An airflow velocity field reconstruction method with sparse or incomplete data using physics-informed neural network

Abstract

Velocity field is critical for creating a healthy and comfortable indoor environment but it is difficult to be measured experimentally. Due to the limitations of measurement instruments and site conditions, the captured data are usually sparse, incomplete, and low-dimensional, which limits further research on indoor velocity field. This study proposes a practical reconstruction and super-resolution method based on physics-informed learning that seamlessly integrates imperfect measured data and physical laws. The ability to reconstruct the airflow velocity field with imperfect data as the training dataset was validated using a three-dimensional and three-component airflow field generated with an axial fan. This method successfully recovered the airflow field from sparse or insufficient data (even when the amount of data was only 5 % of the original vector fields) and incomplete data. This indicates that the proposed method can reconstruct and super-resolve the airflow velocity vector fields from sparse, incomplete, and low-dimensional data.