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

Vortex beams carrying orbital angular momentum enable high-capacity optical communication and imaging, yet multiple scattering in dynamic media such as biological tissues disrupts their wavefront. Brownian motion decorrelates the scattered field and invalidates conventional methods. To address this, we propose a high-fidelity deep-learning demodulation that integrates data-driven and physical priors. Specifically, we directly utilize experimentally acquired full-field speckle patterns from a dynamic aqueous milk scattering system as the training dataset. We introduce an improved quantum-limited-fidelity residual network, QLF-ResNet, that uses a forced L2 normalization layer at the output to hard-code energy conservation and applies rotation-based data augmentation for end-to-end training. Experiments achieve an average classification accuracy of [91.25]% ± [4.15]% for OAM modes l = 1-4. The model resolves complex coefficients, suppresses crosstalk, and analytically renders the donut intensity and helical phase. By hard-coding physical priors, our method avoids artifacts common in pure data-driven models and offers a robust, interpretable decoding scheme for complex time-varying scattering environments.