Physics-guided residual Kolmogorov-Arnold network equalizer for high-speed dual-polarization coherent optical transmission.

Abstract

We propose a physics-guided residual Kolmogorov-Arnold network (RN-KAN) equalizer for nonlinear impairment compensation in high-speed dual-polarization (DP) 16-QAM coherent transmission. RN-KAN combines a multiple-input multiple-output finite-impulse-response (MIMO-FIR) backbone with spline-gated residual branches constructed from self-power and cross-polarization interaction features. These branches provide compact representations of self-phase modulation (SPM)- and cross-polarization modulation (XPolM)-related distortions. Experiments over a 60-km standard single-mode fiber (SSMF) link from 119 to 147 GBaud show lower bit-error rates (BERs) than the evaluated third-order Volterra nonlinear equalizer (VNLE), fully connected deep neural network (FC-DNN), and one-dimensional convolutional neural network (1D-CNN) equalizers. With 1,100 trainable parameters and 178.25 real multiplications per recovered bit (RMPB), RN-KAN has the fewest parameters and the lowest multiplication complexity among the evaluated nonlinear equalizers.