Application of Multi-Fidelity Transfer Learning with Autoencoders for Efficient Construction of Surrogate Models

作者：Yiren Shen, Harsh C. Patel, Zan Xu, Juan J. Alonso · 年份：2024 · DOI：10.2514/6.2024-0013 · 被引用次数：10 · 研究领域：Model Reduction and Neural Networks、Ultrasonics and Acoustic Wave Propagation、Structural Health Monitoring Techniques

The infusion of machine learning techniques into aerodynamic surrogate modeling stands out for its considerable reduction in computational costs, effectively capturing intricate flow dynamics, and complementing costly partial differential equation simulations. Despite their effectiveness, these machine learning-based models grapple with challenges related to generalization and computational demands. Ongoing research endeavors are concentrated on refining their accuracy and efficiency, with a specific emphasis on techniques like transfer learning and the integration of multi-fidelity data. The primary objective of this study is to craft a surrogate model for aerodynamic prediction that strikes a balance between efficiency and accuracy. The strategy involves harnessing the power of transfer learning along with multi-fidelity data, aiming to curtail computational expenses while preserving precision. This is achieved by leveraging abundant low-fidelity data for initial training and subsequently fine-tuning the model with scarce high-fidelity data. The surrogate network, adopting an autoencoder-decoder architecture with super-resolution augmentation, is designed to map low-fidelity flow variables to their high-fidelity counterparts. The model is meticulously trained on a dataset encompassing Euler and Reynolds-averaged Navier-Stokes simulations at three different fidelity levels. The results showcase the model’s adeptness in predicting with efficiency and significant computational...