Machine-learning-assisted loss prediction and analysis of photonic lantern MUX/DEMUX for SDM systems

Abstract Photonic lanterns (PLs) are pivotal passive components for mode-division multiplexing (MDM) in space-division multiplexed (SDM) optical interconnects, enabling low-loss, adiabatic coupling between multiple single-mode and multimode fiber. However, their design relies on computationally intensive full-wave electromagnetic solvers, such as the finite-element method (FEM), beam propagation method (BPM), and finite-difference time-domain (FDTD), which hinder efficient multi-objective optimization across high-dimensional parameter spaces. To overcome this bottleneck, we propose an XGBoost-based supervised surrogate modeling framework for simultaneous prediction of insertion loss (IL), mode-dependent loss (MDL), and polarization-dependent loss (PDL), trained on a physics-validated 2,000-sample FEM dataset spanning seven multicore geometries across the full C-band (1530–1590 nm). The surrogate achieves $$R^2> 0.98$$ , $$\\textrm{RMSE} < 0.13\\,\\textrm{dB}$$ , and $$\\textrm{MAE} < 0.075\\,\\textrm{dB}$$ across all three targets, reducing per-design evaluation time from $$\\sim 6.8\\,\\textrm{h}$$ (FEM/BPM) to $$<1\\,\\textrm{ms}$$ —a speedup of $$\\sim 2.45 \\times 10^{7}$$ —while maintaining sub- $$0.1\\,\\textrm{dB}$$ spectral stability across the full C-band. This capability enables rapid parametric sweeps of PL configurations in under $$10\\,\\textrm{s}$$ , supporting automated multi-objective optimization for MIMO-free MDM intra-data-center links at link lengths below $$2\\,\\textrm{km}$$ .