Reproduction of local turbulent flow characteristics within urban district using supervised machine learning model

It is crucial to know the airflows within the urban district for the improvement of the urban atmospheric environment such as air pollution, thermal comfort, wind load, etc. However, it is not easy to know their spatial distribution in observation. And numerical simulation still requires a high computational cost. Instead, this study attempts to reproduce the local velocity distribution within the urban canopy layer by means of a supervised machine learning model. In order to learn the relationship between urban structures and turbulent velocity distributions, the model is trained on large datasets of turbulent flow simulation data and detailed urban geometries of the Tokyo districts. The simulation cases of urban turbulent airflows have been conducted using lattice Boltzmann equation model, which was selected thanks to its high-computation performance on a massive GPU cluster.

The developed model was applied onto several cities other than Tokyo. Irrespective of the building geometries and arrangement, some general characteristics of the urban airflows have been confirmed; such as, an acceleration of the flows in streets parallel to the main flow, and a down-draft and divergent flows in front of tall buildings. The results were also compared with observation datasets taken directly from the urban districts examined.