Exchange functionals trained on exact exchange for molecular and solid-state systems.

Abstract

Self-interaction error (SIE) in density functional theory (DFT) leads to significant inaccuracies in the calculation of barrier heights of chemical reactions and band gaps of solid-state systems. In this work, we develop a neural-network-based exchange functional aimed at reducing SIE-related errors by training on exact exchange data. The proposed functional achieves improved accuracy in predicting barrier heights (BH) and band gaps compared with the PBE, SCAN, M06L, and revM06L functionals. In addition, tests on other quantities, including atomization energies (AE), ionization potentials (IP), and vibrational frequencies (VF), show that the developed functional also provides reliable performance for these properties.