Agnostic Multi-Fidelity Regression for Aerospace Design Applications

Presentation
Tuesday 12.4.2022, h.11.00 Eastern Time (UTC-4) (17.00 in Italy)

Minisymposium
MS10 – Metamodel-Based Approaches for Robust (Stochastic) Inversion and Optimization – Part I of III

www.siam.org/conferences/cm/conference/uq22

Giulio Gori (presenter), Dipartimento di Scienze e Tecnologie Aerospaziali, Politecnico di Milano, via La Masa 34, 20156, Milano, Italy
Olivier Le Maître, CNRS/INRIA/CMAP, École Polytechnique, IPP, Route de Saclay, 91128 Palaiseau, France
Pietro Marco Congedo, INRIA/CMAP, École Polytechnique, IPP, Route de Saclay, 91128 Palaiseau, France

Abstract
Multi-fidelity regression models bring substantial advantages to the aircraft preliminary design phase, when data from computer simulations and preliminary experimental tests are exploited to define the best feasible configuration. In the multi-fidelity approach to the modeling issue, pieces of information of diverse fidelity and complexity are leveraged concurrently, to improve estimate accuracy and to reduce the burden associated to parametrization. In such setting, it is fundamental to establish the correct hierarchy in terms of data credibility w.r.t. the targeted application. Unfortunately, the complexity challenging aerospace design problems generally makes the direct estimation of data fidelity difficult, if not intractable. Therefore, the fidelity hierarchy is often established according to a hypotheses-driven process, hence leaving ground to modeling biases. We aim at developing an agnostic multi-fidelity regression framework robust to possible modeling biases due to a corrupted Modeler’s prior belief concerning the alleged fidelity of the available data sets. In particular, we focus on multi-fidelity co-kriging methods, proposing an extended formulation capable of mitigating the drawbacks of an ill-chosen fidelity hierarchy.