Nonlinear unsteady aerodynamic modeling at high angles of attack is critical for high-precision control law design of modern aircraft. Current modeling approaches primarily fall into two categories: expert’s experience-informed models and data-driven models. The accuracy of expert’s experience-informed models is limited by the a priori expression terms. Data-driven model has a strong nonlinear mapping ability, but its performance depends on sample size and has insufficient generalization ability in small samples. To address these limitations, this paper proposes a physics-informed data-driven modeling framework, in which a Long Short-Term Memory (LSTM) neural network is trained to reconstruct the a priori expression terms in the differential equation model. While retaining the physical mechanism of the expert’s experience-informed model, the data-driven method is utilized to enhance the prediction accuracy of the model. To validate the model, this paper conducts missile single-degree-of-freedom pitching and fighter two-degree-of-freedom aerodynamics modeling at high angles of attack. Results show that, compared to a traditional differential equation model, a standalone LSTM network, and a hybrid multi-fidelity neural network, the proposed method achieves superior accuracy and generalizability in both cases, providing an effective solution for modeling complex nonlinear unsteady aerodynamic behaviors.