MIT researchers say they have developed a technique to streamline the production of AI models to enable faster, more trustworthy and adaptable projections of the effects of climate change.

Known as climate emulators, these specialized AI models can rapidly simulate the effects of human activities on future climate. They can run thousands of possible scenarios in a matter of minutes, including regional effects. Full-scale Earth system models (ESMs), by contrast, can take months on supercomputers to simulate a single scenario.

The new MIT technique aims to accelerate the training of climate emulators while improving their ability to predict unfamiliar scenarios.

“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,” says Christopher Womack (PhD ’26), a lead author of the paper describing the new technique. The research appeared in the Proceedings of the National Academy of Sciences.

Up-to-date climate projections will be vital tools in the coming years when it comes to testing the resilience of power grids, protecting public health and improving agricultural yields, for example. Researchers at Climate Action Tracker project that long-term global warming will exceed the 1.5-degree Celsius (2.7-degree Fahrenheit) threshold established by the 2015 Paris Agreement around 2030 as emissions continue to rise. The consequences of exceeding that threshold are a major topic at the annual UN climate change meeting in Turkey.

In the short term, continued warming increases the risks of more harmful storms, droughts, heat waves, flooding and other disasters. An “overshoot” occurs when global warming exceeds the 1.5-degree Celsius threshold before potentially falling below it again, making the need for reliable climate emulators that much more important for policy decisions.

The MIT breakthrough allows a climate emulator to be trained with a smaller amount of carefully selected synthetic data than the extensive datasets previously used in training. Perhaps most importantly, the approach allows climate emulators to predict outcomes for scenarios they did not encounter during training.

The MIT algorithm generates training scenarios that might include dramatic increases in carbon dioxide emissions in 2050 or sharp decreases in methane emissions in 2037, for example. These synthetic scenarios contain “high-frequency oscillations” that may seem unrealistic but help the emulator learn more about how the climate system responds to changing emissions.

“In the future, we are definitely going to see climates and scenarios we haven’t seen before,” says Noelle Selin, an MIT professor and co-author of the paper. “If we can simulate these different scenarios more robustly, we can provide more reliable information to the stakeholders who need it.”

The researchers demonstrated the technique using simplified and intermediate-complexity climate models, with further testing involving full-scale Earth system models still planned.

Selin adds that current climate emulators are often out-of-date because training takes months or even years. “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.”