Machine learning density functional compatible with dispersion correction for non-covalent interactions

Machine learning (ML) has demonstrated significant potential in enhancing the predictive capabilities of density functional theory methods. In this study, we develop an ML model for correcting B3LYP-D, a density functional approximation that incorporates dispersion corrections for non-covalent interactions. This model utilizes semilocal electron density descriptors, and is trained with accurate reference data for both relative and absolute energies. Extensive benchmark tests reveal that the ML correction substantially enhances the generalization ability of the B3LYP-D functional, improving the predictions of atomization and dissociation energies for complex molecular systems. It retains the accuracy of B3LYP-D in predicting reaction barrier heights and non-covalent interactions while enabling efficient, fully self-consistent field calculations. This work signifies a promising advancement in the development of ML-corrected functionals that surpass the performance of traditional B3LYP-D.