CrystalReasoner: Reasoning and RL for Property-Conditioned Crystal Structure Generation

2026-05-14

AI Summary

Existing methods for crystal structure generation struggle to simultaneously achieve atomic-level precision and integrate scientific knowledge, often yielding invalid or unstable structures. This work proposes CrystalReasoner, the first end-to-end framework that embeds physical priors into a large language model through “reasoning tokens,” enabling knowledge-guided, adaptive generation. It introduces a multi-objective reinforcement learning reward mechanism tailored to both discrete and continuous crystal properties and explicitly models crystal symmetry and local coordination environments. The approach substantially improves generation quality, tripling the S.U.N. ratio and consistently outperforming existing baselines across diverse property-constrained generation tasks.

Abstract

Generative modeling has emerged as a promising approach for crystal structure discovery. However, existing LLM-based generative models struggle with low-level atomic precision, while diffusion-based methods fall short in integrating high-level scientific knowledge. As a result, generated structures are often invalid, unstable, or do not possess desirable properties. To address this gap, we propose CrystalReasoner (\method), an end-to-end LLM framework that generates crystal structures from natural language instructions through reasoning and alignment. \method introduces physical priors as thinking tokens, which include crystallographic symmetry, local coordination environments and predicted physical properties before generating atomic coordinates. This bridges the gap between natural language and 3D structures. \method then employs reinforcement learning (RL) with a multi-objective, dense reward function to align generation with physical validity, chemical consistency, and thermodynamic stability. For property-conditioned tasks, we design task-specific reward functions and train specialized models for discrete constraints (e.g., space group) and continuous properties (e.g., elasticity, thermal expansion). Empirical results demonstrate that compared to prior works and baselines without thinking traces or RL, \method obtains better performance on diverse metrics, triples S.U.N. ratio, and achieves better performance for property conditioned generation. \method also exhibits adaptive reasoning, increasing reasoning lengths as the number of atoms increases. Our work demonstrates the potential of leveraging thinking traces and RL for generating valid, stable, and property-conditioned crystal structures. Please see our work at https://crystalreasoner.github.io/ .

Problem

• crystal structure generation
• property conditioning
• atomic precision
• scientific knowledge integration
• structural validity

Innovation

• thinking tokens
• reinforcement learning
• property-conditioned generation
• physical priors
• crystal structure generation