Parallel Multifidelity Design of Experiment Strategy Considering Low-Fidelity Simulation Feasibility

Abstract

For computationally demanding engineering applications with variable-fidelity numerical simulators, multifidelity surrogate modeling has emerged as a promising approach to mitigate the computational cost. This paper proposes a novel sequential design of experiments (DoE) approach to efficiently enhance the global predictive accuracy of multifidelity co-kriging surrogate models. To leverage the capabilities of high-performance computing platforms, a parallel updating scheme is proposed for simultaneously identifying multiple experiments. To balance high-fidelity and low-fidelity data acquisition, our proposed DoE simultaneously determines the optimal location and fidelity for each experiment. To address low-fidelity simulation failures due to factors such as coarse mesh and modeling errors, a probabilistic binary classifier is introduced to identify undesirable low-fidelity input regions. Through a series of academic benchmark examples and practical computational fluid dynamics (CFD)–enabled aerodynamic building shape design, the proposed sequential strategy significantly reduces the number of expensive model evaluations while maintaining accurate approximations.