Machine-learning-enhanced density functional theory calculations

Abstract

Machine learning has been widely applied to improve accuracy in computational chemistry. Here, we present a simple yet efficient machine-learning post-correction model that can calibrate the total energy from density functional theory’s (DFT) value to the coupled cluster’s one by training on energy differences between them across 56 small molecules from the G2 dataset. Our approach has significantly reduced the error of absolute energy from 358.7 kcal/mol of DFT calculations to 1.3 kcal/mol on that dataset. Moreover, a reduction in errors of relative energies, including atomization energies, ionization potentials, electron affinities, noncovalent interactions, reaction energies, and barrier heights, on dozens of other datasets demonstrates the strong transferability and applicability of our model. In addition, our method only performs a single post-processing correction step following standard DFT calculations, thus incurring a minor additional time consumption of 0.69 s on average for G2 molecules. This study thus elucidates a systematic and efficient approach for enhancing the accuracy of DFT energy-related calculations.