Work

AI-enhanced mmWave propagation modeling for terrestrial links in Mediterranean climates: Validation with real 80 GHz frequency bands for 5 G and 6 G communication links

Year: 2026

Type: article

Abstract: Accurate channel modeling is essential for developing the high-capacity wireless backhaul required for 5 G and emerging 6 G networks. A significant challenge in large-scale link design is the inadequacy of traditional path loss (PL) models. As established in prior literature, these empirical and deterministic models often lack the fidelity to represent the complex, non-linear attenuation introduced by atmospheric and environmental interactions. This challenge is particularly acute in the Mediterranean climate, characterized by diverse topography and significant rain-induced attenuation, which introduces substantial variability in signal propagation. Additionally, developing a prediction model (PL) for diverse wireless propagation environments remains a significant challenge. This paper proposes a framework that supports the broader vision of intelligent and adaptive wireless networks, where AI-driven channel characterization can facilitate dynamic resource management and enhanced network resilience. Two machine-learning-based ensemble methods for E-band path loss prediction have been proposed: the Generalized Ensemble Method (GEM) and its enhanced variant, the Generalized Weighted Ensemble with Internally Tuned Hyperparameters (GEM-ITH), which are validated using real-world measurements from the Mediterranean climate region. Each method was evaluated under two configurations by varying the number of base learners and the hyperparameter search space. A grid-based deterministic search was used to generate diverse hyperparameter combinations for five base learners (XGBoost, LightGBM, Random Forest, SVR, and Gradient Boosting). Using measurements from four active E-band terrestrial links in Libya—spanning a wide range of environments (mountainous, coastal, and urban) and covering distances between 0.77 and 1.55 km—the models were trained on a feature set that combined geometric and environmental parameters. We employed a 4-fold GroupKFold cross-validation, where each fold corresponds to a measurement site. This leave-one-site-out approach ensures that the model is evaluated on unseen sites, prevents data leakage, and provides a reliable estimate of generalization performance. The GEM-ITH optimization framework is implemented using Sequential Least Squares Programming (SLSQP) to determine the optimal ensemble weights that minimize mean squared error. Experimental validation was conducted using real-world measurements. Statistical analysis demonstrates that the ensemble models outperform individual base learners in path loss prediction and that GEM-ITH, with internally tuned hyperparameters, achieves a minimum MSE of 2.11 and a higher R² of 0.9361. The findings demonstrate the potential of data-driven approaches to outperform traditional models, thereby enabling more reliable link-budget calculations, network planning, and performance optimization for future 5 G and 6 G deployments.

Source: Results in Engineering

DOI: 10.1016/j.rineng.2026.112809