Ephemeral data derived potentials for random structure search

Structure prediction has become a key task of the modern atomistic sciences, and depends on rapid reliable computation energy landscape. First principles density functional based calculations are highly reliable, faithfully describing entire They are, however, computationally intensive slow compared to interatomic potentials. Great progress been made in development machine learning, or data derived, potentials, which promise describe landscape at first quality. However, approaches, their preparation can be time consuming delay searching. Ab initio random structure searching (AIRSS) is straightforward powerful approach prediction, stochastic generation sensible initial structures, repeated local optimisation. Here, scheme, compatible with AIRSS, for construction disposable, ephemeral, derived potentials (EDDPs) described. These constructed using homogeneous, separable manybody environment vector, iterative neural network fits, sparsely combined through non-negative least squares. The tested methane, boron nitride, elemental urea. In case boron, an EDDP generated from small unit cells used rediscover complex $\gamma$-boron without recourse symmetry fragments. Finally, silane (SiH$_4$) 500 GPa enables discovery extremely complex, dense, significantly modifies silane's high pressure phase diagram. This implications theoretical exploration temperature superconductivity dense hydrides, have so far largely depended searches smaller cells.