Atmospheric pollution models are central to understand and predict local high concentration episodes that may occur in a defined area (e.g. cities, sensitive areas like industrial sites, etc.). These models generally involve complex, heterogeneous and multi-scales dynamics and physics, leading to an important computational time.
In view of fastening the simulations, performing predictions and estimations, etc. it is relevant to construct surrogate dynamical models that can be use quickly and easily in place of complex and expensive simulators. Typical applications are the observer design, parametric simulation and
potentially feedback control design.
The purpose of the thesis is to investigate different model approximation and identification methods based on sparse time-domain data. More specifically, the candidate will run different simulations with different excitations on the considered atmospheric model. Then, on the basis of these data, different model structure will be investigated: linear, bilinear, quadratic.
The considered data will be based on the simulations outputs of a meso-scale meteorological research model designed and performed by the candidate or previous works and based on observations of pollutants dispersion.
An illustration is given in the below frames presenting a plume dispersion obtained with a high fidelity LES simulator (left frames) and with a low complexity linear and quadratic model (right frames). More specifically, snapshots of a pollutant dispersion configuration obtained with a highly consuming LES simulation (left) and with a reduced quadratic model (right). It is interesting to notice that the low complexity model catches well the plume trend.