Time-domain sound field estimation for arbitrary 2-D room shapes using deep operator network

Sound field simulation is widely used in various applications, including analyzing sound propagation, designing acoustic spaces, and generating sound fields in virtual environments. In particular, simulations based on wave equation are known for their high accuracy, as these methods can model physical phenomena such as scattering and interference. However, these methods typically involve solving a large number of simultaneous equations, which results in significant computational costs. In recent years, deep learning-based methods for sound field simulations have been proposed, achieving notable reductions in computation time. However, most previous approaches assume a fixed room geometry. In this study, we propose a data-driven deep learning model capable of estimating sound fields for arbitrary room shapes. By employing deep operator networks to separately process room geometry and receiver position information, the model improves generalization performance across different room configurations. In our experiments, the proposed method is used to estimate 2-D time-domain sound fields in rooms with varying shapes. Work partially supported by Research Institute for Science and Technology of Tokyo Denki University Grant No. Q25DS-14/Japan.