Neural network classification of beams carrying orbital angular momentum after propagating through controlled experimentally generated optical turbulence

We generate an alphabet of spatially multiplexed Laguerre–Gaussian beams carrying orbital angular momentum, which are demultiplexed at reception by a convolutional neural network (CNN). In this investigation, a methodology for optimizing alphabet design for best classification rates is proposed, and three 256-symbol alphabets are designed for performance evaluation in optical turbulence. The beams were propagated in three environments: through underwater optical turbulence generated by Rayleigh–Bénard (RB) convection (C2n≅10−11m−2/3${C}_{n}^{2}\cong 1{0}^{-11}\;{\rm m}^{-2/3}$), through a simulated propagation path derived from the Nikishov spectrum (C2n≅10−13m−2/3${C}_{n}^{2}\cong 1{0}^{-13}\; {\rm m}^{-2/3}$), and through optical turbulence from a thermal point source located in a water tank (C2n≅10−10m−2/3${C}_{n}^{2}\cong 1{0}^{-10}\;{\rm m}^{-2/3}$). We report a classification accuracy of 93.1% for the RB environment, 99.99% in simulation, and 48.5% in the point source environment. The project demonstrates that the CNN can classify the complex alphabet symbols in a practical turbulent flow that exhibits strong optical turbulence, provided sufficient training data is available and testing data is representative of the specific environment. We find the most important factor in a high classification accuracy is a diversification in the intensity profiles of the alphabet symbols.