CrystalAgent: Towards Autonomous Crystal Generation via Agentic Reasoning

ICLR 2026 Conference Submission22594 Authors

Keywords: Crystal Generation, Large Language Model

Abstract: Recent advances in large language models (LLMs) have demonstrated remarkable generalization capabilities across diverse domains, and recent studies have begun to explore their application in crystal generation. Nevertheless, most of these approaches rely heavily on extensive fine-tuning with large-scale datasets, which often limits their adaptability and generality when applied to real-world crystal discovery. To overcome these limitations, we propose CrystalAgent, an LLM-based agent that eliminates the need for additional training and adapts flexibly to diverse crystal discovery scenarios. Specifically, we decompose the crystal generation process into four key stages: Extract, Retrieval, Generation, and Optimization. The Extract stage involves extracting crystal design constraints from user inputs. In the Retrieval stage, based on the extracted constraints, the system automatically selects few-shot examples from the database to inform subsequent processes. The Generation stage leverages LLMs to generate crystal structures by learning atomic distribution patterns from the selected examples, and the Optimization stage refines the generated structure by using crystal structure optimization tools and energy evaluation tools to select the optimal structure as the final output. Extensive experiments across various crystal generation tasks highlight the flexibility, controllability, and versatility of our framework, underscoring the substantial potential of LLM agents in automating the generation of crystal materials and advancing the field of materials discovery.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 22594