Machine Learning-Driven Multi-Objective Optimization for Weather-Resilient Resource Allocation in Ultra-Dense WDM-FSO/xPON Systems

Ultra-Dense Wavelength Division Multiplexing (UDWDM) Free Space Optical/Fiber-to-the-× (FSO/FTTx) hybrid networks suffer from drastic changes in communication system performance due to dynamic atmospheric changes, thus affecting the dependability of 5 G/6 G communication systems. In this paper, we propose a novel machine learning-driven multi-objective optimization framework to realize weather-resilient resource allocation. Our approach leveraged two-phase co-design: first, by ensembling pre-trained Extreme Gradient Boosting (XGBoost) and Random Forest (RF) models, to reduce the prediction variance, serving as high-fidelity surrogates for critical cross-layer performance metrics bit error rate (BER), optical signal-to-noise ratio (OSNR), and quality factor (QF), yielding exceptional accuracy; second, the obtained surrogates were integrated within an Non-dominated Sorting Genetic Algorithm II (NSGA-II) optimizer by replacing computationally intensive simulations, such that the multi-objective problem can be efficiently solved in this stage. In this stage, we perform simultaneous minimization of BER while maximizing OSNR and QF, to establish trade-offs among reliability, signal integrity, and noise resilience. This approach yields Pareto-optimal solutions, enabling dynamic power control along with modulation switching over various atmospheric conditions and link lengths, 0.5–12.5 km. The validation confirmed resiliency, while maintaining BER ≤ 3.21 × 10−6, QF stabilized at 24–30.61 dB, OSNR sustained at 48.23–57.69 dB.