Speaker
Description
Estimating acoustic eigenfrequencies is essential in the design of rooms and buildings, particularly at low frequencies and/or small rooms. While analytical solutions exist for simple geometries, more general shapes and materials require numerical algorithms. In recent years, machine learning approaches for eigenvalue problems have been proposed, but their accuracy and robustness remain limited and unexplored, leaving finite element methods as the preferred standard. Although, a neural network architecture called HergNet has been recently introduced which leverages the plane wave decomposition to achieve superior performance over previous physics-based machine learning approaches. In this work, we present a framework that employs HergNet as an eigenvalue solver, obtaining substantially reduced training times while maintaining high accuracy and robustness. The learning mechanism further allows the search to be restricted to a specified eigenfrequency range. The method is tested against FEM across different geometries and boundary conditions, demonstrating its applicability to a wide range of scenarios.