The Pulse

MIT team tests climate AI training method using less data

Researchers used optimized emissions scenarios to train climate emulators on smaller datasets. In a simple climate-model test, an emulator trained on one optimized scenario outperformed one trained on six standard pathways.

AI.info Team · · fact-checked · against 4 sources ↓ · Published 12 Nov 2026

An AI climate emulator trained on one specially designed scenario outperformed an emulator trained on six standard pathways in a simple climate-model test, according to a study in the Proceedings of the National Academy of Sciences. The researchers developed a way to optimize training data so that an emulator can handle scenarios unlike those it saw during training.

The study appeared online on November 9. Lead author Christopher B. Womack is now a postdoctoral researcher at ETH Zurich, while senior author Noelle E. Selin is an MIT professor and director of its Center for Sustainability Science and Strategy.

Replacing costly training runs

Climate emulators are AI models that imitate the behavior of larger earth system models. Scientists use them to explore how human activity might affect future temperatures, sea levels and other aspects of climate.

A physics-based earth system model can take months on a cluster of supercomputers to project a single future scenario. Emulators, by contrast, can run in seconds, allowing scientists to test many different scenarios.

Training the emulators still requires data from climate-model runs. Researchers often use six to eight scenarios based on plausible future emissions trajectories, running each through a full earth system model to create the training set, according to MIT News. The new approach is intended to reduce that requirement: the researchers say a user would need to run the full model once to generate training data.

Designing scenarios for learning

The method begins with an arbitrary scenario generated by a fast, simple climate model. The researchers measure how well an emulator reproduces the model’s results, then repeatedly adjust the training data to improve its predictive skill.

Those adjustments can produce scenarios that look unlike plausible emissions pathways. Instead of smooth trajectories, the synthetic data contain rapid changes, such as a sharp rise in carbon dioxide emissions or a quick fall in methane emissions.

The paper reports that, in the simple-model test, the emulator trained on an optimized scenario distinguished the different physical behaviors of climate forcing agents, including greenhouse gases and aerosols, without separate runs for individual forcing agents. It also performed better despite being trained on a smaller dataset than the comparison emulator.

The researchers also report that scenarios optimized with the simple model produced a more capable emulator when used with an intermediate-complexity climate model. Their comparison in that test was an emulator trained on outputs from standard scenario pathways.

Policy tests and next steps

In a test described by MIT News, an emulator trained with the new technique captured the effects of a local pollution-control policy that reduced one pollutant while other emissions continued to rise. Traditional emulators evaluated by the team failed that test.

The team plans to test the technique using an emissions scenario generated by a full earth system model. It also wants to apply the approach to other climate models and investigate potential uses in broader machine learning and computer vision.

The paper says the code and data needed to reproduce the work are publicly available on Zenodo. The research was part of the Bringing Computation to the Climate Challenge project, with support from Schmidt Sciences through the MIT Grand Challenges.