Physical Sciences Public Lecture Series: Opportunities and Quandaries for Next-Generation Earth System Modeling in the Machine Learning Era
Numerically predicting the future of Earth’s climate is a horsepower-limited problem since the relevant physical systems span 8-10 orders of magnitude in space and time. Doing justice to the true governing equations of atmosphere-ocean fluid dynamics is not possible on planetary scales, even when using the nation’s best supercomputers, due to the associated computational demands. Instead, climate scientists are forced to make predictions on coarse planetary grids, relying on unsatisfying physical approximations of unresolved sub-grid processes like cloud physics and oceanic turbulence, leading to large error bars on future predictions of regional climate. Enticingly, machine learning may disrupt this state of affairs: If neural networks can be successfully trained to “emulate” the high-fidelity physics that are missing from climate models, the calculations may be “outsourced” to computationally efficient machine learning sub-units. Professor Mike Pritchard will discuss recent advances, opportunities, and emerging challenges at this new frontier of earth system modeling augmented by artificial intelligence.
Mike Pritchard is an Associate Professor, and an interdisciplinary climate scientist with expertise in tropical climate dynamics, land-atmosphere interaction, precipitation and drought extremes, and next-gen climate simulation. Recent collaborations between his Computational Clouds and Climate lab in the department of Earth System Science and UCI computer scientists have led to new energy-conserving machine learning emulators of sub-grid cloud physics for dynamical prediction, and strategies for self-supervised analysis of atmospheric turbulence using emerging data science tools such as latent space inquiry. He is a DOE Early Career Award recipient and co-leads a national working group about emerging data science tools for studying climate variability and change.