Unusual Emissions Scenarios Help Climate Models Learn From Fewer Simulations
Optimized training scenarios helped climate emulators handle unfamiliar emissions patterns; evaluation with a full Earth system model remains ahead.
Published November 14, 2026Intelligence / Artificial Intelligence / Predictive Analytics
A government can cut methane emissions while carbon dioxide emissions keep rising. A model predicting the resulting climate must separate the pollutants’ effects and calculate how they interact. Training software to handle unfamiliar combinations is a practical challenge for climate planning.
Researchers at the Massachusetts Institute of Technology (MIT) have developed a way to better train fast climate models. Their method changes the examples used to train the software, deliberately creating emissions patterns with sharp rises and falls. Those patterns expose relationships that smoother scenarios can conceal.
In a comparison using a simple climate model, an emulator trained on one carefully optimized scenario predicted unfamiliar scenarios better than an emulator trained on six standard future-emissions pathways. The team also demonstrated improved performance using an intermediate-complexity climate model. The study appeared November 9 in the Proceedings of the National Academy of Sciences, or PNAS.
The potential benefit is the ability to examine more combinations of emissions policies with limited computing resources. A government comparing several interventions needs estimates of their combined consequences as well as their individual effects.
Choosing better lessons
A climate emulator is a compact mathematical stand-in for a larger simulator. It learns relationships from the simulator’s results, then produces estimates more quickly. “Climate emulators can run in just a few seconds,” lead author Christopher B. Womack said in MIT’s November 12 account. Womack completed his doctorate at MIT and is now a postdoctoral researcher at ETH Zurich.
Producing the training examples can be expensive. Researchers commonly run detailed simulations along plausible future emissions pathways, then use the results as lessons for the smaller model. The new method asks which pathways teach the emulator the most.
The team starts with a simpler climate model and repeatedly adjusts a proposed training scenario. The training pipeline can calculate how a small change in that scenario affects the emulator’s prediction error. The algorithm uses that information to improve its examples. Abrupt variations help the emulator distinguish different climate responses.
The logic resembles experimental design. When several inputs always change together, separating their effects becomes difficult. Deliberately varying them gives the model more information about each relationship.
Preparing for larger climate simulations
Testing the approach with a full Earth system model is a next step, along with extending it to other model types. That work will establish how well the demonstrated improvement transfers to the greater complexity of detailed climate simulations.
The study points to a choice for research centers allocating scarce computing time: devote some simulations specifically to teaching fast models. Carefully designed examples could allow each expensive simulation to support a wider range of policy questions.