Predicting eigenvalues and eigenmodes in non-rectangular rooms with machine learning techniques

Sammanfattning

Knowing the eigenfrequencies and eigenmodes is of great importance to interior design and a common acoustic engineering problem. Challenges in noise control makes knowing the lower eigenfrequencies particularly important. Analytical solutions exist for simple room shapes while methods such as the finite element method (FEM) provide a more general numerical solution to the problem. This thesis outlines the process of generating a dataset of 2D images represent-ing pseudorandom rooms and calculating the eigenfrequencies and eigenmodes using FEM with COMSOL. A machine learning model is presented based on con- volutional neural networks (CNN) that predicts the first ten eigenfrequencies from an image input of the room’s shape with a normalized surface area. To estimate the mode shapes a plane wave decomposition is presented and evaluated com- paring the resulting sound field of the FEM calculated eigenfrequencies and the eigenfrequencies from the machine learning model. The thesis presents a proof- of-concept which predicts the eigenfrequencies with a resulting error between −4 and +6 percent for 90 % of the test set. Furthermore, the potential for predicting eigenmodes is demonstrated.