Tuesday, November 17, 2026

Top 5 This Week

Related Posts

Climate: MIT makes AI emulators more reliable with unrealistic emissions scenarios


To decide on a climate policy, we must be able to quickly answer questions like: what happens if we significantly reduce methane in ten years, but CO2 continues to increase? The major climate models know how to answer this, but at the cost of months of calculations. Researchers from MIT propose a method to more quickly produce climate emulators reliable, these ultra-fast AI models that imitate large models. Their work is based on a counterintuitive idea: training these AIs with emission scenarios that do not resemble any realistic future.

What is a climate emulator?

To project changes in temperatures or precipitation, climatologists use Earth system models. These gigantic software programs describe in detail the atmosphere, oceans, ice and their interactions, based on the laws of physics. Their power comes at a cost: according to MIT, they can mobilize a cluster of supercomputers for months, and for a single future scenario.

THE emulators are AI models trained to reproduce the behavior of these large models. Once ready, they turn a few secondswhich makes it possible to test thousands of scenarios, estimate uncertainties and identify unexpected effects of a regional policy.

This MIT team, led by Noelle Selin, had already shown in 2025 that simple physical models could outperform deep learning for certain climate forecasts. This time she attacks the fuel of these tools: training data.

The problem: expensive and too wise data

Typically, an emulator is trained on six to eight future scenariosbased on realistic emissions trajectories. Each must be calculated by a full Earth system model, making these datasets expensive and time-consuming to produce.

The MIT researchers noted a second flaw: trained on these realistic scenarios alone, emulators can fail when faced with new situations. However, the future holds precisely combinations of emissions never before observed, and these are often those that decision-makers want to explore.

The solution: “bizarre” but instructive scenarios

The team developed an algorithm that produces a synthetic datasetoptimized to improve the predictions of a given emulator. The process takes place in several stages:

  1. a first arbitrary scenario is generated with a simple and rapid climate model;
  2. we measure to what extent the emulator manages to reproduce the results of the reference model;
  3. the algorithm modifies the scenario to reduce this error, for example by causing CO2 emissions to jump in 2050 or methane emissions to fall in 2037;
  4. the operation is repeated until the error falls below a fixed threshold.

Result: instead of smooth emissions curves, the scenarios obtained present sudden variations. “By introducing these rapid oscillations, our data allows the emulator to extract much more information about the entire system,” explains Christopher Womack, lead author of the study, now a postdoctoral researcher at ETH Zurich. “Our algorithm can generate strange scenarios that bear no resemblance to a realistic future, but that doesn’t matter: we’re just looking for the best possible dataset to train an emulator. »

Another major advantage: instead of rotating a model of the Earth system six or eight times, a single complete simulation is enough.

Two ways to train a climate emulator

Classic method

  • Six to eight realistic emission scenarios
  • A complete simulation for every scenario
  • Months, even years of preparation
  • Failures faced with new scenarios

MIT method

  • Synthetic scenarios optimized by an algorithm
  • Only one full simulation required
  • A fraction of the usual data
  • Better predictions on cases never seen before

More reliable predictions on new cases

According to MIT, the emulator trained with this method, using a much smaller amount of data, correctly predicts scenarios he has never seenwhere the traditional emulators tested were unable to adapt.

Speaking example: he manages to capture the effects of a local policy for reducing a pollutantwhile other emissions continue to increase. Classic emulators evaluated by the team failed this test. However, it is precisely this type of question, targeted and nuanced, that public decision-makers are asking themselves.

“In the future, we are bound to see climates and scenarios that we have never seen. If we can simulate them more robustly, we can provide more reliable information to those who need it,” said Noelle Selin, professor at MIT and lead author of the study.

Why it matters for climate decisions

Saving time has a very concrete consequence. According to Noelle Selin, training an emulator can take months or even years, so the tools available are often outdated at the time they are used. Shortening this step would make it possible to update the tools more quickly in the hands of researchers and practitioners.

A reliable emulator also changes the way we work: rather than choosing a few scenarios to calculate in advance, researchers can explore thousands of them, identify those that produce surprising effects, then reserve expensive full simulations for the most interesting cases. The applications are numerous: anticipating the needs of electricity networks, protecting public health during heat waves, adapting agricultural crops or evaluating the effect of regulations on a specific pollutant.

A team between climatology and computer science

The study brings together specialists from several disciplines. In addition to Christopher Womack and Noelle Selin, a professor in the Institute for Data, Systems and Society (IDSS) and the Department of Earth, Atmospheric and Planetary Sciences (EAPS) at MIT, it includes researchers from MIT’s Center for Sustainability Science and Strategy, an oceanographer and a professor from Imperial College London.

This intersection of skills is characteristic of current research on climate and AI: it is necessary to understand the physics of the climate system, master optimization techniques and know how to evaluate the reliability of a model in the face of unforeseen situations. Christopher Womack’s thesis, defended at MIT in 2026, already defends the idea that training data is one of the main obstacles to the quality of emulators.

Limitations and next steps

The researchers remain cautious about the scope of their results. In this study, the initial scenario was produced by a simple climate model; the team now wants test the method with a scenario from a model of the complete Earth system. She also wants to apply it to other types of climate models, and explore its interest in other areas of machine learning, such as computer vision.

The project is part of the Bringing Computation to the Climate Challenge (BC3) program, supported by Schmidt Sciences as part of the MIT Climate Grand Challenges. If the method keeps its promises on a larger scale, it could accelerate the production of tools capable of clarifying, in a few seconds, climate policy choices which until now required months of calculation.

Frequently asked questions

What is a climate emulator?

A climate emulator is an AI model trained to reproduce the behavior of a large model of the Earth system. Where the latter can require months of calculation on supercomputers for a single scenario, the emulator gives an answer in a few seconds. It is used to quickly test thousands of emissions scenarios.

Why does MIT use unrealistic scenarios?

Because they learn more at the emulator. Realistic scenarios follow smooth curves and look the same. MIT’s synthetic scenarios, optimized by an algorithm, contain sudden variations that better reveal how the climate system works. The emulator then better predicts situations it has never encountered.

Do emulators replace classic climate models?

No. They are rapid and simplified copies, whose reliability depends on the major models of the Earth system on which they are based. MIT’s method reduces the number of full simulations needed to just one, but doesn’t do without them. Emulators are mainly used to quickly explore numerous scenarios.

What will these climate emulators be used for?

They should help decision-makers quickly assess the effect of climate policies, such as the reduction of a specific pollutant in a region. According to MIT, they can inform the resilience of electricity grids, the protection of public health or the adaptation of agriculture. Updated more often, they remain more relevant.

Sources

References consulted for the writing of this article. The addresses are provided for information purposes only and are not links.

  1. MIT News via WebWire, New research could lead to faster, trustworthy AI models for climate prediction, November 2026www.webwire.com/ViewPressRel.asp?aId=364090
  2. Phys.org, Unrealistic emissions scenarios help AI climate models prepare for unfamiliar futures, November 2026phys.org/news/2026-11-unrealistic-emissions-scenarios-ai-climate.html
  3. MIT Center for Sustainability Science and Strategy, thesis by Christopher Womack on the optimal design of climate emulatorscs3.mit.edu/publication/118865

Report an error or inaccuracy in this article

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles