Two-dimensional acoustic analysis using a physics-informed neural operator

Surrogate modeling using machine learning has recently attracted significant attention as an alternative to conventional numerical simulations. Compared to high-fidelity methods such as the finite element method (FEM) and the finite-difference time-domain (FDTD) method, surrogate models can provide approximate solutions much faster and are expected to enable interactive design and rapid optimization. One such model, the Neural Operator, can simulate various conditions by taking analysis settings such as boundary conditions as input variables. However, data-driven Neural Operators require training datasets generated by traditional solvers, which can be computationally expensive. To overcome this limitation, the Physics-informed Neural Operator (PINO) has been proposed, which incorporates the governing equations into the loss function, eliminating the need for training data. While PINO has been applied to various physical problems, its use in acoustic wave simulations remains limited due to complex phenomena such as reflection, diffraction, and broadband behavior. This study presents a 2-D acoustic simulation using PINO without training data, based on the wave equation. The accuracy of the results is validated through comparison with FDTD simulations.