An efficient neural network modeling method for rapid prediction of propeller aerodynamic characteristics

Artificial intelligence algorithms have been extensively utilized for the rapid prediction of aerodynamic characteristics, where the fidelity of predictions is predominantly influenced by the fidelity and distribution of aerodynamic datasets within the training set. However, acquiring high-fidelity datasets across the entire dataset range often incurs significant computational cost. To alleviate this computational cost, this study introduces an efficient sampling method grounded in micro-element partitioning, coupled with an adaptive datasets fusion method, to develop an efficient neural network modeling method. The present method not only optimizes the utilization of high-fidelity datasets but also minimizes the impact of low-fidelity datasets on prior models on high-fidelity datasets, facilitating the efficient initialization and precise prediction capabilities of the radial basis function (RBF) variable-fidelity model. The efficacy of the present method is substantiated through statistical outcomes derived from modeling analyses on benchmark functions. Additionally, when applied to the engineering scenario of predicting the aerodynamic characteristics of propellers, the present method demonstrates a notable enhancement in the convergence fidelity and modeling efficiency compared to other variable-fidelity modeling methods, utilizing only a minimal quantity of high-fidelity samples, thereby significantly reducing sampling cost. Notably, under the premise of constant fidelity, the computational cost for training samples is diminished by as much as 72.4%, while the prediction error is reduced by up to 39.7% compared to the traditional neural network models under the condition of consistent computational cost.