The Deep Learning Earth System Model (DL ESy M) represents a breakthrough approach in Earth system modeling that aims to address key challenges associated with computationally intensive traditional models. Its main goal is accurately simulate Earth's current climate over 1,000-year periods, with minimal smoothing and no drift, while being able to distinguish global warming signals from interannual variability. This deep learning model is designed to be a highly efficient and accurate tool that can produce thousands of years of realistic atmospheric and ocean states, including seasonal cycles and interannual variability.

Methodology and training DL ESy M is connected deep learning model, which combines a deep learning-based weather prediction model (DLWP) with a deep learning-based ocean model (DLOM). It is trained on historical data from the ERA5 reanalysis and satellite observations, such as outgoing longwave radiation (OLR) from ISCCP, covering the period from 1983 to 2017. Although the atmosphere and ocean modules are trained separately, they are coupled during inference. The coupling between the atmosphere and ocean is asynchronous, with the atmosphere module updated every 12 hours and the ocean module updated every 4 days, while the DLOM only predicts sea surface temperature (SST). Precipitation is not simulated directly, but is diagnosed by a separate DL module.

Key features and performance

One of the most significant benefits of DL ESy M is its computational efficiencyIt can simulate 1,000 years of equilibrium climate in less than 12 hours on a single NVIDIA A100 GPU. This represents a radical departure from traditional numerical Earth system models, which would require months and vast computational resources, often available only at large national and international centers, to perform a similar simulation.

Regarding simulation fidelity, DL ESy M matches or outperforms leading CMIP6 models in key metrics of seasonal and interannual variability:

• Tropical cyclogenesis: The model realistically simulates tropical cyclones (TCs) over the range of observed intensities, particularly in the western North Pacific (WNP), where its average annual TC count closely matches ERA5. It outperforms CMIP6 models.
• Indian Summer Monsoon (ISM): It accurately reproduces the annual ISM cycle and its spatial distribution. Quantitative comparisons show that the ISM climatology in DL ESy M has better pattern correlation and centered RMSE compared to CMIP6 models and ISCCP satellite observations.
• Atmospheric blocking events: DL ESy M approximates the spatial distribution and frequency of blocking events, achieving better pattern correlation and centered RMSE than all four CMIP6 models compared.
• Annual modes (NAM and SAM): The model exhibits realistic Northern and Southern annual modes with similar spatial structure and magnitude as in ERA5, outperforming most CMIP6 models.
• El Niño Southern Oscillation (ENSO): Spontaneously generates temporal variations in the Niño 3.4 region comparable to observations, although with smaller amplitude.

The model shows negligible drift in globally averaged air temperature values at 2 m SST during long-term simulations, which is crucial for climate modeling and distinguishes it from traditional models that require spin-up and parameter tuning.

Limitations and future directions

Despite its impressive performance, the DL ESy M also has some limitations. overestimates precipitation in the Intertropical Convergence Zone area of the equatorial Pacific. Its simple ocean module, which only predicts SST, leads to a weaker amplitude of the ENSO signal and overall interannual variability. It also shows cold bias in Atlantic SSTs, which leads to an incorrect reduction in the number of Atlantic hurricanes in the simulations. In the current configuration, the model is only suitable for simulations of the current climate, and not for projections of future scenarios that would require the inclusion of greenhouse gas and anthropogenic aerosol impacts.

DL ESy M represents powerful yet affordable tool for studying weather and climate variability, significantly increasing the availability of a high-fidelity model of the Earth system. With its ability to accurately capture patterns of variability across time scales, it has great potential for subseasonal and seasonal forecasts (S2S)Future research will focus on developing a more comprehensive ocean module and exploring options for simulating future climate conditions, for example by using physical constraints in the training process. JRi

Study published in in the journal AGU Advances