Data-efficient learning of exchange-correlation functionals with differentiable DFT

Abstract Machine learning (ML) density functional approximations (DFAs) have seen a lot of interest in recent years, often being touted as the replacement for well-established non-empirical DFAs, which still dominate the field. Although highly accurate, ML-DFAs typically rely on large amounts of data, are computationally expensive, and fail to generalize beyond their training domain. In this work we show that differentiable DFT with Kohn–Sham regularization can be used to accurately capture the behavior of known local density approximations from small sets of synthetic data without using localized density information. At the same time our analysis shows a strong dependence of the learning on both the amount and type of data as well as on model initialization. By enabling accurate learning from sparse energy data, this approach paves the way towards the development of custom ML-DFAs trained directly on limited experimental or high-level quantum chemistry datasets.