This study presents a new framework for decomposing predictive uncertainty of an ensemble of probabilistic surrogate models, with a particular focus on Bayesian neural networks (BNNs). When existing uncertainty quantification methods distinguish between aleatoric and epistemic uncertainties, they often fail to rigorously disentangle model-form errors and data-induced randomness, especially in complex engineering applications. To address this limitation, we introduce an ensemble-based decomposition approach that systematically separates irreducible randomness from reducible uncertainties by leveraging an ensemble of BNNs independently trained with diverse architectures. In comprehensive numerical experiments—ranging from a synthetic one-dimensional example to time history predictions of hysteresis forces under stochastic ground motions—our method demonstrates its ability to yield tighter bounds on inherent variability and provides a more informative measure of epistemic uncertainty—specifically highlighting the reducible component that can be mitigated through additional data or improved modeling— compared to conventional approaches. This refined interpretation of uncertainty enables more robust reliability assessments and supports uncertainty-informed decision-making in engineering applications involving modeling limitations and inherent data noise. • Proposes an ensemble-based framework for uncertainty quantification using Bayesian neural networks (BNNs). • Reveals that conventional aleatoric estimates inadvertently conflate model-form uncertainty and variance discrepancy. • Reattributes model-form uncertainty and variance discrepancy from aleatoric to epistemic component to overcome limitations of model-specific uncertainty decomposition. • Defines ensemble-based aleatoric and epistemic uncertainty estimates based on BNN predictions. • Validates effectiveness on synthetic benchmarks and hysteresis response predictions under stochastic ground motions.
Jeon et al. (2026) studied this question.