Reduced Order Models for 3D Wind Fields across Urban Areas

Abstract

The accurate prediction of three-dimensional wind velocity fields in urban environments is essential for settings such as airborne contaminant dispersion, where rapid response is critical for public safety. Computational fluid dynamics simulations, based on the Reynolds-Averaged Navier-Stokes equations can provide the required accuracy, but their computational cost makes them impractical for real-time or multi-query scenarios. Reduced-Order Models can fulfil this need, by preserving the dominant patterns of the wind flow, while requiring evaluation times orders of magnitude smaller than those of the simulation models.

This thesis develops and compares three non-intrusive ROMs for an urban wind flow, parametrised by the inlet velocity magnitude. The ANSYS Fluent solver was used to produce their training data, using an unstructured polyhedral mesh of 434,388 cells and the k-omega turbulence model. The methods investigated are Proper Orthogonal Decomposition with Interpolation (PODI), serving as the linear baseline and an Autoencoder, both paired with a Radial Basis Function interpolator. The last model was the Neural Implicit Flow network, a mesh-agnostic hypernetwork which maps the spatial coordinates and parameter space directly to the velocity vector. PODI produced the most accurate predictions across the test set in the shortest amount of time, while the Autoencoder followed closely in both regards and NIF maintained a consistent error distribution across the parameter range. The neural network methods however, predicted the steep velocity gradients behind buildings more accurately on a localised level, indicating better suitability when the underlying CFD model produces a finer turbulence resolution.