Evaluation and compensation of the impact of microphone positional errors on sound field modeling using physics-informed neural networks

To accurately obtain acoustic signals over a wide area, it is generally essential to perform measurements using a large number of microphones. Recently, deep learning-based approaches have been proposed to reduce the number of required measurement points. However, purely data-driven methods often yield estimations that deviate from underlying physical principles and require a large amount of training data. To address these limitations, recent studies have explored sound field estimation using Physics-Informed Neural Networks (PINNs), which incorporate physical laws and single-shot measurement data into the loss function of neural networks. Nevertheless, a major challenge remains: the estimation accuracy of PINNs degrades when there are errors in the assumed positions of measurement microphones. In this study, we aim to improve the sound field estimation accuracy of conventional PINNs by introducing a correction method for such positional errors. Specifically, we investigate how microphone position errors affect the training performance and estimation accuracy of PINNs. To evaluate the effectiveness of the proposed method, we conduct simulation experiments using room impulse responses. [Work partially supported by Research Institute for Science and Technology of Tokyo Denki University Grant No. Q24J-04/Japan.]