In this paper
AbstractIntroductionMethodologyResultsConclusionImpact statementReferencesModel and training detailsData and code availabilityTrainingNetwork architectureEvaluationWorksPhysics and Astronomy
FASST: A Compact Stateless Atmosphere Model for Long-Term Ocean Coupling
Abstract
Coupled climate simulations are computationally expensive, largely because of the atmosphere. We introduce FASST, a coupled climate model that retains the prognostic POP2 ocean while replacing the atmosphere and sea ice with lightweight U-Net emulators. The atmospheric emulator is diagnostic, predicting surface fluxes from the evolving ocean-ice state every simulated hours, while the sea-ice emulator advances the ice state monthly. For years of unforced integration, FASST maintains a stable global-mean climate without SST restoring, with post-transient global-mean SST of compared with in CESM piControl. FASST also captures interannual Niño- variability, although with weaker amplitude than CESM ( versus ). These results demonstrate that a stateless learned atmosphere can sustain stable, free-running coupled climate simulations over century-scale integrations.
Introduction
Understanding how the ocean absorbs heat and adjusts to climate change requires coupled ocean–atmosphere models that can be integrated for centuries. However, the atmosphere is a major computational bottleneck. In the CAMulator benchmark manuscript, the Community Atmosphere Model version 6 (a full physics fortran based simulation) achieves only 14 simulated years per day on ten CPU nodes, motivating cheaper atmospheric surrogates for long-duration coupled simulations [1]. The challenge is to reduce atmospheric cost without sacrificing the ocean–atmosphere feedbacks needed to represent climate variability.
A longstanding alternative is to prescribe atmospheric forcing rather than simulate atmospheric dynamics. CORE-I/II and the CMIP6 Ocean Model Intercomparison Project (OMIP), prescribe atmospheric fields from observations or reanalyses, while surface fluxes are computed from bulk formulae using the prescribed atmosphere and evolving ocean–ice state [2, 3]. JRA55-do provides the interannually varying atmospheric forcing for OMIP, with fluxes calculated using the Large–Yeager formulation [4, 5]. In CESM, the Data Atmosphere (DATM) supplies prescribed forcing to components such as POP2 without a responsive atmosphere [6, 7]. Thus, surface fluxes respond to the evolving ocean state, but the prescribed atmosphere does not respond to ocean anomalies, limiting atmospheric feedbacks [2]. Intermediate-complexity models retain atmospheric feedbacks through simplified atmospheric components. Zebiak–Cane coupled a simple atmosphere to a reduced-gravity ocean, while Neelin coupled a simplified tropical-Pacific ocean GCM to a steady-state atmosphere [8, 9]. Later, statistical hybrid models used SVD to relate SST anomalies to wind-stress anomalies,
Tackling Climate Change with Machine Learning: workshop at NeurIPS 2026. followed by U-Net models that learned this mapping directly [10, 11]. These approaches retain key ocean–atmosphere feedbacks at lower cost, but remain largely regional or intermediate-complexity.
Recent global machine-learning emulators offer much larger computational gains. CAMulator achieves a reported 350× speedup over CAM6, while ACE2 remains stable in a 1000-year integration [1, 12]. However, these models are large (450–750M parameters) and require autoregressive atmospheric-state rollouts. They also lack interactive sea ice. When coupled to prognostic oceans, additional stability challenges have been reported: ACE2-SOM develops unphysical transients under non-equilibrium CO forcing, while ACE2-NEMO exhibits sea-ice-volume and sea-surface-height drift that motivated replacing learned ice fluxes with CORE bulk formulae [13, 14]. SamudrACE similarly remains computationally intensive and exhibits weaker ENSO variability [15]. Conversely, fully learned coupled systems such as Ola and DLESyM, as well as learned ocean components such as Samudra, demonstrate impressive long-term speed and stability, but replace the physics-based ocean with a learned representation as well [16, 17, 18].
A central challenge is to replace atmospheric dynamics while retaining surface forcing that responds to the evolving ocean–ice state. FASST addresses this by coupling a learned atmosphere to the physics-based POP2 ocean [19, 6] and a lightweight autoregressive sea-ice emulator. Its three main contributions are: (1) a stateless atmospheric formulation that predicts global surface fluxes directly from the current ocean–ice state without propagating an atmospheric state; (2) a compact and efficient implementation with 7.7M parameters, requiring 10.4 ms of pure network inference per 6-hour step on a single NVIDIA A100 versus 190 ms for CAMulator’s atmospheric inference, yielding an approximately 18× speedup; and (3) long-term coupled stability, demonstrated by a 140-year free-running integration without SST restoring, with global-mean SST equilibrating within 31 years and remaining close to the CESM2 pre-industrial control climatology.
Methodology
FASST couples a diagnostic atmospheric emulator and an autoregressive sea-ice emulator to the physics-based POP2 ocean model at resolution. The two emulators are trained independently with no direct coupling between them and are integrated together only during the coupled simulation through their interaction with the ocean and sea-ice states. Both emulators use the U-Net architecture [20], originally developed for biomedical image segmentation, which we adapt here for global climate fields. We replace the original padding with circular padding in longitude and replicated padding in latitude to preserve periodicity and spatial resolution. Strided convolutions perform downsampling, while bilinear interpolation followed by convolution performs upsampling. Each stage uses convolution, BatchNorm, and nonlinear activation, with a residual connection from the initial projection to the final layer. A final convolution produces the continuous physical fields required for climate model coupling.
Atmospheric emulator: The emulator maps six inputs: sea surface temperature (SST), sea-ice fraction, solar insolation, 24-hour-lagged SST and sea-ice fraction, and SST tendency, defined as , to 13 atmospheric fields spanning air-sea fluxes and near-surface atmospheric variables, including wind stress, sensible and latent heat fluxes, evaporative freshwater flux, radiative fluxes, precipitation, near-surface winds, air temperature, humidity, and surface pressure. The lagged SST and sea-ice fraction are provided as input features and are not predicted by the emulator, so the atmospheric emulator is not autoregressive. The emulator is trained on 6-hourly output from the Community Earth System Model version 2.1.5 (CESM2.1.5), generated using the CREDiT pipeline [21]. The data are remapped from the CAM grid (192 × 288) to the POP2 gx1v7 grid (384 × 320). We use the same chronological 34/1/1-year train/validation/test split as CAMulator [1]. Training for 80 epochs takes 7 hours on an NVIDIA A100-SXM4-80GB. We report the coefficient of determination (see Appendix B), which measures the fraction of target variance explained relative to the pooled test-set mean. is computed by pooling all held-out grid cells and 6-hourly test timesteps for each variable. Most fields achieve , with wind stress at –0.73 and precipitation at .
Sea-ice emulator: autoregressively predicts monthly ice area fraction, thickness, velocity, and ice–ocean stresses from SST, solar insolation, and the previous month’s ice state. It is trained on monthly-mean CESM2-LE2-1231.002 output [22], as these prognostic ice fields are unavailable in the CESM2 CREDiT data. We use a masked MSE loss and AdamW, with an 80/10/10 chronological train/validation/test split. Training for 80 epochs with batch size 8 takes about one hour on an NVIDIA A100-SXM4-40GB. The emulator achieves for ice area fraction and thickness and – for velocity and ice–ocean stress. is evaluated on the monthly-mean fields. Training details and variable-wise scores for both emulators are provided in Appendices A and B.
Coupling: POP2 provides the prognostic ocean dynamics, while the emulators represent the atmospheric and sea-ice processes needed to close surface exchanges. POP2 runs in ocean-only configuration, with the standard CPL6 exchange (the CESM coupler framework) replaced by direct coupling to the emulators. Every 6 simulated hours, POP2 supplies SST to the atmospheric emulator, which predicts surface fluxes for the next interval. The sea-ice emulator predicts monthly ice fields from SST and the previous ice state; these are linearly interpolated to modulate the atmospheric fluxes between updates, allowing ocean SST to feed back through the atmosphere and sea ice onto POP2.
Results
We evaluated FASST using an unforced cold-start simulation of 140 years and compared its climate variability and mean state with CESM2 piControl as a reference for internally generated variability [6]. As shown in Figure 1a, FASST undergoes an initial warm adjustment, reaching nearly 293 K before settling into the piControl range. We define the end of this transient as the first year after which the trailing 15-year rolling trend in global-mean SST remains below 0.02 K yr, which occurs after 31 years. Following this adjustment, global-mean SST is 291.30 0.22 K, close to 291.25 0.10 K in piControl, indicating a stable global-mean climate without SST restoring. The tropical Pacific exhibits weaker Niño-3.4 variability (Figure 1b), with a standard deviation of 0.62 K versus 1.02 K in piControl, consistent with weaker tropical variability reported for other coupled ML atmosphere emulators, including ACE2-NEMO and CAMulator [1, 14].
Figure 1. Long-term behavior of FASST (orange, solid) and CESM2 piControl (grey, dashed) over 140 years. The shaded region denotes the initial spin-up, and the vertical dashed line marks its end at year 31. (a) Global-mean SST. Following spin-up, FASST has a mean of 291.30 0.22 K compared with 291.25 0.10 K for piControl. (b) Niño-3.4 SST anomaly. FASST captures interannual variability but with lower amplitude, with standard deviations of 0.62 and 1.02 K, respectively.
Model Speed Comparison: Across three runs, FASST achieves 160.9 SYPD (simulation years per day) for atmosphere-only inference on a single A100. The pure network inference time is 10.4 ms per 6-hour step, compared with 190 ms for CAMulator’s atmospheric inference [1, 23], corresponding to an 18× speedup for the atmospheric component. In the full coupled configuration, using 2 CPU nodes and 1 A100, FASST achieves 35 SYPD, compared with 42 SYPD reported for CAMulator on the same hardware configuration [23]. We hypothesize that the lower end-to-end throughput is partly attributable to POP2 handling, further system-level optimization may improve overall throughput.
Temperature Perturbation: To test whether FASST damps an imposed ocean heat anomaly, we perturbed the initial POP2 ocean state by +2 K over the upper 100 m and integrated for five model years with all other settings unchanged. The anomaly decayed from +1.90 K at year 0 to +0.98 K at year 4, corresponding to a 48% reduction with no runaway amplification observed over the five-year integration (Figure 2). This test probes FASST’s coupled response on a shorter timescale than the multi-decadal adjustment from a cold start.
Figure 2. Annual-mean global SST anomaly following a +2 K perturbation to the top 100 m of the initial ocean state, relative to an identical unperturbed control. The anomaly decays from +1.90 K at model year 0 to +0.98 K at model year 4, a 48% reduction, with no runaway amplification.
Conclusion
FASST demonstrates a stateless approach to learned atmospheric coupling, in which surface fluxes are predicted directly from the current ocean–ice state and prescribed insolation. Coupled to POP2 and a learned sea-ice component, FASST has been demonstrated to run stably for 140 years without SST restoring, maintains a global-mean SST close to CESM2 piControl, and reproduces interannual Niño-3.4 variability, although with weaker amplitude. Regional SST biases remain, most notably a colder tropical Pacific (Figure 7). A perturbation experiment further shows that an imposed ocean heat anomaly decays without runaway amplification. Together, these results show that a stateless learned atmosphere can support stable, long-duration coupled climate simulations. Future work will evaluate FASST beyond SST, including ocean heat content, salinity, the Atlantic Meridional Overturning Circulation (AMOC), and the spatial structure of these fields. Moreover, we plan to include CO as an input to the model to enable more diverse climate change experiments.
Figure 7. FASST versus CESM2 piControl sea-surface temperature climatology. Columns show CESM2 piControl (truth), FASST (predicted), and the corresponding error (FASST minus truth). Rows show the global domain, tropical Pacific, North Atlantic, Southern Ocean, and Arctic. Climatologies are computed over the post-transient period (model years 31–140) using elapsed-year-matched intervals.
Impact statement
FASST addresses a key computational bottleneck in long-duration climate experiments. Its compact 7.7M-parameter, non-autoregressive atmosphere trains in seven hours on a single A100, reducing the resources needed for model development. This makes longer simulations, larger ensembles, and uncertainty studies of ocean heat uptake, variability, and sea-ice processes more accessible. Because the atmosphere is cheap to fine-tune or retrain, researchers can run controlled experiments by modifying it and coupling it to the ocean. Hence, we hope this simple, stateless, yet realistic atmosphere model will be a useful tool for probing coupled atmosphere–ocean interactions.
References
Model and training details
Data and code availability
The atmospheric emulator is trained on 6-hourly output from Community Earth System Model version 2.1.5, produced for the CREDIT pipeline (https://www.mmm.ucar.edu/opensky-publications/articles:43854/community-research-earth-digital-intelligence-twin-a-scalable-framework-for-ai-driven-earth-system-modeling) [21]. The sea-ice emulator is instead trained on monthly output from a CESM2 Large Ensemble member (1231.002) (https://www.cesm.ucar.edu/community-projects/lens2) [22], since the prognostic sea-ice fields it requires are not available in the CESM2 archive used for the atmospheric emulator. As an unforced baseline, we use the CESM2 piControl simulation (https://data.ucar.edu/dataset/cmip6-cesm2-pi-control-experiment-with-cam6-interactive-land-clm5-coupled-ocean-pop2-with-biogeo) for comparison against the FASST integration [6]. The code used to develop and evaluate FASST is available at https://github.com/pratham-aggr/couple_modeling.git
Training
Both emulators are trained using a masked mean-squared error (MSE) over ocean grid points,
where and are the reference and predicted values, respectively, and denotes the set of points in the ocean grid. All input and target variables are standardized using the training-set mean and standard deviation and transformed back to physical units for evaluation.
| Hyperparameter | Atmospheric | Sea-ice |
| Loss | MSE | MSE |
| Optimizer | AdamW | AdamW |
| Learning rate | ||
| Weight decay | ||
| LR schedule | Cosine | Cosine |
| Epochs | 80 | 80 |
| Batch size | 8 | 8 |
| Base width | 64 | 64 |
| Dropout | 0.1 | 0.1 |
| Training time | 7 h | 1 h |
| Hardware | A100-SXM4-80GB | A100-SXM4-40GB |
Table 1. Training hyperparameters and computational cost for the atmospheric and sea-ice emulators.
Network architecture
Both emulators use a five-level U-Net but differ in their inputs, outputs, and coupling: the atmospheric emulator is diagnostic, whereas the sea-ice emulator is autoregressive. Input and output variables are listed in Table 2. The atmospheric emulator has 7,726,861 trainable parameters, a base width of 64, and 12 convolutional and 10 batch-normalization layers.
| Role | Variable | Description | Units |
| Atmosphere emulator (6-hourly, diagnostic) | |||
| Input | SST | Sea-surface temperature (from POP2) | K |
| Input | ICEFRAC | Sea-ice fraction (from sea-ice emulator) | 0–1 |
| Input | SOLIN | Incoming solar radiation, top of atmosphere | W m |
| Input | SST_prev | Sea-surface temperature, 24 h lag | K |
| Input | ICEFRAC_prev | Sea-ice fraction, 24 h lag | 0–1 |
| Input | dSST_dt | SST tendency | K s |
| Target | TAUX | Zonal wind stress | N m |
| Target | TAUY | Meridional wind stress | N m |
| Target | SHFLX | Sensible heat flux | J m |
| Target | LHFLX | Latent heat flux | J m |
| Target | QFLX | Surface freshwater flux | kg m s |
| Target | FSDS_J | Downwelling shortwave radiation | J m |
| Target | FLDS_J | Downwelling longwave radiation | J m |
| Target | PRECT | Total precipitation | m |
| Target | Ubot | Near-surface zonal wind | m s |
| Target | Vbot | Near-surface meridional wind | m s |
| Target | Tbot | Near-surface air temperature | K |
| Target | Qbot | Near-surface specific humidity | kg kg |
| Target | PS | Surface pressure | Pa |
| Sea-ice emulator (monthly, autoregressive) | |||
| Input | SST | Sea-surface temperature (from POP2) | K |
| Input | SOLIN | Top-of-atmosphere insolation (analytic, date+lat) | W m |
| Input | aice_prev | Sea-ice area fraction, previous month | 0–1 |
| Input | hi_prev | Sea-ice thickness, previous month | m |
| Input | uvel_prev | Sea-ice zonal velocity, previous month | m s |
| Input | vvel_prev | Sea-ice meridional velocity, previous month | m s |
| Target | aice | Sea-ice area fraction | 0–1 |
| Target | hi | Sea-ice thickness | m |
| Target | uvel | Sea-ice zonal velocity | m s |
| Target | vvel | Sea-ice meridional velocity | m s |
| Target | strocnx | Zonal ice-to-ocean stress | N m |
| Target | strocny | Meridional ice-to-ocean stress | N m |
Table 2. Input and target variables for the two emulators. The atmosphere emulator is diagnostic and operates every 6 hours; the sea-ice emulator is autoregressive and operates monthly.
The network begins with a convolution that projects the six input channels to 64 feature channels. The encoder comprises five stride-2 convolutional blocks with channel widths of 512, 256, 128, 128, and 64. The first four blocks use kernels, while the deepest uses a kernel. Each block consists of a convolution, batch normalization, LeakyReLU activation with slope 0.2, and dropout with rate 0.1.
The decoder mirrors the encoder with channel widths of 512, 256, 128, 128, and 64. Each block bilinearly upsamples the feature map by a factor of two followed by a stride-1 convolution. Decoder convolutions use kernels, except at the deepest level, where a kernel is used. Each is followed by batch normalization, ReLU activation, and dropout with rate 0.1. Skip connections concatenate encoder and decoder features at matching resolutions.
The final 64-channel representation is concatenated with the features from the initial projection layer and passed through a convolutional head to produce the 13 output fields. All spatial convolutions use circular padding in longitude and replication padding in latitude, preserving the periodic zonal boundary of the tripole grid without an artificial seam. Convolutional weights are initialized from , batch-normalization scales from , and batch-normalization biases to zero.
Evaluation
We evaluate emulator performance on held-out test data using the coefficient of determination () and root mean squared error (RMSE), computed in physical (de-normalized) units. measures how well the emulator reproduces spatial and temporal variability, while RMSE measures the magnitude of prediction errors in the variable’s native units. The two metrics are complementary: is scale-independent, whereas RMSE reflects the absolute physical error.
The metrics are defined as
where and are the reference and predicted values, is the test-set reference mean, and is the number of evaluated points. For each variable, all unmasked valid grid cells and timesteps are pooled, with no spatial area weighting.
RMSE is reported in each variable’s native units. SHFLX and LHFLX are evaluated as energy accumulated over the 6-hour coupling interval, consistent with the FSDS_J and FLDS_J targets. For visualization, the accumulated energy fluxes are converted to mean fluxes in W m by dividing by the 6-hour coupling interval. Table 3 reports skill for both emulators.
| Atmospheric emulator | ||
| Variable | RMSE | |
| PS | 0.9996 | |
| Tbot | 0.9994 | 1.8736 |
| FLDS_J | 0.9840 | |
| Qbot | 0.9678 | |
| QFLX | 0.9243 | |
| FSDS_J | 0.9316 | |
| LHFLX | 0.9186 | |
| SHFLX | 0.8655 | |
| Ubot | 0.8042 | 1.8729 |
| Vbot | 0.7179 | 1.8921 |
| TAUX | 0.7338 | |
| TAUY | 0.6275 | |
| PRECT | 0.4751 | |
| Sea-ice emulator | ||
| Variable | RMSE | |
| aice | 0.974 | 0.0341 |
| hi | 0.974 | 0.0424 |
| uvel | 0.653 | 0.0127 |
| strocnx | 0.576 | 0.0075 |
| strocny | 0.570 | 0.00562 |
| vvel | 0.531 | 0.00966 |
Table 3. Test-set performance of the atmospheric emulator against CREDIT data and the sea-ice emulator against CESM2 Large Ensemble member 1231.002.
Figure 7 compares the post-transient SST climatology of FASST with CESM2 piControl over matched elapsed years (31–140). FASST captures the broad large-scale spatial structure of SST, while regional biases remain, most notably a colder tropical Pacific, as well as differences in the North Atlantic, Southern Ocean, and Arctic.
Figure 3. Test-period spatial maps of the atmospheric emulator’s boundary-layer state outputs (Ubot, Vbot, Tbot, Qbot, PS; –1.00): reference (left), FASST prediction (middle), and error (right), averaged over the held-out test period.
Figure 4. Test-period spatial maps of the atmospheric emulator’s turbulent and moisture flux outputs (SHFLX, LHFLX, QFLX; –): reference (left), FASST prediction (middle), and error (right), averaged over the held-out test period.
Figure 5. Test-period spatial maps of the atmospheric emulator’s radiation and precipitation outputs (FSDS, FLDS, PRECT): reference (left), FASST prediction (middle), and error (right), averaged over the held-out test period. Radiative fluxes are well reproduced (); PRECT is the weakest-performing variable in the atmospheric emulator ().
Figure 6. Test-period spatial maps of the atmospheric emulator’s wind-stress outputs (TAUX, TAUY; –): reference (left), FASST prediction (middle), and error (right), averaged over the held-out test period. These are the weakest-performing forced fields, of interest given their direct role in ocean momentum forcing.
References
- [1]William E Chapman, John S Schreck, Yingkai Sha, David John Gagne II, Dhamma Kimpara, Laure Zanna, Kirsten J Mayer, and Judith Berner. Calulator: Fast emulation of the community atmosphere model. arXiv preprint arXiv:2504.06007, 2025.arxiv.org/abs/2504.06007
- [2]Stephen M Griffies, Arne Biastoch, Claus Böning, Frank Bryan, Gokhan Danabasoglu, Eric P Chassignet, Matthew H England, Rüdiger Gerdes, Helmuth Haak, Robert W Hallberg, et al. Coordinated ocean-ice reference experiments (cores). Ocean modelling, 26(1-2):1–46, 2009.DOI
- [3]Stephen M Griffies, Gokhan Danabasoglu, Paul J Durack, Alistair J Adcroft, V Balaji, Claus W Böning, Eric P Chassignet, Enrique Curchitser, Julie Deshayes, Helge Drange, et al. Omip contribution to cmip6: Experimental and diagnostic protocol for the physical component of the ocean model intercomparison project. Geoscientific Model Development, 9:3231–3296, 2016.DOI
- [4]Hiroyuki Tsujino, Shogo Urakawa, Hideyuki Nakano, R Justin Small, Who M Kim, Stephen G Yeager, Gokhan Danabasoglu, Tatsuo Suzuki, Jonathan L Bamber, Mats Bentsen, et al. Jra-55-based surface dataset for driving ocean–sea-ice models (jra55-do). Ocean Modelling, 130:79–139, 2018.DOI
- [5]WoG Large and SG Yeager. The global climatology of an interannually varying air–sea flux data set. Climate dynamics, 33(2):341–364, 2009.
- [6]Gokhan Danabasoglu, J-F Lamarque, J Bacmeister, DA Bailey, AK DuVivier, Jim Edwards, LK Emmons, John Fasullo, R Garcia, Andrew Gettelman, et al. The community earth system model version 2 (cesm2). Journal of Advances in Modeling Earth Systems, 12(2):e2019MS001916, 2020.
- [7]ESCOMP. CDEPS: Community data models for earth prediction systems. https://github .com/ESCOMP/CDEPS, 2024. Software.
- [8]Stephen E Zebiak and Mark A Cane. A model for el niño–southern oscillation. Monthly Weather Review, 115(10):2262–2278, 1987.
- [9]J David Neelin. A hybrid coupled general circulation model for el nino studies. Journal of the Atmospheric Sciences, 47(5):674–693, 1990.doi.org/10.1175/1520-0469(1990)047<0674:ahcgcm>2.0.co;2
- [10]Rong-Hua Zhang. A hybrid coupled model for the pacific ocean-atmosphere system. part i: Description and basic performance. Advances in Atmospheric Sciences, 32(3):301–318, 2015.
- [11]Shuangying Du and Rong-Hua Zhang. U-net models for representing wind stress anomalies over the tropical pacific and their integrations with an intermediate coupled model for enso studies. Advances in Atmospheric Sciences, 41(7):1403–1416, 2024.DOI
- [12]Oliver Watt-Meyer, Brian Henn, Jeremy McGibbon, Spencer K Clark, Anna Kwa, W Andre Perkins, Elynn Wu, Lucas Harris, and Christopher S Bretherton. Ace2: accurately learning sub-seasonal to decadal atmospheric variability and forced responses. npj Climate and Atmospheric Science, 8(1):205, 2025.
- [13]Spencer K Clark, Oliver Watt-Meyer, Anna Kwa, Jeremy McGibbon, Brian Henn, W Andre Perkins, Elynn Wu, Lucas M Harris, and Christopher S Bretherton. Ace2-som: Coupling an ml atmospheric emulator to a slab ocean and learning the sensitivity of climate to changed co2. Journal of Geophysical Research: Machine Learning and Computation, 2(4):e2024JH000575, 2025.arxiv.org/abs/2412.04418
- [14]Bobby Antonio, Kristian Strommen, Pablo Ortega, and Hannah M Christensen. Ace2-nemo: Coupling an ml atmospheric emulator to a full-depth dynamical ocean model. arXiv preprint arXiv:2603.28704, 2026.arxiv.org/abs/2603.28704
- [15]James PC Duncan, Elynn Wu, Surya Dheeshjith, Adam Subel, Troy Arcomano, Spencer K Clark, Brian Henn, Anna Kwa, Jeremy McGibbon, W Andre Perkins, et al. Samudrace: Fast and accurate coupled climate modeling with 3d ocean and atmosphere emulators. Geophysical Research Letters, 53(11):e2025GL119340, 2026.arxiv.org/abs/2509.12490
- [16]Chenggong Wang, Michael S Pritchard, Noah Brenowitz, Yair Cohen, Boris Bonev, Thorsten Kurth, Dale Durran, and Jaideep Pathak. Coupled ocean-atmosphere dynamics in a machine learning earth system model. arXiv preprint arXiv:2406.08632, 2024.arxiv.org/abs/2406.08632
- [17]Nathaniel Cresswell-Clay, Bowen Liu, Dale R Durran, Zihui Liu, Zachary I Espinosa, Raul A Moreno, and Matthias Karlbauer. A deep learning earth system model for efficient simulation of the observed climate. AGU Advances, 6(4):e2025AV001706, 2025.
- [18]Surya Dheeshjith, Adam Subel, Alistair Adcroft, Julius Busecke, Carlos Fernandez-Granda, Shubham Gupta, and Laure Zanna. Samudra: An ai global ocean emulator for climate. Geophysical Research Letters, 52(10):e2024GL114318, 2025.DOI
- [19]R Smith, P Jones, B Briegleb, Frank Bryan, Gokhan Danabasoglu, J Dennis, J Dukowicz, C Eden, B Fox-Kemper, P Gent, et al. The parallel ocean program (pop) reference manual: Ocean component of the community climate system model (cesm). LAUR-10-01853, 141:1–141, 2010.
- [20]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015.arxiv.org/abs/1505.04597
- [21]John S Schreck, Yingkai Sha, William Chapman, Dhamma Kimpara, Judith Berner, Seth McGinnis, Arnold Kazadi, Negin Sobhani, Ben Kirk, Charlie Becker, et al. Community research earth digital intelligence twin: a scalable framework for ai-driven earth system modeling. npj Climate and Atmospheric Science, 8(1):239, 2025.DOI
- [22]Keith B Rodgers, Sun-Seon Lee, Nan Rosenbloom, Axel Timmermann, Gokhan Danabasoglu, Clara Deser, Jim Edwards, Ji-Eun Kim, Isla R Simpson, Karl Stein, et al. Ubiquity of human-induced changes in climate variability. Earth System Dynamics, 12(4):1393–1411, 2021.DOI
- [23]William E. Chapman. Towards coupled data-driven Earth system prediction. Invited talk, AI for Good (ITU), Discovery – AI and Climate Science, March 2026. Presented 30 March 2026.






