Atmosphere Ocean Science Colloquium

Toward a better representation of atmospheric processes using machine learning

Speaker: Sara Shamekh, Columbia

Location: Warren Weaver Hall 1302

Date: Wednesday, March 8, 2023, 3:30 p.m.


As the impact of climate change poses an urgent concern for the future of our planet and all its inhabitants, it's crucial that we accurately understand and model this complex system. By doing so, we can develop the ability to take action towards adapting to the impact of climatechange. Climate models are important tools for understanding and predicting global andregional climate change, yet they exhibit key uncertainties that limit their applicability to future projections. Uncertainties in climate models partly originate from a poor or lacking representation of physical processes too small to be resolved by models, such as atmospheric boundary layer turbulence or clouds. Machine learning has the ability to capture nonlinear structures and relationships within complex data and, when combined with traditional physical models, can lead to a better representation of physical processes and provide new insights into atmospheric processes. 

In this talk, I will discuss two examples that highlight the potential of machine learning (ML) combined with physics and the new discoveries made possible through this framework. The first example uses reduced-order models to accurately represent vertical turbulent fluxes in the atmospheric boundary layer across turbulent regimes. The architecture of this model, in which I enforce a physical constraint, allows clear interpretability and discovery of the main modes of turbulent transport. The second example proposes an ML approach to implicitly learn information (degree of clustering) relevant to precipitation prediction from a complex moisture field. I show that this information significantly improves precipitation prediction, also explaining most of precipitation's stochasticity. The latter is crucial for reducing bias in weather and climate models. 

These examples show the promise of ML in advancing our understanding and modeling of physical processes in the earth system. Despite these successes, many challenges remain to be addressed and many questions to be answered, making the future of this interdisciplinary area exciting. I will end my talk with a discussion on some of these challenges and opportunities.