Explainable machine learning for path loss modeling in mmWave and Sub-THz communication systems

Machine learning (ML) has emerged as a promising approach for large-scale path loss modeling in millimeter-wave (mmWave) and sub-terahertz (sub-THz) wireless communication systems because of its ability to capture complex propagation behavior beyond conventional empirical models. However, most existing ML-based path loss predictors operate as black-box models, limiting their interpretability, physical reliability, and practical adoption in deployment-critical wireless system design and standardization. To address these limitations, this paper proposes an explainable machine learning (XML) framework for physically interpretable path loss modeling across heterogeneous mmWave and sub-THz propagation environments. The proposed framework integrates physics-guided deep neural network (DNN) regression with SHapley Additive exPlanations, physical consistency verification, and explanation stability analysis to achieve trustworthy and deployment-oriented wireless propagation modeling. Using publicly available multi-band propagation datasets spanning 28–140 GHz, the framework systematically analyzes the influence of key propagation parameters, including transmitter–receiver separation distance, carrier frequency, antenna heights, and line-of-sight (LOS)/non-line-of-sight (NLOS) conditions. In addition to global and local explainability analysis, the proposed framework introduces quantitative explanation stability evaluation using an Explanation Stability Index and validates whether learned feature relationships remain consistent with established wireless propagation principles. Comparative experiments against conventional empirical models and baseline ML approaches, including the close-in, alpha–beta–gamma, support vector regression, random forest, and conventional black-box DNN models, demonstrate that the proposed XML framework achieves competitive prediction accuracy while providing substantially improved interpretability and physical transparency. Experimental results show that distance and carrier frequency consistently emerge as dominant propagation drivers across mmWave and sub-THz frequency regimes, while LOS/NLOS conditions and antenna heights provide environment-dependent corrective effects. Furthermore, the proposed framework preserves physically meaningful monotonic attenuation behavior with respect to propagation distance and carrier frequency for the vast majority of evaluated samples. Unlike conventional black-box propagation models that prioritize predictive performance alone, the proposed approach emphasizes interpretability, trustworthiness, reproducibility, and deployment reliability. The presented framework provides actionable insights for network planning, digital-twin-assisted wireless optimization, and data-driven calibration of analytical channel models, thereby supporting explainable and trustworthy ML-enabled propagation modeling for future 6 G wireless communication systems.