Hybrid Statistical and Machine Learning Models for Atmospheric Refractivity Prediction in Wireless Channels

Atmospheric ducts, formed by sharp refractive index gradients in the lower atmosphere, play a crucial role in signal propagation, particularly for next-generation communication systems operating in high-frequency bands (sub-6 GHz and beyond), where ducts can significantly modify signal paths. This study refines the standard ITU refractivity model through a region-specific correction and subsequently develops a multiple linear regression (MLR) framework to adapt the model using real meteorological data. The MLR model evaluates predictor significance, addresses multicollinearity, and applies statistical criteria to produce a localized refractivity estimate. To address residual nonlinear patterns not captured by MLR, a machine learning (ML) model is trained on the residuals and integrated with the statistical output to form a hybrid prediction. The region-specific (RS) model shows a systematic deviation from the ITU-based modified refractivity profile, with a mean difference of 4.67 M-units, a standard deviation of 3.49 M-units, and an RMSE of 5.83 M-units. These differences suggest that the ITU profile does not adequately represent local atmospheric conditions. Furthermore, the hybrid model reduces residual error relative to MLR alone. The results highlight the importance of incorporating RS adjustments and data-driven methods in modeling atmospheric ducts for propagation analysis.