Transferable machine learning model for the aerodynamic prediction of swept wings.

This article focuses on a novel machine learning framework designed to predict surface pressure and friction distributions on swept wings with various planform geometries, including complex kink wings, using fewer training samples than traditional methods. Instead of relying on global planform parameters, the model inputs distributed local geometric parameters along the wing span, processed via convolutional neural networks (CNNs) or recurrent neural networks (RNNs), followed by a ResNet-based decoder to reconstruct aerodynamic surface quantities. Trained on datasets of simple single-segment wings and transferred to kink wings through fine-tuning, the model significantly reduces prediction errors—by up to 57.6% for lift coefficient—and lowers the sample size needed for accurate predictions, demonstrating improved generalization and transferability. This approach offers a promising surrogate modeling technique to accelerate aerodynamic shape optimization for engineering-practical wings with complex geometries, although current work assumes fixed cross-sectional airfoil shapes.