High-fidelity demodulation of vortex beams through dynamic scattering media using a physically constrained deep...

Area of Science:

• Optics and Photonics
• Machine Learning
• Biomedical Imaging

Background:

• Vortex beams with orbital angular momentum (OAM) are crucial for high-capacity optical applications.
• Dynamic scattering in media like biological tissues disrupts vortex beam wavefronts, hindering applications.
• Conventional methods fail due to decorrelation caused by Brownian motion in dynamic media.

Purpose of the Study:

• To develop a high-fidelity deep-learning demodulation technique for vortex beams scattered in dynamic media.
• To integrate data-driven approaches with physical principles for robust wavefront reconstruction.
• To enable reliable decoding of OAM states in complex, time-varying scattering environments.

Main Methods:

• Utilized experimentally acquired full-field speckle patterns from a dynamic milk scattering system for training.
• Introduced a quantum-limited-fidelity residual network (QLF-ResNet) incorporating a forced L2 normalization layer for energy conservation.
• Applied rotation-based data augmentation for end-to-end model training.

Main Results:

• Achieved an average classification accuracy of 91.25% ± 4.15% for OAM modes l = 1-4.
• Successfully resolved complex coefficients and suppressed crosstalk between OAM modes.
• Analytically rendered donut intensity profiles and helical phase structures of the vortex beams.

Conclusions:

• The proposed deep-learning method, integrating physical priors, offers a robust and interpretable decoding scheme.
• Hard-coding physical principles like energy conservation mitigates artifacts common in purely data-driven models.
• This technique provides a promising solution for optical communication and imaging in dynamic scattering environments.