Machine Learning methods for interatomic potentials: application to boron carbide

Total energies of crystal structures can be calculated to high precision using quantum-based density functional theory (DFT) methods, but the calculations can be time consuming and scale badly with system size. Cluster expansions of total energy as a linear superposition of pair, triplet and higher interactions can efficiently approximate the total energies but are best suited to simple lattice structures. To model the total energy of boron carbide, with a complex crystal structure, we explore the utility of machine learning methods (L1-penalized regression, neural network, Gaussian process and support vector regression) that capture certain non-linear effects associated with manybody interactions despite requiring only pair frequencies as input. Our interaction models are combined with Monte Carlo simulations to evaluate the thermodynamics of chemical ordering.