Refining Potential Energy Surface through Dynamical Properties via Differentiable Molecular Simulation

TL;DR

This paper introduces a method to refine machine learning potentials for molecular dynamics by leveraging dynamical properties like spectroscopy, using differentiable simulation techniques to improve accuracy beyond traditional thermodynamic-based refinement.

Contribution

It demonstrates that dynamical data can be effectively used for potential refinement, overcoming memory and gradient issues, thus enhancing the accuracy of molecular simulations.

Findings

Dynamical property differentiation is feasible with adjoint and gradient truncation methods.

Transport coefficients and spectroscopic data improve the density functional theory-based potentials.

The approach addresses the inverse problem of spectroscopy by extracting microscopic interactions.

Abstract

Recently, machine learning potentials (MLP) largely enhances the reliability of molecular dynamics, but its accuracy is limited by the underlying $\textit{ab initio}$ methods. A viable approach to overcome this limitation is to refine the potential by learning from experimental data, which now can be done efficiently using modern automatic differentiation technique. However, potential refinement is mostly performed using thermodynamic properties, leaving the most accessible and informative dynamical data (like spectroscopy) unexploited. In this work, through a comprehensive application of adjoint and gradient truncation methods, we show that both memory and gradient explosion issues can be circumvented in many situations, so the dynamical property differentiation is well-behaved. Consequently, both transport coefficients and spectroscopic data can be used to improve the density functional theory based MLP towards higher accuracy. Essentially, this work contributes to the solution of the inverse problem of spectroscopy by extracting microscopic interactions from vibrational spectroscopic data.