Research On Modeling Methods For High Discrepant Aerodynamic Data Based On Multi-Task Learning

Deep learning has demonstrated significant potential in various fields as a machine learning technique in recent times.Traditional aerodynamics research requires a large amount of experimentation and computing resources,making traditional methods inefficient and costly.In contrast,deep learning is a data-driven approach that offers both efficiency and accuracy,making it a popular choice in the field of aerodynamics.Despite the wide-ranging potential of deep learning in aerodynamics,current neural network methods in aerodynamics have some shortcomings,particularly in considering the impact of data variability on prediction results.Data discrepancy refers to the differences between data from different sources or collection areas,which manifest as type discrepancy and distribution discrepancy,such as the external parameters of an airfoil shape and the state parameters of a airfoil during flight.Existing methods in aerodynamics only consider the mean or variance of the data,focusing on solving distribution discrepancy but failing to consider the impact of type discrepancy on model accuracy and robustness.This thesis suggests two techniques to evaluate the effect of data discrepancy on neural network models,in order to tackle these matters and enhance prediction proficiency.The first method is Large-Discrepancy Multi-Task Learning(LD-MTL)approach that combines multi-task learning and cluster networks to address type discrepancy,with multiple function networks for different input data types and a shared context network to learn the impact of each function network’s output on the final result.Experimental results show that LD-MTL achieves higher accuracy in aerodynamic prediction and can analyze the impact of various data types on the result.The second method is Multi-dimension GAN with discrepant inputs(MDGAN),which supplements sample data using a generative adversarial network to address distribution discrepancy before feeding the fused data into the multi-task learning module.Experimental results show that although MDGAN increases training and aerodynamic prediction times,it improves prediction accuracy compared to LD-MTL when accuracy is prioritized over efficiency.Overall,this article proposes two solutions for modeling large discrepancy aerodynamic data using neural networks,providing more effective methods and support for neural network research and application in the field of aerodynamics.