LECTURE 9/27: “Recent Advances in Applying Machine Learning to Weather Modeling at NVIDIA + Beyond” – Frost Institute for Data Science & Computing

LECTURE 9/27: “Recent Advances in Applying Machine Learning to…

The Department of Atmospheric Sciences is pleased to host Senior Research Scientist Yair Cohen, PhD, of NVIDIA Research, in Santa Clara, California, for a guest lecture on machine learning and weather modeling on Friday, 9/27/2024 at 11:00 AM (ET).

Attend in Person | or via ZOOM Link

Rosenstiel School of Marine, Atmospheric, and Earth Science
AUDITORIUM (Doherty Marine Science Center | Campus Map)
4600 Rickenbacker Causeway, Virginia Key, FL 33149

Parking Information

TALK TITLE: “Recent Advances in Applying Machine Learning to Weather Modeling at NVIDIA and Beyond”

Recent years have witnessed a paradigm shift in weather and climate prediction, driven by rapid advancements in Machine Learning (ML). The application of cutting-edge ML architectures such as transformers, Fourier neural operators, and diffusion models to train on state-of-the-art, data-assimilating weather products has shown remarkable promise. As major meteorological institutions like NOAA and ECMWF incorporate ML-based forecasts into their daily operations, these methods are poised to revolutionize weather modeling in several ways. On the global scale, ML models trained on ERA5 reanalysis data (~25 km resolution) have demonstrated forecast skills rivaling those of operational numerical models, while offering an unprecedented four-orders-of-magnitude speedup. This breakthrough unlocks new possibilities for large ensemble forecasts, potentially enhancing our ability to quantify uncertainties and predict extreme events with greater accuracy. Recent research has explored coupling these weather models with parsimonious ocean models, aiming to extend forecast skill beyond the medium range and bridge the gap between weather and seasonal prediction. Concurrently, on regional scales, ML-based downscaling approaches have shown remarkable promise in super-resolving and forecasting kilometer-scale cloud properties, such as radar reflectivity. These downscaling techniques represent a crucial step towards the operational deployment of global ML forecasts. The inherent stochasticity of atmospheric processes at these fine scales necessitates the use of advanced generative ML methods, capable of capturing the probabilistic nature of local weather phenomena. Ultimately, the potential of better and more adaptive, ML-based data assimilation techniques highlights the potential of such models to surpass the skill of their numerical counterparts.

This talk will present a comprehensive overview of recent advancements in ML weather modeling, focusing on three key areas. First, I will present recent works on Ensemble Forecasting using global ML models, discussing their successes challenges and potential pitfalls, biases, and strategies for mitigation. Second, I will discuss two recent weather downscaling works that used generative ML models for regional forecasts. Third, I will discuss the possibility of using generative ML methods (diffusion models) for data assimilation. The presentation will highlight key findings from several groundbreaking studies conducted at NVIDIA and other leading institutions in the field. By synthesizing these recent developments, we aim to provide a comprehensive view of the current state of ML applications in weather and climate science, and offer insights into how these technologies are poised to transform our forecasting capabilities in the near future.

Some relevant publications can be found here:

https://arxiv.org/pdf/2401.15305
https://arxiv.org/abs/2408.03100
https://www.arxiv.org/abs/2408.01581
https://arxiv.org/abs/2309.15214
https://d1qx31qr3h6wln.cloudfront.net/publications/StormCast.pdf
https://arxiv.org/abs/2406.16947

About Dr. Yair Cohen

Dr. Yair Cohen has experience in software and model development, machine learning, data science and applied mathematics. Additionally, he has experience leading a team of researchers and software engineers in agile development of a data driven model and working with technical and non-technical stakeholders.