MIT Method Trains Climate Emulators on Far Less Data

MIT researchers have developed a method for designing optimised training scenarios so climate emulators can be trained on far less data than the usual suite of standard emissions pathways. The MIT News account is dated 12 November 2026. MIT says the paper appears this week in the Proceedings of the National Academy of Sciences.

Climate emulators are specialised AI models that mimic slower, physics-based climate models so that many future scenarios can be tested quickly. According to MIT, they can run in a matter of seconds, which lets scientists explore a much wider set of cases than a full Earth system model would allow. The usual way to train them is to put a handful of realistic future emissions pathways through a large model and use those runs as training data. The new work asks whether the training scenarios themselves can be designed so that a smaller set of runs teaches the emulator more.

The lead author is Christopher B. Womack, who completed his MIT PhD in 2026 and is now a postdoc at ETH Zurich. The senior author is Noelle E. Selin, a professor in MIT’s Institute for Data, Systems, and Society and in Earth, Atmospheric and Planetary Sciences, and director of the Center for Sustainability Science and Strategy. They are joined by Shahine Bouabid, Andrei Sokolov, Popat Salunke and Glenn Flierl at MIT, and Sebastian D. Eastham at Imperial College London. The work was part of the Bringing Computation to the Climate Challenge project and, MIT says, was supported by Schmidt Sciences through the MIT Grand Challenges.

One optimised scenario beat six standard pathways

The method is set out in the authors’ preprint, “Optimal scenario design for climate emulation”, first posted on 17 June 2026 and last revised on 29 September 2026. They use a differentiable simple climate model to measure how emulator error changes when the training data are perturbed, then update those data to improve the emulator’s skill. In that simple-model test, an emulator trained on one scenario optimised in this way outperformed one trained on six standard ScenarioMIP pathways, even though the training set was smaller. The authors say the resulting emulator could distinguish the behaviour of different climate forcing agents, such as greenhouse gases and aerosols, without dedicated single-forcing runs.

They then took scenarios optimised on the simple model and used them to drive an intermediate-complexity climate model. That produced a training set which, they report, yielded a more skilful emulator than training on ScenarioMIP outputs. MIT says the traditional emulators the team evaluated failed to capture a local pollutant-control policy that cuts one pollutant while other emissions keep rising, while the new emulator did.

“Climate emulators are especially well-suited for modeling policy-relevant scenarios. By using our method to generate optimized training data for an emulator, we can speed up the production of much more reliable emulators, so we can better trust their outputs for making policy decisions,” Womack said.

The training scenarios that come out of the algorithm need not look like a plausible future. MIT quotes Womack as saying the method can introduce rapid variations, or “high-frequency wiggles and oscillations,” so the emulator can extract more information about the system. Code and data for the study are available on Zenodo, where the record was made available on 24 September 2026.

This is not MIT’s first recent appearance in TSN’s coverage of machine learning for climate and weather: MIT and Microsoft researchers have also used machine learning to improve forecasts two to six weeks out. Other recent work on faster climate and weather tools includes the University of Chicago’s AI Weather Almanac and, on the observational side, UNEP’s use of AI with methane satellites to flag major leaks.

The method has not yet been tried on a full Earth system model

The results so far come from a simple climate model and an intermediate-complexity model. They have not yet been tested with a full Earth system model. MIT says the researchers want, in future work, to test the technique using an emissions scenario generated by a full Earth system model. The preprint likewise treats a full-scale Earth system model run as future work, noting that the cost of running thousands of such simulations is what stops the authors from searching for an optimal training scenario directly on a full-scale model.

That limit matters because the practical prize the authors describe is exactly that setting: a world in which even a small number of large-model runs is expensive, and the question is how to spend them. Until the method is tried on a full Earth system model, the one-versus-six result and the more skilful intermediate-complexity emulator remain findings on simpler models, not a completed demonstration on the models that take months to run.

“In the future, we are definitely going to see climates and scenarios we haven’t seen before. If we can simulate these different scenarios more robustly, we can provide more reliable information to the stakeholders who need it,” Selin said.

A smaller, richer training set is not the same as a finished operational tool

The claim worth holding onto is modest and specific. On a differentiable simple climate model, one optimised training scenario beat six standard ScenarioMIP pathways. On an intermediate-complexity model, scenarios designed the same way produced a more skilful emulator than ScenarioMIP outputs. MIT says that difference showed up on a local pollutant-control policy that the traditional emulators they tested missed. None of those results is reported here as a percentage for speed or accuracy, because no such figure is given in the sources.

The method is a way of spending scarce climate-model runs more carefully, by shaping the training scenarios rather than stacking more of the usual pathways. That is a useful idea for policy-relevant experiments, which is the use Womack emphasises. It is also unfinished business until someone drives a full Earth system model with an optimised scenario and checks whether the same gain appears. The paper’s appearance in PNAS, if it is published as MIT says it will be this week, will be the formal version of a study whose preprint, code and data are already public. The next result to watch is whether the method still holds when the training data come from a full-scale model.

Sources

1. Adam Zewe, MIT News, “New research could lead to faster, trustworthy AI models for climate prediction,” 12 November 2026. https://news.mit.edu/2026/new-research-could-lead-faster-trustworthy-ai-models-climate-prediction-1112
2. Christopher B. Womack et al., “Optimal scenario design for climate emulation,” arXiv:2606.19302, v2, 29 September 2026. https://arxiv.org/abs/2606.19302
3. Christopher Womack, “Source code and data for Optimal scenario design for climate emulation,” Zenodo, 24 September 2026, DOI 10.5281/zenodo.22875972. https://doi.org/10.5281/zenodo.22875972

Related stories

MIT and Microsoft Researchers Use Machine Learning to Improve Weather Forecasts Two to Six Weeks Out is the nearest TSN account of MIT researchers using machine learning on climate and weather prediction.

Also relevant:

• UChicago Wins $10m Laude Grant for AI Weather Almanac
• UNEP Says Methane Satellites Helped Close 46 Major Leaks