A Semilocal Machine-Learning Correction for Density Functional Approximations

Abstract

Machine learning (ML) algorithms have shown to be potentially effective in the development of density functional theory (DFT) methods.In this work, we have developed a semilocal ML correction for density functional approximations.The correction adopts simple descriptors of electron density and density derivative, and is trained upon the combination of relative and absolute reference energies. The ML-corrected B3LYP functional has been tested on a comprehensive set of various chemical properties, among which the accuracy of the predictions for atomization energies, heats of formation, and total energies of atoms and molecules is significantly improved, and the accuracy of the predictions for ionization potentials, electron affinities, and bond dissociation energies is slightly improved over the original B3LYP, while the same level of accuracy is maintained for isomerization energies and reaction barrier heights.The ML-corrected functional can be used in self-consistent-field calculations of energetic properties and electron density with similar computational efficiency as the original functional.The performance of the ML-corrected functional highlights the potential of the semilocal correction scheme to provide a functional that performs uniformly better than B3LYP.