High-fidelity aerodynamic prediction via a physics-informed transformer framework

Accurate prediction of aerodynamic coefficients for systems with strongly coupled and nonlinear dynamics, such as microspoiler-controlled vehicles, remains challenging in aerospace engineering. Purely data-driven models often lack physical consistency, while classical Physics-Informed Neural Networks are difficult to apply directly to complex three-dimensional aerodynamic flows. This paper proposes the Aero Physics-Informed Network, a data-assisted physics-guided surrogate framework for aerodynamic coefficient prediction. The framework first constructs physically meaningful aerodynamic descriptors from raw operating parameters through physics-guided feature engineering, thereby improving the physical relevance of the input representation without direct PDE-residual optimization. These features are then processed by an Ensemble Kalman Filter-based Physics-Informed Attention mechanism, which refines feature weighting through physically guided updates during learning. The proposed model is trained on high-fidelity CFD data benchmarked against wind tunnel measurements. Comparative, ablation, and robustness studies show that APINN achieves improved predictive accuracy and maintains stronger performance than conventional deep learning baselines under noisy, sparse, and perturbed data conditions. These results indicate that embedding aerodynamic prior knowledge into both feature construction and attention refinement provides an effective and robust surrogate modeling strategy for complex aerodynamic prediction tasks.