Data-Driven Modeling of Aerodynamic Loadings for Tiltrotor Pylon using Multi-Fidelity CFD Data

Abstract

Data-driven modeling for aerodynamic loads has been used to provide relatively fast predictions of aerodynamic loads for airfoils, wings, rotors, and other applications. The data-driven models use data from physics-based or reduced-order models to develop surrogate models that can be used to provide fast/accurate predictions of aerodynamic loads in subsequent simulations. Most of these models use a single fidelity data, while some employ data fusion techniques on multi-fidelity data to improve the efficiency/accuracy of the surrogate models. In this study a multi-fidelity approach based on Gaussian process regression (GPR) is used to provide data-driven multi-fidelity models (MFM) for fast prediction of aerodynamic loadings on a tiltrotor pylon at various pitch/yaw angles. Three sets of training data are obtained using HPCMP CREATE-AV Helios software using FUN3D (medium- and high-fidelity data sets) and a Reduced Order Aerodynamic Model (low-fidelity data set). A total of 106 CFD simulations (50 low-fidelity, 42 medium-fidelity, and 14 high-fidelity results) is obtained using a total of 84K, 149K, and 628K core hours, respectively, for use as training and validation data. The outputs of the MFM models are the predicted values and variances of the pylon lift, drag, and moment at specific pitch/yaw angles. Three MFM models are explored using training data from – (i) low/high-fidelity, (ii) medium/high-fidelity, and (iii) low/med-ium/high-fidelity data sets. The MFM models using low/high-fidelity and low/medium/high-fidelity CFD data as training data produce the MFM models with reasonable accuracy. The low accuracy of the predicted MFM model using medium/high-fidelity training data is due to the large variance of the data at high pylon pitch angle (with substantial separated flow) caused by the coarseness of the volumetric medium-fidelity grid. When the numbers of high-fidelity CFD training data is increased from 3 to 4 and 5 data points, the accuracy of the MFM model as measured by two-norm of the error is increased by 8.5% and 18.1%, respectively. The average computation time for the resulting MFM predictions (after fully trained) is in the order of 5 milliseconds for two-level MFM and 11 milliseconds for three-level MFM.