A regression approach to accurate interaction energies using topological descriptors

Machine learning has a wide range of applications in chemistry, encompassing the prediction of the structure and properties of a variety of chemical systems (molecules, macromolecules and solids). The idea of using a self-learning algorithm to explore chemical problems is particularly alluring when facing open challenges in theoretical chemistry, such as non-covalent interactions, which are known to be critical for density functional methods. Additionally, the difficulty of predicting accurate non-covalent interaction (NCI) energies lies in their small absolute values, which make even the slightest absolute error severely affect the quality of the calculated result. In this work, we test the possibility of computing accurate interaction energies (at the golden standard CCSD(T)/CBS level) of small non-covalent complexes starting from a DFTD/DZ energy and using descriptors derived from the promolecular density. Calculations on the S66x8 dataset of molecular complexes show that these local descriptors can reduce to one third the mean absolute error of DFT results at a virtually negligible computational cost.