Simulation-based inference for the in situ estimation of the acoustic boundary admittance

Achieving accurate results in interior acoustic simulations relies on precise knowledge of the boundary properties of all interacting surfaces. Typically, the boundary admittance, which fully characterizes the acoustic properties of a surface, is determined under laboratory conditions such as the impedance tube. Yet, this approach has limitations, motivating the exploration of in situ methods to characterize materials in real-world conditions. In this work, we present a Bayesian approach to determine the acoustic boundary admittance in situ based on a limited number of measurement points. The method utilizes simulation-based inference, where a neural network is trained to approximate the posterior probability distributions of the unknown boundary admittances. The core of the approach is a finite element model used to generate sound pressure data, which also acts as the forward model during the inference process. Consequently, this technique is especially well-suited for applications involving pre-existing geometrical models, such as digital twin applications or model updating. By adopting simulation-based inference, we gain advantages over sampling-based Bayesian approaches, as it effectively handles complex and computationally expensive forward models.