DFT-accurate simple neural networks that predict ground state energies throughout unconstrained composition space: What they are good for

Unconstrained composition space (UCS) consists of the ground state atomic arrangements for all possible compositions of elements Hx1Hex2Lix3 …Ux92 for all possible mole fractions (x1, … x92). We first demonstrate that given only the stoichiometry of a material, it is possible to train a very simple neural network architecture to reproduce within DFT accuracy (mean absolute error of 0.037 eV on test dataset) the formation energy per atom E(x1, …, x92) of the most stable structures appearing in the Materials Project. This talk addresses the important question of what can be done with access to such a function. We will present results for a wide variety of applications, from predicting stable material phase diagrams and electrochemical Pourbaix diagrams, to predicting interstitial and vacancy energies when no such defect energies appear in the Materials Project training set.