SamudrACE: A Coupled Climate in a Single Day ✛
Context
For decades, the climate community has built global models like modular construction kits: ocean, atmosphere, ice, and land are developed by separate teams and then coupled through flux exchanges. This approach allowed efforts to be distributed and validated components to be accumulated. Neural network emulators, which learn from the outputs of these models, have already proven their power in weather forecasting and in individual environmental components. However, coupling three-dimensional emulators of the atmosphere and ocean into a single working climate had not been achieved until now. The SamudrACE work stands at this crossroads: fast atmospheric emulators, simplified ocean models, and early attempts to link them via slab approximations are all nearby. But a full three-dimensional coupling that reproduces El Niño remained a missing piece.
Methods
The authors took two existing neural network emulators: ACE2 for the atmosphere and Samudra for the ocean. Both were trained independently on 155 years of simulation from the GFDL CM4 climate model under pre-industrial conditions. They then coupled them through a physical interface: every 6 hours, the atmosphere outputs heat, moisture, and wind stress fluxes; the ocean averages them over 5 days and returns surface temperature and sea ice concentration. After coupling, the models were fine-tuned in two stages: first only the ocean, then both components together. A key detail: the exchange occurs in physical space, not in the hidden representations of the neural networks, preserving interpretability and stability. To run on a single H100 GPU, the model uses a 1-degree resolution and 8 atmospheric levels.
Results
SamudrACE runs stably for centuries and reproduces the mean climate with errors comparable to those of the individual components. Global surface temperature, precipitation, sea ice, and ocean currents match the reference within a few percent. The main result: the model reproduces El Niño for the first time, with realistic amplitude and a three-year peak in the spectrum. However, low-frequency variability (periods longer than 4 years) is underestimated: La Niña is weaker than in the reference, and decadal Pacific oscillations are smoothed out. The speed is impressive: 1,500 model years per day on a single GPU, which is 3,750 times more energy-efficient than the original model running on 6,399 CPU cores.
Implications
This work demonstrates that the modular approach to climate modeling transfers to neural networks without losing stability. Unlike neighboring works, which simplify the ocean to a thin layer or replace it with statistics, here the ocean is three-dimensional and fully interacts with the atmosphere. This is the first demonstration that emergent phenomena like El Niño can be obtained in a fully trainable coupled system. For the community, this means that fast climate emulators are no longer a toy for weather forecasting but become a tool for studying climate dynamics on centennial timescales.
Future development
The next step is obvious: add emulators for biogeochemistry, the carbon cycle, and land to move from pre-industrial control to scenarios with rising CO2. A separate challenge is teaching the model to respond to abrupt changes in external forcing, such as a sudden doubling of greenhouse gases. Also, the problem of underestimated low-frequency variability needs to be addressed: longer training series, a different loss function, or an architecture with explicit memory for decadal processes might help.
Impact
Climatology will gain a cheap tool for large ensembles: thousands of runs that previously required supercomputer months now fit into weeks on a single GPU. This will accelerate testing hypotheses about internal climate variability and sensitivity to external forcing.
Next steps
Immediate experiments: train SamudrACE on data with historical greenhouse gas increases and check whether the model reproduces the observed warming. In parallel, improving the representation of extreme precipitation through stochastic loss functions is worth pursuing.
Key open problems
An open question for climate modeling remains: how do neural network emulators conserve physical laws of energy and mass over long timescales? SamudrACE shows that even without explicit constraints, the drift in the energy balance is small, but the mechanism behind this stability is not fully understood. Another unresolved problem is transferring the emulator to conditions far from the training distribution, such as a climate with very high CO2.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors coupled neural network emulators of the atmosphere (ACE2) and ocean (Samudra) into a single model, SamudrACE, training them on GFDL CM4 data and then fine-tuning them together. The model runs stably for centuries on a single GPU, reproduces the mean climate, and for the first time in this class of emulators exhibits El Niño with realistic amplitude and periodicity.
Where its strength is
The strong point is the full three-dimensional coupling of atmosphere and ocean with exchange in physical space, which ensures interpretability and stability. This allows reproducing emergent phenomena such as El Niño, which simplified models cannot capture, and opens the way to fast climate experiments.
Where it could go next
- It might be interesting to look at the STRATA experience, where global weather is assembled from local pieces: the idea of modular training with spatial decomposition could help improve the reproduction of regional variability, for example La Niña, which is currently underestimated.based on: Storm Mosaic: AI Model STRATA Pieces Together Global Weather arXiv:2606.31248
- There is a related work on predicting El Niño one and a half years ahead; perhaps its techniques could help improve low-frequency variability in SamudrACE if used to tune the ocean-atmosphere interaction on interannual scales.based on: El Niño Can Be Predicted a Year and a Half in Advance arXiv:2601.12251
- The archive contains an example of accelerating calculations via neural networks in molecular dynamics; similar approaches to optimizing the flux exchange interface could further increase SamudrACE's speed without losing accuracy.based on: How a Neural Network Speeded Up Molecular Collision Calculations arXiv:2602.13601
Worth trying
- Test the spatial mosaic technique from STRATA to improve regional ocean characteristics.
- Use El Niño prediction methods to correct low-frequency variability in the coupled model.
- Apply the exchange interface optimization from the molecular collisions work to speed up the model.
Why this may matter
If low-frequency variability can be improved, SamudrACE will become a full-fledged tool for studying decadal climate dynamics, which is important for long-term forecasts and climate risk assessments.
Nearby works in our archive
- Weather forecast on a single GPUsimilarity 0.50
- Neutrinos from around the world pieced into one picturesimilarity 0.50
- Storm Mosaic: AI Model STRATA Pieces Together Global Weathersimilarity 0.49
- The Climate Swings of Ancient Marssimilarity 0.48
- The Sun Conducts Ice Agessimilarity 0.48
- The Fourth Spring of Climatesimilarity 0.48
- How a Neural Network Speeded Up Molecular Collision Calculationssimilarity 0.48
- El Niño Can Be Predicted a Year and a Half in Advancesimilarity 0.47
The archive is sparse around this work: the closest review is only 0.50, against a usual 0.64. It looks like few have drilled nearby.
The model is named SamudrACE as a hybrid of Samudra (a Sanskrit word meaning ocean) and ACE (Ai2 Climate Emulator). On a single GPU, it runs climate faster than real time: for each second of computer time, 6.4 model days pass.
The idea of a fast climate simulator that lets us peek into Earth's future echoes Kim Stanley Robinson's novel "The Ministry for the Future," where climate models become tools for political decisions. SamudrACE does not predict the future yet, but it makes such experiments practically accessible.
Key numbers
- simulation speed: 1,500 model years per day
- speedup relative to CM4: 3,750 times
- training dataset: 155 years of simulation
- atmospheric resolution: 1 degree, 6 hours
- ocean resolution: 1 degree, 5 days
- number of parameters: about 600 million