Séminaire

We propose to analyze how new types of data, those provided by weather satellites, have engendered radical changes in numerical weather prediction that constitutes the central tool of modern meteorology (Nebeker, 1995). Indeed, until the end of the 80’s there was a dominant design in what is called “data assimilation” techniques whose goal is to determine as accurately as possible the initial conditions on which a forecast is built. Data assimilation was based on a statistical technique called “optimal interpolation” (OI) which “became the operational analysis scheme of choice during the 1980’s and early 1990’s” (kalnay, 2003). Its principle was to estimate the parameters of the forecasting model at every gridpoint by relying on the observations available in its neighborhood. This techniques works with the traditional, synoptic and direct measurements (typically balloons to measure the temperature from the ground to the top of the atmosphere) used by meteorologist. However, in the eighties they were confronted to several weaknesses of OI, particularly 1) its inability to handle uncertainty dynamically and 2) its difficulty to integrate data provided by space satellites. Indeed from 1981 to 1991 scientific publications demonstrates a negligible or negative impact of space data on the accuracy of weather forecasts. This situation was a nightmare for meteorologist since satellite constitutes unquestionably the only solution to reach the global coverage needed to improve weather prediction. This paradox, a tool providing global data but without impact on weather forecast, leads two of the major research and operational meteorological center, the European Center for Medium-range Weather Forecast (ECMWF) and Meteo-France, to undertake a major project, called IFS/ARPEGE, to develop and implement a new assimilation technique : 4D variational assimilation (or 4D-VAR). In 4D-VAR the statistical techniques underlying OI were replaced by the mathematics of optimal control (Lions, 1971; Le Dimet & Talagrand, 1986) that allow to optimize globally the trajectory of the model over a time windows in order to minimize the distance (least-square) between the model and the available observations. However, as we will discuss, this was a ten-year effort since it was fundamentally a paradigmatic shift in data assimilation that suppose, for example, to completely recode weather prediction model in order to accommodate 4D-VAR techniques (particularly what is called adjoint equations).

mpllmd@lmd.jussieu.fr