How do we forecast weather?
Weather predictions are made using Numerical Weather Prediction (NWP) models that simulate the atmosphere based on the fundamental physical laws of nature.
NWP models use powerful supercomputers and high-quality data to forecast weather, typically up to a week in advance.
In ensemble forecasting, multiple forecasts are made using slightly different starting conditions to show a range of possible weather outcomes.
How does GenCast perform?
Google’s GenCast uses ensemble forecasting with AI instead of traditional NWP, trained on 40 years of reanalysis data.
It outperforms NWP models (like European Centre for Medium-Range Weather Forecasts (ECMWF) ensemble forecasts) on 97.2% of 1,320 targets evaluated and is better at predicting extreme weather.
GenCast is also more accurate than NWP at predicting weather more than 36 hours in advance, achieving accuracy on 99.8% of targets.
How does GenCast work?
GenCast uses a neural network with 41,162 nodes and 240,000 edges to process data.
It takes noisy weather data, refines it through multiple iterations (30 refinements), and generates a prediction for the next weather moment.
This process is faster than traditional NWP, taking only 8 minutes compared to several hours for NWP.
Will GenCast replace NWP?
GenCast provides probabilistic forecasts (e.g., 25% chance of rain), while NWP gives deterministic forecasts (e.g., exact mm of rain).
Experts argue probabilistic forecasts are better for extreme events and give more lead time for preparation.
GenCast doesn’t replace NWP but complements it, as both use fundamental weather data and rely on laws of physics.
GenCast needs continuous reanalysis data to improve, and NWP models still provide essential training data for AI models like GenCast.
COMMENTS