Discovering Reaction Mechanisms with Transition Path Sampling-Based Active Learning of Machine-Learned Potentials

TL;DR

This paper presents an active-learning framework using Transition Path Sampling to efficiently develop accurate machine-learned potentials for simulating reactive molecular processes, especially in transition states.

Contribution

It introduces a novel TPS-based active learning method that improves MLP accuracy in barrier regions without prior mechanistic knowledge.

Findings

Removes nonphysical artifacts in early models

Achieves near-DFT accuracy in energy and forces

Reveals multiple protonation mechanisms in CO2 reduction

Abstract

Machine-learned interatomic potentials (MLPs) provide near density functional theory (DFT) accuracy at reduced computational cost, but their reliability depends on representative training data and often deteriorates in transition-state regions governing rare events. We introduce an active-learning framework in which Transition Path Sampling (TPS) serves as a targeted data-generation engine for constructing MLPs accurate in barrier regions. TPS generates ensembles of unbiased reactive trajectories, and a committee-based uncertainty estimate identifies configurations for selective DFT labeling and retraining. Iterating this cycle systematically refines the potential energy surface in dynamically relevant regions, without the need of prior knowledge of the mechanism. Applied to electrochemical CO$_2$ reduction to CO on copper in explicit water, the approach removes nonphysical artifacts present in early models, achieves near-DFT energy and force accuracy, and enables stable long-time sampling of reactive pathways. Extended TPS simulations reveal multiple dynamically accessible protonation mechanisms. This work establishes TPS as an efficient and principled active-learning strategy for reactive molecular simulations at electrochemical interfaces.