Generative adversarial networks with physical sound field priors

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Accurate sound field reconstruction is crucial for applications like acoustic simulations and virtual reality.
• Traditional methods often struggle with limited measurements, high frequencies, and extrapolation.
• Generative models offer a new paradigm for incorporating physical priors into acoustic problems.

Purpose of the Study:

• To develop a deep learning-based method for spatiotemporal sound field reconstruction.
• To leverage generative adversarial networks (GANs) and plane wave basis for improved accuracy.
• To evaluate the method's performance against state-of-the-art techniques using established datasets.

Main Methods:

• Utilized a generative adversarial network (GAN) architecture for sound field reconstruction.
• Employed a plane wave basis and learned statistical distributions of acoustic pressure.
• Trained and validated the model on two established acoustic datasets.
• Compared performance against existing state-of-the-art reconstruction methods.

Main Results:

• Achieved improved reconstruction performance in terms of accuracy and energy retention.
• Demonstrated superior performance in the high-frequency range and for extrapolation beyond measurement regions.
• Showcased robustness to varying numbers of measurement positions and configurations without performance degradation.
• Validated the GAN-based approach on diverse acoustic scenarios.

Conclusions:

• The proposed deep learning approach offers a promising solution for sound field reconstruction.
• Generative models provide a physically informed prior, enhancing acoustic problem-solving.
• The method demonstrates significant potential for real-world acoustic applications requiring accurate field reconstruction from sparse data.