Deep learning-assisted recognition of space-time wave packets carrying orbital angular momentum in scattering media.

In this paper, we investigate the robustness and classification performance of conventional Laguerre-Gaussian orbital angular momentum (LG-OAM) beams and space-time wave packet orbital angular momentum (ST-OAM) beams in a scattering medium. Experimental interference patterns are analyzed using the proposed VortexNet deep-learning framework, which performs OAM-state classification while quantitatively evaluating image quality through the peak signal-to-noise ratio (PSNR) and the structural similarity index measure (SSIM) under identical scattering conditions. Our results show that LG-OAM beams degrade significantly during propagation, with low image quality (PSNR ≈ 15 dB, SSIM ≈ 0.26). In contrast. the ST-OAM beams preserve their structural integrity, achieving higher fidelity (PSNR ≈ 21-25 dB, SSIM ≈ 0.75), corresponding to an improvement of ∼6-10 dB in PSNR. This enhanced physical robustness leads to improved learning performance, with VortexNet reaching ∼99-100% accuracy for ST-OAM with fast convergence (∼5 epochs), compared with ∼85% validation accuracy and ∼25 epochs required for LG-OAM. The results demonstrate that ST-OAM significantly enhances resilience against scattering, improving both signal fidelity and deep-learning-based classification performance. These findings highlight ST-OAM as a promising platform for robust optical communication and imaging in complex dynamic environments.