Climate Modeling and Data Assimilation with Generative Machine Learning

Philip Brohan

UK Met Office

Tuesday, Apr 09, 2024, 2:00 pm MT Add to Calendar

JOIN THE WEBINAR:

Abstract

Machine Learning (ML) is a powerful technology, but a very general one. How, exactly, do we use it to make a climate model?

The key technology is not simply ML, but Deep Learning (DL) and in particular generative modelling: A generative model generates a complex, detailed output from a much simpler input.

A general circulation model (GCM) can also be thought of as a generative model, outputting a complex, detailed realization of the state of the atmosphere, but GCMs, while awesome, are very slow and very expensive. We can make a fast, simple ML model that is similarly capable of generating detailed atmospheric fields as outputs, by training a deep convolutional variational auto-encoder on monthly observations.

The ML model has the important advantage that it is possible to find output fields that have any desired property by searching its input space. So we can find (complete) output states that match (limited) available observations – effectively doing data assimilation. Or we can find changes in output state that match changes in Sea Surface temperature (SST) – effectively running climate change experiments.

Bio: Philip Brohan is a climate scientist at the UK Met Office, where he has worked since 2002. He received his PhD in 1994 from the University of Cambridge in the Theory of Condensed Matter group and subsequently worked for Nuclear Electric/British Energy modelling reactor core and fuel behaviour.

At the Met Office, Philip works mostly with weather data, particularly historical observations, to reconstruct past weather. Much of his work finding and digitising historical observations, particularly from ship records, has supported development of the NOAA/CIRES 20th Century Reanalysis. Recently he has also become more interested in using machine learning to build models for reanalysis.