At a glance
MIT researchers developed a training technique that builds synthetic, deliberately unrealistic emissions scenarios to train climate emulators. Emulators trained this way accurately predict outcomes for scenarios they never saw during training, while traditional emulators failed. The method needs only a single run of a full climate model instead of six to eight, cutting production time from months or years to a fraction of that. The research appears in the Proceedings of the National Academy of Sciences.
- Climate emulators are fast AI surrogate models that mimic full earth system models and run in seconds instead of months
- Traditional emulators are trained on six to eight realistic emissions scenarios and fail when facing unfamiliar futures
- The MIT algorithm generates synthetic scenarios with rapid variations that teach the emulator far more about the system
- An emulator trained on the smaller synthetic dataset predicted unseen scenarios accurately while traditional emulators failed
- Only one run of a full earth system model is needed to generate training data, instead of six to eight costly runs
MIT researchers have found a way to make AI climate models both faster to build and more trustworthy when the future surprises us. Published this week in the Proceedings of the National Academy of Sciences, their training technique produces climate emulators, fast AI surrogates that mimic enormous physics based climate models, using only a fraction of the data normally required. It matters because planners who size power grids, protect public health, and plan harvests depend on climate projections, and today those projections take too long and fail when reality departs from scripted scenarios.
The problem the MIT team attacked sits at the heart of climate modeling. Earth system models, the gold standard for projecting future temperature, precipitation, and sea level, are so detailed that a single run can take months on a cluster of supercomputers. And that single run simulates just one possible future. Testing thousands of policy choices, from emissions cuts to regional pollution controls, is simply impossible at that speed.
What a Climate Emulator Is
A climate emulator is a compact AI model taught to reproduce the behavior of a full earth system model. Feed it an emissions trajectory and it returns the projected climate response, temperatures, sea level, pollutant concentrations, in a matter of seconds. Because they are so fast, emulators let scientists test thousands of varied scenarios, produce more reliable estimates, and catch unexpected outcomes of regional interventions that would otherwise go unnoticed.
But emulators inherit the limits of their training data. The standard approach is to run a full earth system model six to eight times, once per realistic future emissions pathway, and use those outputs to teach the AI. Each of those runs is expensive and slow. Worse, the MIT researchers found that training on realistic scenarios produces emulators that underperform whenever they face something unfamiliar, which is precisely the situation climate science is heading into.
Training on Deliberately Unrealistic Scenarios
The insight behind the new method is counterintuitive: the training scenarios do not need to be realistic at all. The MIT algorithm generates a synthetic dataset of emissions scenarios specifically optimized to teach one emulator as much as possible. Some generated scenarios look nothing like any plausible future. The algorithm might spike carbon dioxide emissions dramatically in the year 2050, or crash methane emissions in 2037, then measure how well the emulator reproduces the earth system model's response and update the data to shrink the error.
The result is training data full of rapid variations and sharp swings rather than the smooth trajectories of standard emissions pathways. "By introducing these high-frequency wiggles and oscillations, our data enable the emulator to extract a lot more information about the overall system," said Christopher Womack, lead author of the paper, who completed his PhD at MIT in 2026 and is now a postdoc at ETH Zurich.
The efficiency gain is substantial. Instead of running the full climate model six to eight times to build a training set, a user runs it once. That collapses a production pipeline that today takes months or even years, which is why, as senior author Noelle Selin notes, the emulators researchers rely on are usually out of date by the time they are finished.
The Test That Traditional Emulators Failed
The researchers demonstrated the payoff with a hard case: a scenario in which a local pollution control policy cuts one pollutant while other emissions keep rising. This is exactly the messy real world combination that regional regulators create, and it is not something emulators trained on standard smooth pathways usually see. The emulator trained with the MIT synthetic data captured the effect. The traditional emulators the team evaluated failed this test.
"In the future, we are definitely going to see climates and scenarios we haven't seen before," said Selin, who is a professor in the MIT Institute for Data, Systems, and Society and the Department of Earth, Atmospheric and Planetary Sciences, and director of the MIT Center for Sustainability Science and Strategy. "If we can simulate these different scenarios more robustly, we can provide more reliable information to the stakeholders who need it."
Why It Matters for Policy
Climate policy is made under uncertainty, and the tools that quantify that uncertainty shape what governments are willing to do. When an emulator can evaluate a proposed regulation in seconds, decision makers can compare thousands of versions of a policy, including combinations no one has tried before, instead of a handful. Faster production also means the models behind those answers can be refreshed as new science arrives rather than aging for years on a shelf.
The research is part of the Bringing Computation to the Climate Challenge project at MIT and was supported by Schmidt Sciences through the MIT Grand Challenges. The paper's coauthors include Shahine Bouabid, Andrei Sokolov, Popat Salunke, Glenn Flierl, Sebastian D. Eastham, and Noelle E. Selin. Next, the team plans to test the technique with a full earth system model run and apply it to other climate models and to broader machine learning settings.
The larger lesson reaches beyond climate science. AI models are only as good as the data they are trained on, and realistic data is not always the most informative data. Sometimes the best teacher for a model of the future is a future that will never happen.
Common Questions
What is a climate emulator?
A climate emulator is a fast AI surrogate model trained to mimic the behavior of a full earth system model. It simulates how human activities like greenhouse gas emissions affect future climate in seconds rather than months.
What did MIT researchers announce in November 2026?
MIT researchers published a training technique in the Proceedings of the National Academy of Sciences that generates synthetic emissions scenarios to train climate emulators, making them faster to produce and more accurate on unfamiliar scenarios.
Why do traditional climate emulators fail on new scenarios?
Traditional emulators are trained on six to eight realistic emissions pathways. MIT found that this realistic training data leaves the model unprepared for novel situations, such as a regional policy that cuts one pollutant while other emissions keep rising.
How much faster is the MIT training method?
The method needs only a single run of a full earth system model to generate training data instead of six to eight runs, cutting an emulator production pipeline that can take months or years down to a fraction of the time.
Sources: MIT News, "New research could lead to faster, trustworthy AI models for climate prediction" (November 12, 2026); Womack et al., Proceedings of the National Academy of Sciences (2026); MIT Center for Sustainability Science and Strategy; MIT Schwarzman College of Computing; Schmidt Sciences and the MIT Grand Challenges.