Source-linked AI summary
FASST: A Compact Stateless Atmosphere Model for Long-Term Ocean Coupling
Pratham Aggarwal, Duncan Watson-Parris, William E. Chapman
TL;DR
Century-scale ocean adjustment calls for coupled simulations, but atmospheric computation is a major bottleneck. FASST retains the prognostic POP2 ocean while using lightweight, stateless atmospheric and autoregressive sea-ice emulators. Its 140-year unforced run maintains global-mean SST close to CESM2 piControl without restoring, though Niño-3.4 variability is weaker.
Problem
Studying ocean heat uptake and climate adjustment over centuries requires coupled ocean–atmosphere simulations, while atmospheric computation is a major bottleneck.
Source: §1 · p. 1Method
FASST couples the prognostic POP2 ocean to a stateless U-Net atmosphere that predicts fluxes from evolving ocean–ice conditions and an autoregressive sea-ice emulator.
Source: §1 · p. 2, §2 · pp. 2–3Results
291.30 ± 0.22 K versus 291.25 ± 0.10 K: post-transient global-mean SST in FASST and CESM2 piControl, respectively, after a 140-year unforced run without SST restoring.
Source: §3 · p. 3Takeaways & Limitations
The results show that a stateless learned atmosphere can support stable, long-duration coupled climate simulations with a physics-based ocean.
Source: §4 · p. 4
Abstract
from arXiv · showhide
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 $6$ simulated hours, while the sea-ice emulator advances the ice state monthly. For $140$ years of unforced integration, FASST maintains a stable global-mean climate without SST restoring, with post-transient global-mean SST of $291.30 \pm 0.22\mathrm{K}$ compared with $291.25 \pm 0.10\mathrm{K}$ in CESM$2$ piControl. FASST also captures interannual Ni\~no-$3.4$ variability, although with weaker amplitude than CESM ($0.62\mathrm{K}$ versus $1.02\mathrm{K}$). These results demonstrate that a stateless learned atmosphere can sustain stable, free-running coupled climate simulations over century-scale integrations.
1 Introduction
Century-scale ocean adjustment requires coupled simulations, but atmospheric computation is a bottleneck. FASST replaces atmospheric dynamics with a compact, stateless emulator while retaining ocean-responsive surface forcing and the physics-based POP2 ocean.
1 Introduction
- CAM6 achieves only 14 simulated years per day on ten CPU nodes, motivating cheaper atmospheric surrogates for integrations spanning centuries.The challenge is reducing atmospheric cost while retaining ocean–atmosphere feedbacks needed to represent climate variability.Source: §1 · p. 1
- Prescribed-atmosphere approaches let fluxes respond to evolving ocean conditions but do not let the atmosphere respond to ocean anomalies; intermediate-complexity hybrids retain feedbacks but remain largely regional or intermediate-complexity.Source: §1 · pp. 1–2
- Recent global ML emulators offer large speed gains or long integrations, but may be large, require autoregressive atmospheric rollouts, lack interactive sea ice, or show coupled stability challenges.Fully learned coupled systems demonstrate long-term speed and stability but also replace the physics-based ocean with a learned representation.Source: §1 · p. 2
- FASST couples a learned atmosphere to the physics-based POP2 ocean and a lightweight autoregressive sea-ice emulator, targeting long-duration simulation without sacrificing ocean-responsive forcing.The atmospheric emulator predicts global surface fluxes directly from the current ocean–ice state without propagating an atmospheric state.Source: §1 · pp. 1–2
2 Methodology
FASST combines independently trained U-Net atmosphere and sea-ice emulators with the prognostic POP2 ocean. The atmosphere diagnoses fluxes from ocean–ice conditions, while the ice emulator advances monthly and modulates fluxes between updates.
2 Methodology
- FASST couples a diagnostic atmospheric emulator and autoregressive sea-ice emulator to the physics-based POP2 ocean at 1° resolution; the emulators are trained independently and interact through ocean and sea-ice states during simulation.Both use U-Net architectures adapted for global climate fields, including longitude-circular and latitude-replicated padding.Source: §2 · p. 2
- The atmospheric emulator maps six ocean–ice and insolation inputs, including 24-hour-lagged SST and sea-ice fraction and SST tendency, to 13 atmospheric fields without autoregressively predicting its lagged inputs.It is trained on CESM2.1.5 6-hourly output remapped to the POP2 grid using a chronological 34/1/1-year train/validation/test split.Source: §2 · p. 2
- Most atmospheric fields achieve R2 > 0.9, while wind stress reaches R2 = 0.63–0.73 and precipitation R2 = 0.48 on pooled held-out grid cells and timesteps.Source: §2 · p. 2
- The sea-ice emulator autoregressively predicts monthly ice area fraction, thickness, velocity, and ice–ocean stresses from SST, insolation, and the previous month’s ice state.It is trained on monthly CESM2-LE2 output because the required prognostic ice fields are unavailable in the atmospheric training archive; R2 is 0.97 for area fraction and thickness and 0.53–0.65 for velocity and stress.Source: §2 · pp. 2–3
- After year 31 spin-up, FASST’s global-mean SST is 291.30 ± 0.22 K versus 291.25 ± 0.10 K in piControl, while Niño-3.4 variability is weaker at 0.62 versus 1.02 K standard deviation.Source: Figure 1 · p. 3
- Every 6 simulated hours, POP2 supplies SST for atmospheric flux prediction; monthly predicted ice fields are linearly interpolated to modulate fluxes between updates and complete the coupled feedback.Source: §2 · p. 3
3 Results
In a 140-year unforced cold-start run, FASST settles to a global-mean SST close to CESM2 piControl without restoring, while Niño-3.4 variability is weaker. Atmospheric inference is faster than CAMulator’s, and an imposed ocean heat anomaly decays over five years.
3 Results
- 291.30 ± 0.22 K versus 291.25 ± 0.10 K: post-transient global-mean SST in FASST and CESM2 piControl, respectively, after a 140-year unforced run without SST restoring.The transient ends after 31 years; Niño-3.4 standard deviation is 0.62 K in FASST versus 1.02 K in piControl.Source: §3 · p. 3
- +1.90 K at model year 0 to +0.98 K at year 4: the annual-mean global SST anomaly after a +2 K upper-100-m ocean perturbation, a 48% reduction without runaway amplification.The perturbation run lasted five model years and was compared with an identical unperturbed control.Source: Figure 2 · p. 4, §3 · p. 4
- 10.4 ms per 6-hour step versus 190 ms: FASST and CAMulator atmospheric inference times, respectively, corresponding to an 18× atmospheric-component speedup.In the full coupled configuration, FASST reaches 35 SYPD versus 42 SYPD reported for CAMulator on the same hardware; the authors hypothesize POP2 handling partly accounts for the lower throughput.Source: §3 · pp. 3–4
4 Conclusion
FASST demonstrates that a stateless learned atmosphere can support stable, long-duration coupled climate simulations. The conclusion identifies regional SST bias and broader ocean-state and CO2 evaluations as remaining scope.
4 Conclusion
- FASST’s stateless learned atmosphere, coupled with POP2 and learned sea ice, supports a stable 140-year run without SST restoring, with global-mean SST close to CESM2 piControl and weaker-amplitude interannual Niño-3.4 variability.Source: §4 · p. 4
- Regional SST biases remain, most notably a colder tropical Pacific.Source: §4 · p. 4
- Future evaluation will extend beyond SST to ocean heat content, salinity, AMOC, and spatial field structure; adding CO2 as an input is planned for more diverse climate-change experiments.Source: §4 · p. 4
5 Impact statement
FASST’s compact atmosphere is intended to reduce the computational burden of long-duration climate experiments and make broader ocean–climate studies more accessible.
5 Impact statement
- FASST’s 7.7M-parameter atmosphere trains in seven hours on a single A100, and its low fine-tuning or retraining cost supports controlled experiments and longer simulations, larger ensembles, and uncertainty studies.The stated areas of study include ocean heat uptake, variability, and sea-ice processes.Source: §5 · p. 4
A.2 Training
Both emulators are trained with masked mean-squared error over ocean grid points, with variables standardized using training-set statistics.
A.2 Training
- Both emulators use masked mean-squared error over ocean grid points, and input and target variables are standardized using training-set means and standard deviations before evaluation in physical units.Table 1 lists training hyperparameters and computational cost for both emulators.Source: §A.2 · p. 6, Table 1 · p. 7
A.3 Network architecture
FASST uses five-level U-Nets for both emulators, with distinct input-output mappings and update cadences for atmosphere and sea ice.
A.3 Network architecture
- The diagnostic atmospheric emulator has 7,726,861 trainable parameters and produces 13 output fields; it uses a base width of 64, with 12 convolutional and 10 batch-normalization layers.Source: §A.3 · p. 7
- The atmospheric U-Net uses an encoder-decoder with skip connections, bilinear upsampling, and padding that preserves the tripole grid’s periodic zonal boundary without an artificial seam.The decoder concatenates features from matching encoder resolutions; convolutions use circular longitude padding and replicated latitude padding.Source: §A.3 · p. 7
- Both emulators use five-level U-Nets, but the atmosphere is diagnostic and operates every six hours while sea ice is autoregressive and operates monthly.The atmospheric emulator predicts fields from current inputs; the sea-ice emulator uses prior-month ice variables to predict ice fields.Source: §A.3 · pp. 7–8, Table 2 · p. 8
B Evaluation
Evaluation uses held-out data and complementary skill and error metrics; FASST reproduces broad SST structure while retaining regional biases.
B Evaluation
- R2 and RMSE are computed on held-out test data in physical units; R2 measures reproduced spatial and temporal variability, while RMSE measures absolute error in native units.For each variable, valid unmasked grid cells and timesteps are pooled without spatial area weighting; SHFLX and LHFLX RMSE uses energy accumulated over the six-hour coupling interval.Source: B Evaluation · p. 8
- Table 3 reports test-set performance for the atmospheric emulator against CREDIT data and the sea-ice emulator against CESM2 Large Ensemble member 1231.002.The supplied table caption identifies the evaluation references but does not provide cell values.Source: Table 3 · p. 9
- Boundary-layer state outputs have R2 values of 0.72–1.00, while turbulent and moisture flux outputs have R2 values of 0.87–0.92.Figures 3 and 4 show reference, FASST prediction, and error maps averaged over the held-out test period.Source: Figure 3 · p. 10, Figure 4 · p. 11
- Radiative fluxes are reproduced with R2 ≥0.93, while precipitation is the atmospheric emulator’s weakest-performing variable at R2 = 0.48.Figure 5 maps reference, prediction, and error averaged over the held-out test period.Source: Figure 5 · p. 11
- Wind-stress outputs have R2 values of 0.63–0.73 and are the weakest-performing forced fields, despite their direct role in ocean momentum forcing.Figure 6 shows reference, prediction, and error maps averaged over the held-out test period.Source: Figure 6 · p. 12
- FASST captures broad SST spatial structure, with a colder tropical Pacific and differences in the North Atlantic, Southern Ocean, and Arctic relative to CESM2 piControl.Figure 7 compares post-transient climatologies over matched elapsed years 31–140; its columns show reference, prediction, and error, and its rows cover the global domain and named regions.Source: B Evaluation · p. 9, Figure 7 · p. 13