Uncertainty in neural network potential predictions of interfacial dynamics

Density functional theory (DFT) simulations of heterogeneous catalytic interfaces offer atomic-level insights to reactivity, complementing experiments. Although often reasonably accurate, the limited time and length scales accessible in DFT are better suited to model static snapshots of atomic systems. Machine learning interatomic potentials trained to approximate DFT potential energy surfaces (PES) are promising to simulate extended-scale dynamics, key to tuning catalytic performance. Neural networks (NNs) effectively learn the complex reactivity patterns in DFT data, but lack inherent uncertainty measures. By using an ensemble of NNs, prediction uncertainties can be quantified to assess model confidence and improve the fitted PES – an area of ongoing research.

This work presents advancements to machine learning (ML) model development based on artificial and message passing NNs to simulate Pt(111)/H interface dynamics, for renewable hydrogen technologies. A specialized training dataset captures key catalytic effects. Ensembles trained via random subsampling demonstrate a correlation between uncertainty and generalization error. We propose computational frameworks to (1) apply ML ensembles to sample the configuration space, and (2) analyze global and local uncertainties to identify learning gaps, for active learning of the PES. Correlations of local uncertainty with unfamiliar atomic events reveal how the models learn the underlying physical and chemical behavior.