Do numerical weather forecasting models learn weather patterns?

Numerical weather forecasting models are first order differential systems that need to be fed with initial conditions.  Those initial conditions are derived from observations provided by all available meteorological systems, e.g. ground stations, buoys, balloons, weather radars, aircraft, satellites. Two major techniques have been implemented in the last decades in order to address this challenge. The former, called 4DVar, proceeds through the retro-propagation of a gradient, and can be compared with supervised learning in perceptrons. The latter, called "nudging", shares striking similarities with non supervised learning algorithms implemented in Kohonen networks.

Both relationships will briefly be presented, as well as an operational system based on Kohonen network, aimed at assessing the chaotic character of so called "ensemble weather predictions".

About the Speaker : Dr. Jacques Ambühl is Head of research (prognostics division) MeteoSwiss, Federal Office of Climatology and Meterology, Switzerland.