A Machine Learning Surrogate Framework for Wing Geometric Reverse Prediction and Closed-Loop Aerodynamic Optimization

Abstract High-fidelity computational fluid dynamics (CFD) simulation is widely adopted in traditional aircraft aerodynamic design,yet it suffers from excessive computational overhead and long iterative cycles when exploring massive wing geometric configurations.To alleviate the heavy reliance on repeated CFD calls,this work proposes a lightweight machine learning surrogate framework for aircraft design space exploration,which realizes reverse prediction of wing geometric parameters from unsteady lift and drag time-series aerodynamic signals. Two time-series preprocessing strategies are constructed to eliminate CFD inherent noise:the first only removes geometric outliers via the \(\:3{\sigma\:}\) criterion,while the second combines outlier elimination with low-pass smoothing to suppress high-frequency aerodynamic fluctuations.Two feedforward neural networks,including a shallow cascade network and a deep baseline network,are established for ablation comparative experiments on a parametric wing CFD dataset with 5000 samples.A complete closed-loop design pipeline integrating parametric geometric modeling,lightweight aerodynamic proxy simulation,and AI-driven geometric optimization is further developed for iterative wing shape adjustment. Experimental results demonstrate that the combination of smoothed aerodynamic time series and deep feedforward network achieves the best generalization performance with a test set \(\:{R}^{2}\)of 0.374.The proposed closed-loop workflow effectively narrows the deviation between actual lift and the predefined target value,drastically reducing repeated high-cost CFD simulations during design iteration.This study provides a feasible lightweight surrogate modeling solution for rapid configuration optimization of civil aircraft wings.