Rapid Convergence and Low Data Dependency: Physics-Informed Tri-Module Multiplicative Network for DML-DD Link in Analog RoF Fronthaul

The conventional analog radio-over-fiber (A-RoF) scheme offers high spectral efficiency, which can meet the growing capacity demand in 6G cloud radio access networks (C-RAN). With the trend towards dense small cells, the directly modulated laser with direct detection (DML-DD) link stands out as a cost-effective option. The main challenge in the DML-DD link-based A-RoF fronthaul is the nonlinear chirp-dispersion interaction including composite second-order (CSO) and composite triple beat (CTB) distortion. In this work, we propose and experimentally demonstrate a fast-convergence and low data dependency physics-informed tri-module multiplicative network (TriMNet) for C-band DML-DD link. Neural networks (NN) usually require substantial training set to learn and achieve high performance in a data-driven mode. By introducing the analytical form of the CSO and CTB models, TriMNet can extract nonlinear distortion features more quickly and accurately, offering significantly reduced training data requirements. Compared to the fully connected neural network (FNN), TriMNet reduces training symbols by 50% and training epochs by 62.5%, without extra computational complexity. Furthermore, TriMNet enhances the system capacity, achieving a common public radio interface (CPRI)-equivalent rate of 1.02 Tb/s from 0.9 Tb/s in 10-km transmission and 0.84 Tb/s from 0.72 Tb/s in 15-km transmission at 8% error vector magnitude (EVM) threshold compared to full-tap 3^rd order Volterra-based feed-forward equalizer (VFE).