Reproduce “HergNet: a Fast Neural Surrogate Model for Sound Field Predictions via Superposition of Plane Waves” on OpenResearch

About this paper

Matteo Calafà, Yuanxin Xia, Cheol-Ho Jeong

Researchers at the Technical University of Denmark developed HergNet, a neural surrogate model for sound field predictions that utilizes a Herglotz wave function representation to inherently satisfy the Helmholtz equation. The model accurately predicts sound pressure levels and phases, particularly at mid-to-high frequencies up to 6000 Hz, demonstrating a training time of 124 seconds for complex room acoustics, which is a notable improvement over conventional numerical methods and classical physics-informed neural networks.

Key insights

• Embedding the Helmholtz equation directly through the Herglotz representation ensures intrinsic physical validity of solutions, offering a 'hard constraint' approach superior to penalty-based PDE enforcement.
• Restricting the learning process to boundary conditions via a boundary-only loss function substantially enhances computational efficiency and training speed compared to volume-based methods.
• The model demonstrates strong stability and fast convergence, especially in the challenging mid-to-high frequency range, where traditional solvers are most computationally demanding.

Summary by alphaXiv.