Abstract A novel hybrid photonic neural network (PNN) equalizer for coherent hybrid photonic neural network (PNN) equalizer (WDM) optical communication systems is introduced and tested via differentiable simulations. The architecture features a physical 4 × 4 and extendable 16 × 16 Clements interferometer mesh responsible for broadband linear MIMO equalization and per-channel nonlinear microring resonator neurons possessing a Kerr-like corrective phase response. In addition to effectively train the PNN, an end-to-end differentiable WDM fiber channel simulator is provided that encompassing all major physical impairments like chromatic dispersion, polarization-mode dispersion, self-phase modulation, cross-phase modulation, four-wave mixing, and ASE noise. Through physics-informed optimization, the hybrid PNN learns linear and nonlinear compensations for the channel response to achieve full optical domain equalization minimizing expected DSP footprint. Reliability is confirmed through simulation results demonstrating large differences in constellation recovery and error vector magnitude (EVM) such that physical PNN-channel equalization should be possible down the line for truly high-baud rate WDM systems.
Deshmukh et al. (Fri,) studied this question.