Enhancing Uncertainty Quantification in Molecular Machine Learning: A Comparative Study of Deep Evidential Regression and Ensembles with Posthoc Calibration

Machine learning (ML) models are revolutionizing computational chemistry by enabling accurate predictions across diverse applications, including force field generation and chemical space exploration. However, the quality of these models heavily depends on the training data, necessitating a smart data approach to balance data quality and quantity. Uncertainty quantification (UQ) is pivotal in optimizing ML workflows, guiding data selection, and improving model predictions. In this study, we assessed calibration and UQ in molecular property prediction by modifying an equivariant graph neural network with a deep evidential layer (EGNN-DER) and using an ensemble of ANI models. Initial results revealed underconfident uncertainties, which were corrected using posthoc calibration methods (e.g., isotonic regression, standard scaling, and GPNormal). Adversarial group calibration demonstrated the robustness of the recalibrated uncertainties, enhancing active learning by reducing redundancy and targeting underrepresented data regions. Our findings demonstrate the importance of calibration for refining ML-based active learning strategies in chemistry.