Weather forecasts have historically relied on physics-based simulations powered by the world’s largest supercomputers. Such methods, called Numerical Weather Prediction models, are constrained by long computational time, and are sensitive to approximations of the physical laws on which they are based.
Deep learning offers a new approach to computing forecasts that stands to revolutionise weather forecasting: rather than incorporating explicit physical laws and attempting to simulate weather in silico, deep learning models learn to predict weather patterns directly from observed data.
Google DeepMind has multiple cutting-edge weather prediction models – covering both short-term (MetNet-3) and medium term (GraphCast) weather forecasts – which are able to generate forecasts faster and with higher accuracy than current industry standards. That means they are able to compute forecasts in as little as just a few seconds, compared with hours, and they can deliver results at both higher temporal and spatial resolutions.
Learn about GraphCast on our blog:
Discover the capabilities of MetNet-3 from GoogleAI:
Subscribe to our channel for more: @googledeepmind
source
