Revolutionising climate modelling through mathematical machine learning

Written by Josh Stevens
November 5, 2020

In a new discussion series with CCIMI students, we sit down with Rachel to chat about her area of research and PhD journey so far.

Can you introduce yourself and describe your PhD project?
My name is Rachel Furner, and I’m a third year PhD student, based between DAMTP at Cambridge University, and the British Antarctic Survey. My research looks at ways to make climate models more computationally efficient using methods from statistics and machine learning.
Traditional climate models are ‘process based’, meaning they solve physical equations on a discretised grid. These work very well, but take huge amounts of compute power and time to run. Recent work has indicated that machine learning methods could ‘learn’ the dynamics covered in these models, in order to replicate their results, and these methods would likely be far more efficient to use for predicting, meaning we could run more models, and increase our understanding of future climate change.

What excites you the most about your project?
I’ve always found the real world applications of mathematical techniques the most exciting, and with climate change being such a threat to us all, its great to be working on this incredibly relevant and important area. I hope that this work can lead to real improvements in the efficiency of climate models, enabling us to run more experiments and better understand potential changes and the impacts of various mitigation options. The idea that this project could lead to real change and improvements is really exciting.
Alongside this, learning the new techniques of machine learning, and seeing ways to apply methods from other fields to climate modelling is fascinating. Finding similarities between very different application fields has been full of great surprises, and enabled applications of techniques from these fields to my work.

What was the biggest challenge and what did you find most enjoyable in the first year of your PhD?
When I began my PhD the project plans were very broad. This meant I had a lot of freedom to explore, which was great, but it was also very challenging to try and focus the many options and ideas into a concrete and feasible plan. Through this period I also tried various methods and ideas which were not as successful as we’d hoped. This was a difficult process at the time, but with hindsight even these unsuccessful attempts were an opportunity to learn and develop new skills, and helped to shape my project.
For me one the most enjoyable aspects was having time to do a lot of reading in this area and exploring the wealth of ideas and work carried out by others. There are so many exciting and interesting approaches looking at using machine learning in environmental sciences, and discovering some of these was fascinating. I also really enjoyed developing new skills, especially coding up machine learning techniques and seeing these work!

What advice would you have for students interested in the Mathematics of Information PhD?
I’d really recommend the Mathematics of Information PhD to anyone who is interested in applying or developing data-driven techniques. There is a huge breadth of subjects covered by the programme, giving a rare opportunity to belong to a group looking at a wide variety of projects. For me, learning about this wide spectrum of work carried out by fellow students has not only been really interesting and inspiring, but also meant I’ve learned about useful techniques or ideas from these other fields, which I otherwise might not have been aware of.