Enhancing surrogate-based aerodynamic shape optimization via dataset-independent generative modelling and dimensionality reduction

Surrogate-based optimization (SBO) has emerged as a critical tool for efficient global optimization of aircraft aerodynamic configurations driven by expensive computational fluid dynamics (CFD) simulations. However, its application is typically limited to problems with fewer than 100 design variables due to the ‘curse of dimensionality’. While data-driven dimensionality reduction techniques, such as mode decomposition and deep learning, offer potential solutions, they generally require extensive pre-existing high-quality geometric datasets, which are scarce for 3D configurations, especially unconventional aircraft. To address this data scarcity and dimensionality challenge, this study introduces a novel methodology that uses a dataset-independent generative model to assist dimensionality reduction of 3D geometric parameterization via free-form deformation (FFD). Central to this approach is to train a deep neural network to predict displacements of FFD control points, enabling the deformation of a baseline geometry to automatically generate diverse novel shapes. A Laplace energy-based smoothness metric is proposed and incorporated into the network’s loss function, guiding the generation towards physically realistic and aerodynamically smooth shapes. Subsequently, proper orthogonal decomposition (POD) is employed to derive a compact low-dimensional subspace that accurately captures the essence of the original high-dimensional design space. The proposed method is integrated within an SBO framework and validated through aerodynamic shape optimization case studies: the RAE2822 airfoil and the NASA Common Research Model (CRM) wing in transonic flow conditions. Results demonstrate that, compared to conventional FFD-based SBO approaches, it achieves significantly accelerated optimization convergence and produces optimal designs with superior geometric smoothness and quality.