Unsupervised learning to indentify structures in turbulent flows

Coherent structures are ubiquitous in turbulent flows. These structures play a key role in heat and momentum transfer. However, identifying these structures can be difficult as in some cases their characteristics change with properties of the flow. This is where unsupervised methods can help. The goal of these methods is to recognize and learn patterns in large datasets, without a set of ground truth labels. The image below is taken from a direct numerical simulation of a turbulent boundary layer (thesis work of John Lyne). The blue tones indicate vortical structures while the red tones indicate zones of turbulent production where the structures are identified using a self-organized map [Kohonen, 1998]. This is an unsupervised method that is computationally efficient, which is key when using large 4-dimensional datasets. We are working on bringing this into more complex architectures that can be used to predict state transitions in stratified turbulent boundary layers

Credit: John Lyne

Stably stratified flows

Stably stratified flows occur in a wide range of environmental and industrial processes. For example, thermal storage tanks and atmospheric boundary layers. They can influence the energy extracted by a wind turbine or the wind load experience by a building.  Nocturnal boundary layers are often stable due to radiative cooling, which can help fog formation and slippery conditions that impact driving. 

These boundary layers are difficult to capture in forecasting models due to their small scales and intermittent physics. We have been working on improving understanding of the physics of these boundary layers in both transient and steady states using high resolution numerical simulations. Results from these studies can be used for development of better models, We are keen to explore the use of thermal and/or optical imagery to improve our models with data.

As an example, below two images are shown. The top image visualizes the flow structures in an unstratified boundary layer, whereas the bottom image visualizes the flow structures in a strongly stabie boundary layer. It can be seen that for the unstratified case, the flow structures occupy the entire width of the channel, and are accompanied by vigorous mixing; whereas for the strongly stable boundary layer the turbulence is confined to a spanwise strip, and mixing is supressed.

. Credit Amir Atoufi