Challenges for Machine Learning Force Fields in Reproducing Potential Energy Surfaces of Flexible Molecules

TL;DR

This paper evaluates the ability of current machine learning models to accurately reproduce complex potential energy surfaces of flexible molecules, highlighting their limitations and proposing a shift towards localized modeling approaches.

Contribution

It provides a comprehensive assessment of state-of-the-art ML models on complex PES and suggests using multiple local models instead of a single global model for better accuracy.

Findings

GAP/SOAP, SchNet, and sGDML achieve chemical accuracy with <1000 training points

Prediction accuracy varies significantly between equilibrium and transition regions

Current models face challenges in learning long-range interactions and using effective descriptors

Abstract

Dynamics of flexible molecules are often determined by an interplay between local chemical bond fluctuations and conformational changes driven by long-range electrostatics and van der Waals interactions. This interplay between interactions yields complex potential-energy surfaces (PES) with multiple minima and transition paths between them. In this work, we assess the performance of state-of-the-art Machine Learning (ML) models, namely sGDML, SchNet, GAP/SOAP, and BPNN for reproducing such PES, while using limited amounts of reference data. As a benchmark, we use the cis to trans thermal relaxation in an azobenzene molecule, where at least three different transition mechanisms should be considered. Although GAP/SOAP, SchNet, and sGDML models can globally achieve chemical accuracy of 1 kcal mol-1 with fewer than 1000 training points, predictions greatly depend on the ML method used as well as the local region of the PES being sampled. Within a given ML method, large differences can be found between predictions of close-to-equilibrium and transition regions, as well as for different transition mechanisms. We identify key challenges that the ML models face in learning long-range interactions and the intrinsic limitations of commonly used atom-based descriptors. All in all, our results suggest switching from learning the entire PES within a single model to using multiple local models with optimized descriptors, training sets, and architectures for different parts of complex PES.