SamudrACE: Fast and Accurate Coupled Climate Modeling with 3D Ocean and Atmosphere Emulators

TL;DR

SamudrACE is a coupled global climate model emulator which produces centuries-long simulations at 1-degree horizontal, 6-hourly atmospheric, and 5-daily oceanic resolution, with 145 2D fields spanning 8 atmospheric and 19 oceanic vertical levels, plus sea ice, surface, and top-of-atmosphere variables.

Abstract

Traditional numerical global climate models simulate the full Earth system by exchanging boundary conditions between separate simulators of the atmosphere, ocean, sea ice, land surface, and other geophysical processes. This paradigm allows for distributed development of individual components within a common framework, unified by a coupler that handles translation between realms via spatial or temporal alignment and flux exchange. Following a similar approach adapted for machine learning-based emulators, we present SamudrACE: a coupled global climate model emulator which produces centuries-long simulations at 1-degree horizontal, 6-hourly atmospheric, and 5-daily oceanic resolution, with 145 2D fields spanning 8 atmospheric and 19 oceanic vertical levels, plus sea ice, surface, and top-of-atmosphere variables. SamudrACE is highly stable and has low climate biases comparable to those of its components with prescribed boundary forcing, with realistic variability in coupled climate phenomena such as ENSO that is not possible to simulate in uncoupled mode.

Figures (26)

• Figure 1: A single 5-day forward step in uncoupled (a, b) and coupled (c) modes. In uncoupled mode, ACE and Samudra are forced by the appropriate CM4 reference fields. Uncoupled ACE is forced by the reference 5-day average SST and sea ice concentration, stepping forward 6 hours at a time until reaching 5 days, at which point the next 5-day average forcing is prescribed. For uncoupled Samudra, we use the 5-day average of the CM4 reference wind stress, precipitation, and surface fluxes as prescribed forcing inputs. In coupled mode, aggregation of the diagnostic 6-hour average surface boundary conditions to 5-day averages is done online as ACE completes 20 forward steps. The generated 5-day average is then passed as input to Samudra, which generates a single 5-day forward step. Samudra's generated SST and sea ice states will then be used to force ACE in the next iteration of the coupler loop.
• Figure 2: Spatial bias patterns of the generated 40-year time-mean precipitation and surface temperature, computed as the difference between the time mean of the emulated output and the time mean of the reference simulation over the held-out inference period.
• Figure 3: Monthly mean over the 40-year held-out period of (a) Northern and (b) Southern Hemisphere sea ice extent. Shading denotes the interannual standard deviation over 40 years. Panel c-d) shows the time mean sea ice fraction over the same time period for the CM4 target, SamudrACE, and its bias.
• Figure 4: ENSO characteristics in the 200-year CM4 simulation (black) and 5 separate rollouts of SamudrACE starting from different initial conditions (colors): (a) time series of monthly mean Niño 3.4 index; (b) corresponding temporal power spectra (K$^2$/octave) computed using Welch's method with 15-year segments welch1967useharris2005use; (c) regression of the spatial pattern of precipitation on the Niño 3.4 index in CM4 and SamudrACE for the held-out 40-year inference period.
• Figure S1: Training and validation loss and channel-mean RMSE. Samudra and ACE2 are pretrained in uncoupled mode for 150 epochs (106,050 steps) and 50 epochs (707,200 steps), respectively, and both achieve minimum channel-mean RMSE and validation loss late in training. After pretraining, ocean-only coupled fine-tuning ("SamudrACE (FTO)") was carried out for 20 epochs (14,140 steps), during which the Samudra checkpoint with the lowest RMSE was coupled to the pretrained and fixed ACE2 checkpoint with lowest RMSE. At initialization the coupled pretrained emulators result in large validation loss and coupled RMSE (0.13 and 0.23, respectively, at training step 0). After 2 completed epochs the ocean-only coupled fine-tuning achieved its lowest coupled RMSE of 0.052, averaged over all ocean and atmosphere channels. This checkpoint was then further fine-tuned for an additional 20 epochs, updating both Samudra and ACE2 ("SamudrACE"), and reached the lowest coupeld RMSE of 0.048 after 9 completed epochs. RMSEs are averaged across all channels, where the "Coupled" RMSE (solid black line) is the weighted average of the "Ocean" (dashed black line) and "Atmosphere" (dotted black line) channel-mean RMSEs in coupled fine-tuning runs.