Room Impulse response reconstruction using physics-constrained neural networks

In this paper, we present a method for reconstructing the sound field in a room using physics-informed neural networks. Our approach employs a limited set of room impulse responses as training data for the network, while also incorporating the fundamental physical principles of sound propagation in space through the use of the wave equation. This allows the network to not only learn the underlying physics of sound propagation but also utilize the nonlinear mapping capabilities of neural networks to adjust for any inhomogeneities in the room and measurement artifacts. Furthermore, the network can determine particle velocity and intensity through the use of autodifferentiation. The results indicate the effectiveness of the approach in terms of reconstruction accuracy and computational efficiency. This work presents a promising approach for sound field reconstruction and has potential for improving the representation of sound fields in various acoustic settings, including rooms and other complex environments, particularly for the synthesis of room impulse responses.