PaperlayerView exact revision on arXiv

Source-linked AI summary

Tri-Band Channel Measurement-Enabled Multi-Layer Digital Twin for Terahertz Wireless Data Centers

Mingjie Zhu, Ziming Yu, Guangjian Wang, Chong Han

arXiv:2609.01699v1cs.LGcs.IT

TL;DR

Existing optimization-oriented digital twins struggle to combine high-fidelity channel modeling with real-time responsiveness, especially across thousands of wireless links. This paper proposes a measurement-driven, multi-layer twin with an AI acceleration engine and a hybrid propagation model; measurements show the hybrid model reduces RMS delay-spread error to below 1%.

  • Problem

    Existing optimization-oriented digital twins rely on computationally intensive ray-tracing simulations or simplified channel abstractions, limiting high fidelity and real-time responsiveness in complex indoor environments.

  • Method

    The framework combines tri-band measurements, a calibrated physical and ray-tracing twin, a blockage-aware LoS-aware implicit neural-field AI twin, and evaluation layers for coverage and interference analysis.

  • Results

    The hybrid propagation model predicts an RMS delay spread of 0.3957 ns versus the 0.3923 ns measurement, differing by less than 1%.

  • Takeaways & Limitations

    The framework establishes an end-to-end pipeline from channel measurement to network optimization, bridging physical fidelity and computational efficiency.

Abstract

from arXiv · show

The rapid growth of AI computing has driven increasing demands for flexible and high-capacity data-center interconnections. Owing to its ultra-wide bandwidth and high spatial reuse capability, terahertz (THz) communication has emerged as a promising solution for future wireless data centers, while digital twins (DTs) enable efficient wireless planning and real-time optimization. In this work, a measurement-driven multi-layer DT framework is proposed for THz wireless data centers, where the physical, channel, evaluation, and manipulation layers are progressively constructed from bottom to top. First, extensive channel measurements are conducted at 140, 220, and 300 GHz to characterize frequency-dependent propagation behaviors. Based on the tri-band measurements, a measurement-calibrated physical twin is established by jointly optimizing the geometry, material, antenna, and hybrid propagation models. On top of the physical twin, a line-of-sight (LoS)-aware implicit neural field is developed to construct an AI channel twin for efficient channel reconstruction. The proposed AI twin learns location-dependent channel statistics from the calibrated twin, enabling real-time prediction of received power and LoS probability. Building upon the reconstructed channel field, a system-level evaluation layer is derived to analyze coverage and interference for both AP-to-rack and rack-to-rack communications. Experimental results show that the proposed AI twin achieves lower power reconstruction error than existing neural-field baselines while maintaining real-time inference capability. Moreover, the ceiling-mounted AP deployment achieves over 90% coverage under a 10 dB signal-to-interference-plus-noise ratio (SINR) threshold, demonstrating the effectiveness of the proposed DT framework for THz wireless data-center planning and optimization.

I. INTRODUCTION

THz wireless data centers offer high-capacity, directional links but require accurate, adaptive digital twins for planning, interference management, and real-time optimization. This paper proposes a measurement-driven multi-layer framework combining calibrated physical modeling, AI channel reconstruction, and system-level evaluation.

  • THz wireless data centers are motivated by rising data-center traffic and the need for flexible, low-latency interconnections.
  • Reliable THz network planning requires a digital twin that perceives the environment, reconstructs propagation, evaluates performance, and supports optimization.
  • Measurements in wireless data-center environments remain limited, especially for multi-band characterization and digital-twin applications.
  • The proposed framework integrates LiDAR-based geometry, THz-TDS-calibrated materials, and a measurement-calibrated channel twin combining ray tracing and channel-parameter extraction.
  • A blockage-aware implicit neural field uses propagation priors, LoS maps, and dual experts to enable accurate, lower-complexity channel reconstruction than full ray tracing.
  • Analytical coverage and rate models connect channel prediction with coverage, interference, deployment evaluation, and closed-loop network optimization.

II. MEASUREMENT LAYER

The measurement layer establishes the physical foundation of the digital twin through multi-source measurements and tri-band channel characterization. The setup is designed to support consistent comparison across 140, 220, and 300 GHz.

  • The physical twin uses channel, environment, and material measurements to characterize the wireless data-center environment.
  • The campaign measures propagation at 140, 220, and 300 GHz as the physical foundation of the multi-layer digital twin.
  • The channel measurement setup is deployed inside the wireless data-center environment, with transmitter, receiver, antenna, and scanning configurations maintained across bands.
  • The 300 GHz sounder provides 20 GHz bandwidth, yielding 0.05 ns delay resolution and approximately 1.5 cm distance resolution.

2) Measurement Scenario:

Measurements cover LoS and NLoS conditions for both rack-to-rack and AP-to-user links across the data-center scenario. The campaign includes 29 transmitter–receiver locations.

  • The measurement scenario includes four cases: LoS and NLoS conditions for rack-to-rack and AP-to-user links.
  • 29 transmitter–receiver locations are measured across the data-center scenario.
  • Tx 1 and Tx 2 represent rack-to-rack communication from two directions, while Tx 3 simulates a ceiling-centered access point.

B. Channel Parameter Extraction

The channel-parameter extraction procedure converts measured directional responses into delay-angle representations, identifies multipath components, and compares path-loss behavior across frequency bands.

  • Directional channel responses are represented in the delay-angle domain to form the power-angle-delay profile.
  • An iterative peak-search procedure extracts multipath components by identifying the strongest local maxima above a predefined threshold.
  • Each detected path is represented by a parameter vector after its corresponding multipath component is initially estimated.
  • The cross-band similarity of path-loss measurements is quantified using a correlation coefficient for pairs of frequency bands.
  • The resulting correlation matrix assesses propagation consistency across frequency bands and reveals frequency dependence in the data-center environment.

C. Physical Twin Observations

The physical twin reveals dominant LoS, reflected, and diffuse propagation mechanisms across 300, 220, and 140 GHz, while measured large-scale channel variations remain strongly correlated across bands.

  • The physical twin projects MPC trajectories onto the horizontal plane to characterize dominant propagation mechanisms between one transmitter and all receiver locations.
  • 52% of measured receiver locations exhibit LoS propagation, while rack blockage causes NLoS propagation to dominate the remaining 48%.
  • LoS consistently dominates the strongest MPC across all three bands, with additional reflections from the rear glass wall and metallic rack surfaces.
  • Specular reflections and diffuse scattering both contribute significantly to practical THz data-center channel characteristics.
  • The three bands show strong path-loss correlations, including 0.981 between 220 and 300 GHz, indicating highly similar spatial propagation patterns for the higher-frequency pair.

III. CONSTRUCTION LAYER

The construction layer builds the channel twin from tri-band measurements, combining a measurement-calibrated RT twin for physical fidelity with an AI twin for efficient channel reconstruction.

  • Tri-band measurements and extracted channel characteristics support a channel twin that represents THz propagation accurately while reducing computational cost.
  • The channel twin contains complementary calibrated RT and AI components for improving physical fidelity and enabling learning-based channel reconstruction.

A. Measurement-calibrated RT Twin

The measurement-calibrated RT twin jointly incorporates geometry, materials, antenna patterns, and hybrid propagation to reproduce measured THz channels. It captures dominant paths accurately, while hybrid modeling recovers diffuse components that deterministic RT misses.

  • Four complementary calibrations mitigate discrepancies between RT and measurements: geometry, material, antenna, and propagation calibration.
  • LiDAR reconstruction captures racks, corridors, cable distributions, and irregular objects that affect THz propagation, while triangular meshes represent the RT environment.
  • THz-TDS extracts frequency-dependent dielectric properties across 140/220/300 GHz, and measured directional antenna patterns reproduce practical beamwidth and side-lobe characteristics.
  • The hybrid propagation model combines deterministic specular components with diffuse scattering to compensate for unresolved multipath and rough-surface effects.
  • The calibrated RT twin reproduces dominant propagation paths and arrival directions at all three frequency bands, but conventional specular-only RT misses low-power diffuse components.
  • 0.072 ns, 0.350 ns, and 0.241 ns are the delay errors of three dominant MPCs, while corresponding power differences remain within 1 dB.
  • 0.3923 ns measured RMS delay spread versus 0.3691 ns deterministic-RT prediction indicates a 5.9% underestimation; hybrid RT predicts 0.3957 ns, within 1% of measurement.
  • The calibrated RT twin captures dominant LoS and specular paths, while hybrid propagation recovers missing diffuse scattering for a high-fidelity physical twin.

B. AI Twin

The AI twin accelerates channel reconstruction by combining calibrated RT data with a LoS-aware implicit neural field. It preserves blockage-sensitive propagation structures while achieving lower reconstruction error and real-time inference.

  • Architecture: The AI twin uses transmitter–receiver geometry, distance, and grid-based LoS probability to model blockage-sensitive propagation.A shared encoder feeds dedicated LoS and NLoS experts, whose outputs are combined by a visibility-aware gating module.
  • Dataset construction: Training data are generated by sampling receiver locations, running RT simulations, extracting channel parameters, and determining LoS/NLoS conditions.The resulting dataset includes positional encodings and grid-based LoS probabilities for training the dual-expert neural field.
  • Training: The model jointly optimizes power prediction error and LoS gating error, with weighting factors balancing the two objectives.The LoS labels come from a grid-based visibility map.
  • Results: Compared with conventional INF and MLP baselines, the proposed AI twin better preserves blockage boundaries, deep-shadow regions, and physical propagation structures.The conventional INF over-smooths discontinuities, while the MLP introduces spatial artifacts and larger reconstruction errors.
  • Results: The proposed model achieves the lowest reconstruction error among the compared learning-based models while retaining real-time inference capability.The model completes dense channel reconstruction in 143 ms, whereas complete RT simulation requires approximately 1.8× 10^6 ms, corresponding to more than four orders of magnitude acceleration.

IV. EVALUATION LAYER

The evaluation layer converts channel-twin outputs into spatially continuous network metrics. Analytical models enable real-time coverage, interference, rate, and deployment evaluation without repeated RT simulations.

  • Layer role: The evaluation layer uses path loss, received power, delay dispersion, angular characteristics, and LoS probability to assess THz wireless-data-center performance.It bridges channel-level predictions and network-level performance assessment.
  • Layer role: It constructs spatial maps of signal strength, interference, SINR, coverage probability, and achievable rate throughout the data-center environment.The same outputs support AP-to-rack and rack-to-rack communication analysis.
  • Analytical evaluation: Closed-form coverage expressions combine predicted LoS probability, received-power statistics, and blockage information for rapid deployment comparison.These expressions avoid computationally expensive RT simulations during system-level evaluation.
  • Optimization: The framework formulates AP deployment and resource allocation using achievable rate and coverage probability under SINR thresholds.The optimization variables include AP deployment positions, while the DT supplies the required performance metrics.
  • Analytical evaluation: SINR evaluation incorporates transmit power, antenna gains, DT-derived path loss and shadowing, small-scale fading, noise spectral density, and bandwidth.Coverage probability is defined as the probability that received SINR exceeds threshold T.

A. AP Deployment

Coverage decreases with stricter SINR thresholds and higher carrier frequencies, while ceiling-mounted APs provide the strongest deployment performance. At 10 dB SINR, the majority of locations lie in a region where small channel improvements can increase coverage.

  • Evaluation procedure: Coverage can be evaluated from AI-twin-predicted local channel statistics through closed-form LoS and NLoS expressions without invoking RT simulations.The evaluation queries LoS probability and received-power statistics at each receiver location.
  • Frequency comparison: Coverage probability monotonically decreases as the SINR threshold increases because fewer receiver locations satisfy stricter requirements.The comparison is made across the 140, 220, and 300 GHz bands.
  • Frequency comparison: The 140 GHz band achieves the highest coverage probability across the SINR range, followed by 220 GHz and 300 GHz.Higher-frequency degradation is attributed to increased free-space path loss, molecular absorption, and blockage sensitivity.
  • AP positions: Ceiling-mounted APs achieve the highest coverage probability over most SINR thresholds among corridor, corner, and ceiling deployments.The centrally elevated position provides shorter average propagation distances and fewer blockage events, maintaining stronger LoS connectivity.

B. Concurrent Rack-to-rack Communications

The rack-to-rack evaluation models channel gains and concurrent transmissions to derive interference, coverage, and average rate. Analytical SINR results are compared with Monte Carlo simulations across three measured frequency bands as active links increase.

  • Channel Modeling: Channel gains between all rack pairs form a channel matrix for modeling concurrent rack-to-rack communications.The gain from rack i to rack j is denoted by g_ij.
  • Interference Modeling: Aggregate interference is modeled from predicted channel gains and binary transmission activity indicators.The individual gains are approximated as lognormal, enabling Fenton–Wilkinson modeling of aggregate interference.
  • Performance Evaluation: The coverage probability and average rate are derived from the interference-aware rack-to-rack channel model.The coverage expression uses the standard Gaussian cumulative distribution function, while the rate follows from the resulting SINR distribution.
  • Performance Evaluation: Rack-to-rack evaluation explicitly captures concurrent-transmission interference, making it more realistic for dense wireless data-center networks than AP-to-rack evaluation.The evaluation layer connects AI channel-twin statistics to achievable rate, interference, and coverage probability for network optimization.
  • Results: Average SINR decreases as simultaneously active links increase because co-channel interference accumulates.Analytical results are compared with Monte Carlo simulations for the three measured frequency bands.

V. CONCLUSION

The paper proposes a measurement-driven multi-layer digital twin for THz wireless data centers. It combines tri-band measurements, calibrated RT modeling, AI channel reconstruction, and system-level evaluation for coverage and interference analysis.

  • Measurement Layer: Tri-band measurements at 140, 220, and 300 GHz capture propagation characteristics for THz wireless data centers.These measurements form the foundation of the proposed framework.
  • Construction Layer: A calibrated RT twin jointly optimizes geometry, material, antenna, and propagation models, while an AI twin provides accurate, low-latency channel reconstruction.The AI twin supports efficient channel-field construction for subsequent evaluation.
  • Evaluation and Manipulation Layers: The reconstructed channel field supports coverage and interference analysis in an evaluation twin.The framework then provides inputs for manipulation technologies such as power allocation and beam manipulation.
Loading 2609.01699v1…