TRANSFER LEARNING FOR REDUCED-ORDER MODELING OF TRANSONIC FLOWS USING MULTIFIDELITY DATA

Reduced-order modeling for multi-fidelity flow reconstruction enhances accuracy while reducing the costs associated with data generation. The success of multi-fidelity models hinges on accurately capturing the correlations between low- and high-fidelity data. In this work, we propose the use of transfer learning to identify representative features that more effectively correlate the multi-fidelity data, thereby improving the accuracy of multi-fidelity flow reconstruction. Essentially, transfer learning aids the modeling process by applying knowledge acquired from related tasks, thus enhancing performance in multi-fidelity scenarios. Specifically, we introduce a common class of transfer learning based on domain adaptation to uncover domain-invariant features from flow data across multiple sources. We establish two transfer learning frameworks, either through transfer component analysis or geodesic flow kernel, each offering distinct approaches to align transferred features across multi-fidelity data. These transferred features are then utilized to construct the bridge function between low- and high-fidelity data for flow reconstruction. The proposed transfer learning methods have been validated by two test cases, including transonic flows past a NACA0012 airfoil and an ONERA M6 wing. We present the advantages of our transfer learning approach in achieving superior feature representations and in facilitating the construction of more accurate multi-fidelity models for flow reconstruction.