An active learning surrogate model-based uncertainty quantification method for CFD simulation under the framework of evidence theory

Computational Fluid Dynamics (CFD) simulations are subject to various uncertainties, stemming from factors such as model parameters, numerical discretization, and boundary conditions. Given the flexibility of the evidence theory in modeling both aleatory and epistemic uncertainties in CFD simulations, this article introduces an active learning surrogate model-driven approach for uncertainty quantification in CFD simulations. This method aims to properly quantify the uncertainty of CFD simulations using minimal simulation model calls while achieving accurate uncertainty quantification results. The method utilizes the optimization-based max-min distance strategy to generate well-distributed candidate sample points. Moreover, it employs a dynamic entropy-weighted TOPSIS multi-criteria decision analysis to balance the surrogate model’s exploration, exploitation, and robustness. Additionally, this article proposes a composite convergence criterion, combining Hartley's measure and Jousselme distance, to formulate the stopping criterion of the surrogate model. Finally, taking the CFD simulation of the flow field of a supercritical wing with a NASASC(2)0410 airfoil as a case study, the uncertainty quantification of lift-to-drag ratio due to uncertainties in inflow and turbulence model parameters is conducted.