Machine learning (ML) offers the potential to greatly accelerate materials discovery and design by providing the means to efficiently and scalably predict properties across vast chemical spaces. Nevertheless, the hetergenous nature of materials data sets pose major challenges to the development of reliable ML models. Furthermore, state-of-the-art models that can operate across the entire universe of known crystal structures typically require an equilibrium crystal structure as an input, which is typically obtained via expensive DFT calculations. In this talk, I will discuss how both challenges can be addressed by designing ML models to operate on multi-fidelity data sources and using those ML models with Bayesian optimization to solve the structure optimization problem in a "DFT-free" manner. The end result is a ML-driven materials discovery platform that can screen millions of candidates for materials - several orders of magnitude larger than traditional DFT-based screening - with exceptional properties. As a proof of concept, we will demonstrate the discovery of completely new ultra-incompressible and optical materials that have been confirmed experimentally.

Accelerating materials discovery with machine learning

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.