GraphCast
A first-generation AI weather forecasting model developed by Google DeepMind. It matches or locally surpasses traditional numerical weather prediction systems in medium-range forecast accuracy (3–10 days) while offering substantially faster inference.
Timeline (last 90 days)
Pure AI weather models such as GraphCast and Pangu-Weather generally encounter bottlenecks in maintaining accuracy at time ranges beyond ten days.
Google's GraphCast is a global medium-range weather forecasting model based on graph neural networks that can generate a ten-day forecast in less than a minute on a single TPU and has surpassed ECMWF's traditional high-resolution model on multiple metrics.
Machine learning models such as GraphCast, Pangu-Weather, and FourCastNet have already demonstrated accuracy comparable to or even better than top numerical models in short- and medium-range forecasting, with inference speeds thousands of times faster than traditional models.
GraphCast is a model designed specifically for climate and weather prediction.
GraphCast is based on graph neural networks, while Pangu-Weather and FengWu are based on Transformers; they generally adopt autoregressive forecasting with a 6-hour time step, with forecast lead times typically within 10 days.
Pangu-Weather employs a 3D Transformer with Earth-specific positional encoding; GraphCast converts the Earth's grid into a graph structure to model long-range atmospheric teleconnections; GenCast supports ensemble forecasting through probabilistic generation.
Google DeepMind's GraphCast and Google Research's NeuralGCM are recent representative achievements in the field of AI climate modeling.
AI weather models such as FourCastNet, Pangu-Weather, and GraphCast learn the statistical patterns of atmospheric dynamics by training on long-term reanalysis datasets such as ERA5
First-generation AI weather models represented by GraphCast and Pangu-Weather have already matched and in some cases locally surpassed NWP in medium-range forecast (3–10 day) accuracy.
DeepMind launched the GraphCast model in 2023, which is based on a graph neural network architecture and surpassed the then state-of-the-art HRES system on more than 90% of metrics in 10-day forecast accuracy on ECMWF's WeatherBench 2 benchmark.
All Facts (10)
Pure AI weather models such as GraphCast and Pangu-Weather generally encounter bottlenecks in maintaining accuracy at time ranges beyond ten days.
50%UnverifiedGoogle's GraphCast is a global medium-range weather forecasting model based on graph neural networks that can generate a ten-day forecast in less than a minute on a single TPU and has surpassed ECMWF's traditional high-resolution model on multiple metrics.
50%UnverifiedMachine learning models such as GraphCast, Pangu-Weather, and FourCastNet have already demonstrated accuracy comparable to or even better than top numerical models in short- and medium-range forecasting, with inference speeds thousands of times faster than traditional models.
50%UnverifiedGraphCast is a model designed specifically for climate and weather prediction.
50%UnverifiedGraphCast is based on graph neural networks, while Pangu-Weather and FengWu are based on Transformers; they generally adopt autoregressive forecasting with a 6-hour time step, with forecast lead times typically within 10 days.
50%UnverifiedPangu-Weather employs a 3D Transformer with Earth-specific positional encoding; GraphCast converts the Earth's grid into a graph structure to model long-range atmospheric teleconnections; GenCast supports ensemble forecasting through probabilistic generation.
50%UnverifiedGoogle DeepMind's GraphCast and Google Research's NeuralGCM are recent representative achievements in the field of AI climate modeling.
50%UnverifiedAI weather models such as FourCastNet, Pangu-Weather, and GraphCast learn the statistical patterns of atmospheric dynamics by training on long-term reanalysis datasets such as ERA5
50%UnverifiedFirst-generation AI weather models represented by GraphCast and Pangu-Weather have already matched and in some cases locally surpassed NWP in medium-range forecast (3–10 day) accuracy.
50%UnverifiedDeepMind launched the GraphCast model in 2023, which is based on a graph neural network architecture and surpassed the then state-of-the-art HRES system on more than 90% of metrics in 10-day forecast accuracy on ECMWF's WeatherBench 2 benchmark.
50%