Deep learning-based Phase Retrieval Scheme for Minimum Phase Signal Recovery

We propose a deep learning-based phase retrieval scheme to recover the phase of a minimum-phase signal after single-photodiode direct-detection. We show that, by properly generating the training data for the deep learning model, the proposed scheme can jointly perform full-field recovery and compensate for propagation-related linear and nonlinear impairments. Simulation results in relevant transmission system settings show that the proposed scheme relaxes the carrier-to-signal power ratio (CSPR) requirements by 2.8-dB and achieves 1.8-dB better receiver sensitivity while being on average 6 times computationally faster than the conventional 4-fold upsampled Kramers-Kronig receiver aided with digital-back-propagation.

Automated summary

In this paper, we propose a deep learning-based phase retrieval scheme that can perform full-field recovery and compensate for propagation-related impairments. Simulation results show that this scheme can relax the requirements for carrier-to-signal power ratio and improve receiver sensitivity.