Benchmarking a Photonic Neural Network for Optical Signal Equalization

Photonic neural networks (NNs) promise energy-efficient, low-latency equalization for high-speed optical links. This paper numerically benchmarks a hybrid photonic NN integrating a passive photonic reservoir, a Mach-Zehnder interferometer (MZI) mesh implementing a fully connected linear layer, and semiconductor optical amplifiers (SOAs) as activation functions. We evaluate the equalization of optical 16-QAM transmissions over dispersion-shifted fiber (DSF) and standard single-mode fiber (SMF) links. Beyond equalization accuracy, we quantify the linear and nonlinear memory capacities (LMC/NMC), Kernel Rank and Generalization Rank (KR/GR), and standard dynamical benchmarks including Mackey-Glass and the Santa Fe laser time series. These benchmarks establish a reproducible baseline for photonic reservoir-based equalizers and relate task performance to intrinsic capacity measures, informing design trade-offs for practical optical receivers.