Reconstructing Sound Fields with Physics-Informed Neural Networks: Applications in Real-World Acoustic Environments

Kotsakis R, Lois S, Thoidis I, Vryzas N, Vrysis L, Kalliris G. Reconstructing Sound Fields with Physics-Informed Neural Networks: Applications in Real-World Acoustic Environments. In: Express Paper. Audio Engineering Society; 2025. Paper 327. Available from: https://aes.org/publications/elibrary-page/?id=22878

@inproceedings{Kotsakis2025_22878,
  author = {Kotsakis, Rigas and Lois, Sotiris and Thoidis, Iordanis and Vryzas, Nikolaos and Vrysis, Lazaros and Kalliris, George},
  title = {{Reconstructing Sound Fields with Physics-Informed Neural Networks: Applications in Real-World Acoustic Environments}},
  note = {Paper 327},
  year = {2025},
  month = may,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=22878}
}

TY  - CPAPER
TI  - Reconstructing Sound Fields with Physics-Informed Neural Networks: Applications in Real-World Acoustic Environments
AU  - Kotsakis, Rigas
AU  - Lois, Sotiris
AU  - Thoidis, Iordanis
AU  - Vryzas, Nikolaos
AU  - Vrysis, Lazaros
AU  - Kalliris, George
M1  - Paper 327
PY  - 2025
DA  - 2025/05/12
UR  - https://aes.org/publications/elibrary-page/?id=22878
PB  - Audio Engineering Society
LA  - en
AB  - The reconstruction of sound fields is a critical component in a range of applications, including spatial audio for augmented, virtual, and mixed reality (AR/VR/XR) environments, as well as for optimizing acoustics in physical spaces. Traditional approaches to sound field reconstruction predominantly rely on interpolation techniques, which estimate sound fields based on a limited number of spatial and temporal measurements. However, these methods often struggle with issues of accuracy and realism, particularly in complex and dynamic environments. Recent advancements in deep learning have provided promising alternatives, particularly with the introduction of Physics-Informed Neural Networks (PINNs), which integrate physical laws directly into the model training process. This study aims to explore the application of PINNs for sound field reconstruction, focusing on the challenge of predicting acoustic fields in unmeasured areas. The experimental setup involved the collection of impulse response data from the Promenadikeskus concert hall in Pori, Finland, using various source and receiver positions. The PINN framework is then utilized to simulate the halls acoustic behavior, with parameters incorporated to model sound propagation across different frequencies and source-receiver configurations. Despite challenges arising from computational load, pre-processing strategies were implemented to optimize the model's efficiency. The results demonstrate that PINNs can accurately reconstruct sound fields in complex acoustic environments, offering significant potential for real-time sound field control and immersive audio applications.
ER  -