Delta-Augmented Subsystem Density Functional Theory: A Study Across Diverse Systems

In this study, we expand upon and benchmark the Kim-Gordon method (KG), a subsystem density functional theory (DFT) approach appended with a machine-learned correction to compensate for errors in the kinetic energy term and thereby match Kohn-Sham (KS) DFT accuracy. This correction is obtained through 'delta- learning' based on KS-DFT data. The method promises sampling of configurations for condensed molecular systems at the Kohn-Sham DFT level of accuracy at a fraction of the computational cost. Despite encouraging results for liquid water, it was not obvious whether the scheme had more general appeal. In this work, we show that the approach allows for a broad range of applications. In particular, we successfully apply it to complex molecular liquids, such as bulk ammonia and methanol. As a bonus, the correction trained on the bulk KS data is applicable to clusters, illustrating its transferability. By focusing on 'delta-learning'-predicting small corrections rather than full Kohn-Sham (KS) energies and forces-we significantly reduce the required training data. This approach, especially when combined with linear-scaling self-consistent field (LS-SCF) techniques, establishes the method as a highly efficient computational tool for molecular dynamics.