PIGNN-CFD: A physics-informed graph neural network for rapid predicting urban wind field defined on unstructured mesh

Urban wind field plays an important role in quantitative assessment of urban environment. Compared to field measurement and wind tunnel experiment, Computational Fluid Dynamics (CFD) simulation having the ad-vantages of low cost, repeatability and reliable precision is becoming a common scheme to model flow field of one fixed urban scenario, but still faces the problems of time-consuming computation and lack of scalability for practical engineering application. This paper proposes PIGNN-CFD, a novel physics-informed graph neural network for rapid predicting urban wind field based on irregular unstructured mesh data of CFD simulation. Specifically, a CFD model employing the unsteady Reynolds-Averaged Navier-Stokes (RANS) equations with the standard k-epsilon turbulence model, is constructed and then numerically solved by OpenFOAM to simulate urban wind field defined on unstructured mesh. After being validated by publicly available wind tunnel test data provided by the Architectural Institute of Japan (AIJ), the proposed CFD model is employed to build the training and test sample sets of urban wind fields by simulating the wind blowing through various randomly generated small-scale urban scenes. A novel physics-informed graph neural network, both approximating the training data and automatically satisfying the RANS equations, is designed and trained to perform wind field inference on unstructured mesh graph, and then scaled up to predict wind fields of arbitrary large-scale urban scenes. The predicted wind field results of two urban environments at different scales show that the well-generalized PIGNN-CFD model runs 1-2 orders of magnitude faster than the CFD model on which it is trained, while obtaining the consistent computational accuracy.