Density functional latent space representations of dynamical densities and potentials

The combinatorically-large phase space of possible interactions among atomic elements represents both a challenge and an opportunity for machine learning the structures and compositions of novel and useful biomolecules and materials. There is growing interest in combining physics-based models with purely data-driven approaches, in order to optimize the fidelity and efficiency of interaction potentials for atomistic molecular dynamics simulations. The ensemble-charge transfer embedded atom method (ECT-EAM)¹ utilizes energy functionals and atomic density latent variables to flexibly represent complex and dynamically-evolving quantum mechanical interactions among atoms, including bond formation and breaking. We describe the principled construction of these latent variables from density functional theory, and show how it effects a rigorous coupling between the electronic and atomistic length scales. We also discuss how a latent variable neural network representation of the electron density can provide surprising insight into the balance between ionic and covalent bonding as a function of molecular geometry.

[1] S. R. Atlas. “Embedding quantum statistical excitations in a classical force field,” J. Phys. Chem. A, 125, 3760 (2021); K. Muralidharan, S. M. Valone, and S. R. Atlas. “Environment dependent charge potential for water,” arXiv:0705.0857 [cond-mat.mtrl-sci] (2007).