Dr Matthew Coburn - University of Southampton

This project has two aims. The first is to use automation to reduce the time taken to conduct high-fidelity modelling of chemical plumes. The second is to reduce inaccuracies from inflow conditions required to maintain accuracy, by using machine learning.

In order to correctly model dispersion, particularly near a pollution source, a good prediction of the flow field is extremely important. Both convection and diffusion processes of the pollutants should be simulated accurately. In urban areas, convection and diffusion are driven by large and small turbulent eddies, respectively. Computational fluid dynamics (CFD), specifically, large eddy simulation (LES), will be used to resolve the energetic eddies generated within the urban environment. This project will make use of Palm4U, an unsteady LES code designed for simulation of urban atmospheric boundary layers.

It is possible to run simulations at near real time with small grids and suitable computing power. To drastically reduce the time required, we will create a tool to automate the entire setup process. This will include downloading terrain and building data, creating the mesh file, setting up boundary conditions and generating the required input files. Finally, we will use machine learning, to reduce the uncertainty of the meteorological wind conditions, due to the lack of data or inaccurate data of upstream wind direction. This could significantly reduce the number of simulations required. Addressing the uncertainty in the upstream boundary conditions is very new and typically used with steady Reynolds-averaged Navier-Stokes (RANS), but not LES (Garcia et. al. (2014)). Automated setup for LES has not been shown to be used anywhere, especially at such high fidelity.