Stock photo of the earth with visualisation of a network overlaid

About Me

Duncan Watson-Parris

I'm an atmospheric physicist working at the interface of climate research and machine learning. Currently, my research focuses on understanding the interactions between aerosols and clouds, and their representation within global climate models. These interactions are numerous and complex, involving non-linearities and feedbacks which make modelling average responses to any perturbation in aerosol extremely challenging. I'm leading the development of a variety of machine learning tools and techniques to alleviate these difficulties and optimally combine a variety of observational datasets, including global satellite and aircraft measurements, to constrain and improve these models.

I'm also keen to foster the application of machine learning to climate science questions more broadly and am Course Director of the iMIRACLI Innovative Training Network, convene the Machine Learning for Climate Science EGU session and co-chaired the recent Climate Informatics 2020 conference.

Recent work

A selection of recent projects I've been working on; a more comprehensive list can be found under Projects.

In the News

This article from the Wall Street Journal explores the role of Amazon, Microsoft and Google in meeting the need for cloud-scale computing to tackle some of the largest climate research problems we face. It highlights my work on detecting ship-tracks in vast quantities satellite imagery using Machine Learning on Amazon Web Services.

In this podcast Philip Stier and I discuss the use of cloud computing for tackling some of the most pressing climate change questions, including for hosting large machine learning workflows such as the detection of pockets of open cells. This was also featured in a blog post by Amazon CTO Werner Vogels.

This article in Science magazine highlights our work using neural architecture search to develop fast, flexible emulators that are applicable to a wide range of physical simulators. These emulators are often orders of magnitude faster than the simulators they replicate and can allow solving inverse problems - that is, determining model inputs based on matching observable outputs.


Please feel free to get in touch if you have any questions about my work or are interested in collaboration.