What is it about?
Scientists use computer simulations called climate models to project future climate change, but they can take months to run. Faster AI "emulators" mimic them in seconds. We developed a method to design custom, synthetic training scenarios that teach these AI models more effectively, outperforming models trained on six standard realistic scenarios.
Featured Image
Photo by Elena Mozhvilo on Unsplash
Why is it important?
Because standard training scenarios follow smooth, predictable paths, AI emulators can struggle to predict very novel scenarios, such as sudden policy shifts. Designing smarter training data cuts supercomputing costs while making fast climate projections more reliable for researchers and policymakers.
Perspectives
I love that this work flips the traditional climate modeling paradigm on its head. By intentionally designing strange, unrealistic futures, we can help AI models learn faster and make more reliable predictions for the real world.
Christopher Womack
ETH Zürich
Read the Original
This page is a summary of: Optimal scenario design for climate emulation, Proceedings of the National Academy of Sciences, November 2026, Proceedings of the National Academy of Sciences,
DOI: 10.1073/pnas.2620731123.
You can read the full text:
Resources
New research could lead to faster, trustworthy AI models for climate prediction
MIT News press release on "Optimal scenario design for climate emulation"
Code and data for optimal scenario design for climate emulation
Codebase to test the influence of training data on climate emulators. Answering the question: what's the optimal scenario to train an emulator?
Contributors
The following have contributed to this page







