Towards Large-Scale Condensed Phase Simulations using Machine Learned Energy Functions

TL;DR

This paper introduces a machine learning-based workflow for creating accurate and efficient energy functions for large-scale molecular dynamics simulations of condensed phase systems, exemplified by water.

Contribution

It combines neural network and empirical models to decompose and fit energy contributions, enabling scalable MD simulations with improved accuracy.

Findings

Achieved good agreement with experimental properties of water.

Performed multi-nanosecond MD simulations with thousands of molecules.

Demonstrated the potential of ML-inspired parametrization for condensed phase simulations.

Abstract

Accurate, yet computationally efficient energy functions are essential for state-of-the art molecular dynamics (MD) studies of condensed phase systems. Here, a generic workflow based on a combination of machine learning-based and empirical representations of intra- and intermolecular interactions is presented. The total energy is decomposed into internal contributions, and electrostatic and van der Waals interactions between monomers. The monomer potential energy surface is described using a neural network, whereas for the electrostatics the flexible minimally distributed charge model is employed. Remaining contributions between reference energies from electronic structure calculations and the model are fitted to standard Lennard-Jones (12-6) terms. For water as a topical example, reference energies for the monomers are determined from CCSD(T)-F12 calculations whereas for an ensemble of cluster structures containing $[2,60]$ and $[2,4]$ monomers DFT and CCSD(T) energies, respectively, were used to best match the van der Waals contributions. Based on the bulk liquid density and heat of vaporization, the best-performing set of LJ(12-6) parameters was selected and a wide range of condensed phase properties were determined and compared with experiment. MD Simulations on the multiple-nanosecond time scale were carried out for water boxes containing 2000 to 8000 monomers, depending on the property considered. The performance of such a generic ML-inspired parametrization scheme is very promising and future improvements and extensions are discussed, also in view of recent advances for water in particular in the literature.