A Comparative Study on Multi-fidelity Machine Learning Modeling for Aerospace Problems

Multi-fidelity modeling seeks to establish an efficient framework that combines mathematical models of varying complexities, providing accurate approximations while minimizing computational burden. This approach is especially appealing for balancing prediction accuracy and computational cost. We present a comparative study of machine learning-based multi-fidelity analysis tailored to aerospace engineering problems, such as sonic boom prediction using the JWB aircraft model and aeroelastic datasets generated parametrically using the uCRM 13.5 wing. A novel 2-step multi-fidelity support vector regression is proposed and compared with existing multi-fidelity deep neural networks and 2-step multi-fidelity deep neural networks from the literature. The main objective of this study is to provide an alternative approach to neural network-based multi-fidelity models. Three mathematical benchmark problems and three aerospace engineering problems are used to evaluate the performance of the proposed multi-fidelity model. Low- and high-fidelity flow solutions are obtained using the PANAIR program and SU2 suite, respectively, while ground loudness values are computed via NASA's sBOOM code. The evaluation metrics are selected as the cross-validation score, physical training time, and the number of hyperparameters. The proposed 2-step MFSVR method reduces the number of hyperparameters and computational cost while maintaining similar accuracy, offering an efficient alternative to traditional MFDNN models.