Plate Nº 81 · recorded November 13, 2026
Earth & ClimateReported finding
MIT Cuts Climate Emulator Training Data to a Single Model Run
MIT researchers train AI climate emulators with one ESM run instead of eight, using deliberately unrealistic emissions scenarios that prepare the models for futures they never saw.
By Marcus Bennett4 min read727 words
In brief
- The technique appears in the Proceedings of the National Academy of Sciences (DOI: 10.1073/pnas.2620731123).
- Training data can come from a single full ESM run instead of the typical six to eight.
- Emulators run in seconds, while a full Earth system model takes months on a supercomputer cluster.
- Traditional emulators failed a test of a policy cutting one pollutant while other emissions rise; the MIT-trained emulator captured it.
- The synthetic training scenarios include spikes such as dramatic CO₂ increases in 2050 and rapid methane cuts in 2037.
MIT researchers have found a way to train AI climate emulators with a single run of a full-scale Earth system model instead of the usual six to eight — and the resulting emulators predict unfamiliar scenarios that break traditional ones. The work appears in the Proceedings of the National Academy of Sciences.
The technique matters for anyone who builds or uses climate projections. Planners rely on these projections to harden power grids, protect public health and maximize agricultural yields. Earth system models (ESMs) — the enormous physics-based simulations that capture the atmosphere, oceans and their interactions — can take months on a supercomputer cluster to project just one possible future. Climate emulators, fast AI surrogates that mimic ESM behavior, run the same kind of test in a few seconds.
"Climate emulators can run in just a few seconds, so it really expands what we can do in terms of exploring these future climate scenarios," said Christopher Womack, Ph.D., lead author of the paper and now a postdoc at ETH Zurich.
Why did realistic training data fail?
Scientists typically train emulators on six to eight scenarios built from realistic emissions trajectories. Each scenario requires a full ESM run, so these datasets cost months of computation.
The MIT team discovered that realistic scenarios can actually handicap an emulator: models trained on them underperform when asked to predict novel futures. The fix runs counter to intuition. The researchers built an algorithm that generates deliberately artificial, optimized scenarios — including futures that "look nothing like a realistic future."
"We don't care so much if the training scenarios are realistic. We are just trying to make the best possible data set to train an emulator," Womack said.
How does the algorithm tune the data?
The pipeline works in a loop:
- It generates an arbitrary climate scenario with a fast, simple climate model.
- It measures how well the emulator reproduces the ESM's result for that scenario, producing an error score.
- A programming technique built into the pipeline lets the researchers iteratively adjust the training data — spiking CO₂ emissions in 2050, for example, or cutting methane sharply in 2037.
- The optimization continues until the emulator's error drops below a set threshold.
The finished synthetic dataset contains rapid swings — "high-frequency wiggles and oscillations" — rather than the smooth emissions curves of standard scenarios. Those variations, Womack explained, let the emulator extract far more information about the overall system.
What changes for practitioners?
The gains are concrete. An emulator trained this way captured the effects of a local policy that cuts one pollutant while other emissions keep rising; every traditional emulator the team evaluated failed that test. And because training data now requires one ESM run rather than eight, production time shrinks sharply.
"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.
That speed addresses a chronic staleness problem. "Because the process of training an emulator can take months or even years, the emulators we have are usually out of date," said senior author Noelle E. Selin, a professor at the MIT Institute for Data, Systems, and Society and director of the MIT Center for Sustainability Science and Strategy. "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."
Selin also framed the longer-term stakes: "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."
The paper's coauthors are Shahine Bouabid (MIT EAPS), Andrei Sokolov and Popat Salunke (CS3), Glenn Flierl (EAPS), Sebastian D. Eastham (Imperial College London), and Selin.
The caveats are worth noting. The researchers have not yet validated the technique against training data generated by a full ESM, and they have tested it on one type of climate model so far. Next steps include that ESM-based test, applying the method to other climate models, and exploring whether it transfers to broader machine learning and computer vision tasks.
Faster emulators that stay accurate on unfamiliar futures would put current tools, not outdated ones, into planners' hands.
via phys.org/rss-feed/ (Source)
More from Marcus Bennett
Show full bio
News editor covering regulation, policy and their impact on general science news at SciBeat.
388 articles
Nearby plates
- 250 Climate Simulations Show Why Cloud Cover Numbers Mislead
- One AI Model Simulates Sand, Fluids and Proteins Alike
- Princeton AI Predicts Fusion Plasma Instabilities 200 Milliseconds Early
- After Math's Upheaval, Physicists Ask Whether Science Is AI's Next Conquest
- Locally Trained AI on Gut Microbe Data Could Shield African Livestock From Heat