rightarrow \Omega$ to synthesize candidate geometries; a Simulator runs JAX-FEM finite-element analysis; a Reporter synthesizes outputs. The Manager gates physically infeasible candidates before they reach the expensive simulator.

Optimization Layer (SAES) — The formal objective is: $$x^* = \arg\max_{x \in \mathcal{X}} \mathbb{E}_{m \sim G_{\text{MIND}}(x)}[S(m)]$$ Since $S$ is a black-box oracle, SAES estimates a local gradient via Weighted Least Squares over a sliding window of $M$ historical evaluations, weighting neighbors by geometric proximity and temporal recency: $$w_j = \frac{1}{1+\|x_j - x_k\|^2} \cdot \exp\!\left(-\lambda_t \frac{N-t_j}{N}\right)$$ Parameter updates are gradient-guided with directional sign-correction: $$x' = x_k + \eta \cdot \text{sgn}(y_{\text{tgt}} - y_k) \cdot \frac{g_k}{\|g_k\|_2} + \beta\xi$$ Pareto-Driven Selection (NSGA-II non-dominated sorting + crowding distance) maintains population diversity. An Adaptive Weight Update detects stagnation and dynamically rebalances scalarization weights $w_j \leftarrow \text{clip}(w_j \cdot (1+\delta(\gamma_j)),\, 0.1,\, 2.0)$ to escape local optima on the Pareto frontier.

Key Contributions¶

AutoMS framework: first system to use LLM agents as semantic evolutionary operators for inverse design under coupled cross-physics constraints, operating end-to-end from natural language to validated geometry.
SAES algorithm: closed-loop numerical engine using local gradient approximation (WLS), gradient-guided parameter navigation, Pareto-driven elite archiving, and adaptive weight rebalancing—all without analytical gradients.
17-task benchmark: covering material diversity (Al6061, Ti6Al4V, ABS, TPU), 1–5 competing objectives, isolated to full tri-field physical coupling, and three difficulty tiers.
Efficiency via pruning: the Orchestration Layer acts as a computational gatekeeper, reducing wall-clock time 23.3% vs. ReAct by preventing invalid simulation calls.

Results¶

Success Rate (SR): AutoMS 83.8% vs. NSGA-II 43.7%, ReAct 53.3%, BayesOpt 65.0%, CMA-ES 66.7%, Single-Agent+SAES 69.2%, CoT+SAES 74.1%.
Mean Relative Error (MRE): AutoMS 0.0140 vs. NSGA-II 0.4276, BayesOpt 0.0194, CMA-ES 0.0241.
Best Property Match (BPM): AutoMS 94.2% vs. CMA-ES 84.7%, BayesOpt 85.6%.
Quality Score (QS): AutoMS 82.4 vs. CoT+SAES 81.9, BayesOpt/CMA-ES 73.7.
Iterations: AutoMS 14.4 avg. vs. NSGA-II 48.8, BayesOpt 39.2; fastest is S-SAES (9.6) but at lower SR.
Wall-clock time: AutoMS 2180.5 s vs. ReAct 2843.9 s (−23.3%); S-SAES fastest (1743.0 s) at lower accuracy.
Multi-agent ablation gap: removing role decomposition (S-SAES) costs 14.6 pp in SR, confirming orchestration is independently load-bearing.

Limitations¶

Depends on MIND as the foundational single-physics generator; generalization to other generative backends is not demonstrated.
Constraint Satisfaction Rate (CSR) of AutoMS (59.6%) is lower than CoT+SAES (70.2%), indicating that full-framework orchestration sometimes sacrifices partial-satisfaction diversity for complete validity.
Benchmark is domain-specific (materials/microstructure) with 17 tasks; breadth of cross-domain applicability is unproven.
High-fidelity FEA simulations remain expensive; scalability to higher-dimensional or real-time design loops is not assessed.
LLM backbone (DeepSeek-V3.2, GPT-4o, etc.) sensitivity is evaluated but not deeply analyzed; results may degrade on weaker backbones.

Relevance to Harnesses / Meta-Harnesses¶

AutoMS is a concrete implementation of a domain-specific meta-harness: a two-layer coordinator that wraps specialized sub-agents (Parser, Generator, Simulator, Reporter) behind a Manager orchestrator and a numerical SAES engine, tightly coupling open-ended LLM reasoning to a closed-loop simulation backend. Its hierarchical architecture—semantic planning decoupled from numerical search—is a direct instantiation of the harness pattern where the meta-layer enforces constraints and routes work, while inner agents handle execution. The use of GraphRAG for structured knowledge retrieval and adaptive weight rebalancing as a dynamic orchestration signal are design choices directly applicable to building robust harnesses in other scientific or engineering domains. The ablation results quantifying the SR contribution of orchestration vs. algorithm (14.6 pp gap) provide empirical justification for the multi-agent harness decomposition that is often assumed but rarely measured.