Measurement points sampling analysis for indoor airflow field reconstruction using physics-informed neural networks

Creating a comfortable and healthy indoor environment is very important for indoor occupants. It is difficult to obtain complete information about the indoor airflow field from experimental measurements only, but traditional computational fluid dynamics methods require a lot of time and complex boundary conditions. With the advent of data-driven methods, the cost of the reconstruction of indoor airflow fields has been greatly reduced. Therefore, a two-dimensional airflow field with only a limited amount of experimental measurement data is reconstructed using a physics-informed neural network (PINN), and the constructed PINN model increases the physical interpretability of the neural network through the Navier–Stokes equations. Subsequently, the influence of the number and spatial location of the experimental measurement points on the accuracy of the PINN model for the reconstruction of the indoor airflow field is discussed, and the influence of different spatial locations of the measurement points on the PINN model for the reconstruction of the indoor airflow field is analyzed in more detail by orthogonal experiments with non-global information and orthogonal experiments with global information. The results show that the indoor airflow field can be reconstructed accurately by using only a limited amount of measurement points to participate in the training of the PINN model, and the measurement points that can reflect the vortex information in the airflow field are more favorable for the reconstruction of the airflow field. Moreover, only orthogonal experiments with global information can effectively infer the optimal spatial location for sampling measurement points in the airflow field.