Multifidelity Prediction Framework with Convolutional Neural Networks Using High-Dimensional Data

This paper proposes two novel multifidelity neural network architectures developed for high-dimensional inputs such as computational flowfields. We employed a two-dimensional flow-varying transonic supercritical airfoil problem while exploring and comparing our methods with the former "multifidelity deep neural networks" from the literature. We call these novel methods "modified multifidelity deep neural networks" and "multifidelity convolutional neural networks." The main objective of this study is to establish an advanced multifidelity prediction framework that can be applied to any computational data; however, here, we applied our methods to the prediction of aerodynamic coefficients using pressure coefficient fields around the airfoil. To generate the dataset, first, a coarse grid is employed using the SU2 Euler solver for low-fidelity data; then, a relatively finer grid is used to obtain the viscous solutions by using the Spalart-Allmaras turbulence model for the high-fidelity data. The performance metrics to compare the methods are determined as the test accuracy, the physical training time, and the size of the high-fidelity samples. The results demonstrate that the proposed multifidelity neural network architectures outperform the multifidelity deep neural networks in predictive modeling using high-dimensional inputs by improving the multifidelity prediction accuracy significantly.