基于表面无黏流动特征的摩阻分布机器学习

Accurate and efficient prediction of skin friction drag is crucial for aircraft design. However, the computation of the skin friction drag distribution is not only costly but also highly dependent on mesh density, turbulence patterns and numerical algorithms, while experimental measurements are more challenging. To this end, this paper proposes a data-driven machine-learning modeling method for skin friction drag distribution with high generalizability. Based on the numerical solution of Euler's equation, the method combines a small number of skin friction distribution samples computed by RANS to construct a model of the correlation relationship between the surface inviscid flow feature and the friction distribution, so as to realize the prediction of friction. Since the physical model of Euler's equation is embedded in this modeling method, the high generalizability and accuracy of the model can be ensured with few samples; on the other hand, compared with the numerical computation of RANS, the amount of computation is reduced by about one order of magnitude because only the Euler's equation is solved. The study demonstrates the effectiveness of the method for predicting variable geometric shapes skin friction in aerodynamic design by means of test cases for typical airfoils and wings. Compared to end-to-end distributed force deep learning modeling, the method achieves high modeling accuracy (drag error of about 2% ~ 3%) despite a fivefold reduction in sample size, and has strong generalization ability for working conditions and shape changes with low dispersion of results. This study provides a new and efficient research tool for the prediction of the friction distribution of attached-flow airfoils and the optimal design of airfoils.