First-principles structure prediction has enabled the computational discovery of materials with extreme, or exotic properties. For example, the dense hydrides, which following computational searches have been found to exhibit high-Tc superconductivity. However, the quantum mechanical calculations that are performed are costly, and limit the number and complexity of the structures that can be investigated.
The rapid construction of ephemeral, or disposable, interatomic potentials assembled from multiple small neural networks, dramatically accelerates the exploration of energy landscapes, even accounting for the time to create the potentials. This approach is used to uncover a complex and overlooked phase of dense silane, suggesting that the binary hydrides should be revisited.
The Ephemeral Data Derived Potential (EDDP) package has been released alongside an updated version of Ab Initio Random Structure Searching (AIRSS), under the GPL2 open source license.
Figure Caption: Complex silane structure uncovered (left) using an ephemeral two and three body (right) potential designed for the search at high pressure.
“Ephemeral data derived potentials for random structure search”, Chris J. Pickard, Phys. Rev. B106, 014102 (2022) Editors’ Suggestion