A climate researcher wants to compare hundreds of possible futures: What happens if carbon dioxide emissions fall quickly? What if methane emissions decline while other pollutants continue to rise? What if a policy is delayed by a decade?
A detailed climate simulation can be expensive to run. A fast artificial intelligence model called an emulator can approximate a larger model’s responses in seconds, making it easier to explore many scenarios. But there is a catch: the emulator can struggle when a new scenario behaves differently from the examples it learned from.
A study published in PNAS on November 9, 2026 offers an unexpected approach. Instead of training an emulator only on plausible future emissions pathways, researchers deliberately constructed unusual, rapidly changing scenarios designed to teach it more about how a climate model responds.
The result is a useful lesson in both climate science and machine learning: a training example can be unrealistic as a forecast while still being informative as an experiment. The researchers tested their method with a simple climate model and an intermediate-complexity model. They have not yet demonstrated the complete workflow on every full-scale Earth system model, and an emulator’s speed does not make its answers independent evidence about the real world.
What is a climate emulator?
Climate models represent interactions among components of the Earth system, including the atmosphere, oceans, land and processes affected by greenhouse gases and other pollutants. Scientists run these models under different assumptions to study possible outcomes.
Detailed simulations are valuable, but they consume computing resources. If researchers want to examine many combinations of emissions or policies, running a complex simulation for every possibility can be impractical.
An emulator is a faster approximation trained on the input-output behavior of another model. Imagine using a detailed flight simulator to generate examples and then training a smaller system to estimate how the simulated aircraft responds to different controls. The smaller system may answer quickly, but its knowledge comes from the examples and the simulator that produced them.
Climate emulators work in a similar way. Their task is not to replace observations or independently discover what the climate must do. They help researchers explore the behavior of a chosen underlying model at much lower computational cost.
Why familiar-looking training data can leave gaps
Researchers often train climate emulators using a set of standard emissions pathways. These pathways are designed to support comparisons of possible futures, so they tend to have coherent assumptions and relatively smooth trajectories.
That is useful for comparing scenarios. It can be less useful for teaching a machine-learning system how to separate the effects of different inputs.
Consider a simplified example. In a set of training examples, carbon dioxide and methane emissions might often rise or fall in similar patterns. An emulator could learn the combined pattern without learning enough about how the simulated climate responds when one changes and the other does not.
Now ask it to estimate an unfamiliar intervention: methane falls quickly, carbon dioxide rises more slowly, and aerosol pollution changes on a different schedule. If the new combination lies outside the patterns it learned, a confident-looking output may still be inaccurate.
This is a version of a familiar problem in artificial intelligence: a model can perform well on data resembling its training examples and much worse when the structure of the question changes. Here the issue is not merely the number of examples, but whether those examples contain enough varied information to distinguish the processes the emulator needs to approximate.
Teaching with deliberately strange scenarios
The researchers asked whether they could design more informative training scenarios. Their method uses a simple climate model to assess how changes to a proposed training scenario affect the emulator’s performance, then iteratively adjusts that scenario.
The resulting emissions pathways can have rapid changes or oscillations that would be unlikely as literal predictions of future human activity. Their role is different: they are experimental inputs used to reveal how the simulated system responds.
A useful analogy is checking a speaker’s sound across many frequencies. A brief test signal need not sound like a song to reveal how the speaker behaves. Similarly, a deliberately varied emissions pathway may help a model learn responses that smooth, correlated pathways reveal less clearly.
There is a limit to the analogy: the climate system includes processes with different timescales and interactions, and a scenario that is useful for one emulator or underlying model may not be optimal for another. The researchers’ point is that training data can be designed for the learning task rather than selected only because they resemble a plausible future.
What the study actually found
In the simple-model experiment, an emulator trained on one optimized scenario performed better on the researchers’ evaluation than an emulator trained on six standard ScenarioMIP pathways. The authors report that the optimized data helped the emulator distinguish responses to different climate forcing agents, including greenhouse gases and aerosols.
The team also used scenarios optimized with the simple model to drive an intermediate-complexity climate model. The resulting training data produced a more skillful emulator than training on the standard pathway outputs in that experiment.
Those are encouraging comparisons, but they have specific boundaries. “One scenario beat six” describes the setups tested in the paper; it does not establish that a single simulation will outperform six in every climate model, for every output or for every possible future. Nor does it mean that all uncertainty about future emissions, climate physics or regional impacts disappears.
The PNAS paper reports both the simple-model and intermediate-complexity results. MIT’s account of the work says the researchers want to test the approach with a full Earth system model in future work. That distinction matters when interpreting claims about reducing the number of expensive simulations.
Three different kinds of uncertainty
Suppose an emulator estimates warming for a proposed emissions pathway. At least three questions deserve separate answers.
First, did the emulator accurately reproduce its parent model? If the approximation fails on the chosen pathway, speed has little value. Testing against held-out scenarios and, when practical, additional runs of the parent model helps measure this error.
Second, how well does the parent model represent the real climate? An emulator can faithfully reproduce a model that still has limitations. Different climate models can represent physical processes differently, and some regional outcomes are harder to estimate than global averages.
Third, will the assumed emissions pathway actually occur? That depends on future human choices, technologies, economies and policies. A projection conditional on a scenario is not a prediction that the scenario itself will happen.
Improving the first question—the fidelity of the emulator—does not automatically resolve the other two. It makes the exploration tool more useful within a larger scientific process.
Why this matters beyond climate
The research connects two areas often discussed separately: the design of scientific experiments and the training of AI models.
In ordinary machine learning, more data can improve a system, but additional examples are not equally informative. If many observations repeat the same relationship, they may add less than a carefully chosen example that exposes a missing distinction.
In science and engineering, researchers routinely choose inputs to learn about a system: vary the temperature in an experiment, test a material under different loads, or change one parameter in a simulation. The climate-emulator study applies that logic to the question of what an AI system should be shown during training.
For a software analogy, a program tested only with typical inputs might look reliable until it encounters an unusual combination. A deliberately chosen test case can expose a weakness that many similar routine cases miss. The analogy is imperfect: here, the scenarios are used to train an approximation, not merely to test it. But both examples show why variety and purposeful design matter.
InsightArea has previously explored a related question about using AI to understand ecological systems. Both areas face a common challenge: the natural world contains interactions that a model may not learn if important combinations are missing from its data.
What would make this useful in practice?
One important next step is testing the approach across different complex climate models, not just the models used in this study. Researchers would also need to examine which outputs benefit, how errors vary across regions and timescales, and whether optimized scenarios remain useful when scientific questions change.
There is a practical trade-off. Designing and running specialized training scenarios takes work and computing time. A modeling center would need to compare that investment with the benefit of a faster emulator that can answer a wider range of questions. The right choice may depend on how often the emulator will be used and what decisions it is intended to inform.
For policymakers and other users, the most important safeguard is to see what a result represents: an estimate from an emulator of a particular model, under particular assumptions, accompanied by uncertainty. A fast answer should make it easier to investigate alternatives, not encourage false precision.
The broader insight is that better AI is not always a matter of adding more layers or collecting more of the same examples. Sometimes the useful question comes earlier: what should we show the model so that it learns the distinctions we will need later?
Explore more ideas about doing difficult things. In How to Do Hard Things: The Discomfort Trap, Costin Liculescu explores how to respond to discomfort and keep working toward meaningful goals.

Be First to Comment