Comparison of ensemble-based uncertainty quantification methods for neural network interatomic potential

Neural network interatomic potentials (NNIPs) has gained traction in atomistic simulations due to their accuracy that rivals first-principle calculations, while requiring considerably lower computational costs. The precision of these models is critical for assessing the reliability of their predictions. However, understanding the sources of uncertainty inherent in their development remains an open question. In this study, we investigate how various sources of uncertainty in NNIP models impact model performance and predictions. We compare several ensemble-based uncertainty quantification (UQ) methods commonly applied to NNIPs. Our focus is on the development of an NNIP for carbon systems, trained on diverse carbon structures such as diamond, graphene, and graphite. Our findings indicate that the dominant sources of uncertainty depend on the atomic structures and material properties being predicted. Additionally, the hyperparameters and characteristics of each UQ method significantly influence the resulting uncertainty estimates. From these observations, we suggest using multiple UQ methods in tandem to capture a more comprehensive understanding of uncertainty, thereby enhancing the reliability and robustness of NNIP predictions across different material systems.