CrystalFormer-CSP: Thinking Fast and Slow for Crystal Structure Prediction

TL;DR

CrystalFormer-CSP is a novel framework that combines data-driven and physics-based methods, including generative models and machine learning force fields, to efficiently predict stable crystal structures, with reinforcement fine-tuning enhancing accuracy.

Contribution

It introduces a unified approach integrating generative models and ML force fields for crystal structure prediction, improving efficiency and accuracy over existing methods.

Findings

Effective on benchmark problems

Web interface and language model integration demonstrated

Reinforcement fine-tuning boosts accuracy

Abstract

Crystal structure prediction is a fundamental problem in materials science. We present CrystalFormer-CSP, an efficient framework that unifies data-driven heuristic and physics-driven optimization approaches to predict stable crystal structures for given chemical compositions. The approach combines pretrained generative models for space-group-informed structure generation and a universal machine learning force field for energy minimization. Reinforcement fine-tuning can be employed to further boost the accuracy of the framework. We demonstrate the effectiveness of CrystalFormer-CSP on benchmark problems and showcase its usage via web interface and language model integration.