Development of Machine Learning-Aided Rapid CFD Prediction for Optimal Urban Wind Environment Design

Abstract

This paper presents a Machine Learning (ML) model based on Computational Fluid Dynamics (CFD), developed to quickly and accurately predict the impact of buildings on the urban wind environment. While CFD simulations are effective for wind studies, such as analyzing wind loads, pedestrian comfort, and pollution dispersion, they require significant computational resources and time. Recently, Machine Learning has demonstrated strong potential in providing accurate and immediate predictions by learning from datasets. By training on CFD-generated data, the ML model can quickly produce accurate and physically consistent results, addressing the limitations of CFD methods. The Reynolds-Averaged Navier-Stokes (RANS) turbulence model was chosen for CFD simulations, which were validated against experimental data, with mesh sensitivity analyzed at a wind speed of 3 m/s. A dataset of 300 cases, involving 100 hypothetical buildings and three wind speeds (3, 4, and 5 m/s), was generated to train the ML model. A multi-output regression model was proposed to effectively predict key parameters—wind velocity, turbulence intensity, and CO₂ mass fraction—in the selected urban domain. The Random Forest algorithm, which best represented the CFD results, was selected for model development. The ML model demonstrated high efficiency on new data, achieving 88-96% accuracy. This work offers a fast and precise prediction tool, valuable for urban design and related applications.