Optimizing Prediction of Chemical Bonds in Interfacial Dynamics through Local Uncertainty Estimates with Neural Network Ensembles
We present a framework for data-efficient training of machine-learning interatomic potentials for interfacial chemistry, especially heterogeneous catalytic systems. We establish strategies for density functional theory training data generation consisting of procedurally generated bulk, surface, and gas-phase atomic geometries, as well as moderately randomized structures. We show how ensembles of neural network machine-learning interatomic potentials trained on different splits of these training structures yield reliable uncertainty estimates at the atomic node energy level. Our models can thus identify which atomic sites and chemical bonds in a system lead to uncertainties in the predicted potential energy surface. Using hydrogen interacting with platinum as a test case, we find that the atomic uncertainty estimates identify both unphysical bonding scenarios and physically relevant interactions that are underrepresented in the original training data, such as surface diffusion, bond breaking, and bond formation. Building on these insights, we propose local uncertainty-informed strategies that flag outliers via statistical correlations, thereby improving active learning efficiency and enhancing the reliability of neural network-based potentials for extended-scale reactive dynamics.