[Paper Review] MatLLMSearch: Crystal Structure Discovery with Evolution-Guided Large Language Models

Jingru Gan, Peichen Zhong+8 more|ArXiv.org|Feb 28, 2025

Machine Learning in Materials Science4 citations

TL;DR

MatLLMSearch uses pre-trained LLMs guided by evolutionary search to generate thermodynamically stable crystal structures without fine-tuning, achieving higher metastability and stability rates than baselines while reducing training overhead.

ABSTRACT

Crystal structure generation is fundamental to materials science, enabling the discovery of novel materials with desired properties. While existing approaches leverage Large Language Models (LLMs) through extensive fine-tuning on materials databases, we show that pre-trained LLMs can inherently generate novel and stable crystal structures without additional fine-tuning. Our framework employs LLMs as intelligent proposal agents within an evolutionary pipeline that guides them to perform implicit crossover and mutation operations while maintaining chemical validity. We demonstrate that MatLLMSearch achieves a 78.38% metastable rate validated by machine learning interatomic potentials and 31.7% DFT-verified stability, outperforming specialized models such as CrystalTextLLM. Beyond crystal structure generation, we further demonstrate that our framework adapts to diverse materials design tasks, including crystal structure prediction and multi-objective optimization of properties such as deformation energy and bulk modulus, all without fine-tuning. These results establish our framework as a versatile and effective framework for consistent high-quality materials discovery, offering training-free generation of novel stable structures with reduced overhead and broader accessibility.

Motivation & Objective

• Demonstrate that pre-trained LLMs can generate thermodynamically stable crystal structures without fine-tuning.
• Integrate LLM-based reproduction with evolutionary selection to explore crystal structure space.
• Validate stability predictions with MLIPs and DFT, comparing to state-of-the-art baselines.
• Showcase adaptability to crystal structure prediction and multi-objective materials design.

Proposed method

• Form an initial population from known stable structures.
• Prompt LLMs to perform implicit crossover and mutations to generate offspring.
• Relax and evaluate offspring with CHGNet and compute decomposition energy E_d and target properties.
• Select top candidates based on objective scores from parents, offspring, and optional extra pool.
• Final validation of evolved structures with DFT (VASP) for the most promising structures.

Experimental results

Research questions

• RQ1 Can pre-trained LLMs generate thermodynamically stable crystal structures without fine-tuning?
• RQ2 How effective is an LLM-guided evolutionary loop at discovering stable and diverse crystal structures?
• RQ3 What is the impact of extra reference pools on metastability and DFT-verified stability?
• RQ4 Can the framework extend to crystal structure prediction and multi-objective optimization of properties?
• RQ5 How does MatLLMSearch compare to fine-tuned baselines like CrystalTextLLM in stability metrics?

Key findings

• MatLLMSearch achieves 78.38% metastable rate by CHGNet and 31.7% DFT-verified stability, outperforming CrystalTextLLM under similar model sizes.
• Using thousands of reference structures, the method surpasses baselines while avoiding large-scale fine-tuning.
• Excluding f-electron-containing structures from parents improves metastability to 78.4% with 27.0% DFT-verified stability; stable structures without f-electrons rise to 24.6%.
• The approach supports crystal structure prediction and multi-objective optimization (e.g., balancing stability and bulk modulus).
• LLMs’ implicit crossover/mutation guided by prompts enables diverse structural motifs across crystal systems, with minimal computational overhead beyond stability evaluation.