Beyond the Black-Box—New Advances in Bayesian Materials Discovery

Over the past decade, Bayesian methods have quickly emerged as some of the best approaches for accelerated materials discovery. These approaches generally frame the process of materials discovery as a 'black box' optimization problem in which the 'black box' function is sequentially and optimally queried. The typical ingredients of Bayesian Optimization (BO) are: (1) a stochastic model of the world that predicts the outcomes of experiments yet to be run as well as the variance in such outcomes; (2) a prescriptive policy that recommends the next best experiment to run given these outcomes. While general purpose BO methods have proven quite useful in solving complex and expensive optimization problems, significant improvements can be achieved if one looks beyond the 'black box' setting. In this talk, I will review different approaches to improving BO applied to materials discovery. Such improvements include taking advantage of multiple information sources, exploiting correlations among materials observables that have different acquisition costs, embedding physics into the BO process itself, embracing knowledge about intermediate materials states, as well as strategies to bringing human expertise into the discovery loop. Several examples will be provided and a perspective for future improvements will be presented.