Speaker
Description
Despite their computational efficiency, Reynolds-Averaged Navier-Stokes (RANS) models often struggle to accurately represent the complex turbulence and flow separation typical of urban environments.
These limitations highlight the need for data-driven correction strategies to improve predictive accuracy.
Enhancing RANS performance is essential for aerodynamic load estimation and for evaluating pedestrian comfort and pollutant dispersion, critical aspects of urban planning and environmental design.
This work presents a machine learning framework for analyzing urban flows, grounded in a prior
data assimilation phase. The study focuses on the Architectural Institute of Japan (AIJ) [1] Case A
dataset, which models flow around a single rectangular building at varying heights (see Fig. 1) under
controlled inlet conditions reproducing the atmospheric boundary layer. Experimental measurements of
velocity and turbulent kinetic energy are available for validation. Baseline steady RANS simulations are
performed in OpenFOAM using the k − ε turbulence model, incorporating the atmospheric boundary
layer profile. The results of this model show discrepancies with experimental data (see Fig. 1).
The same procedure previously applied to rectangular cylinders [2] will be used herein for this
more complex three dimensional case. To perform the data assimilation, the DAFoam library, written
for OpenFOAM, is employed. In this phase field variables extracted from scale-resolving simulations
are utilized to compute correction fields to enhance the accuracy of the k − ε baseline model. More
specifically, an objective functional is chosen to minimize the discrepancy in both the velocity field
within a region around the body and the pressure field on the body’s surface. The corrective fields
obtained for different building heights are used to train a machine learning regression model to predict
the correction terms of the RANS baseline model for unseen configurations, for instance varying the
building heigth or wind intensity and direction.