Physics-guided dual-stage surrogate framework for precisely predicting the behavior and flow field of an airfoil system undergoing flutter

초록

Flutter provides a viable mechanism for wind energy harvesters installed on offshore structures via sustained, large-amplitude oscillations. However, the intrinsic complexity of nonlinearity and modal coupling complicates accurate predictive modeling. This study newly proposes a physics-guided dual-stage surrogate framework to analyze airfoil flutter. In Stage I, a physics-informed neural network resolves time-evolving structural responses and aerodynamic loads. Guided by governing equations of motion, this stage simultaneously identifies critical structural parameters through a tri-phasic learning process. In Stage II, a separable physics-informed neural network reconstructs the unsteady flow field by integrating spatial-temporal coordinates with data from the structural stage. Validation against high-fidelity reference data shows that the proposed framework predicts the responses with small RMSE values, corresponding to relative errors of 0.10% for both cases. Under sparse interior supervision, the Stage II model reconstructs the unsteady flow field with a full-domain combined relative L 2 error of 0.034, with a component-wise maximum relative L 2 error of 0.047, while maintaining physical consistency in both the body-adjacent and wake regions. The total training time of the proposed framework was 1 h 37 min, compared with 8 h 43 min for the reference URANS simulation, indicating its practical efficiency for early-stage prediction and design analysis.

키워드