Imitating beyond-DFT calculations via external on-site potentials

Density functional theory (DFT) based methods are the basis for the high-throughput calculations of structural and electronic properties currently used in most Materials-by-Design approaches for materials discovery. To improve the accuracy of such predictive theory, it is now necessary to bridge the gap between efficient, but approximate, DFT calculations, and accurate, but computationally expensive, “beyond-DFT” approaches, such as GW electronic structure and random phase approximation (RPA) total energy calculations. In this direction, we aim to reproduce the beyond-DFT level by using angular-momentum and energy-dependent external on-site potentials. The problem of fitting the potential parameters is equivalent to solving a multi-objective optimization problem. Using VO2 as a model system, we develop a workflow for potential fitting via machine learning, so to explore the extensive multi-dimensional search space for the highly non-linear objective function.