<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=us-ascii">
</head>
<body style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class="">
<div dir="auto" style="word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class="">
The MET seminar given by <a href="https://climate-dynamics.org/people/ycohen/" class="">Dr. Yair Cohen</a> has been rescheduled to *next Thursday, October 6*.
<div class=""><br class="">
</div>
<div class=""><b class="">DATE:</b> <b class="">Thursday October 6</b></div>
<div class=""><b class="">STUDENT Q&A: </b>2:15 PM, EOA 6067</div>
<div class=""><b class="">SEMINAR TIME:</b> Refreshments at 3 PM, Talk 3:15 PM - 4:15 PM. <br class="">
<b class="">SEMINAR LOCATION:</b> EOA 1044 (speaker remote)<br class="">
<b class="">SPEAKER:</b> <a href="https://climate-dynamics.org/people/ycohen/" class="">Dr. Yair Cohen</a> <br class="">
<br class="">
<b class="">TITLE: </b>Combining physics and machine learning in the turbulence-convection parameterization of the CliMA climate model</div>
<div class=""><br class="">
</div>
<div class=""><font face="Helvetica, Arial" class=""><b class="">ABSTRACT: </b></font><span class="" style="font-family: Roboto, RobotoDraft, Helvetica, Arial, sans-serif; white-space: pre-wrap; background-color: rgb(255, 255, 255);">The representation of turbulence
and convection at the subgrid scale of climate models by various parameterization schemes is one of the largest sources of model uncertainty in long term climate predictions. The crux of these parameterization schemes is the formulation of closure functions
for processes that cannot be observed or simulated in high resolution models. In this talk I will present a hybrid approach that combines a physics based parameterization scheme with physical and machine learning (ML) closures. Such a hybrid approach cannot
be trained using backpropagation typical to neutral networks, as this requires partial derivatives of the physics+ML setup in parameter space. Instead we use a gradient free (ensemble) method to train physics+ML setup from high resolution LES data. The hybrid
approach ensures conservation properties, allows for straightforward interpretability of its data driven components and reduces the degrees of freedom to allow us to train from spares data. The physical model is based on the extended Eddy Diffusivity/Mass
Flux (EDMF) scheme, derived by a systematic coarse-graining of the equations of motion and includes subgrid scale memory and prognostic equations for first and second moments. The closures for this scheme combine physical arguments with nondimensional functional
forms that can be learned from data using various ML architectures. Our parameterization, in a single column of a climate models, reproduces well the corresponding LES of the full spectrum of boundary layer and turbulent motions: from polar boundary layers
(vertical scale of 300m), through the stratocumulus-topped boundary layer (vertical scale of 1km), shallow convection (vertical scale of 3km), and deep convection (vertical scale of 15km). Furthermore, using ML closures we show that a hybrid model is able
to extrapolate by training on current climate simulations to a validation set from 4K climate warming scenario.</span></div>
<div class=""><br class="">
<div class="">
<div dir="auto" style="caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0); letter-spacing: normal; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; word-spacing: 0px; -webkit-text-stroke-width: 0px; text-decoration: none; word-wrap: break-word; -webkit-nbsp-mode: space; line-break: after-white-space;" class="">
<div>--------------------------------------------<br class="">
Allison A. Wing, Ph.D.<br class="">
Associate Professor<br class="">
Department of Earth, Ocean, and Atmospheric Science<br class="">
Florida State University<br class="">
<a href="mailto:awing@fsu.edu" class="">awing@fsu.edu</a></div>
</div>
</div>
<div><br class="">
</div>
</div>
</div>
</body>
</html>