Smart Sampling for Chemical Property Landscapes with BOSS

Atomistic structure search for organic/inorganic heterostructures is made complex by the many degrees of freedom and the need for accurate but costly density-functional theory (DFT) simulations. To accelerate and simplify structure determination in such heterogenous functional materials, we developed the Bayesian Optimization Structure Search (BOSS) approach [1].

Bayesian optimization is employed to build N-dimensional surrogate models for the energy or property landscapes and infer global minima. The models are iteratively refined by sequentially sampling DFT data points that are promising and/or have high information content. Representing heterogenous materials with compact chemical ‘building blocks’ allowed us to build in prior knowledge and reduce search dimensionality. The uncertainty-led exploration/exploitation sampling strategy delivers global minima with modest sampling, but also ensures visits to less favorable regions of phase space to gather information on rare events and energy barriers.

We applied this active learning scheme to study adsorption at the organic/inorganic interfaces: C₆₀ on TiO₂(101) anatase and camphor on Cu(111). Global minima in 6 dimensions are identified with reasonable computational efficiency. BOSS produces chemically-intuitive adsorption energy landscapes that are parsed for local minima and the minimum energy paths between them, allowing us to extract complex barrier-related atomistic pathways. With a recent batch implementation for active learning, BOSS can make use of exascale computing resources to solve large-scale structural problems without sacrificing quantum-mechanical accuracy.

[1] M. Todorović, M.U. Gutmann, J. Corander and P. Rinke, npj Comput. Mater. 5 (2019) 35.