Beam-Constrained Kinematics for Frequency-Hop Blockage Escape at Terahertz

Terahertz links are the very high frequency radio links proposed for 6G. They travel almost like light: a pedestrian stepping into the path can cut them off, and at these frequencies they cannot bend around the person. This proposal borrows a trick from particle physics to predict such blockages reliably. In collider experiments, physicists substitute a quantity they know very precisely, the beam energy, for a quantity they measure poorly. This produces a variable in which real events form a sharp peak and random background forms a smooth, fittable floor. We build the wireless analogue. The imprecise quantity is the direction to the user computed from noisy GPS. The precise quantity is the direction of the radio beam the link is actually using, which the transmitter knows exactly from its list of preset beam directions (its codebook). Measured against that beam, true blockers cluster tightly and harmless passers-by spread smoothly. The smooth part can be fitted from nearby "sideband" regions of the plot, which gives a measured false-alarm rate rather than a guessed one.

The second half of the idea is what the prediction triggers. Graphene antennas and reflecting surfaces can be retuned electrically in nanoseconds by changing a gate voltage. A multi-band graphene reflecting surface cannot, however, be set independently in each of its frequency bands. We turn this limitation into a mechanism. The backup path around an obstacle is parked on the reflecting surface in a second frequency band. When a blockage is predicted, the link escapes by hopping its carrier frequency, not by reprogramming thousands of surface cells. The calibrated false-alarm rate matters because every unnecessary hop moves the link into a band with more water-vapour absorption. The predictor therefore directly sets how much signal-to-noise the system sacrifices.

Core Claim

We claim two things. First, a two-variable "beam-constrained" kinematic plane gives blockage prediction whose false-alarm rate can be estimated from unlabeled sideband data. Second, it separates true blockers from background tracks better than raw GPS or GPS-derived geometry fed to a network. In addition, pairing this predictor with graphene carrier-frequency hopping onto a pre-parked reflected path reduces outage time relative to reactive recovery. It does so at lower control cost than keeping the reflecting surface continuously re-steered.

Where the Directions Stand

The kinematic idea comes from electron-positron collider analyses. BESIII measurements of semileptonic charm decays (Ablikim et al.) and observations of $B_s$ production (Bonvicini et al.) rely on two variables. The energy difference $\Delta E$ peaks at zero for correctly reconstructed candidates. The beam-constrained mass $M_{bc}$ replaces the poorly measured candidate energy with the precisely known beam energy. This supplies the method we reuse: substitution of a precise known quantity, a two-dimensional selection plane, and sideband background subtraction.

The same literature also constrains the design. Kinematics alone cannot reject "peaking backgrounds" that share the signal's kinematics, and BESIII must add particle identification to handle them (Ablikim et al.). The wireless analogue is an object that crosses the beam's line in the ground-plane view but passes under or over it in height. The plan below handles this explicitly.

Two works motivate the gap on the wireless side. Mollah et al. predict mmWave beams from GPS, camera and LiDAR on DeepSense 6G. They feed raw coordinates to a one-dimensional convolutional network, and by their own account they restrict evaluation to line-of-sight conditions, leaving blockage open. Ji Wang et al. convert GPS readings into text prompts for a language encoder and align camera and LiDAR by contrastive pre-training. They reach a distance-based accuracy score of 0.9016 with 75.09M parameters and 11.09 GFLOPs. Both predict the best beam, not blockage events, and neither gives a calibrated false-alarm rate. The closest prior work on proactive blockage handling is Dynamic Blockage Pre-Avoidance using Reconfigurable Intelligent Surfaces, which configures reflecting-surface paths before an obstacle arrives. Our difference lies in the physics-derived, sideband-calibrated predictor and in using frequency hopping, rather than surface reprogramming, as the fast switch.

The hardware side supplies the action space. Hosseininejad et al. simulate a graphene Yagi-Uda antenna tunable between 1.5 and 2.3 THz. It switches between 1.6 dB omnidirectional and 5.2 dB directional modes at nanosecond granularity under a small lookup-table controller, but it takes its commands blindly from the protocol layer. Taghvaee et al. design a graphene reflecting surface operating at 0.65, 0.85 and 1.05 THz, with a worst-case reflection of −4 dB and four phase states. They report that biasing for one band detunes the others and that dense bias routing is costly. We exploit exactly these two limits. Binxu Wang et al. supply a single-band alternative with 361° phase range at 1.05 THz and steering to ±23°. The cost of hopping is set by absorption: Sarieddeen's analysis notes a roughly 10 dB signal-to-noise penalty from moving 1 to 3 THz. Faisal et al. propose Fermi-level tuning to dodge absorption lines, but without any sensing loop.

The Beam-Constrained Plane

Let $\mathbf{b}$ be the precisely surveyed transmitter position and $\mathbf{u}_k$ the unit pointing vector of the currently used codebook beam $k$. Let $\mathbf{o}(t)$ be a LiDAR-tracked obstacle with fitted velocity. The first variable is the constrained miss distance:

$$d_k = \min_{s \in [0, r]} \big\| \mathbf{o}(t_\ast) - (\mathbf{b} + s\,\mathbf{u}_k) \big\|,$$

where $r$ is the GPS-estimated user range and $t_\ast$ is the time at which the obstacle's extrapolated path is closest to the beam ray. The second variable is the reconstruction consistency $\Delta\theta = \theta_{\text{GPS}} - \theta_k$, the angle between the GPS-implied user direction and the beam direction. It plays the role of $\Delta E$: it sits near zero when the user position is well reconstructed and has tails when GPS or track association is bad.

Genuine blockers should pile up at $|d_k|$ below a body half-width, with $\Delta\theta \approx 0$ and $t_\ast$ inside the prediction horizon. Pass-by tracks should populate the $d_k$ axis smoothly. We fit that smooth distribution in the sidebands at larger $|d_k|$ and interpolate it under the signal window. This gives an estimate of false alarms per unit time that needs no blockage labels, which we then validate against labels.

The substitution should help when the codebook's angular quantization, evaluated at the obstacle's distance, is finer than the GPS-induced angular error. That condition is testable, and we test it rather than assume it. Peaking backgrounds (a low car crossing under an elevated beam) are removed with LiDAR height, which plays the role of particle identification.

Research Program

We will build the predictor as a small model on the plane variables plus the obstacle's height and time-to-crossing, with an optional light classifier trained on them. Blockage labels are derived from sharp drops in measured received power in DeepSense 6G.

We will then build a link-level emulator of the terahertz system. Its antenna and surface response curves are taken from the reported simulated behaviour of the graphene Yagi-Uda antenna and multi-band surface: gains, band centres, −4 dB worst-case reflection, and nanosecond switching. Absorption per band comes from the transparency-window structure. In the protocol, the direct link runs in one window while the surface holds a backup reflection toward the predicted user position in a second window. A predicted blockage triggers a carrier hop. The surface is re-steered only when the plane predicts the backup geometry has drifted. One caveat is stated plainly: the sensing data is mmWave and the terahertz hardware exists only in simulation. The emulator therefore transplants measured trajectories onto modeled terahertz links and does not claim measured terahertz performance.

Baselines

The naive reference is reactive recovery. The link stays on its direct beam, detects blockage from a received-power drop, and then re-sweeps or hops. It runs in the same emulator on the same trajectories, and we report its outage duration and lost packets as the floor.

The strongest existing learned competitor is the multimodal predictor of Mollah et al.. It reports up to 98.19% top-13 beam accuracy and received-power ratios of 99.92% for vehicle-to-infrastructure and 98.86% for vehicle-to-vehicle links on DeepSense 6G. The model of Ji Wang et al. is a second competitor, with a distance-based accuracy of 0.9016. These are beam-prediction results; there is no reported blockage-prediction result on Lacuna for either model. We will retrain both architectures with a blockage-event head on the same DeepSense 6G scenarios, splits, labels and hyperparameter search budget as our predictor. We will compare precision-recall at equal prediction horizon and parameter count. Beating them means higher recall at a matched false-alarm rate, together with a sideband-predicted false-alarm rate that matches the held-out rate.

The ablation that isolates the key ingredient keeps everything fixed and changes one thing. It replaces the beam-constrained $d_k$ with the same miss distance computed from the GPS-implied direction. If the substitution buys nothing, the physics analogy has failed.

Why not just…?

Why not just run a Kalman-filter tracker and compute time-to-collision with the GPS line of sight? This is the most serious threat, because it captures most of the geometry with no learning. It is exactly our unconstrained ablation plus a tracker, and we will give it the same LiDAR tracks. If it matches the constrained plane, the remaining contribution is the sideband false-alarm calibration and the hop protocol, and we will report it that way.

Why not just train a bigger multimodal network on more data? The existing models were built and evaluated for best-beam prediction in line-of-sight conditions (Mollah et al.). The heavier one costs 11.09 GFLOPs per inference (Ji Wang et al.), and neither provides a label-free estimate of its false-alarm rate. The false-alarm rate is what sets the absorption cost of hopping. The matched-budget retraining in the baselines will show whether capacity alone closes the gap.

Why not just keep the backup path permanently configured and hop reactively, since hopping takes nanoseconds? This is also a real threat. We will run it as a third system. The predictor must then justify itself in two ways. It must avoid the packets lost before a power drop is detected, and it must reduce how often the densely wired surface has to be re-steered to follow a moving user. Surface re-steering is the cost Taghvaee et al. flag as the open problem. If always-on reactive hopping matches us on outage at similar re-steering counts, the predictor is unnecessary for the hop, and we will say so.

Why not just reprogram the surface instead of hopping frequency? Reprogramming requires re-biasing many cells through costly routing. The hop changes only the antenna's driven element. We will count bias updates and switching latency for both options.

Decisive Checks

On DeepSense 6G, the plane must show a visible peak over a smooth sideband in $d_k$ for power-drop events. The sideband-predicted false-alarm rate must agree with the held-out rate within its statistical uncertainty. The constrained variable must beat the GPS-derived one, most clearly where GPS error is largest (the nighttime and vehicle-to-vehicle scenarios).

In the emulator, the hop protocol must reduce outage time relative to reactive recovery and to always-on backup at an equal or smaller number of surface re-steerings. It must also stay net-positive in average signal-to-noise once the absorption penalty of false hops is charged. Failure of the first check kills the kinematic claim. Failure of the second reduces the work to a better blockage predictor, with frequency hopping left as an unneeded extra.

References

  1. Improved measurements of <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mo>ℓ</mml:mo><mml:mo>+</mml:mo></mml:msup><mml:msub><mml:mi>ν</mml:mi><mml:mo>ℓ</mml:mo></mml:msub></mml:math> and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mover accent="true"><mml:mi>K</mml:mi><mml:mo stretchy="false">¯</mml:mo></mml:mover><mml:mn>0</mml:mn></mml:msup><mml:msup><mml:mo>ℓ</mml:mo><mml:mo>+</mml:mo></mml:msup><mml:msub><mml:mi>ν</mml:mi><mml:mo>ℓ</mml:mo></mml:msub></mml:math>
  2. Observation of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>B</mml:mi><mml:mi>s</mml:mi></mml:msub></mml:math>Production at the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>Υ</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mn>5</mml:mn><mml:mi>S</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math>Resonance
  3. Multi-Modality Sensing in mmWave Beamforming for Connected Vehicles Using Deep Learning
  4. When Vision-Language Model (VLM) Meets Beam Prediction: A Multimodal Contrastive Learning Framework
  5. Dynamic Blockage Pre-Avoidance using Reconfigurable Intelligent Surfaces
  6. MAC-oriented programmable terahertz PHY via graphene-based Yagi-Uda antennas
  7. Multiwideband Terahertz Communications Via Tunable Graphene-Based Metasurfaces in 6G Networks: Graphene Enables Ultimate Multiwideband THz Wavefront Control
  8. Full 360° Terahertz Dynamic Phase Modulation Based on Doubly Resonant Graphene–Metal Hybrid Metasurfaces
  9. Terahertz-Band Ultra-Massive Spatial Modulation MIMO
  10. Ultramassive MIMO Systems at Terahertz Bands: Prospects and Challenges