Atomic Energies from a Convolutional Neural Network

Abstract

Understanding interactions and structural properties at the atomic level is often a prerequisite to the design of novel materials. Theoretical studies based on quantum-mechanical first-principles calculations can provide this knowledge but at an immense computational cost. In recent years, machine learning has been successful in predicting structural properties at a much lower cost. Here we propose a simplified structure descriptor with no empirical parameters, "k-Bags", together with a scalable and comprehensive machine learning framework that can deepen our understanding of atomic properties of structures. This model can readily predict structure-energy relations that can provide results close to the accuracy of ab initio methods. The model provides chemically meaningful atomic energies enabling theoretical analysis of organic and inorganic molecular structures. Utilization of the local information provided by the atomic energies significantly improves upon the stochastic steps in our evolutionary global structure optimization, resulting in a much faster global minimum search of molecules, clusters, and surfaced supported species.