Researchers at the Massachusetts Institute of Technology have developed an artificial intelligence training method that generates climate emulators capable of simulating the effects of human activities on climate much more quickly than traditional approaches. Published in the Proceedings of the National Academy of Sciences, the study demonstrates that training an emulator with a smaller amount of carefully selected synthetic data allows the model to accurately predict outcomes of climate scenarios it did not encounter during training. These specialized artificial intelligence models act as surrogates that predict future impacts, such as changes in global temperature and sea level, in a matter of seconds.
Accurate and up-to-date climate projections provide critical guidance for long-term planning decisions, including efforts to enhance the resilience of electrical power grids, protect public health, and maximize agricultural yields. Because climate emulators execute in seconds rather than months, scientists can deploy them to evaluate thousands of diverse climate scenarios. This rapid turnaround provides more reliable estimates of potential outcomes and helps researchers identify unexpected regional consequences resulting from specific interventions.
Conventional climate projections rely primarily on full-scale earth system models, which are complex physics-based simulations that capture interactions among the major components of the global climate system in extreme detail. While powerful, earth system models require clusters of supercomputers and can take several months of computation to project the progression of a single future climate scenario. To bypass these extensive computational bottlenecks, scientists design climate emulators as fast surrogate models that mimic the behavior of full-scale earth system models, enabling rapid testing of multiple scenarios, refined forecasts, and localized analyses.
Standard emulator development typically requires training an artificial intelligence model on six to eight realistic future scenarios based on projected greenhouse gas emission trajectories. Producing this training data demands that each realistic scenario be simulated through a full-scale earth system model, making dataset generation expensive and time-consuming. Furthermore, the MIT research team identified that training models exclusively on smooth, realistic trajectories leads to underperforming emulators that struggle or fail when tasked with predicting unfamiliar or novel climate scenarios.
To overcome these limitations, lead author Christoher Womack, a former MIT doctoral student now at ETH Zurich, and senior author Noelle E. Selin, a professor in MIT’s Institute for Data, Systems, and Society and Department of Earth, Atmospheric and Planetary Sciences, devised an optimization algorithm that generates synthetic training scenarios. Rather than running a full earth system model six or eight times, the new pipeline requires running the full-scale model only once, drastically reducing the data volume and computational investment needed to train an emulator.
The training process begins with the generation of an arbitrary climate scenario using a simplified, fast climate model. The researchers evaluate the emulator’s ability to reproduce the model’s outputs, deriving a quantitative measure of error. Utilizing their specialized programming pipeline, the algorithm iteratively adjusts the synthetic training data—for instance, by abruptly elevating carbon dioxide levels in 2050 or sharply cutting methane emissions in 2037—until the emulator’s predictive error falls below a specified target threshold.
The resulting synthetic training datasets do not follow the smooth, realistic curves characteristic of projected human emissions, but instead feature deliberate, rapid variations. Womack noted that introducing these high-frequency oscillations and variations forces the emulator to extract significantly more structural information about the climate system. By learning from these synthetic fluctuations, the emulator achieves greater overall predictive accuracy and adaptability across unseen conditions.
In validation tests, an emulator trained via this synthetic approach successfully captured complex regional effects, such as local pollutant control policies where one specific emission decreases while other emissions continue to rise. In contrast, conventional emulators evaluated by the researchers failed to adapt to these divergent conditions. Selin emphasized that because future decades will inevitably bring unprecedented climate conditions, robustly simulating novel scenarios ensures that policymakers and stakeholders receive more reliable projections.
The co-authors on the study include Shahine Bouabid and Glenn Flierl from the MIT Department of Earth, Atmospheric and Planetary Sciences; Andrei Sokolov and Popat Salunke from MIT’s Center for Sustainability Science and Strategy; and Sebastian D. Eastham of Imperial College London. The project was conducted as part of the Bringing Computation to the Climate Challenge initiative and received funding support from Schmidt Sciences through the MIT Grand Challenges.
Looking forward, the research team plans to test their optimization technique using emissions scenarios generated directly by full-scale earth system models. They also intend to apply the synthetic data methodology to alternative climate modeling frameworks, as well as investigate its utility across broader machine learning and computer vision applications. By cutting emulator training times down from months or years, the researchers aim to provide practitioners with consistently updated tools for climate policy decisions.
A central scientific problem driving the research stems from the trade-offs inherent in contemporary climate simulation. Full-scale earth system models incorporate extensive physical relationships governing atmosphere, ocean, and terrestrial dynamics. Because these models track complex, high-dimensional physical equations across fine grids, simulating a single climate pathway across decades or centuries demands substantial high-performance computing allocations. In practical terms, this intensive computational requirement restricts researchers to examining only a handful of predetermined pathways, limiting the capacity to explore a wide distribution of localized interventions or unexpected policy decisions. Climate emulators serve as computationally agile surrogates designed to approximate these comprehensive outputs, generating projections within seconds and enabling researchers to test thousands of varied scenarios that would otherwise be computationally intractable.
Despite their speed, existing climate emulators encounter severe operational bottlenecks during training. Conventional emulator construction relies on supervised training using full earth system model runs driven by smooth, plausible emissions scenarios. Because running a full earth system model through six to eight distinct emissions pathways requires months of supercomputing time, the resulting training datasets are costly and slow to acquire. Consequently, climate emulators frequently lag behind modern data needs, often taking months or even years to build and leaving scientific and policy communities with outdated predictive models. Moreover, emulators optimized solely on smooth, realistic trajectories often overfit to those standard curves, leaving them vulnerable to significant projection errors when confronted with novel, unobserved conditions such as unconventional regional mitigation strategies.
To overcome these challenges, the investigators introduced a synthetic data generation framework that reformulates how climate emulators learn physical system behavior. In this framework, the team initializes the training process by running an arbitrary scenario through a simplified, rapid climate model. Rather than attempting to match realistic emissions trajectories, the algorithm compares the emulator’s projected response directly against the output of the reference model to compute an explicit error signal. By using a specialized programmable training pipeline, the researchers iteratively perturb emissions pathways to maximize the information content delivered to the neural surrogate. Through successive iterations, the optimization framework dynamically adjusts emissions inputs until the emulator’s predictive error decreases below a predetermined threshold.
The resulting synthetic scenarios diverge sharply from standard historical or projected emissions paths. Instead of smooth, monotonic increases or gradual decreases, the generated datasets introduce abrupt, high-frequency oscillations. For example, the algorithm introduces synthetic scenarios where carbon dioxide emissions surge precipitously in 2050 or methane emissions decline steeply in 2037. The researchers found that these non-smooth fluctuations force the surrogate network to explore system boundaries and capture underlying non-linear responses across varied timescales. By training on these rapid oscillations rather than realistic, smooth trends, the emulator learns generalized systemic dynamics rather than memorizing particular trajectory profiles, allowing it to predict outcomes across scenarios that lie well outside standard projections.
This synthetic training design significantly lowers the computational barrier required to build accurate climate surrogates. Rather than executing six to eight separate full-scale earth system model simulations to construct an effective training suite, the new methodology requires running a full-scale model only once. This reduction in required training volume sharply accelerates development, transforming a multi-month or multi-year pipeline into a fast, repeatable procedure. Consequently, researchers can continuously produce tailored emulators that integrate fresh assumptions or specific research questions without tying up supercomputing clusters for prolonged operational windows.
The researchers demonstrated that this increased adaptability translates directly into higher predictive performance during policy-relevant stress tests. In experimental evaluations, the team tested emulators on scenarios representing targeted local pollutant controls, where one particular atmospheric pollutant is sharply reduced while other greenhouse gas emissions continue to climb. Standard emulators trained on traditional, realistic emission pathways failed when tasked with projecting the consequences of these decoupled conditions. In contrast, the emulator trained on synthetically generated, oscillating pathways successfully and accurately predicted the climate impacts of these asymmetric regional policies, demonstrating robust generalization to unobserved atmospheric mixtures.
Looking forward, the authors identified several key next steps to broaden the scope and operational utility of their scenario optimization framework. The team plans to evaluate the method using synthetic emissions scenarios derived directly from full-scale earth system models rather than simplified climate representations. Additionally, the researchers plan to extend the training framework to diverse climate modeling architectures and investigate whether the synthetic data optimization methodology can improve generalization in broader machine learning contexts, including complex computer vision systems. By streamlining the production of robust, adaptable surrogates, the methodology aims to deliver timely, reliable climate projections to planners managing public health, agricultural resilience, and energy infrastructure.
Subject of Research: Artificial intelligence training technique for climate emulators
Article Title: New research could lead to faster, trustworthy AI models for climate prediction
Article References: New research could lead to faster, trustworthy AI models for climate prediction. Original publication (Bibliographic details unavailable.)
Image Credits: AI Generated
DOI: Not provided
Keywords: climate emulators, artificial intelligence, earth system models, climate prediction, synthetic data, scenario optimization, greenhouse gas emissions, carbon dioxide emissions, methane emissions, climate policy, Massachusetts Institute of Technology, ETH Zurich, Imperial College London, Proceedings of the National Academy of Sciences, Center for Sustainability Science and Strategy, Bringing Computation to the Climate Challenge
News Source: Denise Maddox. (November 12, 2026). MIT Researchers Develop Synthetic Training Method for Faster, Adaptable Climate Emulators. Scienmag.
