Multi-fidelity Bayesian neural networks for complex aircraft configurations: CRM and M6 case studies

Abstract

In practical engineering, the fidelity of aerodynamic data for aircraft is often proportional to its acquisition cost, leading to a scarcity of high-fidelity (HF) data. Data fusion addresses this challenge by strategically integrating abundant low-fidelity (LF) data with limited HF data, enabling high-accuracy predictions at reduced costs. This approach effectively balances cost and fidelity trade-offs in diverse engineering applications. Building upon our previous work that introduced a Multi-Fidelity Bayesian Neural Network (MFBNN) model for aerodynamic data fusion, the present study introduces several key innovations to enhance its practicality. This work extends MFBNN’s applicability, elucidates the effects of data quality on model performance, and integrates transfer learning to improve generalization and efficiency in practical aerodynamic modeling. Specifically, we investigate key technical challenges in practical MFBNN deployment, including (1) the impact of HF dataset size on model performance, providing guidance for optimal HF data selection; (2) the use of transfer learning (TL) to expedite model adaptation for new flow conditions; and (3) the extension of MFBNN’s input dimensions to four and five by incorporating inflow parameters (Mach number and angle of attack). The results demonstrate the MFBNN’s capability to construct robust, generalized data fusion models adaptable to varying flow conditions, highlighting its potential for real-world aerodynamic design and analysis.