Augmenting the discovery of computationally complex ceramics for extreme environments with machine learning

We present a high-throughput, material-agnostic strategy to discover new compositionally complex ceramics (C 3 ) for extreme environments by utilizing machine learning (ML) techniques to predict the stoichiometries and properties of structures within a given design space. This example study focuses on a well-understood design space (Si–C–N) so that predictions may be validated. Evolutionary structure searches coupled with density functional theory (DFT) calculations are applied to find structures with low energies (i.e., lying on or close to the convex hull), while also maximizing a targeted property (in this case, hardness). The structure–property relationship data obtained throughout these searches are exploited in ML algorithms to create an accurate and efficient surrogate model of the energy and hardness landscapes. The ML models serve to screen structures with optimal attributes and reduce computational costs associated with the property calculations, thereby accelerating the discovery of new structures and stoichiometries with desired traits. Graphical abstract