Strange emissions scenarios help fast AI climate models handle new futures

In its own tests, an MIT team in the United States found that climate emulators (fast AI stand-ins for full climate models) trained on deliberately odd scenarios predicted unseen ones. The study is published in the Proceedings of the National Academy of Sciences.

Share
Strange emissions scenarios help fast AI climate models handle new futures
Illustration: AI-generated

Fast AI stand-ins for climate models predicted climate scenarios they had not seen in training, after learning from deliberately odd emissions scenarios, according to a team at the Massachusetts Institute of Technology (MIT) in the United States. MIT's announcement, dated 12 November 2026, calls these stand-ins climate emulators.

A climate emulator is a specialised AI model that quickly simulates how human activities affect a future climate, such as temperatures or sea level. Full Earth system models (ESMs) can take months on a cluster of supercomputers to project one possible future. Emulators run in seconds, so scientists use them to test thousands of scenarios.

Emulators are usually trained on six to eight realistic emissions scenarios, and each must be run through a full ESM, which makes the data costly and slow to produce. Noelle E. Selin, senior author and a professor in the MIT Institute for Data, Systems, and Society, says training an emulator can take months or even years, so existing emulators are usually out of date.

The new method starts with an arbitrary scenario from a fast, simple climate model. An algorithm measures the emulator's error, then changes the training scenarios again and again until the error falls below a threshold. For instance, it could raise carbon dioxide (CO2) emissions sharply in 2050 or cut methane emissions rapidly in 2037. With the new method, the team says, a user would need to run the full ESM only once.

The resulting scenarios look unrealistic: emissions rise and fall rapidly instead of following smooth paths. Lead author Christoher Womack PhD '26, now a postdoc at ETH Zurich, says in the announcement: “By introducing these high-frequency wiggles and oscillations, our data enable the emulator to extract a lot more information about the overall system.”

In the researchers' own tests, an emulator trained on a much smaller amount of carefully selected made-up (synthetic) data accurately predicted scenarios it had not seen in training, while traditional emulators failed to adapt. In one example from those tests, the new emulator captured the effects of local pollution rules that cut one pollutant while other emissions keep rising. The traditional emulators the team evaluated failed this test. The work is published in the Proceedings of the National Academy of Sciences, in a paper titled “Optimal scenario design for climate emulation”.

So far, the training scenarios start from a fast, simple climate model, not a full ESM. Next, the researchers say they want to test the technique with an emissions scenario generated by a full ESM, and to apply it to other types of climate models and to other computer tasks, such as those that work with images.

Sources