Abstract

In this work we explore how the contributions of exact exchange significantly impact electronic states, which are key in forming and breaking covalent bonds in chemical species. Traditionally, hybrid density functional approximations have been effective by averaging the exact exchange admixture across various compositions. However, they’ve struggled to achieve high-level quantum chemistry accuracy due to delocalization errors.

To address this, we introduce a novel approach to dynamically adjust hybrid functionals by generating optimal admixture ratios of exact exchange for any given chemical compound. This is achieved using highly data-efficient quantum machine learning models with minimal computational overhead.

Our adaptive Perdew-Burke-Ernzerhof based hybrid density functional (aPBE0) demonstrates remarkable accuracy in calculating atomization energies. Additionally, aPBE0 enhances the energetics, electron densities, and HOMO -LUMO gaps in organic molecules from the QM9 and QM7b datasets. We’ve also used aPBE0 with a large basis set to revise the entire QM9 dataset, resulting in more accurate quantum properties, including stronger covalent binding, larger band gaps, more localized electron densities, and greater dipole moments. While aPBE0 performs exceptionally well in equilibrium conditions, it does face limitations in accurately dissociating covalent bonds beyond the Coulson-Fisher point.