High-fidelity modeling of unsteady aerodynamic loads under structural vibration using dual modal spaces and LSTM networks

Rapid and accurate prediction of aerodynamic load distribution induced by structural vibrations is critical for aircraft performance evaluation and safety assurance. However, for this kind of strong unsteady flow prediction problems, existing proper orthogonal decomposition (POD) based reduced-order models (ROMs) face significant challenges, including difficulty in mapping physical inputs (e.g., structural deformations) to modal coefficients due to modal mixing phenomena, and convergence issues in surrogate modeling caused by high-frequency oscillations of higher-order modal coefficients. This study proposes a Dual Modal Space (DMS) model that integrates structural modes and aerodynamic POD modes into a unified reduced low-dimensional space, establishing mappings between high-dimensional aerodynamic load field and low-dimensional generalized aerodynamic forces. Since the generalized aerodynamic forces retain explicit physical property and are obtained through projection onto structural modes, the spatial integration filters out small-scale, high-frequency flow structures that are poorly correlated with structural responses. As a result, the modal mixing phenomenon is avoided and the high-frequency oscillations present in higher-order aerodynamic POD coefficients are suppressed. Combining DMS with Long Short-Term Memory (LSTM) neural networks yields an efficient surrogate model for predicting generalized aerodynamic forces under structural vibrations. Using the dual modal mapping matrix created in DMS, the aerodynamic load distributions can be precisely reconstructed from predicted generalized aerodynamic forces. Numerical validation under subsonic conditions shows that DMS-LSTM maintains over 97.2 % prediction accuracy of aerodynamic load distribution. Compared to traditional POD-LSTM models, prediction errors were reduced by 50.762 % for a 2D NACA65A004 airfoil, and 50.134 % for a 3D AGARD445.6 wing. The proposed model effectively mitigates the accuracy limitations of traditional POD-ROM under structural vibration conditions, and has potential to facilitate open-loop unsteady aerodynamic load prediction for prescribed structural motions in the subsonic regime, within the tested parametric range.