SuperConvNet: Super Resolution Orchestrated Deep Convolutional Neural Network for Spatial Beam Prediction in B5G and 6G Systems

Abstract

Beam management is one of the most critical tech- nologies in cellular wireless networks, as it directly influences the trade-offs between coverage, capacity, and initial access latency. These trade-offs become even more stringent in massive multiple input multiple output systems operating at millimeter-wave and terahertz frequency bands. In current 5G networks, the minimum beam acquisition latency required to identify the best beam is approximately 360 ms, with an associated overhead ranging from 1.3 Mbps to 160 Mbps. Also existing beam selection algorithms inherently suffer from a trade-off, where reducing the beam acquisition latency typically results in a proportional degradation in network throughput. To address this, we propose a deep learning based beam prediction model, termed SuperConvNet, which reduces beam acquisition latency by more than 80% while simultaneously improving network throughput compared to traditional non deep learning based beam selection methods in urban macro-cell deployments. The SuperConvNet model also demonstrates robustness to practical quantized reference signal received power reporting, accurately predicting the best beam in approximately 84% of the cases, and the top-3 beam in nearly 100% of the cases. Moreover, it minimizes both feedback and training data collection overhead, offering a scalable and efficient solution for real world deployments.