A Multi-fidelity Prediction with Convolutional Neural Networks Using High-Dimensional Data

This paper proposes two novel multi-fidelity neural network architectures tailored for high-dimensional inputs such as computational flow fields. The proposed methods are compared with the multi-fidelity deep neural networks from the literature using a 2-dimensional flow-varying supercritical airfoil problem. The main objective of this study is to generate a multi-fidelity prediction of aerodynamic coefficients using pressure coefficient fields around the airfoil. To generate the dataset, a coarse grid is solved using SU2 Euler solver for low-fidelity data whereas a relatively finer grid is utilized for high-fidelity data to obtain viscous solutions using the Spallart-Allmaras turbulence model. The performance metrics to compare the methods are determined as the test accuracy, physical training time, and the size of the high-fidelity samples. Results demonstrate that the proposed multi-fidelity neural network architectures outperform the multi-fidelity deep neural networks in predictive modeling using high dimensional inputs by improving the multi-fidelity prediction accuracy up to 78.7%.