Accelerating Materials Design via LLM-Guided Evolutionary Search

TL;DR

This work tackles the real-world challenge of discovering materials that simultaneously satisfy multiple, potentially conflicting objectives while ensuring synthesizability. It introduces LLEMA, a framework that fuses large language model reasoning with chemistry-informed design rules, surrogate-based property prediction, and memory-based evolutionary refinement to navigate multi-objective spaces. Across 14 industrially relevant tasks, LLEMA delivers higher hit rates and stronger Pareto fronts than baselines, while generating thermodynamically stable and chemically plausible candidates and reducing memorization. The approach offers a principled, scalable pathway for autonomous materials discovery that integrates domain knowledge, predictive signals, and iterative learning to accelerate practical design.

Abstract

Materials discovery requires navigating vast chemical and structural spaces while satisfying multiple, often conflicting, objectives. We present LLM-guided Evolution for MAterials design (LLEMA), a unified framework that couples the scientific knowledge embedded in large language models with chemistry-informed evolutionary rules and memory-based refinement. At each iteration, an LLM proposes crystallographically specified candidates under explicit property constraints; a surrogate-augmented oracle estimates physicochemical properties; and a multi-objective scorer updates success/failure memories to guide subsequent generations. Evaluated on 14 realistic tasks spanning electronics, energy, coatings, optics, and aerospace, LLEMA discovers candidates that are chemically plausible, thermodynamically stable, and property-aligned, achieving higher hit-rates and stronger Pareto fronts than generative and LLM-only baselines. Ablation studies confirm the importance of rule-guided generation, memory-based refinement, and surrogate prediction. By enforcing synthesizability and multi-objective trade-offs, LLEMA delivers a principled pathway to accelerate practical materials discovery. Code: https://github.com/scientific-discovery/LLEMA

Figures (15)

• Figure 1: Overview of our multi-objective material discovery benchmark. The benchmark spans six diverse disciplines encompassing fourteen tasks with thermodynamic, electrical, physical, and chemical properties of materials.
• Figure 2: LLEMA Framework, consisting of four main components: (A) Material Candidate Generation, where an LLM generates candidates based on task descriptions and property constraints; (B) Crystallographic Representation, which converts generated materials into structured crystallographic information files (CIFs); (C) Physicochemical Property Prediction, to predict material properties such as formation energy and band gap, etc; and (D) Fitness Assessment and Feedback, which evaluates constraint satisfaction and provides iterative feedback through success/failure memory pools.
• Figure 3: Evolution of candidates for the Stable Wide-Bandgap Semiconductor at different stages of LLEMA.
• Figure 4: Pareto front analysis of candidate materials for two design tasks. (a) Wide-Bandgap Semiconductors; (b) Hard–Stiff Ceramics.
• Figure 5: Percentage of generated candidates from the Materials Project across four domains for different baselines. Lower values indicate less memorization.