Shotgun crystal structure prediction using machine-learned formation energies

TL;DR

This paper introduces ShotgunCSP, a machine learning-based method for crystal structure prediction that significantly reduces computational costs while maintaining high accuracy by using transfer learning and generative models.

Contribution

The paper presents a novel shotgun approach combining transfer learning and generative models for efficient, accurate crystal structure prediction with minimal first-principles calculations.

Findings

Achieved 93.3% accuracy in benchmark tests.

Reduced computational effort compared to traditional methods.

Successfully predicted diverse crystal structures with minimal first-principles calculations.

Abstract

Stable or metastable crystal structures of assembled atoms can be predicted by finding the global or local minima of the energy surface within a broad space of atomic configurations. Generally, this requires repeated first-principles energy calculations, which is often impractical for large crystalline systems. Here, we present significant progress toward solving the crystal structure prediction problem: we performed noniterative, single-shot screening using a large library of virtually created crystal structures with a machine-learning energy predictor. This shotgun method (ShotgunCSP) has two key technical components: transfer learning for accurate energy prediction of pre-relaxed crystalline states, and two generative models based on element substitution and symmetry-restricted structure generation to produce promising and diverse crystal structures. First-principles calculations were performed only to generate the training samples and to refine a few selected pre-relaxed crystal structures. The ShotunCSP method is computationally less intensive than conventional methods and exhibits exceptional prediction accuracy, reaching 93.3% in benchmark tests with 90 different crystal structures.