A Genetic Algorithm For Convex Hull Optimisation

TL;DR

This paper introduces a genetic algorithm that efficiently optimizes convex hulls for multi-species crystal structures, aiding high-throughput materials discovery with machine learning potentials.

Contribution

It presents a novel genetic algorithm approach that directly optimizes convex hulls, improving efficiency in discovering stable multi-species crystal structures.

Findings

Successfully discovered known LiSi structures

Identified new potential LiSi candidate structures

Demonstrated efficiency with machine-learned potentials

Abstract

Computationally efficient and automated generation of convex hulls is desirable for high throughput materials discovery of thermodynamically stable multi-species crystal structures. A convex hull genetic algorithm is proposed that uses methodology adapted from multi-objective optimisation techniques to optimise the convex hull itself as an object, enabling efficient discovery of convex hulls for N >= 2 species. This method, when tested on a LiSi system utilising pre-trained machine learned potentials, was found to be able to efficiently discover reported structures as well as new potential LiSi candidate structures.