A 7.7-million-parameter atmosphere ran a coupled climate for 140 years

A climate model that replaced its atmosphere with a small neural network ran for 140 years without having its sea-surface temperature forcibly corrected. After an initial 31-year adjustment, its global mean was 291.30 K—within 0.05 K of the benchmark CESM2 climate model. That is a meaningful step toward making long ocean experiments cheaper. It is not a replacement for a full climate model: the tropical Pacific ran too quietly, and regional biases remain. But FASST shows that an atmosphere can be simplified dramatically without immediately sending a coupled ocean climate off course.

The distinction matters because the ocean is slow. Heat absorbed at the surface can take decades or centuries to move through the deep ocean, and the questions scientists ask about that process often demand long integrations and many repeated simulations. Yet the atmosphere—the part that exchanges heat, water, and momentum with the ocean—can dominate the computing cost. FASST tries a striking compromise: keep the physically based ocean, replace the atmosphere and sea ice with compact learned models, and see whether the parts still work together over a century.

Global mean sea-surface temperature after spin-up

Label	Value
FASST (years 31–140)	291.3 K; ± reported variability
CESM2 piControl	291.25 K; ± reported variability

• FASST (years 31–140) 291.3 K; ± reported variability
• CESM2 piControl 291.25 K; ± reported variability

FASST post-transient mean: 291.30 ± 0.22 K; CESM2 piControl: 291.25 ± 0.10 K.

The Science

FASST couples the Parallel Ocean Program version 2 (POP2), a physics-based ocean model, to two machine-learning emulators. An emulator is a model trained to reproduce the outputs of a more detailed system. Here, a compact U-Net—a neural network architecture that processes spatial patterns at multiple scales—takes the place of atmospheric dynamics, while a second U-Net predicts the evolving sea-ice state. The ocean remains prognostic: it evolves through time according to its own physical equations rather than being replaced by a learned approximation.

The key design choice is that FASST’s atmospheric emulator is stateless. It does not generate a sequence of atmospheric states, such as winds and temperatures that feed into the next atmospheric time step. Instead, every six simulated hours it looks at the current ocean–ice conditions and predicts the surface exchanges for the coming interval. Inputs include sea-surface temperature (SST), sea-ice fraction, sunlight, the SST and ice fraction from 24 hours earlier, and the recent rate of SST change. Outputs include heat and freshwater fluxes, wind stress, precipitation, radiation, and near-surface atmospheric conditions.

That matters for stability and cost. Many learned weather or climate systems must repeatedly feed their own predictions back into themselves; small errors can build from one step to the next. FASST’s atmosphere avoids that particular chain: each prediction is anchored to the evolving ocean and ice state. This does not make the predictions error-free, nor does it remove feedback. The ocean changes the conditions presented to the emulator, and the emulator’s fluxes in turn affect the ocean. It is a simplified feedback loop, not a one-way forcing file.

Sea ice has a different rhythm. Its emulator advances monthly, predicting ice area, thickness, motion, and stress on the ocean from the current SST, sunlight, and previous month’s ice. The model interpolates the monthly ice fields between updates so they can shape the more frequent atmospheric exchanges. The atmosphere and ice emulators were trained separately, then coupled together through their shared interaction with POP2.

The networks learned from CESM2 data: six-hourly atmospheric output for the atmosphere and monthly sea-ice output for the ice model. The atmospheric network has about 7.7 million trainable parameters. Training it took seven hours on one NVIDIA A100 graphics processor; the sea-ice model took about an hour. These are compact models by the standards of global climate emulation, though training cost is only one part of the expense of running a complete coupled simulation.

For the long test, the authors started from a cold ocean state and ran FASST without restoring SST toward a prescribed target. Restoring is a corrective procedure that nudges a model’s temperatures back toward a reference. It can help stabilize a simulation, but it can also hide whether the coupled system can maintain a climate on its own. The authors compared the run with CESM2’s pre-industrial control simulation, which represents an unforced climate used to study internally generated variability.

What They Found

FASST’s global mean SST initially warmed, reaching nearly 293 K before settling. The team defined the end of this transient as the point after which a rolling 15-year temperature trend stayed below 0.02 K per year. That occurred after year 31. Over the remaining 109 years, the global mean was 291.30 ± 0.22 K, compared with 291.25 ± 0.10 K in CESM2’s control. The close averages are encouraging; the larger spread in FASST also signals that its climate varies more around that average.

This was not a case of matching one global number while reproducing nothing else. FASST generated interannual variability in the tropical Pacific, including changes in Niño-3.4, a standard measure of sea-surface temperature anomalies across a central-eastern Pacific region associated with El Niño and La Niña. But the amplitude was weaker: the reported standard deviation was 0.62 K in FASST and 1.02 K in CESM2. A model that makes this variability too subdued could understate some important swings in ocean heat and climate patterns, even if its global average looks sound.

Niño-3.4 variability is weaker in FASST

Label	Value
FASST	0.62 K standard deviation
CESM2 piControl	1.02 K standard deviation

Reported standard deviations: 0.62 K for FASST and 1.02 K for CESM2 piControl.

The pattern of errors helps explain the gap. The model captures broad spatial features of SST, but the paper reports a colder tropical Pacific and differences in the North Atlantic, Southern Ocean, and Arctic. Those regions are not decorative details: they are places where ocean circulation, sea ice, and exchanges of heat between ocean and atmosphere shape the climate signal. The authors identify these as important targets for future evaluation, rather than claiming the global average proves the regional climate is solved.

The component emulators also have uneven strengths. On held-out atmospheric data, the model’s coefficient of determination, or R², was 0.93 for downwelling shortwave radiation and 0.98 for downwelling longwave radiation. R² summarizes how much of the reference data’s variation the predictions explain; a score nearer 1 indicates stronger agreement under this test. Precipitation was harder, at R² = 0.48. Wind stress, which pushes directly on the ocean surface, scored 0.63–0.73. The sea-ice emulator did well on ice area and thickness (R² = 0.974 for each), while ice velocity and ocean stress were less well captured, with scores from 0.53 to 0.65.

That unevenness is scientifically consequential. Heat and moisture exchanges can be well represented in aggregate while a weaker wind-stress prediction alters the ocean’s momentum or circulation. A high score on a field-by-field test also does not, by itself, guarantee that the coupled simulation reproduces the right long-term feedbacks. FASST’s 140-year integration is therefore a more demanding test than a snapshot comparison—but it is still an initial demonstration, not a full validation of every ocean and climate process.

A shorter experiment tested how the coupled system responded to an imposed heat anomaly. The researchers warmed the upper 100 metres of the ocean by 2 K and ran the model for five years. Relative to an otherwise identical control, the global SST anomaly fell from +1.90 K at year 0 to +0.98 K at year 4, a reported reduction of 48%. There was no runaway amplification over that interval. This suggests the system dissipated that particular perturbation rather than amplifying it, but a five-year test cannot establish how it handles every kind of forcing or long-term climate change.

An imposed ocean heat anomaly declined over four years

Label	Value
Year 0	1.9 K anomaly
Year 4	0.98 K anomaly

The paper reports a 48% reduction from +1.90 K at year 0 to +0.98 K at year 4; no runaway amplification was observed during the five-year test.

FASST is also fast at the atmospheric-network level. On one A100, the paper reports 10.4 milliseconds of pure network inference per six-hour step, compared with 190 milliseconds for CAMulator’s atmospheric inference—about an 18-fold difference. Yet the full coupled setup ran at 35 simulation years per day on two CPU nodes plus one A100. That is lower than the 42 years per day reported for CAMulator on the same hardware configuration. The authors suggest the POP2 ocean component is part of the bottleneck. In other words, making the atmosphere cheap does not make the whole Earth system equally cheap; the remaining physics and the data exchange between components still count.

Why This Changes Things

Climate science needs long runs for a basic reason: many relevant ocean processes unfold slowly, and a single simulation cannot disentangle a forced response from natural variability. A lower-cost atmosphere could make it more practical to run larger ensembles—sets of simulations with small differences in initial conditions—or to test many model configurations. More runs can help researchers map uncertainty rather than relying on one expensive trajectory. FASST’s significance is not that it predicts a particular future climate outcome, but that it offers a potential tool for asking more questions about ocean adjustment and feedback.

Its approach also occupies a useful middle ground. Traditional ocean-only experiments often prescribe the atmosphere from observations or reanalysis. Surface fluxes can respond to the ocean’s changing state, but the prescribed atmospheric conditions themselves do not react to ocean anomalies. Fully learned climate systems can be highly efficient, but may emulate the ocean as well as the atmosphere. FASST keeps a dynamical, physics-based ocean while making the atmosphere responsive to the evolving ocean–ice state. That is a practical architectural choice for researchers who want to study ocean behavior without paying the full cost of atmospheric dynamics.

The word “stateless” should not be mistaken for “feedback-free.” FASST’s atmosphere responds diagnostically to current ocean conditions, and its inputs include recent SST history and tendency. What it omits is a separately evolving atmospheric state. This reduces the burden of maintaining an autoregressive atmospheric rollout, but it may also limit the representation of atmospheric memory and dynamics. The paper’s weaker wind-stress and precipitation skill, along with muted Niño-3.4 variability, makes clear that compactness has trade-offs.

The speed numbers, too, need their scale attached. An 18-fold inference advantage compares the atmospheric neural-network computation, not every part of the coupled model or the cost of a climate experiment from start to finish. The end-to-end throughput indicates that the ocean and system integration matter. Even so, a smaller and faster atmosphere can make model development and controlled experiments easier, especially if later work improves the surrounding workflow.

What’s Next

The authors plan to evaluate more than SST: ocean heat content, salinity, the Atlantic Meridional Overturning Circulation (AMOC), and the spatial structure of these fields. Those tests will help determine whether the right global surface temperature conceals errors in the ocean’s stored heat or large-scale circulation. They also plan to add carbon dioxide as an input, a necessary step if FASST is to examine a broader range of climate-change scenarios rather than only an unforced control climate.

Several limits remain. The 140-year run demonstrates stability in one unforced setup, not robustness under rising greenhouse gases or other external forcing. The regional temperature biases and weaker tropical variability need attention, and the less accurate wind-stress, precipitation, ice-motion, and ice-stress predictions may affect coupled behavior in ways that simple test scores cannot fully capture. More validation against observations and independent model experiments will be needed before using the system to support specific projections.

Still, the central result is worth taking seriously: a small, diagnostic neural atmosphere coupled to a physics-based ocean maintained a broadly stable climate for more than a century without temperature nudging. That makes a new kind of experiment more plausible—not a shortcut around climate physics, but a way to spend less computation on atmospheric machinery and more on exploring the slow, consequential questions of how the ocean stores heat, moves it around, and responds to a changing world.