Multifidelity flutter prediction using regression cokriging with adaptive sampling

Abstract

This work presents a flutter prediction approach that uses regression cokriging metamodels of generalized aerodynamic influence coefficients with adaptive sampling based on propagated model uncertainty along the flutter boundary. The use of regression cokriging models is compared to cokriging and regression cokriging with reinterpolation, as well as their single-fidelity counterparts. Comparisons to direct quantity-of-interest-based metamodeling are also shown. Several infill criteria based on the propagated flutter speed uncertainty are demonstrated on common flutter test cases. The value of adaptive sampling, multiple fidelity levels, and metamodeling of intermediate quantities is investigated by quantifying average cost and error metrics for the cases. Scalability with the number of structural modes is also investigated to gauge how the approach might fare for more conventional aircraft. Overall, the main benefits seen in this work stem from modeling intermediate quantities, with direct modeling costing six to eight times as much for multifidelity approaches, and three to five times as much for the single-fidelity comparators. In addition, using multiple fidelities was more accurate and required fewer infill points for convergence, leading to a cost savings of roughly 25% to 70%, depending on the case.