A Deep-Learning Approach for Turbulence Correction in Free Space Optical Communication with Laguerre-Gaussian Modes

Abstract

Free space optical communication has become increasingly popular in the past decade due to its terahertz bandwidth, unlicensed spectrum, and enhanced security features. This technology has been attracting significant interest from the network community due to its broad applications in future military and civilian protocols. Recently, the use of orthogonal spatial states of the light field, including those with nonzero orbital angular momentum, has increased communication system capacities and bit-transfer rates per photon. Here we introduce a machine-learning approach to address the impact of atmospheric turbulence on spatial multiplexing of Laguerre-Gaussian structured light modes under the adaptive optics family of correction methods. A new deep-learning architecture, named cGULnet, is proposed. It is a U-shaped architecture trained adversarially to a patch discriminator similar to conditional generative adversarial neural-networks. The network is trained for exact phase recovery systems without approximating Zernike polynomials, enabling its use in highly phase-sensitive but high-speed spatial multiplexed communication systems. The proposed architecture can predict the successive three frames in a single shot using the previous ten frames. It reduces the bit error rate to the International Telecommunication Union recommended forward error correction limit of 3.8 × 10−3 under high turbulence conditions. The results were achieved for a wide range of beam waist sizes in a simulated Laguerre-Gaussian spatial mode multiplexed 1 km free space optical communication link and at 20 dB optical signal-to-noise ratio, with turbulence conditions parameterized by the refractive index structure parameter of Cn2 ≤ 5 × 10−15 m−2/3.

Keywords: adaptive optics, Deep-learning, CGAN, LSTM, Laguerre-Gaussian Modes, Free space optical communication