A dual-network framework integrating adaptive data refinement BPNN for aeroelastic stability prediction and inverse design in tilt rotor/wing system

To address the challenges of aeroelastic stability prediction and generative design for tilt rotor/wing system, this study proposes a dual-network framework based on a Back-Propagation Neural Network (BPNN) for forward prediction of whirl flutter boundaries and inverse design of dynamic parameters. A multi-modal coupled aeroelastic stability analysis model is established using multi-body dynamics methods. Datasets are generated for parameters strongly correlated with system stability to develop a Forward Prediction Neural Network (FPNN). An adaptive data refinement technique is introduced and integrated into the Inverse Prediction Neural Network (IPNN) to enhance critical-regions accuracy. The results demonstrate that within the training range, maximum prediction errors for the whirl flutter boundary and critical speed are 3.231% and 0.00343%, respectively, while the FPNN reduces computational time by four orders of magnitude (0.0012 s vs 15.35 s). Outside the range, errors remain as low as 5.858% and 0.0151%, indicating good generalization ability. The refinement technique effectively suppresses peak-to-peak fluctuations during modal transitions and near critical points, improving both accuracy and reliability. Furthermore, the dual-network inverse design framework achieves dynamic parameters prediction with relative errors below 1% and enables closed-loop validation. This neural-network-based approach establishes a novel paradigm for rapid and cost-effective design of tilt rotor/wing system.