Enhancing turbulence modeling: Machine learning for pressure-strain correlation and uncertainty quantification in the Reynolds stress model

Abstract

This research study delves into the realm of enhancing the Reynolds stress model through the remarkable capabilities of machine learning. Focusing on two key aspects, we investigate the modeling of fluctuating pressure-rate-of-strain tensor and uncertainty quantification in turbulent fluid flows. In the first part of our investigation, a neural network is employed to predict the pressure-strain correlation term for the frictional Reynolds number, even in cases where it was not trained. Through the use of high-fidelity datasets, we surpass the limitations of traditional numerical methods, achieving more accurate results with reduced computational costs. Going beyond conventional analysis, we explore the isotropic and anisotropic components of the pressure-strain correlation term. Intriguingly, our findings unveil the significant influence of production and redistribution terms on the anisotropic component by 50% and viscous dissipation impacts the isotropic part by 65% in wall-bounded turbulent flows. The second part introduces a novel approach for uncertainty quantification in the Reynolds stress model. Leveraging a binary classification model, we identify error-prone points within the simulation domain with an impressive accuracy of up to 95%. This critical step paves the way for a deeper understanding of uncertainties in turbulent flows for more reliable predictions. By harnessing the power of machine learning, our study provides two compelling ways to enhance the Reynolds stress model, contributing to a more efficient and dependable prediction of turbulent fluid dynamics.