Differentiable physics for sound field reconstruction.

Area of Science:

• Acoustics
• Computational Physics
• Machine Learning

Background:

• Sound field reconstruction estimates sound fields from limited spatial observations.
• Conventional methods like physics-informed neural networks (PINNs) incorporate physics into the loss function.
• Severe undersampling poses challenges for accurate sound field reconstruction.

Purpose of the Study:

• To introduce a differentiable physics approach for robust sound field reconstruction.
• To enhance accuracy and convergence under data-scarce conditions.
• To enforce physics as a strong constraint during network training.

Main Methods:

• Approximating wave equation initial conditions with a neural network.
• Utilizing a differentiable numerical solver for the differential operator.
• Incorporating a sparsity-promoting constraint for improved reconstruction.

Main Results:

• The proposed method achieves stable network training by enforcing physics as a strong constraint.
• Demonstrated successful sound field reconstruction under extreme data scarcity.
• Outperformed conventional physics-informed neural networks in accuracy and convergence.

Conclusions:

• The differentiable physics approach offers a stable and effective method for sound field reconstruction.
• This technique significantly improves performance in undersampled scenarios.
• It provides a strong alternative to existing physics-informed neural network methods.