The exposure to poor air quality is highly variable within cities and between adjacent neighbourhoods and even adjacent streets, yet environmental monitoring of cities is typically done by use of only a small number of fixed monitoring stations high above street level and at low spatial resolution. Similarly, modelling of urban meteorology is carried out at a large scale, with full details of the urban geometry seen as a roughness feature of the atmospheric boundary layer. Yet high variability can be found in pollution levels within street canyons as well. These need to be modelled at a greater level of accuracy and higher spatial resolution, with more attention to the details of street geometry and local parameters, such as location of traffic lanes and position of trees. Complex geometries of street canyons lead to significant differences in local wind regimes, especially at ground level, and these in turn have a profound effect on the dispersion of locally emitted pollutants within the street and on their rate of transport out of the street. The use of Computational Fluid Dynamics (CFD) models has become well established in engineering and architectural practice for applications related to the indoor environment. For complex case studies of the outdoor environment such as ground level air pollutant dispersion in street canyons and around large urban environments, Atmospheric boundary layer wind tunnels are still more often used. However, there are several disadvantages to wind tunnel modelling, such as the requirement to adhere to similarity criteria, which is problematic when modelling buoyant flows, and the difficulty in obtaining measurements of the full flow field in 3D or even in 2D slices through the flow (Blocken et al, 2011). Numerical modelling of the outdoor environment with CFD is becoming increasingly attractive - as computing power increases - as it can provide detailed information on the flow field, thus avoiding some of these limitations. There are a great deal of careful choices that need to be made in the modelling process, such as in implementing the geometry in the model, generating the mesh, and selecting the appropriate models and turbulence models for the solution.

As Direct Numerical Simulation (DNS) is prohibitively expensive to run at this scale, turbulence is modelled within the CFD simulation using Reynolds-Averaged Navier Stokes (RANS) turbulence models, which capture the main properties of the flow without modelling them explicitly. The Reynolds stress is then approximated by a turbulence closure scheme. One popular choice in CFD modelling is the standard k - &epsilon model, which is cheap and fast to run, is robust and has been well validated against a number of different flow cases. It has been found to have problems in predicting flow separation, and can under-predict turbulent kinetic energy values within street canyons. However, the performance of this model can be improved by the appropriate choice of model parameters, wall functions for roughness parameters, and inlet profiles, based on validation from wind tunnel experiments.

This paper reviews published guidelines for use of CFD modelling in studies of the built environment. It looks at the use of standard turbulence models and discusses which are most appropriate for simulation of urban micro-environments and the various validation schemes. We then present an advanced statistical calibration technique for validation of CFD models, relying on the adaptation of the k - &epsilon model constants following the Bayesian calibration framework of Kennedy and O'Hagan (2001). We use observations of turbulent kinetic energy vertical profiles from laboratory experiments for simple street canyons, to calibrate the CFD model. This process is shown to narrow down the set of parameter values that provide the best match between the CFD model outputs and the observations, whilst quantifying the uncertainty of these parameters and of the turbulent kinetic energy outputs. The procedure allows us also to evaluate the systematic bias in the CFD model itself, which is found to be very small using ANSYS CFX for the simulation. The paper shows that following the Bayesian calibration procedure, satisfactory results can be obtained quite quickly, even with a fast and less precise model. The implications of this process for modelling of more complex urban flows are discussed and recommendations made for air pollution modelling efforts in urban micro-environments.