Site-Specific Channel Generation Integrating ML-Predicted Propagation Parameters in the FR3 Upper Mid-Band

Abstract

Toward the realization of 6G networks and wireless digital twins, site-specific channel modeling that accurately replicates real-world environments is indispensable. While conventional map-based hybrid models capture spatial structures, they struggle to accurately reproduce local shadowing and microscopic fading characteristics due to inherent ray-tracing computational errors and the limitations of purely statistical models. In this paper, we propose a novel framework for generating high-fidelity channels by integrating site-specific propagation parameters predicted via machine learning (ML). The proposed method employs a Residual Network (ResNet) to predict path loss (PL) and delay spread (DS) from spatial features and uses these values to appropriately scale and correct the cluster powers in the hybrid model. To validate its generalizability, we conducted a rigorous ablation study comparing five parameter configurations using urban measurement data at 8.45 GHz in the FR3 upper mid-band—a key candidate spectrum for 6G—across two continuous routes with distinct propagation topologies. The evaluation results demonstrate that the proposed method accurately tracks environment-specific macroscopic trajectories while eliminating the inherent overestimation bias of conventional ray-tracing. Furthermore, in evaluating microscopic small-scale fading, the proposed method improved the Kolmogorov-Smirnov (KS) distance by approximately 9–17% compared with conventional methods. Most notably, spatial autocorrelation analysis demonstrated that our method successfully mitigates the critical spatial inconsistency of conventional models, deterministically reproducing decorrelation distances of 35.4 m and 46.0 m on the two evaluation routes under the route-exclusive split, in close agreement with the measured values (36.8 m and 46.0 m), which themselves align with the 3GPP UMa standard parameters (37 m for LoS / 50 m for NLoS). These findings establish the proposed framework as a powerful next-generation methodology for highly realistic wireless digital twins.