INDOOR AIRFLOW PREDICTION

It is common to use Computational Fluid Dynamics (CFD) to understand natural ventilation. However, the heavy computation load, time, and detailed information for boundary condition usually limit its application in early design stage. To overcome such limitations of CFD simulation, many studies have adopted machine learning models which can make approximations instantaneously with reasonable accuracy. In this paper, we propose a method using a conditional Generative Adversarial Network (CGAN) based image-to-image translation (pix2pix) to predict the indoor airflow condition. Compared to other traditional machine learning models which usually predict a single value (e.g., averaged indoor air velocity), pix2pix can predict contour plot of indoor airflow for the given building floor plan image, which provides more intuitive feedback to designers. The proposed method was tested to understand indoor air movement caused by wing-walls attached to the windows. The test showed overall prediction accuracy of 94%, taking less than a second to predict. The main contribution of this study is to demonstrate the possibility of using the pix2pix as a proxy model of conventional CFD simulations. Since this model can provide instant ventilation prediction without complex inputs, it can help designers to optimize the ventilation performance in early design stage.

Team: You-Jeong Kim, Manal Anis, Yun Kyu Yi1