Ensemble adaptive gated multi-fidelity neural network for Bayesian optimization: Application to hydrofoil design

Journal
Ocean engineering  
Volume
343
Article Number
123314
Citation
Ocean engineering 343 (2): 123314 (2025)
Publisher DOI
10.1016/j.oceaneng.2025.123314
Publisher
Elsevier BV
High-fidelity computational fluid dynamics (CFD) simulations provide critical predictive accuracy in marine and ocean engineering design; however, their substantial computational expense often renders direct optimization infeasible. To alleviate this limitation, surrogate models approximate expensive objective functions from a finite set of observations, thereby enabling more tractable design exploration and optimization. Our objective is to build a novel, general-purpose multi-fidelity surrogate modeling approach that integrates seamlessly into a Bayesian optimization framework and remains robust under sparse high-fidelity data. We propose an adaptive gated multi-fidelity neural network (AGMF-Net), which incorporates three specialized expert subnetworks—linear, nonlinear, and residual—combined through a deep Mixture-of-Experts gating network that dynamically adjusts their contributions based on the input. To improve predictive uncertainty estimation, we ensemble multiple independently initialized AGMF-Net instances and use the resulting variance to guide sampling decisions. We embed this surrogate into a Bayesian optimization workflow driven by the logarithmic expected improvement acquisition function, which balances exploration and exploitation while maintaining numerical stability. We evaluated the proposed method against co-Kriging and the multi-fidelity neural network baseline on benchmark functions. AGMF-Net achieved higher initial predictive accuracy, rapidly converged to global optima, and maintained lower mean absolute relative error during optimization iterations. Finally, we applied the framework to a hydrofoil design optimization. The model successfully identified a subtle camber modification that improved the lift-to-drag ratio by 41.6 % compared to the baseline geometry, demonstrating that AGMF-Net can accelerate CFD-driven hydrodynamic design scenarios that combine sparse high-fidelity data with cheaper simulations. These results highlight the potential of adaptive gating and ensemble uncertainty quantification to accelerate design exploration and improve solution quality when only limited high-fidelity evaluations are feasible.
DDC Class
600: Technology