Machine learning of potential-energy landscapes with support-vector machines and neural networks
Stephen R. Xie, Shreyas Honrao, and Richard G. Hennig
University of Florida, Department of Materials Science and Engineering, Gainesville, FL, 32611-8435
We present two machine learning approaches to predict the formation energies of compounds in multicomponent materials systems. We compare models generated by support vector regression (SVR) and artificial neural networks (ANN) to approximate the overall energy landscapes as a function of composition and structure. We demonstrate that atomic structure descriptors based on partial radial and angular distribution functions and the Chebyshev descriptor encode the relevant physical information for machine learning to lead to prediction errors better than chemical accuracy. We test the approaches on the energy landscapes of the Cd-Te and Li-Ge binary systems, which we explore with genetic algorithms. To overcome the small-data problem that plagues surrogate model development in the early stages of genetic algorithm structure searches, we show that augmenting the data using local information such as local energies and local descriptors significantly accelerates the learning. These machine-learning models have the potential to accelerate crystal structures prediction in multicomponent systems.