Dispersion-Compensating Method for High-Capacity Fiber-Optic Communication System Using Machine Learning Optimization

Simulation-driven optimization of complex engineering systems increasingly demands intelligent frameworks capable of replacing exhaustive parameter sweeps with rapid, learned predictions. This paper presents a general surrogate optimization framework that couples high-fidelity simulation with a machine learning regression model to predict system performance and select optimal configurations without exhaustive re-simulation. Fiber-optic dispersion compensation is adopted as a representative benchmark: selecting the optimal placement strategy for dispersion-compensating fiber (DCF)-pre-, post-, or symmetrical-across varying system parameters is computationally expensive, making it an ideal testbed for surrogate-assisted design. A structured dataset is constructed from systematic OptiSystem simulation of DCF-compensated single-mode fiber (SMF) links across bit rates of 2.5-10~Gbps and fiber lengths of 80--160~km, with Q-factor and bit error rate (BER) serving as ground-truth labels. A Random Forest (RF) regression model is trained on this dataset to learn the nonlinear mapping from system parameters to performance outcomes. The framework achieves a prediction accuracy of 97.7% while reducing configuration search time by 91.2% relative to exhaustive simulation. Pre-compensation is consistently identified as the optimal DCF scheme, providing data-driven confirmation of this design preference. The proposed framework is domain-agnostic and generalizes to any simulation-intensive configuration task where labeled performance data can be systematically generated.