Machine learning for weather and climate

We work on applying the latest AI and machine learning methods to tackle weather and climate prediction.

These methods show high skill for improving representation of the very complex processes involved and improving simulation of phenomena that our best physical models still struggle with. One particular challenge is representing complex fields like precipitation and surface winds. Another is producing information at the local scales required for impacts assessments, with it being unaffordable to run the models at such high spatial resolution and also comprehensively sample climate scenarios and weather events.

New AI and machine learning methods are very promising for deriving realistic representations of these complex processes, based on data from decades of observations and the high resolution simulations that are available. They enable us to produce predictions at a fraction of the cost of running traditional numerical models, allowing study of a much wider range of climate projections and rare extreme events.

In our group, we have applied state-of-the-art machine learning methods to problems such as high-resolution rainfall prediction and to weather situations including hurricanes, mid-latitude cyclones and tropical storms. Understanding how to better model climate dynamics and impacts on society are other key interests.

We collaborate with experts in computer science and statistics. We also work with the Met Office, transferring knowledge for producing practical tools and benefitting from their expertise in observations and high-resolution modelling.

Realistic rainfall simulations generated by a machine learning model over England and Wales (Addison et al).