Climate projections drive decisions about power grids, public health and farming, yet the physics-based models behind them can take months on a supercomputer to simulate a single future. A new technique from MIT aims to shrink that bottleneck by changing what AI models learn from.
In research announced by MIT News on November 12, 2026 and published in the Proceedings of the National Academy of Sciences, a team describes a way to train climate emulators with a fraction of the usual data while making them more adaptable.
The problem with slow climate models
Scientists typically use earth system models (ESMs) for long-term projections of temperature or precipitation under different greenhouse gas emissions. These models capture the major parts of the climate in great detail, but they are heavy: one scenario can take months on a cluster of supercomputers.
Climate emulators are fast AI surrogates that mimic an ESM. They run in a few seconds, so researchers can test thousands of scenarios, explore specific pollutants and regional effects, and spot unexpected outcomes of local interventions.
The catch is training. Developers often train an emulator on six to eight future scenarios based on realistic emissions paths, and each must be run through a full ESM. The MIT team had recently found that realistic scenarios can produce emulators that fail on unfamiliar situations.
A smarter way to build training data
The researchers developed an algorithm that generates a synthetic dataset of optimized climate scenarios tailored to one emulator. It starts with an arbitrary scenario from a fast, simple climate model and measures how well the emulator reproduces the ESM’s results. That gives an error measure.
Because of the programming technique used in the training pipeline, the training data can then be updated iteratively to raise the emulator’s skill. The algorithm might sharply raise CO2 emissions in 2050 or quickly cut methane in 2037. It repeats until the error falls below a threshold.
The result looks nothing like a smooth, realistic emissions curve. “We are just trying to make the best possible dataset to train an emulator,” says lead author Christopher Womack, now a postdoc at ETH Zurich. Those rapid “wiggles and oscillations,” he explains, let the emulator extract more information about the overall system. Importantly, a user only needs to run the full ESM once rather than six or eight times.
What the tests showed
MIT reports that an emulator trained on a smaller amount of carefully selected synthetic data accurately predicted outcomes for scenarios it never saw in training, while traditional emulators failed to adapt. One example involved local pollutant control policies that reduce one pollutant while other emissions keep rising, a case the conventional emulators in the evaluation could not handle.
The paper, titled “Optimal scenario design for climate emulation,” also lists Shahine Bouabid, Andrei Sokolov, Popat Salunke, Glenn Flierl and Sebastian Eastham as co-authors, with Noelle Selin as senior author. The work was part of the Bringing Computation to the Climate Challenge project and was supported by Schmidt Sciences through the MIT Grand Challenges.
What it means
Selin notes that training an emulator can take months or even years, so the tools available are often out of date. Cutting production time could put updated tools into the hands of researchers and practitioners sooner, and stronger generalization makes outputs easier to trust for policy-relevant scenarios.
The approach is also a reminder that in scientific AI, the design of the training data can matter as much as the model. Next, the team wants to test the technique with an emissions scenario generated by a full ESM, apply it to other kinds of climate models, and explore whether it helps in broader machine learning and computer vision tasks.
Conclusion
MIT’s method does not replace big climate models; it makes their fast imitators cheaper to build and more dependable on new situations. If it holds up in the follow-up tests, planners could get faster answers to what-if questions about emissions.