Complex-valued physics-informed neural networks for sound field estimation

Paul VS, Hahn N, Nelson P. Complex-valued physics-informed neural networks for sound field estimation. In: 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio. Audio Engineering Society; 2025. Paper 11. Available from: https://aes.org/publications/elibrary-page/?id=23000

@inproceedings{Paul2025_23000,
  author = {Paul, Vlad-Stefan and Hahn, Nara and Nelson, Philip},
  title = {{Complex-valued physics-informed neural networks for sound field estimation}},
  booktitle = {2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio},
  note = {Paper 11},
  year = {2025},
  month = sep,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=23000}
}

TY  - CPAPER
TI  - Complex-valued physics-informed neural networks for sound field estimation
AU  - Paul, Vlad-Stefan
AU  - Hahn, Nara
AU  - Nelson, Philip
T2  - 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio
M1  - Paper 11
PY  - 2025
DA  - 2025/09/02
UR  - https://aes.org/publications/elibrary-page/?id=23000
PB  - Audio Engineering Society
LA  - en
AB  - Sound field reconstruction from sparse measurements is a challenging problem in acoustics, with applications in room acoustics, spatial audio capture, and active noise control. In this work, a complex-valued physics-informed neural network (PINN) is proposed for reconstructing sound fields from data collected by an open-sphere microphone array at control points distributed on a 2D grid. Unlike conventional PINNs that use only spatial coordinates as input, our model additionally incorporates measured pressure values, enabling it to learn the mapping between observed fields and their spatial distributions. In order to investigate the impact of the physical constraints on the model's performance, the network is trained on sound fields generated from superposed single-frequency plane waves and its performance is compared to that of a complex-valued multilayer perceptron (MLP). Results show that the complex-valued PINN yields smoother training curves and extended regions of low reconstruction error, highlighting its enhanced spatial consistency. Future work will focus on investigating the models generalization to unseen sound fields and varying control point grids, as well as optimizing the placement and number of training points.
ER  -