Rapid mapping of alloy surface phase diagrams via Bayesian evolutionary multitasking

Many global optimization tasks have partially or completely shared search spaces. This is particularly common in computational surface science, where surfaces are optimized to their lowest-energy configurations under various idealized chemical conditions to construct surface phase diagrams. If each task represents a global optimization of the surface system under a unique reaction condition, all tasks essentially shares the same configurational search space as well as the same black-box function that takes as input the configuration and returns the electronic energy of the relaxed structure. However, the configurational search space can be huge for multi-component systems such as alloys, while each energy evaluation also requires an expensive electronic structure calculation. Bayesian optimization represents an efficient paradigm for optimizing an expensive black-box function. Evolutionary multitasking is a newly emerging paradigm for solving multiple global optimization problems simultaneously. Here we present a novel Bayesian evolutionary multitasking (BEM) framework combining the best of the two. As a case study, we have used our method to derive surface phase diagrams of Pt-Ni alloy catalysts under a wide range of reaction conditions for steam methane reforming. Integrating knowledge such as graph theory, lattice symmetry and active learning, the BEM framework as a whole can significantly accelerate the mapping of surface phase diagrams for very complex heterogeneous systems.