Atmosphere Ocean Science Colloquium
Integrating Physics, Data, and Scientific ML to Better Understand and Model Climate Variability and Extremes
Speaker: Pedram Hassanzadeh, Rice U
Location: 60 Fifth Avenue 7th Floor Open Space
Date: Monday, March 27, 2023, 2:30 p.m.
The Earth system is a complex, spatiotemporal, multi-scale, multi-physics, and nonlinear
dynamical system. As a result, analyzing, understanding, and simulating the Earth system is a
major challenge. Still, in the past century, integrating observations and physics-based models
(theory and hierarchies of computer models) has resulted in significant advances. However, this
“conventional approach” alone still cannot provide reliable short-term forecasts and long-term
climate change projections, in particular for extreme weather events and regional scales. In this
talk, I will first present a brief overview of our work on using the conventional approach to
improve the understanding of climate variability and extremes (e.g., blocking events, heat
waves), mainly via better quantifying the multi-scale, nonlinear eddy-mean flow interactions.
Then, I will argue that integrating scientific machine learning (ML) with the conventional
approach could potentially open new avenues to substantially accelerate climate research, e.g.,
via developing better and faster weather/climate models, extracting more information from
observational data, and even (potentially) improving our fundamental knowledge. I will present
some of the promising results around subgrid-scale parameterization (closure modeling) of
geophysical turbulence and fully data-driven spatiotemporal forecasting of extreme events.
However, as scientific ML is in its infancy, there are significant challenges for the operational
use of ML in climate applications that need to be addressed first. These challenges include
interpretability, stability, out-of-distribution generalization, and learning in the small-data
regime. I will highlight some of our recent work on addressing these challenges by closely
combining tools and insight from physics, applied math, numerical analysis, and ML. In
particular, I will discuss a new framework, based on integrating the Fourier analyses of neural
networks and nonlinear physics, which enables us to explain and connect the learned physics
and inner workings of the network, particularly for transfer learning. In the end, I will discuss
our aim at developing rigorous frameworks for applying ML techniques to improve scientific
computing in general and climate modeling in particular, via interdisciplinary collaborations.