Aerodynamic data predictions based on multi-task learning

The quality of datasets is one the key factors that affect accuracy aerodynamic data models. For example, in uniformly sampled Burgers’ dataset, insufficient high-speed overwhelmed by massive low-speed data. Predicting more difficult than predicting data, owing to fact number limited, i.e. dataset not satisfactory. To improve datasets, traditional methods usually employ resampling technology produce enough for parts original before modeling, which increases computational costs. Motivated mixtures experts natural language processing, we propose a multi-task learning (MTL) scheme field predictions eliminate need resampling. Our MTL quality-adaptive scheme, combines task allocation and characteristic together disperse pressure an entire task. divides whole into several independent subtasks, while learns these subtasks simultaneously achieve better precisions. Two experiments with poor are conducted verify quality-adaptivity datasets. results show whose divided K-means accurate fully connected networks (FCNs), generative adversarial (GANs) radical basis function neural (RBFNNs)