A Gradient-based Sequential Multifidelity Approach to Multidisciplinary Design Optimization

Multifidelity design optimization is a strategy that can reduce the high computational cost in cases where the high-fidelity model is too expensive to use directly in optimization. However, current multifidelity approaches cannot handle the high-dimensional problems commonly encountered in industrial settings. Furthermore, they cannot accommodate arbitrary analysis fidelities, directly handle multidisciplinary problems, or provably converge to the high-fidelity optimum. In this paper, we present a practical multifidelity approach that leverages the advantages of conventional gradient-based approaches. Rather than constructing a multifidelity surrogate, we perform a sequence of single-fidelity gradient-based optimizations. The framework determines the appropriate fidelity and updates it during the optimization process. Finally, we demonstrate the proposed approach on a multipoint aerostructural wing optimization problem with over a hundred design variables. The multifidelity approach reduces the computational cost by 59% compared to the high-fidelity approach while obtaining the same numerical optimum.