Google DeepMind's AI Weather Forecaster Handily Beats a Global Standard

weather-forecastinggraph-neural-networksdeepmindgraphcastclimate

Abstraction: DeepMind's GraphCast beats ECMWF across 90% of atmospheric variables

Key points:

GraphCast outperformed ECMWF forecasts on over 90% of 1,300+ atmospheric variables; results published in Science in November 2023
The model uses graph neural networks (GNNs) trained on 39 years of ECMWF observations; runs on a laptop in under a minute vs. hours on a supercomputer
Predicted Hurricane Lee's landfall in Nova Scotia 10+ days out when official models were still uncertain
Key weakness: does not produce ensemble forecasts with probability ranges, and tends to underestimate extreme events like Category 5 storms
Historical-data training is a potential fragility as climate shifts; however the model's broad generalization suggests it has "internalized physics" of the atmosphere
ECMWF is building its own AI model inspired by GraphCast, aiming to launch within 1-2 years

Connections: Google Deepmind · Ecmwf · Graphcast · Graph Neural Networks · AI Weather Forecasting