Data-Driven Physics-Informed Neural Network for Sound Field Estimation in Rooms of Arbitrary Size
Gen Sato, Yusuke Ikeda · 2024
In recent years, many methods for simulating acoustic fields using deep learning have been proposed. In particular, neural networks using physical knowledge (PINN: Physics-Informed Neural Network) have been proposed for the purpose of solving various acoustic problems with higher accuracy. Many of the proposed PINN-based methods have achieved the estimation of a specific sound field. In this study, we propose a data-driven PINN for the estimation of simulated sound fields in rooms of arbitrary size with a single model. By using the coordinates and room size information as the input of PINN, the proposed method demonstrated a notable improvement in estimation accuracy by approximately 5.4 dB compared to the simple data-driven model.