A U-Net based surrogate model for fast prediction of indoor airflow under variable geometries

Abstract Machine learning (ML) based surrogate models have been widely applied to accelerate the simulation process of indoor airflow distribution. However, current ML methods only perform well under indoor geometries which have appeared in the training set. The lack of spatial generalization ability limits its applications in scenarios requiring abundant and diverse simulations, such as the energy-saving design of indoor airflow. To address this challenge, this study proposes a U-Net based framework aims to maintain high precision under variant indoor geometries. A grid-based feature transformation approach is proposed to transform indoor geometries adaptively. A domain customized U-Net neural network is proposed to effectively capture spatial features and predict target airflow distribution. Evaluations are made on a dataset contains 3000 indoor airflow velocity fields with various air inlet conditions. Results show that the proposed surrogate model has higher accuracy than conventional ML methods, the average R 2 is 0.973 on unseen indoor geometries, while the baseline model’s average R 2 is 0.216 in this situation. The proposed method proves the spatial generalization ability of U-Net architecture in this domain, and this method can be further applied in many downstream tasks, such as energy-saving design and real-time control of indoor environment.