Résumé : NOWADAYS, a number of studies keep on demonstrating the existence of a strong relation between high concentrations of particulate matter (PM) and the prevalence of human morbidity and mortality. Large particles can be filtered in the nose or in the throat, while fine particles (about10 micrometer) can settle in the bronchi and lungs, leading to more serious consequences. According to Karagulian et al., the major sources of urban air pollution are traffic (25%), combustion and agriculture (22%), domestic fuel burning (20%), natural dust (18%) and industrial activities (15%).As a consequence, the detailed study of dispersion phenomena within the urban canopy becomes a target of great interest. To this end, Computational Fluid Dynamics (CFD) can be successfully employed to predict turbulence and dispersion patterns, accounting for a detailed characterization of the pollutant sources, complex obstacles and atmospheric stability classes.Despite being intrinsically different phenomena, turbulence and dispersion are closely related. It is universally accepted that, to reach accurate prediction of the concentration field, it is necessary to properly reproduce the turbulence one. For this reason, the present PhD thesis is split into two main Sections: one focused on turbulence modelling and the subsequent, centered on the dispersion modelling.Thanks to its good compromise between accuracy of results and calculation time, Reynolds-averaged Navier-Stokes (RANS) still represents a valid alternative to more resource-demanding methods. However, focusing on the models’ performance in urban studies, Large Eddy Simulation (LES) generally outperforms RANS results, even if the former is at least one order of magnitude more expensive. Stemming from this consideration, the aim of this work is to propose a variety of approaches meant to solve some of the major limitations linked to standard RANS simulation and to further improve its accuracy in disturbed flow fields, without renouncing to its intrinsic feasibility. The proposed models are suitable for the urban context, being capable of automatically switching from a formulation proper for undisturbed flow fields to one suitable for disturbed areas. For neutral homogeneous atmospheric boundary layer (ABL), a comprehensive approach is adopted, solving the issue of the erroneous stream-wise gradients affecting the turbulent profiles and able to correctly represent the various roughness elements. Around obstacles, more performing closures are employed. The transition between the two treatments is achieved through the definition of a Building Influence Area (BIA). The finalgoal is to offer more affordable alternatives to LES simulations without sacrificing a good grade of accuracy.Focusing on the dispersion modelling framework, there exists a number of parameters which have to be properly specified. In particular, the definition of the turbulent Schmidt number Sct, expressing the ratio of turbulent viscosity to turbulent mass diffusivity, is imperative. Despite its relevance, the literature does not report a clear guideline on the definition of this quantity. Nevertheless, the importance of Sct with respect to dispersion is undoubted and further demonstrated in the works of different authors. For atmospheric boundary layer flows, typical constant values range between 0.2 and 1.3. As a matter of fact, the local variability of Sct is supported by experimental evidence and by direct numerical simulations (DNS). These observations further suggest that the turbulent Schmidt number should be prescribed as a dynamic variable. Following these observations a variable turbulent Schmidt number formulation is proposed in this work. The latter stems from the same hypothesis of the variable formulation developed by Gorlé et al.. Moreover, the relevant uncertain model parameters are optimized through uncertainty quantification (UQ). This formulation further increased the accuracy of the predictions, and was successfully verified by Di Bernardino et al. However, the turbulent Schmidt number resulting from this formulation is still intrinsically linked to the turbulence model employed, i.e. to the Cμ coefficient. To overcome this constraint, the nature and the dependencies of Sct were further analyzed through correlation studies and employing principal component analysis (PCA) on data obtained through the proposed ABL RANS model. Subsequently, the same data-driven technique was employed based on the high-fidelity outcomes of a delayed Detached Eddy Simulation (dDES) to derive a generalized turbulentSchmidt number formulation. The latter can be employed within a wide range of turbulence models, without limiting its variability.