2:00 PM in NSII 1201

Assoc. Prof. Michael Pritchard of Earth System Science will present the inaugural seminar for all Physical Sciences researchers interested in machine learning (title and abstract below). The talk will include fascinating insight into use of machine learning and also present topics which will lead immediately at the end of the seminar into general discussion of machine learning needs for Physical Sciences research. This seminar series will be offered every other Tuesday at 2 in NSII 1201 to inform, engage, and facilitate artificial intelligence research activities throughout all four PS departments.

Come to these events to help make the Machine Learning Nexus in Physical Sciences a useful tool for your research and career. And also contribute and keep track of the Machine Learning Nexus at https://ps.uci.edu/psml/.

Seminar Information
Speaker: Prof. Michael Pritchard, Earth System Science

Title: Advances from machine learning in the climate sciences – from emulating turbulent physics to interpreting multi-scale dynamics.

Abstract: I will discuss my 3-year journey from skeptic to machine learning enthusiast in the context of a long-standing challenge in numerical climate simulation – understanding how clouds and boundary layer turbulence interact with global climate dynamics, which is one of the largest uncertainties dictating future climate vulnerability. The deep learning story begins as a pragmatic exercise in computational efficiency – it turns out neural network emulation of explicit cloud physics can be made surprisingly skillful, including enforcing exact analytic constraints like energy and mass conservation.  This already seems poised to usher in an age of explicit turbulence in climate simulation ahead of computational schedule. I will discuss the current outlook and challenges for a next generation of ML-based climate models, and especially how this disruption is raising new philosophical tensions in the community due to trade-offs intrinsic to outsourcing actual physics to neural networks. More interestingly, this seeming act of engineering turns out to be deeply physical --  I will conclude by discussing some recent adventures in neural network assisted dynamical inquiry into the complex multi-scale physics of the atmosphere. Preliminary results from interpretable machine learning applied to the climate are increasingly enticing in ways that should be broadly relevant across the physical sciences when the system has overwhelming degrees of freedom. A few illustrative examples will be discussed ranging from understanding interactions of atmospheric moist convection with its thermodynaminc environment, to equation discovery of multi-scale dynamical relationships in ocean turbulence, as well as discriminating forced climate change from internal chaotic variability.