Graph Neural Networks Revolutionize Global Weather Forecasting: From Graph Weather to Open-Source Practice of Multi-Model Fusion

The OpenClimateFix open-source project graph_weather introduces graph neural networks into the field of weather forecasting, achieving data-driven forecasting capabilities comparable to traditional physical models, and continuously integrating cutting-edge models such as DeepMind GenCast and Microsoft Aurora.

• Original Author/Maintainer: OpenClimateFix Organization (Jacob Bieker et al.)
• Source Platform: GitHub
• Original Title: graph_weather
• Original Link: https://github.com/openclimatefix/graph_weather
• Paper Source: Ryan Keisler, 2022, "Forecasting Global Weather with Graph Neural Networks" (arXiv:2202.07575)
• Inclusion Date: 2026-06-04

Traditional Numerical Weather Prediction (NWP) relies on solving complex fluid dynamics equations, requiring supercomputers for massive calculations. Major global meteorological agencies such as ECMWF and GFS consume large amounts of computing resources every day to generate forecast results for the next few days. With the intensification of climate change and the frequent occurrence of extreme weather events, the demand for faster, more accurate, and more energy-efficient forecasting methods is becoming increasingly urgent.

In 2022, Ryan Keisler from Google Research published the landmark paper "Forecasting Global Weather with Graph Neural Networks", which first proved that purely data-driven graph neural network methods can match or even surpass traditional physical models in key metrics. This breakthrough opened up a new path for weather forecasting.

The graph_weather project by OpenClimateFix is a PyTorch implementation of the paper, but it goes far beyond that—it has evolved into an open-source toolkit integrating multiple cutting-edge graph neural network weather models. The project adopts a modular design, allowing researchers to flexibly combine different model components for experiments.

The project models the Earth's atmosphere as a graph structure: each latitude and longitude grid point serves as a graph node, and the spatial relationships between nodes as edges. This representation is naturally suitable for capturing non-local interactions in the atmosphere and avoids the grid bias of traditional convolutional neural networks.

The model input is the current three-dimensional atmospheric state (including multiple variables such as temperature, pressure, humidity, wind speed, etc.), and the output is the state prediction for the next six hours. Through recursive application, continuous forecasts up to several days can be generated.

graph_weather has evolved from a single model implementation to a multi-model fusion platform, currently supporting the following cutting-edge methods:

Based on the original implementation of Keisler's paper, it uses graph neural networks to learn the temporal evolution of atmospheric states. In key metrics such as Z500 (500 hPa geopotential height) and T850 (850 hPa temperature), its performance is comparable to the operational models of GFS and ECMWF.

A graph diffusion model published by DeepMind in Nature, which uses generative methods for probabilistic ensemble forecasting. Unlike traditional deterministic forecasting, GenCast outputs a forecast distribution, which can better quantify uncertainty and is particularly important for extreme weather warnings.