Multi-fidelity aerodynamic data fusion by autoencoder transfer learning
Researchers have developed a novel deep learning framework for aerodynamic data fusion, combining autoencoder transfer learning with a Multi-Split Conformal Prediction (MSCP) strategy. This approach effectively utilizes abundant low-fidelity data to learn a physics representation, which is then fine-tuned with minimal high-fidelity samples. The method has demonstrated success in predicting surface pressures for airfoils and wings with high accuracy and providing robust uncertainty quantification, exceeding 95% pointwise coverage. AI