Multimodal sensing of underwater optical turbulence using beams carrying orbital angular momentum, part 2: deep learning classification

We present a convolutional neural network (CNN) framework for classifying underwater optical turbulence strength from multimodal measurements of laser beams carrying orbital angular momentum propagated through a Rayleigh–Bénard convection tank. Three modalities were evaluated for four beam types ( ℓ ∈{−2,0,2,4}): pupil-plane intensity achieved the highest accuracy (96.1%–98.1%), followed by focal-plane (85.0%–90.0%), mode sorter (78.5%–79.3%), and Gaussian beam phase (79.3%). Classification accuracy was comparable across beam topological charges, consistent with findings from our companion paper which analyzed traditional optical turbulence quantification statistics such as scintillation, beam wander, and mode sorter crosstalk. Dual-input fusion architectures that paired the pupil-plane with a second modality did not improve over pupil-only performance. Unlike classical metrics that require multiple frames, the CNN classifies turbulence from individual frames, offering a path toward real-time sensing.