Uncertainty Quantification of Aerodynamic Coefficients Based on Improved Kolmogorov–Arnold Networks

To address insufficient accuracy in uncertainty quantification (UQ) and the limitation of existing studies, which focus mainly on 2D airfoils and fail to adapt to 3D flight vehicles, an aerodynamic coefficient UQ method based on the improved Kolmogorov–Arnold network (KAN) is proposed. Targeted improvements are made to the original KAN to construct a high-precision surrogate model, which is coupled with the Monte Carlo (MC) method to form the KAN-MC UQ framework. Validations are conducted on 2D airfoil and 3D unmanned aerial vehicles (UAVs) datasets, with multidimensional comparisons to multilayer perceptron coupled with MC (MLP-MC) and polynomial chaos expansion methods. Results show that the improved KAN exhibits excellent prediction performance: coefficients of determination for lift, drag, and pitching moment coefficients reach 0.9892, 0.9527, and 0.9657 on the airfoil dataset and 0.9977, 0.9621, and 0.9492 on the UAV dataset, significantly outperforming the hyperparameter-optimized MLP, especially for the small-magnitude drag coefficient. UQ analysis shows that KAN-MC accurately identifies dominant parameters and that its calculated aerodynamic coefficient statistical characteristics reflect uncertainty fluctuations. This method expands aerodynamic UQ from 2D airfoils to 3D UAVs, providing an efficient tool for 3D flight vehicle aerodynamic design and robust optimization.