Machine Learning of Surface Adsorbate Structures

COMP Seminar (Otakaari 1). Speaker: Dr. Milica Todorović (Computational Electronic Structure Theory group).

The adsorption and self-organisation of molecules at inorganic surfaces is central to many industrial processes from catalysis and coatings to organic electronics and solar cells. Computer simulations can help identify interface morphology and functionality, but sampling many atomic configurations over large length scales is prohibitively costly. We combined Bayesian optimisation with accurate atomistic simulations in our efficient structure search tool BOSS, designed for intelligent probabilistic sampling of atomic configurations. The nearly parameter-free framework relies on Gaussian processes to construct a probable potential energy surface, which is then iteratively refined by the input of energy data points from selected configurations. We report a dramatic speed-up in identifying optimal structures, compared to the traditional chemical intuition technique, without loss of accuracy.