Abstract
The objective of this study is to simplify the process of generating airflow in turbulent boundary layer wind tunnels using machine learning. A total of 55 sets of roughness element configurations, combining spires, barriers, and roughness blocks, were prepared. The airflow characteristics resulting from these configurations were organized into five target parameters for prediction: the gradient of the mean wind speed, the gradient and magnitude of turbulence intensity, and the gradient and magnitude of the turbulence integral scale. The dataset used comprised airflow data collected in a single wind tunnel, created in the past to target ground surface roughness categories I to IV and model scales ranging from 1/100 to 1/500. Random Forest (RF) and Support Vector Regression (SVR) were employed as the machine learning methods. There was no significant difference in prediction accuracy between the two methods, allowing users to choose whichever is more convenient for them. While prediction accuracy decreases with extreme roughness element configurations, the target airflow characteristics were obtained within a ±20% range of the parameters in the training data, making the method practically sufficient for application.