MIT Technique Trains Climate Emulators With Less Data
The method lets fast climate emulators predict scenarios they never saw during training, using only a fraction of the data typically required.
Updated on Nov. 12, 2026 in Environmental

Live Poll
Do you trust AI-generated climate predictions to guide major policy decisions?
MIT researchers developed a training technique for climate emulators that uses only a fraction of the data typically required. Emulators trained with the method accurately predicted outcomes of scenarios not seen in training, while traditional emulators failed to adapt to the same new situations.
Why it matters
Earth system models can take months to run on supercomputer clusters and project only one possible future per run. A faster, data-efficient emulator could let scientists explore many more climate scenarios in the time it takes a full model to finish one.
Training an emulator with realistic scenarios typically requires six to eight scenarios run through a full-scale earth system model, while a trained emulator runs a scenario in a matter of seconds rather than months on a supercomputer cluster. The reported tests compare the new technique against traditional emulators, but the extraction does not specify test sizes beyond this comparison.
The players
Massachusetts Institute of Technology (MIT)
A private research university in Cambridge, Massachusetts, whose researchers developed the climate emulator training technique.
ETH Zurich
A Swiss public research university where the lead author of the study currently holds a postdoctoral position.
Imperial College London
A public research university in London whose researcher Sebastian Eastham co-authored the study.
Schmidt Sciences
A philanthropic science organization that supported the research through the MIT Grand Challenges program.
The details
The technique trains a climate emulator on a synthetic dataset of optimized climate scenarios generated by an algorithm that iteratively updates the training data until the emulator's error reaches a set threshold. The researchers found that realistic future scenarios used as training data had led to underperforming emulators that fail on novel scenarios, which the new approach avoids.
Timeline
2026: Womack received his PhD from MIT.
November 12, 2026: The research was published in the Proceedings of the National Academy of Sciences this week.
Roadmap
Full earth system models remain the standard for climate projection but are slow and yield a single future per run, which is why emulators have become a growing research focus. This MIT approach extends that line of work by making emulators reliable on scenarios outside their training set.
Faster climate emulators could let researchers test many more emissions and warming scenarios than full models allow, which may sharpen planning for future climate risks. The technique is still at the research stage and the extraction does not give a timeline for its use in practice.
The takeaway
Climate models that once took months to produce one projection could soon be supplemented by emulators that answer many questions in seconds. Such tools may help researchers explore more of the uncertain range of possible futures.
Further reading
For broader background on how scientists model future climate, see the Environmental section.
Live Poll
Do you trust AI-generated climate predictions to guide major policy decisions?







