Physics Guided Machine Learning For Uncertainty Quantification In Turbulence Models

Predicting the evolution of turbulent fluid flows represents a central need across fields of science and engineering. Almost all investigations use computer simulations using turbulence models for this purpose. Turbulence models rely on empirical simplifications for computational feasibility and in doing so introduce epistemic uncertainties in their predictions. The Eigenspace Perturbation Frame6work (EPM) is the only physics based method to predict these uncertainties. This framework being purely physics based often over predicts the uncertainty. We use Convolutional Neural Network (CNN) based models to modulate the magnitude of perturbations in the EPM leading to better calibrated uncertainty bounds. This is tested and validated across a suite of benchmark turbulent flow cases.