Toward Generalizable mmWave and Terahertz Deep Learning-Based Channel Prediction in 6G Networks

Abstract

The use of millimeter-wave (mmWave) and Terahertz (THz) bands is essential for 6G networks, as they enable ultra-high data rates, increased capacity, and low-latency communication, thereby enabling transformative applications from virtual reality to holographic communications. Channel prediction using deep learning (DL) is essential in both bands to overcome these bands’ challenges, such as sensitivity to blockage. The existing prediction strategies assume static data distributions, where both training and testing datasets are derived from identical or highly similar environmental conditions. However, real-world conditions change over time due to non-stationarity in indoor wireless environments, resulting in data drift from differences between the train and test data distributions. Data drift can impact the generalization of DL channel prediction models by causing model drift. In this paper, the data drift in the mmWave and THz bands is categorized for the first time and represented in three scenarios: user densities, mobility stages, and room configurations. We investigate how data drift affects the generalization of DL models for mmWave and THz channel prediction, and propose a mitigation approach to address model drift and maintain generalization in affected scenarios. Our results show that mmWave models are most affected by mobility-stage changes, with performance degradation ranging from 11% to 68%, whereas THz models are most affected by user-density changes, with degradation ranging from 4% to 62%. To mitigate these effects, we propose a weighted ensemble approach for mmWave, reducing deterioration to $\leq 4\%$ , and a two-stage framework for THz, keeping deterioration $\leq 9\%$ .