MIT researchers have found a way to build fast climate prediction tools using far less costly data. The new method needs only one run of a large climate model, where older approaches often need six to eight.
The tools are called climate emulators. They are AI models that copy the behavior of large physics-based climate models and predict effects such as changes in temperature or sea level.
The research appears this week in the Proceedings of the National Academy of Sciences. The lead author is Christopher Womack, who earned his PhD at MIT in 2026 and is now a postdoc at ETH Zurich.
Why Training Data Costs So Much
Scientists usually make long-term climate forecasts with earth system models. These models show the climate in great detail, but one scenario can take months to run on a group of supercomputers.
Emulators run in a few seconds. That speed lets scientists test thousands of climate scenarios and spot unexpected results from regional actions.
To train an emulator, scientists often use six to eight future scenarios based on realistic emissions paths. Each one must first be run through a full earth system model, which makes the data slow and costly to produce.
The MIT team had earlier found that training on realistic scenarios can lead to weak emulators. These models often fail when asked to predict situations they have not seen before.
Strange Scenarios, Better Results
The new algorithm builds a synthetic set of climate scenarios designed to improve one emulator’s performance. The scenarios do not need to look like a real future.
“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,” Womack said. He added that the goal is to make the best possible dataset to train an emulator.
The process starts with a random scenario made by a fast, simple climate model. The team measures how well the emulator copies the results, then updates the training data again and again until the error falls below a set level.
The algorithm may, for example, sharply raise carbon dioxide emissions in 2050 or quickly cut methane emissions in 2037. The final data contain rapid swings instead of the smooth emissions paths seen in realistic scenarios.
Womack said these “high-frequency wiggles and oscillations” help the emulator learn more about the overall climate system. The emulator trained this way accurately predicted scenarios it did not see during training.
In one test, the new emulator captured the effects of a local policy that cut one pollutant while other emissions kept rising. The traditional emulators the team tested failed this task.
Senior author Noelle Selin, an MIT professor and director of its Center for Sustainability Science and Strategy, said training an emulator can take months or even years. She said this leaves current emulators often out of date, and called the new technique a “game-changer” for getting updated tools to researchers.
The research was part of the Bringing Computation to the Climate Challenge project. It was supported by Schmidt Sciences through the MIT Grand Challenges.
Next, the team plans to test the method using an emissions scenario produced by a full earth system model. They also want to try it on other types of climate models and in wider machine learning and computer vision work.
