Multi-Fidelity Data Fusion for Aerodynamic Prediction Including Parametric Uncertainty

Abstract

Recent developments in data fusion provide an algorithmic means of combining predictions from models of various fidelities into a unified model that provides both the mean and variance of desired outputs over the input space. In this work we show how cokriging can be used to combine aerodynamic predictions from four data sources into a single Gaussian process (GP) regression model. We use the multi-fidelity Gaussian utilities from the Sora flight prediction tool from Lawrence Livermore National Laboratory to perform cokriging and interpolation of total force and moment coefficients as functions of Mach, angle of attack (AoA), and aerodynamic bank angle. We use generalized least squares (LS) regression to convert the total force and moment predictions to polynomials with the sine of AoA as the variable. By propagating the uncertainty from GP predictions through the LS regression, we can choose a polynomial form with minimal coefficient uncertainty. The final model form consists of tables of polynomial coefficients as functions of Mach and aerodynamic bank angle.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Oct 01, 2023
Accession Number
AD1214017

Entities

People

  • Bradley T. Burchett

Organizations

  • United States Army Research Laboratory

Tags

Readers

  • Approximation Theory.
  • Fluid Dynamics.
  • Neural Network Machine Learning.

Technology Areas

  • Space