Development and application of $Δ$-machine-learned interatomic potentials to platinum-Nafion interfaces

Abstract: Machine-learned interatomic potentials (MLPs) have rapidly progressed in accuracy, speed, and data efficiency in recent years. However, training robust MLPs in multicomponent systems still remains a challenge. In this work, we demonstrate a method to physically inform a MLP by $\Delta$-learning the correction to the bonding term in a Dreiding potential for Nafion. The $\Delta$-MLP respects the internal bonds of the polymer but otherwise learns atomic interactions and reactivity from density functional theory training data. We use this $\Delta$-MLP to study the properties of the platinum-Nafion interface, analyzing polymer structure, proton mobility near and far from the platinum catalyst, and reaction pathway performance. The method shows promise in terms of data-efficient treatment of multicomponent systems, but further refinement is needed in the quality of input data to achieve combined insights into structure, transport, and reactivity at polymer-catalyst interfaces.