Deep Learning-Based Beam Training Methods for 6G MIMO Systems: From Far-Field to Near-Field

Abstract

Efficient beam training across the transition from conventional far-field massive multiple-input multiple-output (Massive MIMO) to near-field extremely large-scale MIMO (XL-MIMO) is fundamental to forth-coming sixth-generation (6G) systems, yet remains impeded by severe mobility, spherical-wave propagation and prohibitive training overhead. This paper presents a unified survey of contemporary deep learning (DL)-based beam training techniques, reviewing their evolution from early deep neural networks to more advanced architectures such as convolutional, recurrent, and other specialized models. These approaches exhibit strong adaptability and feature extraction capabilities. However, research in this area is still in its early stages, and such methods have yet to be widely applied in emerging scenarios, such as dynamically varying or multi-user interference-limited near-field communication systems. Motivated by the survey insights, we further introduce a lightweight stacked bidirectional gated recurrent unit (BiGRU) architecture equipped with residual shortcuts and adaptive normalization to capture the temporal evolution of near-field channels, referred to as the stacked residual-BiGRU (SRB). Simulations conducted on a dynamic XL-MIMO testbed demonstrate the competitive achievable rates and low training overhead of SRB, highlighting how the recurrent structure complements spatial feature extractors. Finally, we outline open challenges and future directions toward scalable, energy-efficient and robust beam management frameworks for real-world deployments.