Joel Oskarsson: Frontiers in machine learning for weather forecasting

Recent years have seen rapid progress in using machine learning models for weather forecasting. These models show impressive performance, matching or even outperforming existing physics-based models, while running in a fraction of the time. This is fundamentally and rapidly changing the landscape of weather forecasting today. In this talk I will discuss the factors that enabled this paradigm shift, the core machine learning methods used and the research questions at the bleeding edge of machine learning for weather. 

In particular I will focus on how current methods can be extended to regional and probabilistic forecasting. For regional forecasting I will showcase graph-based methods for building limited area weather forecasting models. I will also discuss how generative machine learning methods can enable probabilistic forecasting, giving much-needed estimates of uncertainty and allowing for predicting extreme weather events.

Joel Oskarsson got his MSc in computer science and engineering from Linköping University in 2020, where he is now a PhD student working on probabilistic machine learning. In his research he develops machine learning methods for data with spatial-, temporal- and graph-structure, including combinations of these. He is also interested in applications of these methods to earth system modeling. https://joeloskarsson.github.io/