Apport des infrasons pour l'assimilation de données dans un modèle global de prévision numérique du temps

Infrasounds are acoustic waves between 0 and 20 Hz emitted by a wide variety of sources. Due to their low frequencies, vertical gradients of winds and temperature lead to the formation of waveguides in which they can propagate over thousands of kilometers with little attenuation. Since this type of propagation is intrinsically linked to the atmospheric circulation, these waves can be used as tracers. The general circulation in the stratosphere and the mesosphere, which forms the middle atmosphere, provides the most efficient acoustic waveguides. It is proposed here to use ground-based infrasound signals to provide new information on the circulation of the middle atmosphere. It can be used to improve the initialization of numerical weather prediction (NWP) models. For an observation to be assimilated, it is necessary to define an observation operator, able to simulate it from model predictions. The aim of this thesis is to examine how to assimilate infrasounds by designing a dedicated observation operator.After presenting the state of knowledge on atmospheric circulation and the principle of data assimilation, we examine the infrasound data of the international monitoring system.This system is maintained by the Comprehensive Nuclear Test Ban Treaty Organization(CTBTO). An observation operator compatible with the needs of assimilation can be designed from a propagation model by taking advantage of fast propagation methods and well-known natural sources such as volcanoes. Given the limitations associated with these methods and sources, it is difficult to design an observation operator allowing the direct assimilation of infrasound data. We propose then to solve the inverse problem by retrieving vertical profiles of winds and temperature consistent with the recorded infrasound signals. Firstly, we have highlighted the need to superimpose a disturbance onto the atmospheric profiles in order to simulate real acoustic signals using ray tracing or modal decomposition methods. Then, within the frame of a simplified atmosphere with two degrees of freedom and simulated signals, we examined the ability of local optimization methods to estimate the atmospheric disturbance minimizing the discrepancies between modelled and observed signals. The results of this study show the difficulty in obtaining an optimal disturbance without prior information on it. This led us to develop a Bayesian approach, taking as prior information an ensemble of analyzes and short-term forecasts of the Météo-France global NWP model. The differences between the simulation results from a ray tracing code and the infrasound measurements are used to compute the likelihood of each member of the ensemble. This method was assessed using infrasound signals emitted by Mount Etna eruptions during May 2016. We show that this method allows to select the most likely members of the ensemble, and thus to provide atmospheric profiles, in agreement with observed infrasounds. These profiles could then be considered as pseudo-observations in a data assimilation system.Cette thèse a pour but d’améliorer la prévision numérique du temps (PNT) dans la moyenne atmosphère (entre 17 km et 100 km) via l’apport d’informations déduites de la propagation d’ondes infrasonores. Ces ondes peuvent être vues comme des traceursde l’état de la moyenne atmosphère. On se propose d’exploiter les bulletins de détections du réseau global d’antennes acoustiques de l’Organisation du Traité d’Interdiction Complètedes Essais nucléaires pour fournir des informations sur la température et les vents. Pour résoudre ce problème inverse, des méthodes d’optimisation locale ont d’abord été testées. Puis, il a été développé une approche bayésienne pour sélectionner les états de l’atmosphère les plus vraisemblables prévus par un modèle de PNT. Cette approche a été appliquée sur une base de détections issues de l’activité volcanique du Mont Etna en mai 2016. La thèse montre qu’à partir des détections infrasonores, il est possible d’identifier, au sein d’un ensemble de prévisions, des états atmosphériques réalistes qui soient en accord avecles détections. Il peut alors être déduit des profils de températures et de vents dans la moyenne atmosphère à assimiler dans les modèles de PNT