A Multi-Task Learning Method Combined with GAN for Aerodynamic Prediction

The geometric parameters of the airfoil and the flight state parameters are two different types of parameters. These two types of parameters have great discrepancies in data types and will have different degrees of influence on the aerodynamic coefficient of the aircraft. If this discrepant is ignored, it will cause the loss of model accuracy while predicting aerodynamic forces by these two kinds of parameters. Inspired by the idea of multi-task learning, multi-task learning – gan (MLG) model is designed to use large-discrepant data for modeling. GAN is adopted by MLG to complement the training set, and then divides the large-discrepant data into different subtasks for learning separately, finally the results of each subtask are fused through the context network. Experiment results prove that MLG has better performance than traditional methods while modeling for aerodynamic prediction based on large-discrepant data.