Forecasting Global Weather with Graph Neural Networks

TL;DR

The paper introduces a data-driven global weather forecaster based on graph neural networks that advances the atmospheric state by six hours per step and chains steps for multi-day forecasts. Trained on ERA5 reanalysis and GFS forecasts, it uses an encoder-processor-decoder architecture on an icosahedral latent grid to handle spherical geometry and dense state representation (78 channels across 13 pressure levels). The approach yields competitive performance with operational models at 1° resolution and demonstrates a live GFS integration as a hybrid physics-ML pipeline, highlighting stability and scalability considerations. Overall, the work shows the potential of dense, GNN-based forecasts to complement traditional NWP, with clear avenues for higher resolution, adaptive meshing, and ensemble applications.

Abstract

We present a data-driven approach for forecasting global weather using graph neural networks. The system learns to step forward the current 3D atmospheric state by six hours, and multiple steps are chained together to produce skillful forecasts going out several days into the future. The underlying model is trained on reanalysis data from ERA5 or forecast data from GFS. Test performance on metrics such as Z500 (geopotential height) and T850 (temperature) improves upon previous data-driven approaches and is comparable to operational, full-resolution, physical models from GFS and ECMWF, at least when evaluated on 1-degree scales and when using reanalysis initial conditions. We also show results from connecting this data-driven model to live, operational forecasts from GFS.

Figures (8)

• Figure 1: Using the current atmospheric state, the model evolves the state forward by 6 hours. The 3D atmospheric state is defined on a uniform latitude/longitude grid, with 78 channels per pixel (6 physical variables $\times$ 13 pressure levels = 78 channels). An Encoder GNN encodes onto latent features defined on a icosahedron grid, a Processor GNN performs additional processing of the latents, and a Decoder GNN maps back to the atmospheric state on a latitude/longitude grid.
• Figure 2: A schematic view of the local graph connectivity in the Encoder, Processor, and Decoder. Left: local spatial and channel information is encoded into an icosahedron node using data from nearby nodes on the input latitude/longitude grid. Center: data on the icosahedron node is further processed using data from nearby icosahedron nodes (including itself, which is not explicitly shown). Right: the output latitude/longitude data is created by decoding data from nearby icosahedron nodes.
• Figure 3: An example of the 6-hour difference in geopotential height, temperature, and humidity in the ERA5 dataset (left column) and the prediction from the machine learning model (right column). The model is able to accurately predict 6-hour changes in these variables using only the initial state. All raster figures in this paper use a $1^\circ$ latitude/longitude grid centered on the prime meridian.
• Figure 4: An example of the 6-hour difference in eastward component of the wind, the northward component of the wind, and the vertical component of the wind in the ERA5 dataset (left column) and the prediction from the machine learning model (right column). The model is able to accurately predict 6-hour changes in these variables using only the initial state.
• Figure 5: An example multi-step rollout of the ML forecast vs reanalysis data from ERA5. Beginning with the ERA5 initial conditions at 0 hours, the ML system steps forward autoregressively in 6-hour steps. While the model evolves 78 separate physical channels, we show only Q850, the specific humidity on the 850 hPa pressure level. The output of the ML forecast generally tracks the large-scale flows seen in ERA5, although the predictions do become smoother over time. Additional media, including videos, can be found at https://rkeisler.github.io/graph_weather