Aerodynamic shape optimization via active learning-driven design space refinement

Abstract

In Aerodynamic Shape Optimization (ASO) studies, the ranges of design variables critically influence both optimization efficiency and solution quality. Traditional approaches typically maintain fixed boundaries based on experiential knowledge, struggling to balance computational efficiency with optimization outcomes. This paper proposes an innovative design space refinement framework driven by active learning. This framework offers a flexible and efficient strategy to identify a compact design space encompassing high-performance designs. Active learning enhances sampling efficiency, while the refinement mechanism progressively directs the optimization focus toward higher-performing designs. The proposed framework is validated through a comparative analysis with Principal Components Analysis (PCA), a prevalent method for reducing the number of design variables, on supercritical airfoil optimization. The results demonstrate that dynamically refining the boundaries of design variables can significantly improve the efficiency and effectiveness of ASO. By integrating active learning with adaptive design space refinement, the proposed approach effectively focuses the optimization process on regions likely to contain superior designs.