Accelerated searches, made possible by machine learning techniques, are of growing interest in materials discovery. A suitable case involves the solution processing of components that ultimately form thin films of solar cell materials known as hybrid organic–inorganic perovskites (HOIPs). The number of molecular species that combine in solution to form these films constitutes an overwhelmingly large “compositional” space (at times, exceeding 500,000 possible combinations). Selecting a HOIP with desirable characteristics involves choosing different cations, halides, and solvent blends from a diverse palette of options. An unguided search by experimental investigations or molecular simulations is prohibitively expensive. In this work, we propose a Bayesian optimization method that uses an application-specific kernel to overcome challenges where data is scarce, and in which the search space is given by binary variables indicating whether a constituent is present or not. We demonstrate that the proposed approach identifies HOIPs with the targeted maximum intermolecular binding energy between HOIP salt and solvent at considerably lower cost than previous state-of-the-art Bayesian optimization methodology and at a fraction of the time (less than 10%) needed to complete an exhaustive search. We find an optimal composition within 15 ± 10 iterations in a HOIP compositional space containing 72 combinations, and within 31 ± 9 iterations when considering mixed halides (240 combinations). Exhaustive quantum mechanical simulations of all possible combinations were used to validate the optimal prediction from a Bayesian optimization approach. This paper demonstrates the potential of the Bayesian optimization methodology reported here for new materials discovery.
Applications of Bayesian optimization to problems in the materials sciences have primarily focused on consideration of a single source of data, such as DFT, MD, or experiments. This work shows how it is possible to incorporate cost-effective sources of information with more accurate, but expensive, sources as a means to significantly accelerate materials discovery in the computational sciences. Specifically, we compare the performance of three surrogate models for multi-information source optimization (MISO) in combination with a cost-sensitive knowledge gradient approach for the acquisition function: a multivariate Gaussian process regression, a cokriging method exemplified by the intrinsic coregionalization model, and a new surrogate model we created, the Pearson-r coregionalization model. To demonstrate the effectiveness of this MISO approach to the study of commonly encountered materials science problems, we show MISO results for three test cases that outperform a standard efficient global optimization (EGO) algorithm: a challenging benchmark function (Rosenbrock), a molecular geometry optimization, and a binding energy maximization. We outline factors that affect the performance of combining different information sources, including one in which a standard EGO approach is preferable to MISO.
27 June 2017 in Journal of Chemical Theory and Computation