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

Beom Seok Kang, Vignesh C Bhethanabotla, Mohammadamin Tavakoli, William Goddard, Anima Anandkumar

Track: Extended abstract

Keywords: Geometry Aware, Quantum Mechanics, Equivariance

Abstract:

The description of potential energy surfaces (PES's), the differentiable function of a chemical system's energy with respect to its nuclear coordinates, is the underlying foundation to many computational chemistry studies. Prediction of such surfaces for open-shell systems (i.e. systems with unpaired electron spins) is essential for many of these calculations.
In this work, we present a proposed extension for utilizing OrbNet-Equi framework, an equivariant deep learning architecture operating on the atomic-orbital basis, to train closed-shell and open-shell systems together, beyond the original work's limit to the training of closed-shell molecules. In a similar treatment to the original work, we use the spin-polarized treatment of GFN-xTB as input features and predict energies and forces in the delta learning approach to a higher level DFT treatment of the potential energy surface of various molecular systems.

Submission Number: 79