A Multi-Fidelity Neural Network Framework for Enhancing Strip-Theory Aerodynamics

Kurzfassung

Accurate yet computationally efficient aerodynamic modeling remains a critical bottleneck in flight dynamics simulations of flexible aircraft. While low-fidelity analytical approaches like strip theory offer rapid inference speeds, they often exhibit significant discrepancies when compared to empirical wind-tunnel data, as observed during the testing of the TU-Flex flying demonstrator. To address this limitation, this paper proposes a multi-fidelity Scientific Machine Learning framework designed to enhance baseline strip theory predictions without the prohibitive computational costs of high-fidelity solvers. The methodology employs a sequential dual-network architecture. The first neural network employs transfer learning to effectively combine the aerodynamic trends captured by Computational Fluid Dynamics simulations with the magnitudes observed in experimental wind-tunnel measurements. This data mixing serves as a high-fidelity baseline to parameterize the physics-informed loss function of a second surrogate network. By embedding the governing equations of generalized aerodynamic forces into its training process, this second model maps operational flight conditions to quasi-steady modal forces, successfully recovering continuous spanwise lift and pitching moment distributions. Results demonstrate that the proposed framework accurately matches target high-fidelity load profiles while requiring minimal computational effort. This surrogate model establishes a pathway for future real-time aeroelastic evaluation and system identification within the TU-Flex flight simulation environment.