GraphCast: graph neural networks for weather forecasting
Around 2023, DeepMind created GraphCast. A deterministic GNN that represents the atmosphere as a multi-scale icosahedral mesh with learned edges. Two 6-hour-apart states in, autoregressive rollout to 10 days out. It matched or beat the physics-based HRES model. It was pretty much the first time in history that an ML model was better than a simulation. In 2024 GenCast was invented, same graph substrate, but the deterministic decoder was swapped for a conditional diffusion process, producing 50+ ensemble scenarios instead of a single forecast. Uncertainty from a graph neural network. Somewhere in 2026 Google released WeatherNext 2. Still a graph mesh but with a GNN encoder/decoder mapping the lat-lon grid onto a icosahedral latent mesh, with a graph transformer operating on the mesh nodes . Noise gets injected directly into the model’s normalization layers rather than through a diffusion process. It’s now live in Google Search, Maps, Gemini, and Pixel Weather. The physical mesh remained the same but the generative mechanism evolved:from deterministic to diffusion to function-space noise injection. The mesh topology turned out to be the durable part of the design while the uncertainty model was the part that got iterated.
Before you ask me to create a GNN for you: I tried to estimate the cost of training graphcast and came up with something ~$400K. Very few companies have the data and the cash to experiment with this kinda technology.
☀️ WeatherNext2: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/weathernext-2 ⛈️ Graphcast: https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/