Distribution-free total uncertainty quantification for aerospace flow field prediction: integrated aleatory and epistemic sources

Deep learning-based reliability analysis of aerospace flow fields encounters significant challenges due to multi-source uncertainties while existing methods primarily address individual types of uncertainty. Therefore, determining the total uncertainty and identifying the dominant sources are crucial. This study proposes a three-stage framework that integrates uncertainty decomposition, quantification, and aggregation. The key innovation is the development of a hybrid Monte Carlo strategy: Monte Carlo Dropout is used to probabilistically characterize epistemic uncertainty arising from model structure defects, while Monte Carlo Quantile Regression is employed to interpret aleatory uncertainty inherent in measurement data. Through systematic source decomposition, the dominant uncertainty contributors are identified. Aggregation via variance summation results in total uncertainty estimates. Validation using an aerospace flow field prediction case reveals two critical findings. First, the proposed method enables comprehensive uncertainty evaluation by simultaneously addressing both aleatory and epistemic components, outperforming conventional single-source approaches. Second, epistemic uncertainty contributes more significantly to the total variance, establishing it as the dominant factor influencing prediction reliability. This quantification provides engineers with valuable insights into uncertainty source identification, thereby enhancing the accuracy of flight state predictions and system reliability analysis.