Physics-Informed Generative Adversarial Networks (GANs) for Fast Prediction of High-Resolution Indoor Air Flow Field.

Computational Fluid Dynamics (CFD) provides high-resolution airflow field prediction, which has been widely utilized for evaluating indoor air flow pattern to support energy-efficient and sustainable building ventilation design. One of the major challenges of using CFD is its high computational costs, which intrigues the idea of developing surrogate models for fast prediction of the indoor air flow field. Inspired by automatic image generations using Generative Adversarial Networks (GANs), this paper proposes a novel surrogate model that combines a convolutional neural network (CNN) and a physics-informed GAN to predict high-resolution indoor airflow field. Specifically, the CNN predicts a low-resolution airflow field for a given boundary condition, and then a physics-informed GAN enriches the prediction to a high-resolution airflow field. To evaluate performance of the proposed model, a case study is conducted with a simple ventilation case. A dataset is generated by performing 480 airflow simulations with various boundary conditions using fast fluid dynamics running on GPU. The proposed model is then trained and evaluated based on the dataset. It is found that the predictions from the proposed surrogate model are generally in good agreement with the ground truth while achieving 480 times of speedups compared to fast fluid dynamics simulation on GPU. In addition, the performance, applicability, and limitations of the proposed surrogate model are also discussed. [ABSTRACT FROM AUTHOR]