Climate: AI trained on invented emission scenarios
· BehindLines newsroom
MIT researchers have developed a method to train artificial intelligence models for climate prediction faster, with a fraction of the usual data. MIT presented the work on 12 November 2026, saying it is due out this week in the journal Proceedings of the National Academy of Sciences. A manuscript also exists on the arXiv preprint server: submitted on 17 June 2026 and revised on 29 September, it predates MIT's announcement by several months, and its arXiv page mentions no journal.
"Emulators" to replace months of computation
Long-term climate projections rely on earth system models, which are very detailed but heavy: MIT says it can take months on a cluster of supercomputers to simulate a single possible future. "Emulators" are AI models that mimic their behavior and run in a few seconds, which, according to MIT, makes it possible to test thousands of scenarios.
The flaw of realistic scenarios
Usually, an emulator is trained with six to eight realistic emission trajectories, each requiring a full run of the reference model. MIT reports that its researchers had found that such scenarios produce emulators that perform worse on unfamiliar situations. The study's abstract on arXiv attributes this performance ceiling to the low structural diversity of the usual scenarios.
A made-to-measure dataset
Christopher Womack, lead author, now a postdoc at ETH Zurich, and his colleagues designed an algorithm that generates a set of optimized synthetic scenarios. According to MIT, it starts from an arbitrary scenario created with a simple, fast climate model, measures the gap between the emulator and the reference model, then updates the data iteratively until the error reaches a threshold. The result does not look like a plausible future: the trajectories contain sudden variations, and the algorithm could, for example, sharply raise CO2 in 2050 or rapidly lower methane in 2037. For Christopher Womack, these rapid oscillations allow the emulator to extract far more information about the system as a whole.
The full model then needs to be run only once, instead of six or eight times. According to the abstract on arXiv, with a simple climate model, an emulator trained on a single optimized scenario does better than an emulator trained on six standard trajectories (ScenarioMIP). The authors add that scenarios optimized with this simple model, then used to drive a model of intermediate complexity, yield a training set that performs better than the ScenarioMIP outputs. According to MIT, an emulator trained this way can also predict the effect of local policies that lower one pollutant while other emissions continue to climb, something the conventional emulators tested did not manage.
Why it matters
According to Noelle Selin, senior author of the study, training an emulator can take months or even years, so the available tools are often out of date; she believes that by shortening this delay, the method could be a game-changer. Christopher Womack hopes these fast, adaptable emulators will better serve the development of public policy. The researchers now want to test the technique with a scenario produced by a full model, and apply it to other types of climate models.