Outlier-aware multi-fidelity surrogate-based hyperparameter optimization for topology optimization

This paper proposes an outlier-aware multi-fidelity topology optimization framework for improving topology optimization under a limited high-fidelity computational budget. Although multi-fidelity surrogate modeling reduces high-fidelity cost using inexpensive low-fidelity data, its direct combination with topology optimization is often hindered by infeasible or physically meaningless designs that distort the surrogate and mislead optimization. To address this issue, the proposed approach establishes a unified pipeline in the hyperparameter design space, where outliers are systematically detected, labeled, and exploited throughout low-fidelity and high-fidelity data generation and multi-fidelity surrogate-based optimization. First, numerous low-fidelity samples are generated using early-terminated topology optimization runs to efficiently explore the design space. Second, low-fidelity outlier labeling identifies infeasible and physically meaningless configurations. Third, high-fidelity samples are allocated using a low-fidelity-guided design of experiments that prioritizes promising regions while avoiding outliers. Fourth, high-fidelity outlier labeling exploits the low-fidelity-guided outlier information to classify limited high-fidelity samples into inliers and outliers. Finally, an outlier-aware multi-fidelity surrogate-based optimization combines an inlier-trained regressor with an outlier-aware acquisition function to efficiently identify high-performing settings. Numerical results on two benchmark problems demonstrate that the proposed method achieves better optimal solutions than conventional methods under comparable high-fidelity budgets.