Machine learning interatomic potentials with and without (much) human labor

In this talk I will show the state of the art in generating an interatomic potential using machine learning techniques (specifically, the Gaussian Approximation Potential (GAP) framework, with the Smooth Overlap of Atomic Positions (SOAP) representation of atomic geometry) that is capable of representing the Born-Oppenheimer potential energy surface of a material based on data (energies and forces) computed using density functional theory. First I will show that a careful and very human labor-intensive process of assembling a training database results in a potential with exquisite accuracy, far surpassing any empirical potential in the literature in predicting material properties. Then I will demonstrate, using several different elemental compounds with quite different bonding chemistry, that the database building process can be largely automated. A combination of iterative training and ab initio random search (AIRSS) can be used to simultaneously "discover and fit" a material without the need for any prior knowledge of what structures are relevant for a given material.