Hypernetwork-augmented LSTM improves unsteady aerodynamic reduced-order modeling

To enhance aerodynamic simulation efficiency, data-driven Reduced-Order Models (ROMs) for unsteady flows have been extensively studied. However, the substantial data requirements of deep neural networks increase modeling costs and undermine the computational advantages of ROMs. We address this issue by proposing a Long Short-Term Memory (LSTM)-ROM architecture integrated with hypernetworks that dynamically generate weights conditioned on velocity characteristics. Trained on limited harmonic excitation cases (as opposed to conventional stochastic samples), our framework achieves frequency-aware generalization through velocity-adaptive parameterization. This innovation significantly improves modeling accuracy while overcoming data scarcity constraints. Validations on unsteady aerodynamic modeling of NACA0012 airfoil demonstrate average 40% higher accuracy versus standard LSTMs. Crucially, this accuracy advancement enabled high-fidelity prediction of NASA Common Research Model (CRM) dynamic stall, with cross-validation confirming average 40% improvement in experimental data modeling capability for lift and moment coefficients under deep stall conditions. Quantitative statistical analysis of the model weight parameters confirms that the hypernetwork architecture enables adaptive parameter generation, which is further validated through cross-frequency predictions of lift and moment coefficients. By extending viable training data into the experimental domain, the proposed hypernetwork approach establishes a new paradigm for efficient data-driven unsteady aerodynamic modeling.