Building a-priori Confidence Strategies for Data-Driven RANS Uncertainty Quantification
Nikita Kozak, Stanford University
This work builds the understanding of the critical machine learning considerations when predicting the uncertainty bounds for a Reynolds Averaged Naiver-Stokes (RANS) model. More specifically, RANS models exhibit superior computational efficiency compared to competing approaches like LES or DNS, by making strong approximations on the universal behaviors seen in turbulence. These strong approximations included the Boussinesq hypothesis which is used to create a linear eddy viscosity model and calibrated from fundamental flows like a wavy wall or a jet. Then, these models are used to predict the flow field of more complex and applied problems. Recent advances in turbulence modeling include the use of data-driven techniques to predict the uncertainty in extrapolating RANS models to more complex problems. This work focuses on the use of eigenvalue perturbations to the Reynolds stress anisotropy tensor to create uncertainty bounds. However, like the approximations, this data-driven technique is again calibrated from a set of flow fields and then extrapolated to other problems. This work seeks to formulate an approach that provides confidence that the data-driven technique is capable of accurately predicting on the non-calibrated problem. By coupling data science, fluid mechanics and machine learning, this work presents an approach utilizing feature engineering, clustering, and probability distributions to determine how much overlapping and relevant information the training and test cases share. Furthermore, this overlap is quantified to present the a-priori confidence that a data-driven technique has for RANS uncertainty quantification.
Abstract Author(s): Nikita Kozak, Stanford University; Jan Heyse, Stanford University; Aaswhin Ananda Mishra, SLAC National Accelerator Laboratory; Gianluca Iaccarino, Stanford University