Taylor Sparks1,2,Hasan Sayeed1,Christopher Collins2,Matthew Rosseinsky2
University of Utah1,University of Liverpool2
Taylor Sparks1,2,Hasan Sayeed1,Christopher Collins2,Matthew Rosseinsky2
University of Utah1,University of Liverpool2
To enable efficient exploration of vast composition spaces and enhance crystal structure prediction, we have extended the capabilities of the Flexible Unit Structure Engine (FUSE) by integrating machine learning (ML) techniques. FUSE is a widely used computational tool for materials discovery across diverse compositions. Our approach consists of two key components. Firstly, we employ classical ML models to accurately predict volume/atom ratios, facilitating the estimation of unit cell sizes for various compositions. Secondly, we integrate a well-established ML-driven crystal structure prediction technique, utilizing advanced methods such as graph networks to establish correlations between crystal structures and formation enthalpies. To expedite the search for crystal structures with the lowest formation enthalpy, this technique incorporates an optimization algorithm. Through the integration of ML-based volume/atom prediction and crystal structure prediction within FUSE, our approach significantly accelerates the discovery of experimentally realizable compounds. This integration offers a promising pathway for efficient materials discovery, pushing the boundaries of knowledge in complex composition spaces.