The Best DFT Functional Is the Ensemble of Functionals

Abstract

Abstract The development of better density functional theory (DFT) methods is one of the most active research areas, given the importance of DFT for ubiquitous molecular and materials simulations. However, this research primarily focuses on improving a specific exchange‐correlation Kohn–Sham density functional. Here, a robust procedure is proposed for constructing transferable ensembles of density functionals that perform superior to any constituent individual density functional. It is shown that such ensembles built only with the density functionals predating the GMTKN55 benchmark of 2017 can reach a record‐low weighted error of 1.62 kcal mol−1 on this benchmark compared to 3.08 kcal mol−1 of the best constituent density functional. The DENS24 density functional ensembles are also introduced as practical DFT methods with consistently accurate performance for various simulations at affordable cost. DENS24 ensembles are open‐source and can be used for simulations online. Additionally, it is shown that the ensembles can be integrated into the SCF procedure by creating mixed DENS24 functionals, which have the same accuracy but are faster than ensembles of independent functionals.

Keywords

• ensemble learning
• density functional theory
• DFT calculations