Abstract: An aerodynamic modeling method based on machine learning theory applying to large airspace and wide speed range is proposed. This method utilizes aerodynamic parameter identification results from flight trajectory simulation data and adopts neural network technology, so the aerodynamic model has a high precision fitting ability for the aerodynamic data with high nonlinear characteristics, regarding variable height, velocity, attitude angle and control rudder angle conditions. Firstly, the effects of the number of neural network and hidden neural unit number on model accuracy are discussed. The proper neural network layer numbers and hidden neural units are obtained through several neural network modeling calculations. Then we use the flight trajectory simulation data to get the whole aerodynamic parameters by identification method. By using BP neural networks, six aerodynamic models including three aerodynamic force coefficients and three aerodynamic moment coefficients are obtained. The models have high precision according to the comparison between model output results and origin data, as well as the aerodynamic prediction of four trajectories which are not applied to the modeling. The results show that the aerodynamic modeling method based on machine learning is feasible and effective. The aerodynamic models have a great fitting ability and some extrapolations in the ranges of data samples including height, Mach number, attack angle, sideslip angle, angular rate, and control surface angle. This study provides a new method on flight test experiment aerodynamic modeling and a good feasibility of the usage in the field of aerodynamic data mining, flight simulation, and aircraft flight evaluation.