At high angles of attack, aircraft airflow exhibits strong nonlinear and unsteady characteristics, where traditional aerodynamic models struggle. This paper proposes a neural network-augmented physics-based model that integrates neural networks’ nonlinear prediction with dynamic derivative models’ physical accuracy, improving high-angle-of-attack aerodynamic prediction accuracy and robustness. Using open-loop excitation and spin flight data for training/simulation, results show the hybrid model has smaller aerodynamic coefficient prediction errors than traditional models and standalone LSTM networks, with better accuracy/robustness in spin simulations. This approach enhances extreme-flight-condition aerodynamic modeling, offering new insights for aircraft design and reliable foundations for flight dynamics and control system development.