MIT researchers have developed a training technique that produces fast, adaptable AI climate models using a fraction of the data normally required, a step they say could cut the production time of these tools from months or years down to something far more usable for policy decisions. The research appears this week in the Proceedings of the National Academy of Sciences.
The models in question are climate emulators: compact AI surrogate systems that mimic the behavior of full-scale earth system models (ESMs). ESMs are the workhorses of climate science — enormous physics-based simulations that project how temperatures, precipitation and sea levels respond to greenhouse gas emissions — but a single future scenario can take months to run on a cluster of supercomputers. Emulators, by contrast, run in seconds, letting scientists test thousands of scenarios, probe regional effects and specific pollutants, and flag unexpected outcomes of policy interventions. For ongoing coverage of AI's expanding role in science, follow our AI research news.
The problem: realistic training data makes fragile models
When researchers build a climate emulator, they typically train it on six to eight future emission scenarios that follow realistic trajectories. Each one must be run through a full ESM to produce training data, which makes dataset generation costly and slow. The MIT team had previously identified a subtle failure mode: emulators trained on realistic scenarios can underperform when asked to predict unfamiliar or novel situations — exactly the situations policymakers increasingly need to explore.
"Our algorithm might generate some weird scenarios which look nothing like a realistic future, but we don't care so much if the training scenarios are realistic," said Christopher Womack PhD '26, now a postdoc at ETH Zurich and the paper's lead author, in MIT News. "We are just trying to make the best possible dataset to train an emulator."
One ESM run instead of eight
The new technique inverts the usual pipeline. Instead of assembling realistic scenarios, an algorithm generates optimized synthetic scenarios tailored to maximize a specific emulator's predictive skill. The process starts with an arbitrary scenario run through a fast, simple climate model, measures how well the emulator reproduces the full ESM's results, then iteratively perturbs the training data — raising CO2 emissions sharply in 2050, say, or cutting methane rapidly in 2037 — until the emulator's error falls below a threshold.
The resulting synthetic scenarios look nothing like smooth real-world trajectories. "By introducing these high-frequency wiggles and oscillations, our data enable the emulator to extract a lot more information about the overall system," Womack said. The practical payoff is substantial: where building a training set once required running a full ESM six to eight times, a user now needs a single run.
Passing the test traditional models failed
The researchers report that an emulator trained with their method accurately predicted outcomes of scenarios it had never seen during training, while traditional emulators failed to adapt. In one test highlighted by MIT News, the synthetic-trained emulator correctly captured the effects of a local pollutant-control policy that reduces one pollutant while other emissions continue rising — a case where the conventionally trained emulators they evaluated failed.
"Climate emulators are especially well-suited for modeling policy-relevant scenarios," Womack said. "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."
The adaptability matters because the future will not resemble the training set. "In the future, we are definitely going to see climates and scenarios we haven't seen before," Selin said. "If we can simulate these different scenarios more robustly, we can provide more reliable information to the stakeholders who need it."
Why timing matters
Senior author Noelle Selin, a professor in MIT's Institute for Data, Systems, and Society and the Department of Earth, Atmospheric and Planetary Sciences, and director of the Center for Sustainability Science and Strategy, framed the work as a supply-chain fix for climate information itself. Because emulator training can take months or years, the tools available to planners are often out of date. "By cutting down that production time, I think this technique will be a game-changer in terms of getting updated tools into the hands of researchers and practitioners," she said.
The applications extend well beyond academia: up-to-date climate projections inform decisions on power-grid resilience, public health protection and agricultural planning. The team also included Shahine Bouabid, Andrei Sokolov, Popat Salunke and Glenn Flierl of MIT, and Sebastian D. Eastham of Imperial College London. The researchers' next target is extending the technique so emulators remain robust as the planet enters climate conditions with no historical precedent.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
