Coupled Ocean-Atmosphere Dynamics in a Machine Learning Earth System Model

TL;DR

The paper addresses the challenge of seasonal climate forecasting by developing Ola, a high-resolution AI/ML coupled Earth System Model with SFNO-based ocean and atmosphere components. Ola demonstrates learned ocean–atmosphere coupling, reproduces tropical wave dynamics, and generates internally consistent ENSO with realistic amplitude and vertical structure, achieving competitive skill to the physics-based SPEAR model in preliminary tests. The study shows that coupling ML components can yield fast, scalable forecasts with reduced tropical biases and realistic subsurface temperature patterns, highlighting the potential of AI/ML for climate-scale prediction. While promising, the work also identifies high-latitude drift and rollout stability as limitations, outlining pathways for improvement through richer state representations, longer training data, and systematic intercomparisons to enable operational deployment and climate projections.

Abstract

Seasonal climate forecasts are socioeconomically important for managing the impacts of extreme weather events and for planning in sectors like agriculture and energy. Climate predictability on seasonal timescales is tied to boundary effects of the ocean on the atmosphere and coupled interactions in the ocean-atmosphere system. We present the Ocean-linked-atmosphere (Ola) model, a high-resolution (0.25°) Artificial Intelligence/ Machine Learning (AI/ML) coupled earth-system model which separately models the ocean and atmosphere dynamics using an autoregressive Spherical Fourier Neural Operator architecture, with a view towards enabling fast, accurate, large ensemble forecasts on the seasonal timescale. We find that Ola exhibits learned characteristics of ocean-atmosphere coupled dynamics including tropical oceanic waves with appropriate phase speeds, and an internally generated El Niño/Southern Oscillation (ENSO) having realistic amplitude, geographic structure, and vertical structure within the ocean mixed layer. We present initial evidence of skill in forecasting the ENSO which compares favorably to the SPEAR model of the Geophysical Fluid Dynamics Laboratory.

Figures (28)

• Figure 1: Sea Surface Temperature (SST) anomaly in the Niño 3.4 region of the Pacific Ocean (5$^{\circ}$N to 5$^{\circ}$S, 170$^{\circ}$-120$^{\circ}$W) from 6-month ensemble forecasts generated by the coupled AI/ML ESM Ola(light red lines) and the physics-based model GFDL-SPEAR (light purple lines). Lagged Ensemble Forecasts (see Sec. \ref{['sec:LEF']} for details) are generated at the beginning of every month and run out to 6 months. The ensemble mean for the 12 ensemble forecasts is indicated by corresponding bold lines. The observed ERA5 re-analysis (ground truth) is indicated by the bold black line.
• Figure 2: Longitude-time diagrams show the simulations (top row) and corresponding ground truth (UFS-replay dataset) verification (bottom row) of equatorial Kelvin and Rossby waves in the Ola ML ESM initialized on January 1 2018 and simulated out to 6 months. The left column ((a), (d)) shows the SSH anomaly averaged from 2$^{\circ}$N to 2$^{\circ}$S. The middle column ((b), (e)) and right columns ((c), (f)) shows the SSH anomaly averaged from 5$^{\circ}$ to 10$^{\circ}$ in Southern and Northern Hemisphere. In each panel, the annual mean climatology is subtracted at each grid point to compute the SSH anomaly.
• Figure 3: Composite Sea Surface Temperature (a-d) and upper ocean temperature (e-h) anomaly visualizations during El Niño and La Niña events from the Ola coupled ML model ensemble simulations ((a), (b), (e), (f)) at a 4-month lead time and the observed ground truth record ((c), (d), (g), (h)). Ensemble forecasts were generated from the coupled ML model every month from 2017 to 2021 with 12 ensembles at each initialization providing a total of 720 model runs (see Sec. \ref{['sec:LEF']} for detail). The Ola model forecasts were bias corrected by computing a lead-time dependant model bias from simulations in the training period of 1994-2016 (see Appendix \ref{['sec:app_bias']} for bias correction details). The criteria for El Niño and La Niña conditions were defined as the SST monthly anomaly in the Niño 3.4 region exceeding 0.5K. There were 136 months with El Niño states and 187 months with La Niña states at a lead time of 4 months in the 720 ensemble simulations while the observed record during the period from 2017 to 2021 had 13 months with El Niño states and 14 months with La Niña states. The hatching indicates statistically significant differences between El Nino/La Nina states and Neutral states indicated by a two-sample T-test with a significance threshold of 0.05.
• Figure 4: Same as Figure \ref{['fig:enso_comp_ocean']} but for Mean Sea Level Pressure (MSLP) anomaly during El Niño and La Niña events. The top row (a,b) shows the average MSLP anomaly from the Ola coupled ESM at a lead time of 4 months, and the bottom row (c,d) shows the groud truth (ERA5).
• Figure 5: The ensemble mean reforecasts by Ola initialized in 2018/05. (a) and (c) show the SST and equatorial Pacific upper ocean temperature anomalies at one month lead time, while (e) and (g) show the same variables but at 6 months lead time. The results are averaged over 12 ensembles initialized with LEF method (see details in Sec. \ref{['sec:LEF']}) with biases correction (see details in Sec. \ref{['sec:app_bias']}). The corresponding observational records ((b), (d), (f) and (h)) are shown in the right column for comparison.