Geometry Aware Deep Learning for Integrated Closed-Shell and Open-Shell Systems

Abstract

Simulations of chemical systems rely on calculation of their potential energy surfaces (PES), i.e., a function which returns the energy of a system under study. The electronic structure of a molecule may be closed-shell or open-shell, where either all electron spins are paired, or one or more electrons are unpaired in spin, respectively. While the cost of quantum-chemistry calculations can be reduced by assuming a closed-shell electronic structure and removing the necessity of the spin degree of freedom, it is often important to consider systems with unpaired spins, i.e. open-shell, such as in radical chemistry or description of chemical reactions. Here, we propose an extension for OrbNet-Equi, an equivariant deep-learning quantum mechanical approach to representing chemical systems at the electronic structure level. By utilizing a spin-polarized treatment of the underlying semi-empirical quantum mechanics featurization, OrbNet-Equi can describe both closed and open-shell electronic structures. We test the efficacy of this new representation with representative datasets.