Efficient Interval-Based Uncertainty Quantification for Model Validation and Predictive Capability

This paper presents a methodological framework for tackling the Second Uncertainty Quantification Challenge Problem for Aerodynamics. The challenge requires assessments of validation and predictive capability for an aerodynamic tool given imperfect experimental measurements that are scarce and imprecise. We propose an efficient interval-based integrated approach that comprehensively accounts for multiple sources of uncertainties including experimental data uncertainties, inputs incertitude, model-form discrepancy, numerical approximation, surrogate model prediction, and, importantly, uncertainties related to extrapolated application domains. These uncertainties are tackled by efficient meta-modelling techniques involving interval predictor models, Bayesian optimization on the basis of Gaussian processes, and stochastic neural network models. Our results underscore the effectiveness and efficiency of the approach in achieving a practical balance between comprehensive uncertainty quantification and computational cost, demonstrating broad applicability for uncertainty-driven validation and assessment of predictive capability in engineering applications.