Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Coupled Earth System Dynamics with GraphDOP

Eulalie Boucher, Mihai Alexe, Peter Lean, Ewan Pinnington, Simon Lang, Patrick Laloyaux, Lorenzo Zampieri, Patricia de Rosnay, Niels Bormann, Anthony McNally

TL;DR

The paper tackles the challenge of fully coupling Earth System components in data-driven forecasting by presenting GraphDOP, an observation-driven graph neural network that embeds heterogeneous satellite and in-situ observations into a shared latent space to learn cross-domain dynamics without explicit coupling. It extends AI-DOP by using an encoder–processor–decoder framework with a latent o96 grid and 1024 features, trained on two decades of observations and evaluated on three case studies spanning cryosphere–ocean, ocean–atmosphere, and land–atmosphere interactions. The results show GraphDOP can reproduce key coupled dynamics such as Arctic sea-ice rapid freezing, hurricane Ian’s atmosphere–ocean–wave response, and European heatwave evolution, albeit with some biases due to observation coverage and limited storm-core data. The findings indicate that end-to-end data-driven Earth System prediction is viable, with potential for scalable, flexible forecasts that integrate diverse observational streams, while highlighting the need for richer observational inputs and probabilistic or memory-enhanced training to improve extremes and deep-ocean coupling.

Abstract

Interactions between different components of the Earth System (e.g. ocean, atmosphere, land and cryosphere) are a crucial driver of global weather patterns. Modern Numerical Weather Prediction (NWP) systems typically run separate models of the different components, explicitly coupled across their interfaces to additionally model exchanges between the different components. Accurately representing these coupled interactions remains a major scientific and technical challenge of weather forecasting. GraphDOP is a graph-based machine learning model that learns to forecast weather directly from raw satellite and in-situ observations, without reliance on reanalysis products or traditional physics-based NWP models. GraphDOP simultaneously embeds information from diverse observation sources spanning the full Earth system into a shared latent space. This enables predictions that implicitly capture cross-domain interactions in a single model without the need for any explicit coupling. Here we present a selection of case studies which illustrate the capability of GraphDOP to forecast events where coupled processes play a particularly key role. These include rapid sea-ice freezing in the Arctic, mixing-induced ocean surface cooling during Hurricane Ian and the severe European heat wave of 2022. The results suggest that learning directly from Earth System observations can successfully characterise and propagate cross-component interactions, offering a promising path towards physically consistent end-to-end data-driven Earth System prediction with a single model.

Learning Coupled Earth System Dynamics with GraphDOP

TL;DR

The paper tackles the challenge of fully coupling Earth System components in data-driven forecasting by presenting GraphDOP, an observation-driven graph neural network that embeds heterogeneous satellite and in-situ observations into a shared latent space to learn cross-domain dynamics without explicit coupling. It extends AI-DOP by using an encoder–processor–decoder framework with a latent o96 grid and 1024 features, trained on two decades of observations and evaluated on three case studies spanning cryosphere–ocean, ocean–atmosphere, and land–atmosphere interactions. The results show GraphDOP can reproduce key coupled dynamics such as Arctic sea-ice rapid freezing, hurricane Ian’s atmosphere–ocean–wave response, and European heatwave evolution, albeit with some biases due to observation coverage and limited storm-core data. The findings indicate that end-to-end data-driven Earth System prediction is viable, with potential for scalable, flexible forecasts that integrate diverse observational streams, while highlighting the need for richer observational inputs and probabilistic or memory-enhanced training to improve extremes and deep-ocean coupling.

Abstract

Interactions between different components of the Earth System (e.g. ocean, atmosphere, land and cryosphere) are a crucial driver of global weather patterns. Modern Numerical Weather Prediction (NWP) systems typically run separate models of the different components, explicitly coupled across their interfaces to additionally model exchanges between the different components. Accurately representing these coupled interactions remains a major scientific and technical challenge of weather forecasting. GraphDOP is a graph-based machine learning model that learns to forecast weather directly from raw satellite and in-situ observations, without reliance on reanalysis products or traditional physics-based NWP models. GraphDOP simultaneously embeds information from diverse observation sources spanning the full Earth system into a shared latent space. This enables predictions that implicitly capture cross-domain interactions in a single model without the need for any explicit coupling. Here we present a selection of case studies which illustrate the capability of GraphDOP to forecast events where coupled processes play a particularly key role. These include rapid sea-ice freezing in the Arctic, mixing-induced ocean surface cooling during Hurricane Ian and the severe European heat wave of 2022. The results suggest that learning directly from Earth System observations can successfully characterise and propagate cross-component interactions, offering a promising path towards physically consistent end-to-end data-driven Earth System prediction with a single model.
Paper Structure (9 sections, 7 figures, 1 table)

This paper contains 9 sections, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. The GraphDOP Model
  3. Observation Datasets
  4. Sea ice rapid freezing case study: cryosphere and ocean
  5. Hurricane Ian case study: ocean and atmosphere
  6. Atmospheric fields
  7. Ocean surface and wave response
  8. European heatwave case study: land surfaces
  9. Discussion and conclusion

Figures (7)

  • Figure 1: GraphDOP alexe2024graphdopskilfuldatadrivenmediumrange with 3-hour autoregressive time-stepping (rollout) in the latent space. During fine-tuning, the forecasted observations are fed back as inputs to produce the next forecast (rollout in observation space).
  • Figure 2: AMSR-2 channel 5 (10v) brightness temperatures (K) from left to right: target (observed), forecasted, difference (target minus forecast). The rightmost column shows the sea ice concentration of the ECMWF ORAS6 reanalysis. The top row shows observations at 24-hour lead time (21 October 2022), the middle row shows the predicted observations at 5-day lead time (25 October 2022), whereas the bottom row shows the predicted observations at 10-day lead time (30 October 2022). The 10-day forecast shown here has been chosen because it features a rapid freezing event, and was initialised using 12 hours of observations between 09:00 and 21:00 UTC on 20 October 2022.
  • Figure 3: Near-surface atmospheric conditions (10 m wind and 2 m temperature) from the GraphDOP forecast (top row) and the ERA5 reanalysis (bottom row) for the three regions that exhibit the largest sea-ice extent mismatches: Baffin Bay, the Kara Sea, and the East Siberian Sea (from left to right, respectively). The colour scale of the 2 m temperature is centred at $-2.5\,^{\circ}\mathrm{C}$ to highlight the likely sea-ice edge position.
  • Figure 4: ERA5 reanalysis (top row of each subfigure) and GraphDOP forecasts (bottom row) for (a) mean sea-level pressure and (b) 10 m wind speed, valid at 00:00 UTC from 26 September to 1 October 2022 (left to right). Forecasts were initialized with observations between 09:00 UTC and 21:00 UTC on 24 September 2022.
  • Figure 5: Observed (top row) and forecasted (bottom row) AVHRR visible reflectance at 2, 3 and 4-day lead times, valid between 26 and 28 September 2022.
  • ...and 2 more figures