Ephemeral data derived potentials for random structure search

Abstract

Structure prediction has become a key task of the modern atomistic sciences and depends on the rapid and reliable computation of energy landscapes. First-principles density functional based calculations are highly reliable, faithfully describing entire energy landscapes. They are, however, computationally intensive and slow compared to interatomic potentials. Great progress has been made in the development of machine learning, or data derived, potentials, which promise to describe entire energy landscapes at first-principles quality. Compared to first-principles approaches, their preparation can be time consuming and delay searching. Ab initio random structure searching (AIRSS) is a straightforward and powerful approach to structure prediction, based on the stochastic generation of sensible initial structures and their repeated local optimization. Here, a scheme, compatible with AIRSS, for the rapid construction of disposable, or ephemeral, data derived potentials (EDDPs) is described. These potentials are constructed using a homogeneous, separable many-body environment vector and iterative neural network fits, sparsely combined through non-negative least squares. The approach is first tested on methane, boron nitride, elemental boron, and urea. In the case of boron, an EDDP generated using data from small unit cells is used to rediscover the complex γ-boron structure without recourse to symmetry or fragments. Finally, an EDDP generated for silane (SiH4) at 500 GPa enables the discovery of an extremely complex, dense structure which significantly modifies silane's high pressure phase diagram. This has implications for the theoretical exploration for high temperature superconductivity in dense hydrides, which have so far largely depended on searches in smaller unit cells.