Development of convolutional neural network-based turbulence modelling framework for outdoor wind field simulation

Accurate flow field estimation is pivotal in urban design, playing a crucial role in enhancing the urban microclimate, but computational fluid dynamics (CFD) based on the widely used Reynolds-averaged Navier-Stokes (RANS) method has limitations in this regard. Previous studies have attempted to improve the RANS method by employing machine learning-based turbulence models. However, most studies trained models without considering the spatial structure and correlation of turbulence, which can result in discontinuities or a serrated distribution in predicted flow fields, making the models difficult to apply practically.

This study aimed to develop a turbulence modeling framework that can integrate the spatial characteristics of the flow field to accurately model turbulence in urban flow fields. To achieve this, a convolutional neural network (CNN) model was constructed to incorporate the spatial characteristics of the flow field and establish the mapping between the RANS flow field and the high-fidelity Reynolds stress.

Then, this model was integrated into the CFD solver, supplanting the original turbulence model to produce accurate and continuous flow fields. The generalization capability and accuracy of the proposed framework was initially demonstrated on simplified benchmark configurations, which exhibited satisfactory performance. The validated framework was then applied to three increasingly complex case studies of urban wind environments: an isolated building, a building array, and actual building complexes, to further assess its performance.

The results showed that the framework was shown to be capable of delivering accurate predictions of the velocity field around an isolated building, including predictions of velocity magnitude and recirculation zones. For a more complex configuration of a building array, the proposed framework performed well in regions where the flow properties were covered by the training dataset.

Moreover, the present framework provided a continuous and smooth velocity field distribution in highly complex urban environments, underscoring the robustness of the proposed turbulence modelling framework. The significance of the proposed framework lies in its potential for practical application across diverse urban scenarios, offering efficient and accurate flow field predictions that supports better urban planning and decision-making, ultimately enhancing urban microclimate conditions.