A Deep Learning Earth System Model for Efficient Simulation of the Observed Climate

TL;DR

This work introduces the Deep Learning Earth System Model (DLESyM), a parsimonious coupled atmosphere-ocean DL framework designed to reproduce the current Earth climate with long, drift-free rollouts. By deploying two U-net based prognostic networks for the atmosphere and ocean, plus a separate precipitation-diagnosis module, and training on short-term data from ERA5 and ISCCP, the model achieves 1000-year equilibrated climate simulations in about 12 hours. DLESyM demonstrates realistic representations of seasonal and interannual variability, including tropical cyclones, atmospheric blocking, NAM/SAM modes, and the Indian Summer Monsoon, with fidelity competitive to CMIP6 models at similar resolution. The approach lowers computational barriers and enables extensive long-range climate analysis and S2S forecasting, while highlighting pathways for future improvements such as a more sophisticated ocean component and forcing schemes for future climate scenarios.

Abstract

A key challenge for computationally intensive state-of-the-art Earth System models is to distinguish global warming signals from interannual variability. Here we introduce DLESyM, a parsimonious deep learning model that accurately simulates the Earth's current climate over 1000-year periods with no smoothing or drift. DLESyM simulations equal or exceed key metrics of seasonal and interannual variability--such as tropical cyclogenesis over the range of observed intensities, the cycle of the Indian Summer monsoon, and the climatology of mid-latitude blocking events--when compared to historical simulations from four leading models from the 6th Climate Model Intercomparison Project. DLESyM, trained on both historical reanalysis data and satellite observations, is an accurate, highly efficient model of the coupled Earth system, empowering long-range sub-seasonal and seasonal forecasts while using a fraction of the energy and computational time required by traditional models.

Figures (14)

• Figure 1: Architecture and coupling of DLESyM. (a) Schematic representation of the DLWP module as a sequence of operations on layers (see legend). U-net levels are labeled by their channel depth, with $D_1=180$ and $D_2=D_3=90$ being associated with the first convolutions in each level. Each ConvNeXt block (blue) is replaced by the layers and operations shown in the inset labeled CNB, with generic embedding depths $D$ and $I$ determined by the channel depth of the input and the labeled value of $D_n$. The purple blocks labeled GRU denote convolutional Gated Recurrent Unit layers, which are implemented with $1\times 1$ spatial convolutions. Other layers evaluated by the encoder are shown as dark green, while those evaluated by the decoder are shown as light green. (b) DL ESyM coupling mechanism. Atmosphere and ocean states are denoted by $A$ and $S$, respectively, with the subscript indicating time in hours with respect to initialization at hour 0. Light blue box represents the atmospheric module (DLWP) while the darker blue represents the ocean (DLOM). Data flow and sequence of calls are given with arrows.
• Figure 2: Multi-century DLESyM simulations continually generate high amplitude features as well as correct representations of current global climate. Contours of $Z_{1000}$ (black) and 10-m wind speed (color fill) for (a) a simulated ETC on January 22, 3016 and (b) observed storm on March 3, 2018. Zonally averaged 3-day mean of $Z_{500}$ plotted as a function of time and latitude: (c) for the first and last 5 years of a recursive 1,000-year model simulation initialized on 1 January 2017 (d) corresponding $Z_{500}$ field from the ERA5 reanalysis for the years 2017-2022. Also plotted are 15-day averaged values for the 560 dam contour for the DL ESyM simulation (white line) in (c) and for ERA5 (black line) as a reference in both (c) and (d). Globally averaged annual mean (e) T$_{2m}$ and (f) SST (black) during the 1,000 year simulation with the linear fit (red) and the trend noted in each panel.
• Figure 3: Climatology of global annual precipitation within DLESyM simulations. Comparison of annual average precipitation (a) diagnosed from the last 41 years of our 100-year simulation, (b) from ERA5 reanalysis for 1979-2020.
• Figure 4: DLESyM tropical cyclone climatology in WNP compared to those of CMIP6 historical simulations. (a) ERA5 reanalysis on November 12, 1996 and (b) DL ESyM on August 19, 2114 showing $Z_{1000}$ (black contours, c.i.=6 dkm, contours $\leq$ 0 dkm are dashed) and 10-m wind speeds (color fill). Both ERA5 and DL ESyM fields are plotted at 1$^\circ$ latitude-longitude resolution. (c)--(h): average frequency of TCs in storms per day from ERA5 and CMIP6 historical runs from 1985-2014, or 2085-2114 from DL ESyM. Average annual number of TCs of Ranks 1-4 is noted in each panel. Black curve in (c)-(h) is the running 20-day average frequency of Rank 2 storms in ERA5.
• Figure 5: Tropical cyclone tracks in the Western North Pacific (WNP) are similar in the ERA5 record and DLESyM simulations. Comparison of TC tracks over a 5-year period from (A) ERA5 (2010-2014), (B) from the DL ESyM rollout (2110--2114). These tracks use data at 6-h time resolution, with ranks 1--4 representing storms that, at their highest intensity exceed thresholds corresponding to approximately the 1st, 0.1st, 0.01st, and 0.001st percentile $Z_{1000}$ values in warm core systems over the western North Pacific region.