01.1 - Aircraft Design and Integrated System (Basics and Theory)
DATA-DRIVEN AERODYNAMIC SHAPE OPTIMIZATION AND MULTI-FIDELITY DESIGN EXPLORATION USING CONDITIONAL DIFFUSION-BASED GEOMETRY SAMPLING METHOD
A. Yang¹, J. Zhang, Shanghai Artificial Intelligence Laboratory, China; J. Li, Institute of High Performance Computing, A*STAR, Singapore; R. Liem¹; ¹Hong Kong University of Science and Technology, Hong Kong SAR of China
Aerodynamic shape optimization plays a crucial role in designing efficient and high-performance aircraft. However, the computational cost associated with high-dimensional data often poses a challenge in achieving accurate and efficient optimization. In this study, we propose a multifidelity Bayesian Neural Networks (BNNs) framework tailored for high-dimensional data-driven aerodynamic shape optimization. The framework combines data from multiple fidelity sources to enhance estimation accuracy while reducing computational expenses. The BNNs are used to capture the inherent uncertainties in modeling the relationship between low- and high-fidelity data. They also provide quantified uncertainty measures to optimize shape predictions. We employ variational inference and Monte Carlo sampling techniques to train and estimate posterior distributions of the BNNs. The proposed framework is validated using a high-dimensional dataset, showcasing its capability in accurately optimizing aerodynamic shapes while accounting for uncertainty. The results demonstrate liem{substantial} improvements in both optimization accuracy and computational efficiency compared to traditional approaches. This multifidelity BNNs framework has the potential to further enhance the capability of aerodynamic shape optimization in the aerospace industry, especially in tackling higher-dimensional problems.