Estimation of time-domain scattered sound fields using physical models and deep learning

Sound field measurement helps us to understand the sound field, which is difficult to understand only by hearing. In particular, in order to accurately understand acoustic scattering, in which sound reflects in irregular direction, it is necessary to require a large number of microphones. Recently, to reduce the measurement points, methods for modeling sound fields, such as the equivalent source method, have been proposed, and methods for estimating sound fields using physics-informed neural networks (PINNs) based on physical models have attracted much attention. We previously proposed a frequency-domain method assuming smooth specific acoustic impedance, which improved estimation accuracy. However, its effectiveness in the time domain has not yet been investigated. In this study, we propose a method for estimating scattered sound fields in the time domain from a small number of microphones, by using deep learning based on physical models. From the simulation experiments, we evaluate the estimation accuracy of the proposed method under various conditions, such as different number of microphones and object shapes, in a 3-D scattered sound field. Work partially supported by Research Institute for Science and Technology of Tokyo Denki University Grant No. Q25DS-17/Japan.