AI Weather Model Outperforms Traditional Forecasts for Extreme Events
Scientists have developed an AI-powered weather model that can predict certain extreme weather events, like major storms, several days in advance, more accurately than traditional forecasting systems. This new machine learning approach also delivers these forecasts significantly faster, using much less computing power. It marks a promising step forward in preparing for severe weather.
More accurate and faster warnings for extreme weather could give communities more time to prepare, potentially saving lives and reducing damage.
Numerical weather prediction relies on solving physical equations on supercomputers. A neural model trained on decades of reanalysis data learned to forecast atmospheric evolution directly, producing 10-day forecasts in under a minute.
In head-to-head evaluation the system matched or beat an established physics-based model on many variables, including the tracks of tropical cyclones — information with direct value for disaster preparedness.
Because it is so cheap to run, the model can generate large ensembles that better capture the probability of rare, high-impact events.
🔬 Read the technical version
Neural weather model outperforms traditional forecasting for some extreme events
A machine-learning forecasting system produced more accurate medium-range predictions than a leading physics-based model for several classes of extreme weather, while running in a fraction of the compute time.