Generalizable Multifidelity Aerodynamic Wing Shape Design Optimization

Despite the demonstrated success of traditional surrogate-based approaches in establishing the relationship between the design space and aerodynamic performance for specific aerodynamic shape optimization (ASO) problems, they still lack generalization for broader optimization tasks. To overcome this limitation, we propose a gradient-based, data-driven ASO framework for multifidelity, high-dimensional optimization. By integrating deep learning and a multifidelity approach with gradient-based optimization, our method efficiently scales to complex design tasks while using a single model to solve multiple ASO problems, thereby enhancing adaptability and efficiency. We use compact modal parameterization to capture the global representation of wing shapes while reducing design dimensionality. To enable fast and efficient design solutions, the multifidelity surrogate models are fine-tuned with residual learning to preserve accuracy and control model complexity before being integrated into gradient-based optimization. In this paper, we showcase the versatility of the derived multifidelity surrogate model by applying it to a range of single-point and multipoint ASO problems, producing aerodynamic optimization results that are closely aligned with those from adjoint-based computational fluid dynamics (CFD) solvers, while substantially reducing wall-clock time during the optimization stage after the initial CFD-driven training data have been prepared. With this approach, we offer a new data-driven pathway for tackling general high-fidelity ASO tasks.