There is a stubborn gap in the middle of every forecast. Meteorologists can tell you with reasonable confidence what the weather will do over the next few days, and climate scientists can describe how a season is likely to trend across months and years. But the window in between — roughly two weeks to two months out — has long resisted reliable prediction. It is close enough that the specific weather of any given day has faded into chaos, yet far enough that the slow, planetary rhythms of the ocean and atmosphere have not yet fully taken over. Researchers at the U.S. National Science Foundation National Center for Atmospheric Research (NSF NCAR) are now testing whether artificial intelligence can help close that gap, and their early work centers on a machine-learning tool with an unassuming name: CAMulator.
CAMulator is what scientists call an emulator. Rather than solving the physics of the atmosphere from first principles, as a conventional weather model does, an emulator studies the enormous archives of output that those physics-based models have already produced and learns to reproduce their behavior. Once trained, it can generate forecasts at a small fraction of the time and computational cost of the original. CAMulator was trained on the sixth version of the Community Atmosphere Model, or CAM6, a widely used simulation of global atmospheric conditions developed at NSF NCAR. According to the team's description of the project, the emulator can generate 480 simulated years in a single day — about 350 times faster than running CAM6 itself. The research is funded by the U.S. Department of Energy and NSF NCAR, and the underlying method is documented in a preprint titled CAMulator: Fast Emulation of the Community Atmosphere Model.
Why It Matters
The appeal of the subseasonal window is not academic. Whole sectors of the economy depend on knowing whether the coming month is likely to be hotter, colder, wetter, or drier than usual. Energy utilities plan around demand spikes that follow heat waves and cold snaps. Reservoir managers decide how much water to hold or release. Farmers weigh planting and irrigation choices against the odds of drought. For all of them, a forecast that reaches a few weeks further out, with even modestly improved skill, can translate into more resilient planning and steadier operations.
What makes the subseasonal range so difficult is that it sits at an awkward crossroads of predictability. The precise state of tomorrow's storm is long gone by week three, but the large, slow-moving signals that shape a season are still gathering. Accurate forecasts at this scale therefore hinge on capturing global atmospheric patterns tightly coupled to the state of the ocean. CAM6 can represent both, but its physics-based architecture is computationally expensive, which limits how many simulations scientists can afford to run and slows the pace of experimentation. By imitating the model at a sliver of the cost, CAMulator loosens that constraint. Crucially, it has shown skill at reproducing the kind of large-scale, slowly varying patterns that genuinely carry subseasonal predictability — the El Niño Southern Oscillation, the Pacific–North American pattern, and the North Atlantic Oscillation among them.
The Reaction
Kirsten Mayer, the NSF NCAR scientist leading the subseasonal research with CAMulator, reaches for a homely image to explain why these patterns matter so much. Predicting subseasonal weather, she suggests, is a little like guessing where a rubber duck will drift in a bathtub churned by a splashing toddler — the motion is chaotic and the outcome nearly impossible to call. But once a strong, organizing signal appears, the problem changes character. It becomes more like predicting the duck's path after someone turns on the faucet at full pressure: uncertainty remains, but a known force is now steering the water. When a signal such as El Niño is present, the atmosphere behaves with more discernible structure, and the emulator's ability to track patterns tied to that signal becomes genuinely promising.
The team has been putting that promise to the test in the open. Over the past year, Mayer and her colleagues have entered CAMulator in the AI Weather Quest, an international competition organized by the European Centre for Medium-Range Weather Forecasts, in which teams vie to produce the most accurate AI-based subseasonal forecasts. Rather than treating the emulator as a finished product, the researchers are using the contest as a proving ground, watching how it performs against rival approaches and folding the lessons — both from their own model and from others' successes — back into the next round of development. It is a notably candid way to build a scientific tool, one that treats limitations as data rather than something to obscure.
What Comes Next
CAMulator is not without weaknesses, and its designers are the first to name them. Because the emulator works by making a six-hour forecast, feeding that result back in to produce the next six hours, and repeating the cycle, small errors tend to compound as the forecast reaches further into the future. That gradual erosion of accuracy is a universal forecasting problem, not a quirk of this particular system, but it sets a clear target for improvement.
The next steps run in two directions. One is a sibling emulator called subCESMulator, built on nearly identical architecture but based on the Community Earth System Model, which brings interactive land and ocean variables into the picture; Mayer hopes those additional ingredients will sharpen the accuracy of subseasonal predictions. The other, and perhaps more quietly consequential, direction is access. Because emulators demand so little computing power, CAMulator and subCESMulator are designed to run on a modern laptop. That portability could let individual researchers run their own simulations without booking time on a supercomputer — a shift Mayer frames as democratizing the science and accelerating discovery across the community. Both emulators are built on a research platform called CREDIT, developed at NSF NCAR to make training and evaluating AI weather models more approachable.
Closing Thoughts
It is tempting to read a project like CAMulator as another entry in the long contest between physics and machine learning, but that framing misses what is actually happening here. The emulator does not replace the Community Atmosphere Model; it is trained on it, learns from it, and exists to let scientists ask more questions of it than brute computation would ever allow. The interesting story is not AI versus physics but AI in service of physics — a way to stretch a finite budget of supercomputer hours into a far larger space of experiments.
.jpg){kind=link}
.jpg){kind=link}
Whether emulators ultimately deliver the reliable subseasonal forecasts that so many industries want remains an open question, and the researchers are careful not to overpromise. What they have shown is a method that captures the atmosphere's most meaningful slow rhythms at a small fraction of the usual cost, tested transparently against peers, and built to run on ordinary hardware. If the coming forecasts do improve, the quiet enabler may turn out to be less a single clever model than the freedom to explore — the ability, as Mayer puts it, to ask questions that were simply out of reach before. In a field defined by uncertainty, that expanded room to experiment may prove as valuable as any single prediction.
한글 요약
미국 국립대기연구센터(NSF NCAR) 연구진이 하루하루의 날씨는 이미 혼돈에 빠졌지만 계절을 좌우하는 느린 대규모 신호는 아직 뚜렷하지 않은 애매한 구간이기 때문이다. CAMulator는 엘니뇨 남방진동(ENSO), 태평양–북미(PNA) 패턴, 북대서양진동(NAO)처럼 예측 가능성을 실제로 지닌 대규모·저속 변동 패턴을 잘 재현하는 것으로 나타나다. 연구를 이끄는 커스틴 메이어 박사는 강한 신호가 나타나면 예측이 훨씬 수원해진다고 설명하며, 연구팀은 유럽중기예보센터(ECMWF)가 주최하는 국제 대회 'AI Weather Quest'에 참가해 성능을 투명하게 검증하고 개선점을 반영하고 있다.
다만 6시간 예보를 반복 입력하는 구조상 시간이 지날수록 오차가 누적되는 한계가 있어, 연구진은 육지·해양 변수를 반영한 후속 에뮬레이터 subCESMulator를 개발 중이다. 특히 이들 에뮬레이터는 연산 부담이 작아 일반 노트북에서도 구동 가능하도록 설계돼, 슈퍼컴퓨터 없이도 개별 연구자가 직접 시뮬레이션을 돌릴 수 있게 함으로써 연구 접근성을 넓히려 한다. 물리 모델을 대체하는 것이 아니라 보완하며 더 많은 실험을 가능케 한다는 점에서, 이 접근의 진짜 가치는 단일 예측보다 '탐구의 폭'을 넓히는 데 있다는 평가다. (참고: NCAR/UCAR News, arXiv 2504.06007, AI Weather Quest)