- The paper introduces SamudrACE, a physically constrained deep learning emulator that efficiently simulates Indian monsoon intraseasonal and interannual variability using a coupled 3D neural framework.
- It employs independent training of atmospheric and oceanic components on 150 years of CM4 data, capturing key features like ITCZ migration and rainfall EOF patterns despite underestimating amplitude.
- The study reveals challenges in replicating equatorial waves and ENSO-monsoon teleconnections, indicating a need for improved resolution, refined coupling strategies, and better training data.
Deep Learning-Based Simulation of Indian Monsoon Variability Using SamudrACE
Introduction
Significant advances in AI/ML-based emulation of atmospheric and climate dynamics have recently extended the frontier for high-resolution, low-cost simulation, especially for weather forecasts. However, realizing analogous capability for sub-seasonal to seasonal (S2S) variability—specifically for coupled ocean-atmosphere processes—remains unresolved. "A Deep Learning Earth System Model Simulation of Indian Monsoon Intraseasonal and Interannual Variability" (2607.01676) provides a comprehensive and critical evaluation of SamudrACE, the first operational, physically-constrained, deep learning (DL) Earth system emulator, focusing on its skill in capturing monsoon intraseasonal oscillations (MISO), interannual Indian summer monsoon rainfall (ISMR) variability, and air-sea interaction processes such as ENSO-monsoon coupling.
SamudrACE: Model Architecture and Experimental Protocol
SamudrACE is a coupled 3D neural emulator of the GFDL-CM4 global climate model. Its atmospheric (ACE2) and oceanic (Samudra) components are trained independently on 150 years of pre-industrial CM4 simulations, then coupled using physically consistent fluxes (heat, momentum, freshwater). The system is highly computationally efficient, achieving 300-year simulations with daily outputs at 100 km resolution in a single day on modern GPU hardware. The time integration handles the faster atmospheric processes (6-hour timestep) and slower ocean dynamics (5-day timestep) with carefully tuned surface coupling.
Three independent 300-year runs were performed for robust assessment. Verification uses both high-resolution reanalysis and satellite rainfall (CMORPH), instrumental ISMR data (IITM RR-65), and sea surface temperature (COBE/ERA5) for process evaluation and spectral analysis.
Intraseasonal (MISO) and Equatorial Wave Representation
Equatorial Wave Diagnostics
Wheeler-Kiladis spectra were computed to diagnose the simulation of convectively coupled equatorial waves. SamudrACE replicates most symmetric wave modes (MJO, Kelvin, westward-propagating long Rossby waves) comparably to the parent CM4, but further exacerbates CM4's underestimation of eastward-propagating atmospheric Kelvin waves (3–10-day period) and antisymmetric mixed Rossby-gravity (MRG) waves. These biases likely stem from limitations in the training data and may affect large-scale air-sea interaction processes critical for ENSO and monsoon variability.
MISO Propagation and Structure
SamudrACE captures the northward migration of the ITCZ and MISO activity up to ~22°N, which is an improvement over many CMIP6-class models (which arrest migration at 13–15°N). However, both SamudrACE and CM4 misplace the secondary equatorial rainband (oceanic ITCZ) north of the equator, contrary to observed positioning south of ~5°S, indicating a bias inherited from training. Spectral analysis of the MISO index (from extended EOF decomposition) reveals a marked model bias: the observed dominant power at 42 days is severely damped (model captures no more than 50% of variance at this period) and instead shifts to a strong, exaggerated ~60-day period. Both amplitude and spatial structure are overly zonal, lacking the observed southeast-to-northwest tilt during northward propagation. Thus, the emulator qualitatively tracks propagation but underestimates amplitude and structural realism, with significant implications for S2S forecasting.
Interannual Variability of Indian Monsoon Rainfall
SamudrACE moderately reproduces the climatological mean and main empirical orthogonal function (EOF) patterns of JJAS rainfall and its leading interannual variability modes over India. The explained variances for the top two modes agree well with both CM4 and ERA5, with EOF1 capturing the canonical east-west dipole and EOF2 reflecting a north-south tripole. The interannual standard deviation simulated (64.0 mm) is below observed (83.2 mm), with a similar deficit in CM4, confirming that amplitude errors are primarily inherited from the training GCM rather than from the DL approximation alone. The model nonetheless reproduces the frequency and statistics of droughts and excess monsoon years reasonably well.
In the frequency domain, SamudrACE replicates multi-decadal (60–80 year) variability but develops a spurious decadal (~20 year) peak absent from observations. For ISMR prediction, skill at reproducing observed statistical properties of variability is necessary but not sufficient—physical linkage to ENSO remains the critical test.
ENSO Simulation and Air-Sea Coupling Biases
ENSO Magnitude and Structure
The emulator substantially underestimates the amplitude of interannual Niño-3.4 SST variability (σ = 0.435°C vs. 0.632°C observed, 0.663°C in CM4). The simulated ENSO is too periodic (~3-year dominant period), lacking realistic intermittency and amplitude of extreme events, and exhibits a substantial cold bias (≥1°C) in tropical SSTs, also inherited from CM4. The dominant EOF mode of SST ("global ENSO mode") is realistic, but secondary modes are poorly simulated.
ENSO–Monsoon Teleconnection
The phase relationship of ISMR-ENSO is notably misrepresented. Observations feature peak negative correlation (about –0.46) when boreal summer monsoon rainfall leads Niño-3.4 by ~3 months; SamudrACE shifts this maximum anticorrelation to a lead of ~12 months, distorting the contemporaneous teleconnection structure and impugning its fidelity for operational seasonal forecast applications.
Air-Sea Instability and Coupled Processes
Process diagnostics of super El Niño events—examining the time-longitude evolution of equatorial SST, surface wind and thermocline anomalies—show that while SamudrACE can simulate the timing and propagation of event “seeds” (e.g., subsurface thermocline anomalies), it fails to amplify air-sea coupled instabilities adequately. Both westerly wind bursts and Kelvin wave propagation occur but with muted feedback, resulting in underpowered ENSO events. This highlights a deficiency of the reduced vertical ocean resolution and the 5-day mean coupling approach in capturing critical nonlinear instability mechanisms.
Model Bias Origins and Implications for Improved Simulation
Biases in SamudrACE originate from two main sources: (i) GCM-based training data biases (systematic mean state, climatology, and variability deficits in CM4), and (ii) errors intrinsic to the DL emulator related to training setup, vertical/horizontal resolution, and coupling strategy. The main weaknesses are:
- Persistent cold SST bias and amplification of global SST errors;
- ENSO energy deficit: amplitude of ENSO is ~50% too weak and events are too periodic, indicating loss of intermittency and stochasticity in the DL system;
- Incorrect lead-lag structure: phase relationship between ENSO and ISMR is improperly simulated, undermining S2S skill;
- Dampened coupled instability: Emulated ocean–atmosphere coupling fails to amplify large ENSO events due to insufficient vertical oceanic resolution and underresolved coupling frequencies.
However, SamudrACE's strengths include computational efficiency, realistic multidecadal ISMR variability (which, intriguingly, is not present in the training GCM), and potential as a framework for further model improvement. The spontaneous generation of physically plausible multidecadal modes raises challenging questions for future work on the nature of emergent variability in machine-learned coupled systems.
Recommendations for Future Development
Substantial improvements in simulation fidelity could be achieved by:
- Retraining SamudrACE with high-resolution, low-bias CMIP6-class GCM outputs, leveraging ensemble diversity;
- Increasing vertical resolution in the ocean module, especially in the upper 300 m, to resolve thermocline processes;
- Enhancing coupling frequency (potentially to subdaily steps) between atmospheric and oceanic modules, as recent studies suggest this increases ENSO simulation skill;
- Applying transfer learning on observational reanalyses post-GCM pretraining, as demonstrated in other seasonal prediction studies;
- Systematic sensitivity analysis of coupling schemes and architectural modifications, particularly focused on improving ENSO-monsoon teleconnections and extremes.
Conclusion
The study offers the first in-depth evaluation of a fully coupled DL-based ESM emulator, SamudrACE, for Indian monsoon simulation at both intraseasonal and interannual scales. While mean state and dominant spatial/temporal patterns are retained from the parent GCM, substantial amplitude, frequency, and air-sea interaction biases undermine S2S forecast potential—particularly for ENSO and its teleconnection to the monsoon. These findings underscore the necessity of upgrading training datasets, increasing model resolution, and refining coupling strategies to realize the practical use of DL-based emulators for operational climate prediction. SamudrACE's computational efficiency and successful emulation of multidecadal natural variability, despite strong bias inheritance, position it as a strategic platform for rapid iteration and improvement of next-generation S2S climate emulators.