Anatase (001) surface from machine learning

Surfaces of semiconductors often exhibit reconstructions that can significantly influence their physical and chemical properties. Especially for catalysts, studying the surface structures at the atomic scale is crucial to gain a better understanding of the catalytic activity and, ultimately, to design improved materials. Atomistic simulations can provide insight, but ab initio structure predictions are rather challenging due to the large supercells required to model 2D surfaces.

To alleviate this issue, we use a machine-learned interatomic potential trained on DFT data in conjunction with a sophisticated structure prediction method to explore low-energy reconstructions of the TiO₂ Anatase (001) surface. We identify a new surface reconstruction that is comparable in energy with the well-known (001)-1x4 ad-molecule model and could explain the recently discovered experimental structures from STM imaging.