Identification of physical properties in acoustic tubes using physics-informed neural networks

Abstract

Physics-informed Neural Networks (PINNs) is a method for numerical simulation by introducing a loss function with respect to the governing equations into a neural network. Although PINNs has been studied for its usefulness in the field of inverse analysis, but there are few examples of PINNs applied to acoustic analysis. In this study, we report a method for identifying loss parameters in acoustic tubes using PINNs. We set the energy loss parameters within the acoustic tube as trainable parameters of the neural network. The problem of identifying the loss parameters was then formulated as an optimization problem for the neural network, and the physical properties were identified. The neural network structure used in this process was based on our previously proposed ResoNet, which is a PINNs to analyze acoustic resonance. The validity of the proposed method is evaluated by forward and inverse analysis (identification of loss parameters). The results showed that when the parameters to be identified have multiple local solutions, the parameters converge to one of these solutions, and not necessarily to the true value. However, in problem settings where there are no local solutions, the parameters can be identified with high accuracy. This method can be applied to various sound fields by simply changing the governing equations in the loss function, and is expected to have a wide range of applications.