🌍 We're hiring a postdoc at the intersection of cutting-edge #AI and #Earth system science. Help us redefine how we build and evaluate the next generation of climate models. Based at @ucsandiego.bsky.social , funded by Google.org. 🧵

All code to reproduce the experiments is available at github.com/treigerm/cli.... Please reach out if you're interested in using it! (/end)

There's lots more work to be done to have principled workflows to calibrate climate models to observational data but Bayesian experimental design provides a rigorous foundation for model calibration!

Work done with the amazing Tom Rainforth and @dwatsonparris.bsky.social . (6/N)

We are able to demonstrate on data from the CESM2 model that BED algorithms can more quickly narrow down the set of plausible input parameters that is consistent with observational data compared to traditional LHS sampling. This scales even to 42-dimensional parameter spaces! (5/N)

Framing the calibration process as a Bayesian experiment design (BED) problem (doi.org/10.1214/23-S...) allows us to derive principled algorithms for finding climate model parameters that leverages the observational data. (4/N)

The cheap emulator model can then be used to investigate what parts of the parameter space are consistent with observational constraints. However, using LHS to explore the parameter space can be wasteful because it is ignoring all information about the observational constraints. (3/N)

A popular workflow to explore the parameter space of climate model parameterizations is to run a perturbed parameter ensemble (PPE) with latin hypercube sampling (LHS) and then train a cheap emulator model on the PPE data. (2/N)

Ever wondered whether there's a principled way to calibrate all those parameters controlling climate models? In our new paper we show how to calibrate climate model parameterizations using ideas from Bayesian experimental design: doi.org/10.1088/3049... . (1/N)

Three panel thing. In the left panel we use error bars. In the second, we take statistical significance as the biggest number but still have error bars. In LLM science, we just have the biggest number

What if we did a single run and declared victory

This work was done as part of the Embed2Scale project (@embed2scale.bsky.social) and with Juniper Tyree, @milank.bsky.social, Peter Dueben, @atmbnl.bsky.social, Dorit Hammerling, Allison Baker, Sara Faghih-Naini, and @philipstier.bsky.social !

I'll be at @egu.eu 2025 next week in Vienna. You can catch me on Monday at 14:05 in Room 2.92 talking about "ClimateBenchPress: A Benchmark for Compression of Climate Data". I'll be at EGU the whole week so let me know if you want to chat about (neural) compression for climate or anything else!

I know @avt.im and @mjhutchinson141.bsky.social have worked on this (e.g. arxiv.org/abs/2110.14423) so they might know more!

physics of climate impacts 101:
- Stuff gets hot (worse heatwaves)
- Hot air holds more water vapor (heavier rainfall)
- Hot air is thirstier air (higher drought risk)
- Warm water is hurricane food (stronger storms)
- Hot water expands and hot ice melts (sea level rise)