Data-Driven Crystal Structure Prediction for Ternary Metal Chalcogenides

TL;DR

This paper explores how generative AI can help discover new crystal structures for materials more efficiently than traditional methods.

Contribution

The study demonstrates that generative artificial intelligence outperforms traditional methods in predicting stable crystal structures for ternary metal chalcogenides.

Findings

Generative AI (genAI) using denoising diffusion matches or exceeds traditional methods in identifying low-energy crystal structures.

Machine-learned interatomic potentials provide reliable energy estimates and uncertainty quantification for candidate structures.

The approach is applied to ternary metal chalcogenides like Na2SiS3 and KMo2S4, showing promise for scalable materials exploration.

Abstract

The efficient design
and discovery of stable inorganic crystal
structures is central to materials innovation. Here, we compare data-driven
approaches for accelerated crystal structure prediction: substitution
into known prototype structures, generative artificial intelligence
(genAI) using denoising diffusion as implemented in Chemeleon, and an evolutionary global optimization search.
Candidate structures are optimized using an ensemble of machine-learned
interatomic potentials, providing both energy estimates and uncertainty
quantification. Applied to a large set of known and hypothetical ternary
metal chalcogenide compositions, including technologically relevant
sulfides such as Na2SiS3, RbPS3,
and KMo2S4, our analysis reveals that the genAI approach not only matches but can surpass traditional
methods in identifying diverse, low-energy structures. These findings
highlight the promise of generative models for scalable structural
exploration of inorganic materials space.

Figures