Identifying and addressing inherent challenges in data-driven turbulence models

What is it about?

Turbulent flow exist all around us - for example in wind flow over buildings, water flow in rivers, and haemodynamics in blood vessels. We can simulate turbulent flow scenarios on computers to predict how they will develop or settle. Computational capabilities have matured to the stage where scale-resolved simulations can simulate turbulent flows to minute resolutions and impressive accuracy. However, they can take on the order of months to run and are therefore computationally expensive to perform. An alternative is to use Reynolds-averaged modelling, which predicts the flow in a settled state. While most engineers are more interested in the settled state of a turbulent flow scenario, Reynolds-averaged modelling is well-known for its inaccuracies in predicting certain flow physics due to its built-in assumptions. Recently, there has been growing interest in using machine learning to develop Reynolds-averaged models that overcome these assumptions. However, little is known about the relationship (or mapping) that is being used to train these models. We shed light on the non-unique mapping problem that is prevalent in these models and how it can be addressed.

Featured Image

Photo by Vansh Juneja on Unsplash

Photo by Vansh Juneja on Unsplash