Abstract
Deep learning has been widely applied in surrogate modeling for airfoil flow field predic-tion.The success of deep learning relies heavily on large-scale,high-quality labeled samples.How-ever,acquiring labeled samples with complete annotations is prohibitively expensive,and the available annotations in practical engineering are often sparse due to limited observation.To lever-age samples with sparse annotations,this paper proposes an uncertainty-based active transfer learn-ing method.The most valuable positions in the flow field are selected based on uncertainty for annotation,effectively improving prediction accuracy and reducing annotation costs.Our method involves a novel active annotation based on synchronous quantile regression,which can mitigate the computational cost of query annotation.Besides,a novel quantile levels-based consistency regular-ization is proposed to constrain the remaining unlabeled regions and further improve the model per-formance.Experiments show that our method can significantly reduce prediction errors with only 1％extra annotations,and is a promising tool for achieving rapid and accurate flow field prediction.
万方数据