Duncan Watson-Parris

@[email protected]
314 Followers
27 Following
14 Posts
Climate scientist using machine learning to improve our understanding of clouds and aerosols: https://duncanwp.github.io
Show threadDuncan Watson-ParrisNov 8, 2022

I think I broke the thread :-(

It carries on here: https://mastodon.online/@duncanwp/109308827893221490

Duncan Watson-Parris (@[email protected])

A key aspect with this approach is that we’re also able to take into account the spatial distribution of aerosol emissions, which can make a crucial difference c.f. https://www.nature.com/articles/s41558-022-01415-4

Mastodon
Show threadDuncan Watson-ParrisNov 8, 2022

All the models and training data are publically available and we very much encourage you to download it and give it a go! Feel free to reach out with ideas / suggestions for future work :-)

Data: https://zenodo.org/record/7064308#.Y2pyRy-l3jg
Code: https://github.com/duncanwp/ClimateBench

ClimateBench

ClimateBench is a benchmark dataset for climate model emulation inspired by WeatherBench. It consists of NorESM2 simulation outputs with asociated forcing data processed in to a consistent format from a variety of experiments performed for CMIP6. Multiple ensemble members are included where available. Data processing scripts and a baseline model example can be found here. Version history 0.1 - Initial release (not including ssp245 data held back for hackathon) 0.2 - Updated release to fix cumulative CO2 concentrations which were calculated incorrectly for future scenarios. 0.3 - Added test data and included a minor fix to hist-aer input data which had misaligned aerosol and GHG data. 0.4 - Fixed an error in the diurnal temperature range due to missing values in some NorESM2 piControl daily temperature data on ESGF. 1.0 - Release version to coincide with accepted version of manuscript. Now includes associated CMIP6 files

Zenodo
Show threadDuncan Watson-ParrisNov 8, 2022
Given the importance of these tools for underpinning the sorts of negotiations happening right now in #COP27, we also touch on the trustworthiness of these approaches in the paper (Section 5.2). I would love to hear other peoples thoughts on this.
Show threadDuncan Watson-ParrisNov 8, 2022
But, are the results physically robust? This is the $1m question, and while the preliminary results showing global conservation of energy are encouraging, there’s certainly room for improvement (watch this space!).
Duncan Watson-ParrisNov 8, 2022
A key aspect with this approach is that we’re also able to take into account the spatial distribution of aerosol emissions, which can make a crucial difference c.f. https://www.nature.com/articles/s41558-022-01415-4
Strong control of effective radiative forcing by the spatial pattern of absorbing aerosol - Nature Climate Change

Changes in the spatial pattern of aerosol could influence climate through effects on radiative forcing. Model experiments show that while aerosol absorption in the midlatitudes and regions of tropical descent can warm the planet, aerosol absorption in regions of tropical ascent can cool the planet.

Nature
Show threadDuncan Watson-ParrisNov 8, 2022
It turns out you can using a variety of machine learning techniques, and pretty well! There are the differences between the emulators and the full model but they’re mostly small
Show threadDuncan Watson-ParrisNov 8, 2022
These are physically very robust (arguably more so than the full models) but only predict global mean temperature. In ClimateBench we ask: can we emulate the full spatial response of the models, and not just for temperature??
Show threadDuncan Watson-ParrisNov 8, 2022
Full complexity climate models are really expensive to run, so are usually only run for a handful of scenarios, including a low end (best case) and a high end (worst case) of emissions, and these simulations are used to fit simple energy balance models such as #FAIR
Show threadDuncan Watson-ParrisNov 8, 2022
I think most people assume that we use complex climate models to try and predict the future warming under all the various different possible future emission scenarios, but this isn’t quite true...
Show threadDuncan Watson-ParrisNov 8, 2022
tl;dr: We used machine learning regression models trained on CMIP6 data to predict maps of temperature and (extreme) precipitation at the end of the century under unseen scenarios with excellent accuracy!