Neural-Network-Backed Evolutionary Search for Novel (2×4) SrTiO₃(110) Surface Reconstructions

Abstract

Understanding the atomistic models underlying experimentally observed surface reconstructions of oxides remains a key challenge, requiring efficient exploration of the high-dimensional potential energy surface. The covariance matrix adaptation evolution strategy (CMA-ES) has been successfully applied to identify SrTiO3(110) surface reconstructions in both the Ti-rich and Ti-poor regimes [1, 2]. However, intermediate regimes remain unresolved. Although designed for global exploration, evolutionary algorithms may in practice miss important regions of the search space. CMA-ES depends on both the initial structure and hyperparameters such as the initial step size. Here, we introduce a strategy that initiates searches from randomly generated structures, enabling a broader and less biased sampling. Using this approach, and by combining the CMA-ES with a neural network force field, we not only successfully reproduce previously known SrTiO3(110) reconstructions but also discover novel (2×4) variants.