This paper reviews graphene-based terahertz (THz) metasurfaces for holographic beamforming in next-generation communication systems, including 6G/7G, UAV, and satellite networks. These metasurfaces provide tunable plasmonic responses, strong electromagnetic confinement, and dynamic control over phase, amplitude, and polarization. THz graphene metasurfaces have tunable plasmonic responses, confinement, and control over phase, amplitude, and polarization that enable functionalities beyond traditional antennas and metasurfaces. This review concerns the mechanisms of dynamic beam manipulation in graphene-based THz metasurfaces, specifically beam scanning, Pancharatnam–Berry (PB) phase-controlled focusing, and advanced holographic representations that can be implemented under realistic biasing and material limitations. This paper reviews the current state of beam scanning, Efficient Pancharatnam–Berry (PB) phase-locked focusing, dual holography, and multi-spin holography, as well as ultra-thin, ultra-Efficient metasurfaces based on graphene and Vanadium dioxide (VO₂). Previous reviews focused on device demonstrations or material property discussions in a disaggregated fashion. This review presents an integrated system-level review and identifies the materials and design, tunability, biasing, and EM performance trade-offs in THz metasurface designs, revealing the dominant performance limitations in phase manipulation, Fermi level control, quality, bandwidth and attenuation. These gaps are attributed to non-uniform bias, THz absorption, slow carrier mobility, and switching limitations, as well as incomplete realization of full 2D holographic beamforming. This review aims to articulate a design framework to illustrate the synthesis of these limitations and integrated system designs focused on scalable system architecture to improve insights on practical system trade-offs. The synthesis in this review should guide metasurface-enabled communications, integrated THz antennas, patterned phase control, and large-scale biasing networks. An analysis of future research directions on scalable fabrication, ultrafast gating, hybrid PB–propagation meta-atoms, and multi-beam holography is presented. This forms an excellent basis for the construction of intelligent THz transceivers for next-generation wireless technologies requiring high mobility and high capacity.

This paper presents two related benchmark studies of AI-enabled security for FR3 (7-24 GHz) ultra-massive MIMO (UM-MIMO) 6G IoT networks, addressing two different layers of the security stack. Study 1 evaluates machine learning approaches for network-layer intrusion detection using a telecom-relevant benchmark dataset (UNSW-NB15), cross-validating multiple classifiers and reporting a systematic literature synthesis, data-quality audit, and hyperparameter-tuned test-set performance comparison. Study 2 evaluates physical-layer angle-of-arrival (AoA) spoofing detection for FR3 UM-MIMO IoT authentication, benchmarking equal and adaptive snapshot-allocation policies for multi-device authentication under a shared sensing budget, validated against closed-form Cramer-Rao bound theory. Together, the two studies show that AI-enabled security benchmarking must be evaluated separately at each layer, network-layer intrusion detection depends primarily on model choice, hyperparameter tuning, and data-quality awareness, while physical-layer spoofing detection depends primarily on how a shared sensing budget is allocated across devices, indicating that a defence-in-depth FR3 UM-MIMO IoT security posture needs distinct, layer-appropriate AI benchmarking rather than a single unified approach.

This paper presents two related benchmark studies of artificial intelligence (AI)-enabled wake scheduling for ultra-low-power devices in Frequency Range 3 (FR3, 7–24 GHz) ultra-massive multiple-input multiple-output (UM-MIMO) Internet of Things (IoT) networks, evaluated as Pareto-frontier trade-offs between energy and access performance. Study 1 benchmarks wake scheduling for narrowband IoT (NB-IoT) devices under simulated bursty traffic, comparing a classical fixed-interval policy, a genie-aided policy with perfect intensity knowledge, and a learned traffic-intensity estimator, and characterizes the energy–latency frontier together with tail (95th-percentile) latency. Study 2 addresses beam-sweep-aware wake scheduling, in which an idle-mode UM-MIMO IoT device must also account for directional beam-sweep paging overhead, and evaluates listen-window policies against a learned arrival-time predictor on an energy–reliability frontier. In both studies, learned, time-varying state estimates did not uniformly improve on well-designed fixed-policy baselines. They helped in specific regimes, substantially reducing tail latency relative to the genie-aided reference in Study 1 and uniquely reaching near-perfect paging reliability in Study 2, whereas simple classical policies remained preferable elsewhere. The mechanism behind each gain differed between the two problems, indicating that gains from learned prediction are regime- and context-dependent rather than universal.

This paper presents two related benchmark studies of AI-enabled wireless power transfer (WPT) for FR3 (7–24 GHz) ultra-massive MIMO (UM-MIMO) 6G base stations. Study 1 benchmarks time-slotted WPT scheduling across multiple IoT devices, comparing naive, sustainability-aware, round-robin and proportional-fair policies with a learned time-to-depletion predictor. Across 80 randomized 15-device populations, naive battery-priority scheduling saved none of the 238 physically sustainable devices, sustainability awareness raised the save rate to 23.5%, and a proportional-fair policy weighted by harvesting surplus performed best (51.7%), ahead of the learned predictor (42.4%). Study 2 asks whether an array should share its aperture between communication and WPT by partitioning its elements (power-splitting) or by alternating between the two functions in time (time-switching) under time-varying battery and queue demand. Power-splitting yielded shorter communication queues at every tested distance, and at 5 m, adaptive power-splitting avoided battery depletion in all 40 trials, whereas fixed allocation depleted the battery in every trial; adaptivity provided no such benefit under time-switching. Together, the studies indicate that the aperture-sharing architecture sets the achievable ceiling and the scheduling policy determines how close a deployment comes to it, so the architecture decision should be made first.