Physics-based generative adversarial network for real-time acoustic holography.

Area of Science:

• Acoustics
• Holography
• Deep Learning

Background:

• Acoustic holography (AH) encodes complex acoustic fields into 2D holograms.
• Phase-only holography (POH) offers superior information volume and storage efficiency.
• Existing POH algorithms lack high fidelity and real-time performance.

Purpose of the Study:

• To develop a deep learning algorithm for accurate and real-time POH calculation.
• To address limitations of current POH methods using a physics-informed approach.

Main Methods:

• A deep learning algorithm combining the Angular Spectrum Method (ASM) with a generative adversarial network (GAN) termed soft-GAN.
• Implementation of a Y-Net structure with dual decoders for frequency and spatial domains to handle high-frequency features.

Main Results:

• Achieved a state-of-the-art Peak Signal-to-Noise Ratio (PSNR) of 24.05 dB.
• Demonstrated accurate and real-time acoustic hologram reconstruction.
• The proposed method shows significant potential for practical applications.

Conclusions:

• The physics-informed deep learning approach effectively learns the inverse mapping for POH generation.
• The soft-GAN and Y-Net structure overcome limitations in fidelity and speed for POH.
• This method paves the way for advanced dynamic acoustic manipulation.