Pilot-assisted optical extreme learning architecture for turbulence-resilient OAM demultiplexing in free-space optical links

This study introduces a new optical machine learning framework for the demodulation of orbital angular momentum (OAM) spatially multiplexed beams affected by atmospheric turbulence in free-space optical communication systems. The proposed architecture employs a pilot-assisted optical extreme learning approach, in which the interference between the data and pilot beams acts as a high-dimensional kernel transformation, enabling turbulence-resilient demultiplexing. Unlike conventional deep-learning-based receivers that rely on computationally intensive electronic processing, the proposed system offloads most of the nonlinear mapping to passive optical propagation, leaving only a simple linear regression layer in the electronic domain. This results in ultra-low latency and energy efficient turbulence correction and symbol recovery. Simulation results show that the system achieves symbol accuracies up to 0.99 under both weak and moderate turbulence and 0.76 under strong turbulence when four OAM modes are transmitted simultaneously. Even when the number of transmitted modes increases to nine, the system maintains a symbol accuracy of 0.55 (considering all possible weather conditions), demonstrating its potential for high-capacity, low-latency optical communication links and for hybrid optical–electronic learning architectures.