Non-linear Equalization in 112 Gb/s PONs Using Kolmogorov-Arnold Networks

TL;DR

The paper addresses non-linear distortions in 112 Gb/s PAM4 PONs caused by EAM and SOA nonlinearities, seeking low-complexity nonlinear equalizers. It introduces Kolmogorov-Arnold Networks (KANs) that use trainable 1D activation functions implemented with linear B-splines, and compares them to CNN and FIR equalizers while applying pruning to derive Pareto fronts. On a 2.2 km C-band PON setup, KANs—especially KAN-2—outperform CNNs across equivalent complexities, with KAN-1 remaining competitive with CNN-2 up to ~121 rvms; pruning reduces the gap further. The study demonstrates that KANs offer a hardware-friendly, multiplier-free path to effective nonlinear equalization, enabling efficient upgrades to 100+ Gb/s PON deployments.

Abstract

We investigate Kolmogorov-Arnold networks (KANs) for non-linear equalization of 112 Gb/s PAM4 passive optical networks (PONs). Using pruning and extensive hyperparameter search, we outperform linear equalizers and convolutional neural networks at low computational complexity.

Figures (5)

• Figure 1: Diagram of a single neuron in (a) a multi-layer perceptron (MLP) network and (b) a Kolmogorov-Arnold network (KAN) network, where boxes represent 1D functions.
• Figure 1: Overview of the Search Space
• Figure 2: Experimental setup for a 112 Gb/s (56 GBd, PAM4) PON upstream through a SSFM in the C-band. In the right, the non-linear EAM transfer function and SOA gain function is shown.
• Figure 3: Pareto fronts for unpruned (a) and pruned (b) networks at $\text{ROP}=-2$ dBm and 2.2 km (C-band) considering the maximum over all training sequences of the mean $\overline{\text{BER}}$ taken at the last ten training iterations.
• Figure 4: Mean $\overline{\text{BER}}$ for a fiber length of 2.2 km (C-band) for models taken from the pruned Pareto curve with different complexities of 21 rvms (a), 51 rvms (b), 121 rvms (c) and 321 rvms (d).