Cross-environment transfer learning for robust mmWave path loss modeling in 6G wireless networks

Accurate millimeter wave (mmWave) and sub-terahertz (sub-THz) path loss modeling is essential for reliable deployment of sixth-generation (6G) wireless networks, yet remains challenging due to severe cross-environment variability, blockage sensitivity, limited measurement data, and computational constraints. This paper proposes a deployment-oriented cross-environment transfer learning framework integrating parameter-based inductive transfer learning, selective layer freezing, optional Maximum Mean Discrepancy (MMD)-based domain adaptation, and structured pruning for robust and computationally efficient path loss prediction across heterogeneous propagation environments. The framework first pretrains a deep neural network on a data-rich source domain, preserves transferable low-level propagation representations through frozen shared layers, and fine-tunes higher adaptive layers for target domain generalization under sparse data conditions. Optional MMD-based feature alignment is incorporated to improve robustness under severe source-target distribution mismatch, while structured pruning reduces inference complexity, memory usage, and energy consumption with minimal performance degradation. The proposed framework is evaluated across urban microcell, indoor office, and sub-THz corridor scenarios using standardized synthetic and measurement-inspired datasets. Experimental results demonstrate that the proposed framework achieves the lowest root mean squared error (RMSE) of approximately $$6.1 \pm 0.2$$ dB, outperforming baseline deep neural networks and conventional analytical models across all evaluated environments. Cross-environment experiments further show that the proposed transfer learning strategy substantially mitigates domain shift-induced degradation, improves convergence stability, and enhances robustness under heterogeneous propagation conditions. In addition, structured pruning reduces training energy consumption by 38%, inference latency by 30%, and memory usage by 35%, enabling deployment in resource-constrained wireless systems. The results confirm that the proposed framework provides an effective, transferable, and scalable solution for mmWave and sub-THz path loss modeling while addressing communication–computation trade-offs relevant to adaptive 6G, UAV-assisted, and edge intelligent wireless networks.