Uncertainty Quantification in Computational Fluid Dynamics: Physics and Machine Learning Based Approaches

Abstract: Turbulent flow has been extensively studied using computational fluid dynamics (CFD) simulations since turbulent flow regime is so frequently encountered in both academic and engineering applications. The high-fidelity simulation of the Direct Numerical Simulation (DNS) requires a sufficiently fine mesh to resolve the smallest Kolmogorov length scale of turbulent motion, which requires a tremendous amount of computational resources, and hence usually prohibited in engineering applications. At the current state of computational power we can only execute DNS simulations for small sections of an aircraft wing at high Reynolds numbers. Even though large-eddy simulation (LES) has reduced the computational overheads by only resolving the large-scale eddies with the small-scale eddies being modelled. However, for extremely high Reynolds number and complex geometry flows LES is still restricted to academic studies. As the result, Reynolds-averaged Navier-Stokes (RANS) based turbulence models have significantly reduced the computational overheads by modeling all scales of turbulent motion. Unlike DNS and LES which are dedicated to representing the true physics of turbulent flows, RANS approach uses simplifying modeling assumptions to describe both lower order and higher order quantities. This makes RANS still remain the most widely used CFD method in engineering applications; however, simplifying assumptions also introduces sources uncertainties during simulation.