Machine Learning of Density Functionals for Accurate, Large-Scale Materials Simulations

We have recently developed the CIDER formalism for machine learning exchange-correlation functionals, with a particular emphasis on using nonlocal features to achieve hybrid density functional theory (DFT) accuracy at semilocal DFT cost for large solid-state simulations. In this talk, we will cover current directions being pursued to further improve CIDER functionals, including training full exchange-correlation functionals for applications to heterogeneous systems and improving the accuracy of CIDER functionals for band gap and charge transfer-related problems. We will also discuss how CIDER can be used to overcome cost-accuracy trade-offs for materials science applications where both large system sizes and hybrid DFT accuracy are required, such as the calculation of charged point defect properties in semiconductors.