Evidential Deep Learning: Enhancing Predictive Uncertainty Estimation for Earth System Science Applications

Uncertainty quantification is critical for reliable weather and climate modeling, yet challenging to compute robustly because of the various sources of uncertainty that propagate through workflows. Here we demonstrate evidential deep learning, integrating probabilistic modeling with deep neural networks, as an effective technique for predictive modeling and calibrated uncertainty estimation. Through classification and regression tasks in atmospheric science, we show evidential deep learning models attaining predictive accuracy rivaling standard methods, while robustly quantifying uncertainty. Uncertainty decomposition into aleatoric and epistemic components provides insights into uncertainty from data variability versus model limitations, respectively. This study compares the uncertainty metrics derived from evidential neural networks to those obtained from calibrated ensembles. Evidential neural networks result in comparable uncertainty estimates with a significant reduction in computational expenses. Analyses reveal links between uncertainties and underlying meteorological processes, facilitating interpretation. This study establishes evidential neural networks as an adaptable tool for augmenting neural network predictions across geoscience disciplines, overcoming the limitations of the prevailing uncertainty approach of ensembles of deterministic models. With the ability to produce robust uncertainties alongside predictions, evidential deep learning has the potential to transform weather and climate modeling, aiding critical analysis and decision-making under uncertainty. The conceptual simplicity, interpretability, and computational efficiency of evidential neural networks make them highly extensible, offering a promising approach for reliable and practical uncertainty quantification in Earth system science modeling.