Presenter: Dr. Hesam Salehipour, University of Toronto
Topic: Deep learning of turbulent mixing
Current global climate models are essentially blind to small-scale physical processes that are not explicitly resolvable by such low spatial resolution structures. For instance, the turbulence related processes in the oceans that mix deep cold waters with shallower warm waters and that control the so-called overturning circulation cannot be explicitly represented. Such models must therefor rely on somewhat ad-hoc “parameterizations” of such processes if they are to adequately represent important climate system processes. Recently, a significantly improved understanding of the turbulence induced by two hydrodynamic instabilities that are known to be omnipresent in the oceans (i.e. two “atoms” of ocean turbulence) has been achieved thanks to the availability of an unprecedented volume of highly-resolved simulations associated with these turbulent processes. In this talk, we will discuss the way in which we have employed the important advances in artificial neural networks to construct predictive models of turbulent ocean mixing. We will demonstrate, for the first time, that a deep neural network trained on one “atom” of turbulence is capable of revealing some of the significant characteristics of the turbulence generated by a dramatically different mechanism, suggesting that through the application of appropriate neural networks, significant universal abstractions of density stratified turbulence have been recognized.
About Dr. Salehipour
Hesam is a fluid dynamicists with a strong passion for multi-disciplinary problems that involve computational analysis of geophysical, biological and engineering flows. Currently, he is a post-doctoral fellow at the University of Toronto and is working in collaboration with Autodesk Research. Hesam has obtained his PhD and MSc degrees in Physics and also has an MSc and BSc in Mechanical Engineering. For his doctoral dissertation at the University of Toronto, he studied ocean mixing and turbulence which led to eight publications in top-tier journals. In addition, he has published on ocean tides and the energetics of flapping flight. Hesam has recently become very interested in the application of deep learning methods to various problems in fluid mechanics, in particular that of turbulence modelling.
© 2018 Compute Ontario