ICCS seminar: Hannah Christensen

Machine-learning weather prediction models have made remarkable advances in short-range weather forecasting, with skill matching or exceeding traditional numerical models. However, it’s unclear whether a similar approach can be useful for climate prediction, where the ability to generalise to unseen climates is key. I will show evidence that current ML atmosphere models struggle to generalise to warmer climates, and explore why some of the features that improve weather prediction can hinder their generalisability. What might this  mean for the design of future climate models? Instead of using ML to replace entire physical models or components, I will argue for a more deliberate approach: retaining the physics we understand, while using machine learning to represent the uncertain parts. We find such an approach has many benefits: reducing biases and computational cost, while improving model stability. Finally, I will discuss why this may also bring an ability to generalise, thereby offering a more robust route from weather prediction to climate modelling.

How to attend

Join on Zoom or in-person at MR13, Centre for Mathematical Sciences, Wilberforce Rd, Cambridge CB3 0WA

About ICCS Seminars

This talk is part of a series of ICCS Seminars. Members of the VESRI teams, ICCS and the wider research community present their work, research papers or concepts significant to their field.

Watch previous seminar recordings