Pith. sign in

REVIEW 3 major objections 4 minor 48 references

The paper claims that a single calibration network, trained at 95 GHz on unpaired data, can make a ray-traced W-band channel twin statistically match lab measurements and carry that calibration to adjacent carriers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 14:28 UTC pith:WCXSXJVH

load-bearing objection Promising unpaired W-band channel calibration with genuine cross-frequency transfer, but the in-domain metrics are selection-biased and the input distribution is a hand-set Gaussian impairment layer. the 3 major comments →

arxiv 2607.26501 v2 pith:WCXSXJVH submitted 2026-07-29 eess.SP

Calibrating the Digital Twin Channel: Statistics-Consistent Sim-to-Lab Adaptation for W-Band Industrial OFDM Links

Pulok Tarafder , Abigail O. Oyekola , Jasni Areepatta Mannil , Imtiaz Ahmed , Zoheb Hassan , Danda B. Rawat , Wenjie Che This is my paper
classification eess.SP
keywords digital twin channelchannel calibrationsim-to-lab adaptationW-bandOFDMray tracingdelay spreadunpaired domain adaptation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that a ray-traced digital-twin channel can be calibrated to match real W-band measurements even though simulated and measured captures are unpaired. It proposes SC-SLA, a non-adversarial, cycle-consistent translator that maps simulated channel frequency responses to testbed-like ones by aligning ensemble statistics—mean power delay profile, sub-band PDPs, distribution of RMS delay spread, and per-subcarrier magnitude. On held-out 95 GHz captures it cuts the RMS-delay-spread KS statistic from 0.86 to 0.0495, beats four supervised baselines, and the same checkpoint keeps the lowest PDP and magnitude errors at 92–94 GHz without retraining. If true, this means one measurement batch at a single carrier could certify a twin over a band, reducing over-the-air validation cost for 6G industrial links.

Core claim

The central claim is that a ray-traced channel twin can be calibrated to laboratory fidelity without paired samples and without per-carrier retraining. SC-SLA learns two generator networks that translate between the simulated and measured CFR domains; the forward generator is trained to match batch-level statistics—the mean truncated power delay profile, ten sub-band delay profiles, the mean, standard deviation, and sorted quantiles of RMS delay spread, and the mean CFR-magnitude profile—using a GAN-inspired but fully non-adversarial objective with cycle and identity regularization. On 2,000 held-out 95 GHz captures, the calibrated forward map lowers the RMS-delay-spread KS statistic from 0.

What carries the argument

The machinery is a pair of 1D residual convolutional generators forming a cycle-consistent translator between simulated and measured CFRs, with adversarial discriminators removed and replaced by explicit batch-level channel-statistics losses: ℓ1 on the mean truncated PDP and ten sub-band PDPs, absolute differences on the mean and standard deviation of τrms, the empirical 1-Wasserstein distance between sorted τrms values, and normalized MSE of the mean CFR-magnitude profile. A tanh-bounded residual head couples each output to its input CFR, and cycle/identity losses keep the unpaired map from drifting into an arbitrary permutation; the statistics losses do the distributional alignment, while

Load-bearing premise

The load-bearing premise is that the synthetic ensemble used to train the calibrator—Gaussian timing jitter of 2.5 ns, Gaussian phase jitter of 0.20 radians, and Gaussian per-capture SNR around 28 dB—accurately represents the gap between ray tracing and the real testbed; if the true hardware gap is non-Gaussian or frequency-dependent, the calibrated twin and its transfer to other carriers inherit that mismatch.

What would settle it

Take the trained 95 GHz forward map and apply it to ray-traced CFRs generated in a different geometry (e.g., TX–RX separation 3 m in the same room, or a second room with different materials), then compare the calibrated τrms CDF and mean PDP to new W-band measurements; if the KS statistic and PDP NMSE are no better than the uncalibrated twin, the learned correction is specific to the training scene rather than a general sim-to-lab calibration. A second, sharper falsifier: replace the Gaussian impairment layer with measured phase-noise and SNR statistics from the actual W-band converters and re

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If SC-SLA is right, a ray-traced twin needs to be calibrated only once per band: the 95 GHz checkpoint transfers to 92–94 GHz with no retraining, giving the lowest PDP and CFR-magnitude errors among all compared models at each held-out carrier.
  • Unpaired measurements are sufficient for channel-twin calibration; the method removes the requirement for sample-wise simulated/measured correspondence, which is usually unavailable in field captures.
  • The calibrated twin reproduces capture-to-capture delay-spread variability (KS 0.05 vs 0.86), so OFDM link studies that depend on cyclic-prefix margin, pilot density, or beam coherence can use the twin instead of repeated over-the-air tests.
  • Calibration is cheap enough to insert into simulation workflows: the deployed forward generator has 2.62 million parameters and converts 50,000 CFRs in about 9.75 seconds (≈5,100 CFRs/s).
  • Cutting measured training data to 25% (11,250 captures) keeps in-domain KS below the full-data value under a single seed, suggesting the measurement burden for a first calibration is modest.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • My reading is that the least-tested boundary is the synthetic impairment layer: training on Gaussian timing jitter (σ=2.5 ns), Gaussian phase jitter (σ=0.20 rad), and Gaussian SNR (mean 28 dB, σ=2 dB) presumes the ray-tracing-to-testbed gap is additive and Gaussian; if a different W-band front end produces non-Gaussian, frequency-correlated impairments, the same pipeline may learn a map that carri
  • A stronger stress test than carrier shift would be geometry shift—keeping the 95 GHz checkpoint and moving transmit/receive locations within the same room, or moving to a different room—since the present evaluation holds the scene, distance, and antenna orientation fixed.
  • Because the statistics losses are carrier-agnostic (they operate on the OFDM grid), the method could likely be extended to other bands or to sub-band scheduling, but the cross-frequency result alone does not establish that; a multi-band evaluation with independent measurements would be needed.
  • The single-seed sensitivity and ablation results should be read as existence proofs, not as claims about expected performance across seeds; repeating the data-fraction and ablation experiments over several seeds would be the natural follow-up (the paper lists this as future work).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes Statistics-Consistent Sim-to-Lab Adaptation (SC-SLA), a non-adversarial, cycle-consistent post-processing layer that maps Sionna ray-traced CFRs to CFRs whose statistics match a 92–95 GHz USRP/WR-10 testbed. The simulated ensemble is produced by augmenting a deterministic RT CFR with hand-set Gaussian timing jitter, common phase jitter, and per-capture SNR (Eq. 4). A two-generator CycleGAN-style translator with ResNet backbones is trained with batch-level losses on truncated PDP, ten sub-band PDPs, τrms moments/quantiles, CFR-magnitude NMSE, and cycle/identity regularization. The paper reports an in-domain held-out 95 GHz τrms KS reduction from 0.861 to 0.0495, the lowest PDP and CFR-magnitude errors among four supervised baselines, and zero-shot transfer of the same checkpoint to 92–94 GHz. It also includes inference scalability, measured-data-fraction sensitivity, and a loss/architecture ablation.

Significance. The unpaired, statistics-based formulation is a useful contribution to digital-twin channel calibration, and the release of code, datasets, and checkpoints is a clear strength. The ablation and cross-frequency stress test are also valuable. However, the headline evaluation is compromised by the model-selection protocol: the checkpoint is chosen by minimizing J (Eq. 17) on reserved captures, and J contains the very statistics that are then reported as evidence (D_KS and the NMSE terms). In addition, all simulated stochasticity comes from the untested hand-set impairment layer of Eq. (4), so the calibration claim is conditional on that synthetic ensemble being a faithful model of the RT-to-testbed discrepancy. These issues are addressable but require additional evaluation and sensitivity experiments before the central claims can be accepted.

major comments (3)
  1. [§III-C, Eqs. (14)–(17); §IV-B] The reported 95 GHz numbers are model-selected rather than independent held-out results. Algorithm 1 selects the checkpoint minimizing J = D_KS + 50·NMSE_PDP + 0.25·NMSE_mag on reserved captures, and Sec. IV-B reports D_KS = 0.0495, NMSE_mag = 0.1077, and NMSE_PDP = 3.02e−5 on 2,000 held-out captures. These are essentially the selection criterion evaluated on (a subset of) the same reserved set: D_KS is the headline metric, NMSE_mag has the same mathematical form as L_mag (Eq. 10), and NMSE_PDP is an ℓ2 variant of L_PDP (Eq. 6). The statement in §IV-A that “the KS statistic is not directly optimized by any training loss” is true, but D_KS is directly optimized by model selection. Please evaluate on a third split never used for training or checkpoint selection, and report multiple seeds or bootstrap intervals; otherwise the 44% improvement and the 0.0495 value are best interpreted as sele
  2. [§II-C, Eq. (4); §V] The only stochastic input to SC-SLA is the synthetic impairment layer of Eq. (4). Because G_S→R is deterministic, the corrected output distribution is a function of this hand-set Gaussian ensemble: timing jitter σ=2.5 ns, phase jitter σ=0.20 rad, and per-capture SNR ~ N(28,2) dB clipped to [8,60] dB. If the true RT-to-testbed residual contains non-Gaussian, carrier-dependent, or geometry-dependent effects (e.g., analog-chain frequency ripple, phase-noise coloring, diffuse scattering), the learned mapping is fitted to the wrong input domain. The cross-frequency test cannot expose this, because the same Eq. (4) parameters are reused at 92–94 GHz. The paper itself lists “quantify sensitivity to the impairment parameters” as future work in Sec. V, confirming that this assumption is untested. I request an explicit sensitivity study over these parameters and/or validation of the augmented ense
  3. [§IV-E, §IV-F] All data-fraction and ablation experiments are single-seed. Reported KS differences are often smaller than 0.05 (e.g., full model 0.0495 vs. “w/o cycle + identity” 0.0515 in Fig. 8; data fractions 0.0250–0.0350 vs. 0.0495 in Table IV), and Table IV shows non-monotonic behavior under increasing data. Without confidence intervals, multiple seeds, or a noise floor estimate, these differences are within plausible training variability, so the claims about component contributions and “preserved calibration quality under fourfold reduction” are not yet supported. Please add seeds/CI or soften the corresponding conclusions.
minor comments (4)
  1. [§II-C] Typo: “while PC-SLA learns the residual” should read “SC-SLA learns the residual.”
  2. [Fig. 3] The caption calls the ensemble “Sionna RT-based CIR ensemble,” but the data are actually impairment-augmented via Eq. (4). Please label the panel “impairment-augmented Sionna” to avoid ambiguity.
  3. [§IV-A] The supervised baselines are trained on index-aligned pairs even though no physical correspondence exists. I recommend describing them as “weakly supervised” or explicitly stating that the arbitrary pairing is a limitation; the current wording “strongest paired-supervision setting supported by the data” may overstate what index-aligned Huber regression can achieve.
  4. [Eq. (9)] The statement that Lτq equals the empirical 1-Wasserstein distance is correct only for equal-size mini-batches and ignores sampling randomness. It would be clearer to say that it is the mini-batch estimate of that distance.

Circularity Check

3 steps flagged

In-domain headline metrics are direct training/selection objectives; only the 92–94 GHz zero-shot transfer is genuinely independent confirmation.

specific steps
  1. fitted input called prediction [Sec. III-C (Algorithm 1, checkpoint selection; Eq. (17)); Sec. IV-B (Fig. 5)]
    "Evaluate D_KS, NMSE_PDP, and NMSE_mag on reserved captures. Compute J = D_KS + ω_PDP NMSE_PDP + ω_mag NMSE_mag. If J < J* then save the current G_S→R checkpoint. ... The reported 95 GHz results are therefore in-domain, model-selected results on non-training captures."

    The headline claim that SC-SLA reduces the τ_rms KS statistic to 0.0495 is obtained by explicitly selecting the checkpoint that minimizes J, a criterion that contains D_KS directly as one of its three terms. Thus the reported KS improvement is an optimized model-selection target on reserved captures, not an independent prediction of the τ_rms distribution. The same holds for the in-domain PDP and CFR-magnitude NMSE values that enter J.

  2. fitted input called prediction [Sec. III-B, Eq. (9); Sec. II-C, Eq. (5); Sec. III-C]
    "For equal-sized, equally weighted mini-batches, L_τq equals the empirical 1-Wasserstein distance between the measured and generated τ_rms samples in Eq. (5)."

    The training loss L_τq is, by the paper's own statement, exactly the empirical 1-Wasserstein distance on the τ_rms samples, and L_τ additionally matches the mean and standard deviation of τ_rms. Therefore the in-domain improvement in any τ_rms distribution statistic is statistically forced by the training objective; it cannot serve as an independent confirmation of the claimed delay-spread calibration. D_KS is a different functional but is targeted indirectly by these losses and directly by checkpoint selection.

  3. fitted input called prediction [Sec. III-B Eqs. (6),(10); Sec. III-C Eqs. (15),(16); Sec. IV-B]
    "NMSE_mag has the same mathematical form as the training loss L_mag, but it is evaluated on the reserved or held-out sets E_g and E_r rather than on training mini-batches. In contrast, NMSE_PDP uses normalized squared ℓ2 error between ensemble-mean PDPs, whereas L_PDP uses mean absolute error. The PDP metric therefore provides an in-objective consistency check without duplicating the training loss."

    The paper concedes that the reported PDP NMSE 'compares the same ensemble-mean PDPs as L_PDP' and that NMSE_mag has the same mathematical form as L_mag. Consequently the in-domain PDP and CFR-magnitude results are the fitted training objectives re-evaluated on held-out batches. They are consistency checks of the optimization, not independent verification of the calibration. The genuinely non-circular evidence is the 92–94 GHz zero-shot transfer, where those carriers never entered training or model selection.

full rationale

The central in-domain evaluation is partially circular by the paper's own equations. The training objective (Eq. 13) directly minimizes L_τq, which is the empirical 1-Wasserstein distance on τ_rms (Eq. 9 = Eq. 5), L_PDP and L_mag (Eqs. 6, 10), and the checkpoint-selection criterion J (Eq. 17) includes D_KS, NMSE_PDP, and NMSE_mag. The reported in-domain numbers — τ_rms KS = 0.0495, PDP NMSE = 3.0e-5, CFR-magnitude NMSE = 0.1077 — are therefore optimized targets rather than independent confirmations. The paper itself flags the PDP metric as 'an in-objective consistency check,' and says NMSE_mag 'has the same mathematical form as the training loss L_mag.' This is not a hidden flaw, but it lowers the evidentiary value of the in-domain headline. The 92–94 GHz zero-shot transfer is genuinely non-circular: those carriers are not used for optimization, model selection, or tuning, and the checkpoint trained at 95 GHz still reduces τ_rms KS relative to the uncalibrated twin and achieves the lowest PDP and magnitude NMSE. The remaining concern — that all stochastic variation in the simulated input is generated from the hand-set impairment parameters of Eq. (4), whose sensitivity the paper lists as future work ('quantify sensitivity to the impairment parameters') — is a correctness/assumption limitation rather than a circularity: it does not reduce a claimed output to an input by construction, but it does bound the external validity of the calibration claim. No load-bearing self-citation or imported uniqueness theorem appears. Because the in-domain predictions reduce by construction while the cross-frequency evidence remains independent, the appropriate score is 6.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

SC-SLA's central claim rests on the constructed simulated input distribution (impairment-augmented Sionna CFRs), the measured testbed as ground truth, the sufficiency of selected channel statistics for fidelity, and the transferability of a single checkpoint across carriers. The paper states most of these assumptions but does not independently justify them (no link-level validation, single scene, single seed, no external data). No new physical entities are introduced.

free parameters (5)
  • Timing-jitter std Δτ_i = 2.5 ns
    Hand-chosen as one eighth of the 20 ns tap spacing to add capture-to-capture variation (Sec. II-C, Eq. 4); not varied or fitted to held-out data, but it shapes the simulator input distribution.
  • Common-phase-jitter std φ_i = 0.20 rad
    Hand-set phase spread of about 11.5 degrees in Eq. (4); contributes to the simulated τrms and CFR variability.
  • Per-capture SNR mean/std = 28 dB / 2 dB
    Approximates the measured 28.98 dB / 2.87 dB (Sec. II-B vs II-C); the simulated SNR distribution defines the input-domain noise level.
  • SC-SLA loss weights λ = {λcyc=10, λid=2, λτ=5, λτq=3, λPDP=15, λsPDP=20, λmag=1}
    Chosen by coarse 95 GHz validation (Sec. III-B, Table II); the balance determines which statistics are matched and therefore the reported fidelity.
  • Model-selection weights ω_PDP, ω_mag = 50, 0.25
    Fixed on 95 GHz validation and used in J (Eq. 17) to choose the checkpoint; directly influences which epoch's metrics are reported.
axioms (5)
  • domain assumption The USRP/WR-10 testbed captures after 8 dB SNR gating are the ground-truth W-band channel distribution.
    The entire calibration target is defined by these measurements (Sec. II-B); there is no independent validation of the testbed's own chain response or capture-selection bias.
  • ad hoc to paper The impairment-augmented Sionna CFR distribution (Eq. 4) is a valid input distribution for the RT twin.
    The paper itself shows this input has KS 0.861 versus measured, and the distribution is constructed from hand-chosen Gaussian jitter/phase/SNR values; SC-SLA is trained and evaluated only on this synthetic ensemble.
  • domain assumption Matching mean PDP, sub-band PDP, τrms moments/quantiles, and CFR-magnitude NMSE is sufficient for channel-twin fidelity.
    No BER, phase/temporal statistics, or independent channel metric is validated; the evaluation metrics overlap the training and selection objectives (Secs. III-B, III-C, IV-B).
  • domain assumption The 95 GHz-trained mapping transfers to 92–94 GHz because the residual correction is carrier-insensitive.
    Empirically tested in Sec. IV-C, but only under the same fixed impairment parameterization, one static scene, and one hardware chain.
  • domain assumption Cycle consistency, identity losses, and the tanh residual head keep the unpaired mapping physically meaningful.
    Sec. IV-F reports mean input–output correlations of 0.75–0.82, so a substantial fraction of sample-specific CFR structure is not explicitly constrained.

pith-pipeline@v1.3.0-daily-deepseek · 22204 in / 19506 out tokens · 164501 ms · 2026-08-01T14:28:09.873148+00:00 · methodology

0 comments
read the original abstract

Digital twins (DTs) can reduce over-the-air validation cost in industrial wireless networks, but their utility depends on the fidelity of the underlying channel twin (CT). At W-band, site-specific ray tracing captures deterministic propagation geometry, yet its channel frequency responses (CFRs) do not reproduce the small-scale impairments and capture-to-capture variability observed in laboratory orthogonal frequency-division multiplexing (OFDM) measurements above 90 GHz. This paper proposes Statistics-Consistent Sim-to-Lab Adaptation (SC-SLA), a calibration framework that improves the fidelity of a 95 GHz Sionna ray-traced CT toward that of the testbed by aligning the mean power delay profile (PDP), the distribution of root-mean-square delay spread ($\tau_{\mathrm{rms}}$), and per-subcarrier statistics at 50 MHz sampling bandwidth. SC-SLA uses a generative adversarial network (GAN)-inspired, cycle-consistent architecture with ResNet generators and batch-level channel-statistics losses on the PDP, sub-band PDP, $\tau_{\mathrm{rms}}$ moments and quantiles, and normalized mean-square error (NMSE) of the mean CFR-magnitude profile. The framework is non-adversarial and requires neither paired simulated/measured samples nor discriminators. On held-out 95 GHz data, SC-SLA reduces the $\tau_{\mathrm{rms}}$ Kolmogorov-Smirnov (KS) statistic from 0.86 to 0.050 relative to the impairment-augmented ray-traced input, and by 44% (from 0.089 to 0.050) relative to the strongest of four supervised baselines (FCNN, CNN1D, BiLSTM, and UNet1D). Without retraining, the same checkpoint also generalizes to 92-94 GHz carriers, where it achieves the lowest PDP and CFR-magnitude errors among all baselines while reducing the $\tau_{\mathrm{rms}}$-distribution mismatch relative to the uncalibrated twin.

Figures

Figures reproduced from arXiv: 2607.26501 by Abigail O. Oyekola, Danda B. Rawat, Imtiaz Ahmed, Jasni Areepatta Mannil, Pulok Tarafder, Wenjie Che, Zoheb Hassan.

Figure 1
Figure 1. Figure 1: W-band OFDM testbed setup. The transmitter and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ray-traced CT construction. (a) Blender reconstruction of the indoor W-band testbed. (b) Imported Sionna RT scene [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Delay-domain comparison of the impairment-augmented Sionna RT and measured testbed channels at 95 GHz using [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: SC-SLA framework. The simulation branch (top left) runs the Sionna RT path solver on an indoor Mitsuba scene, [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: In-domain calibration at 95 GHz on 2,000 held￾out captures. The panels report (a) the τrms KS statistic, (b) truncated PDP NMSE on a logarithmic axis, and (c) per￾subcarrier CFR-magnitude NMSE across SC-SLA and the four supervised baselines. SC-SLA achieves the lowest error on all three metrics, with KS 0.0495 and PDP NMSE 3.0 × 10−5 . 92 93 94 95 Frequency (GHz) 10 4 10 3 10 2 PDP N MSE (log) train (a) Po… view at source ↗
Figure 6
Figure 6. Figure 6: Cross-frequency truncated PDP NMSE at 92–94 GHz [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-frequency per-subcarrier CFR-magnitude NMSE [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: In-domain ablation at 95 GHz. The panels report (a) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cross-frequency ablation using the 95 GHz checkpoint without retraining. The 95 GHz point is the in-domain held-out [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 7 linked inside Pith

  1. [1]

    A survey on information and communication technologies for industry 4.0: State-of-the-art, tax- onomies, perspectives, and challenges,

    G. Aceto, V . Persico, and A. Pescap ´e, “A survey on information and communication technologies for industry 4.0: State-of-the-art, tax- onomies, perspectives, and challenges,”IEEE Communications Surveys & Tutorials, vol. 21, no. 4, pp. 3467–3501, 2019

    2019

  2. [2]

    A functional architecture for 6G special-purpose industrial iot networks,

    N. H. Mahmood, G. Berardinelli, E. J. Khatib, R. Hashemi, C. De Lima, and M. Latva-aho, “A functional architecture for 6G special-purpose industrial iot networks,”IEEE Transactions on Industrial Informatics, vol. 19, no. 3, pp. 2530–2540, 2023

    2023

  3. [3]

    Digital twin: Mitigating unpredictable, un- desirable emergent behavior in complex systems,

    M. Grieves and J. Vickers, “Digital twin: Mitigating unpredictable, un- desirable emergent behavior in complex systems,” inTransdisciplinary Perspectives on Complex Systems. Springer, 2016, pp. 85–113. 15

    2016

  4. [4]

    Digital twin networks: A survey,

    Y . Wu, K. Zhang, and Y . Zhang, “Digital twin networks: A survey,” IEEE Internet Things J., vol. 8, no. 18, pp. 13 789–13 804, 2021

    2021

  5. [5]

    Digital-twin- enabled 6G: Vision, architectural trends, and future directions,

    L. U. Khan, W. Saad, D. Niyato, Z. Han, and C. S. Hong, “Digital-twin- enabled 6G: Vision, architectural trends, and future directions,”IEEE Commun. Mag., vol. 60, no. 1, pp. 74–80, 2022

    2022

  6. [6]

    Digital-twin empowered deep reinforcement learning for site- specific radio resource management in NextG wireless aerial corridor,

    P. Tarafder, Z. Hassan, I. Ahmed, D. B. Rawat, K. Hasan, and C. Pu, “Digital-twin empowered deep reinforcement learning for site- specific radio resource management in NextG wireless aerial corridor,” arXiv:2602.03801, 2026

    arXiv 2026

  7. [7]

    Electromag- netic wave property inspired radio environment knowledge construction and artificial intelligence based verification for 6G digital twin channel,

    J. Wang, J. Zhang, Y . Sun, Y . Zhang, T. Jiang, and L. Xia, “Electromag- netic wave property inspired radio environment knowledge construction and artificial intelligence based verification for 6G digital twin channel,” Frontiers Inf. Technol. Electron. Eng., vol. 26, no. 2, pp. 260–277, 2025

    2025

  8. [8]

    Wireless network digital twin for 6G: Generative AI as a key enabler,

    Z. Tao, W. Xu, Y . Huang, X. Wang, and X. You, “Wireless network digital twin for 6G: Generative AI as a key enabler,”IEEE Wireless Commun., vol. 31, no. 4, pp. 24–31, 2024

    2024

  9. [9]

    Ray tracing for radio propagation modeling: Principles and applications,

    Z. Yun and M. F. Iskander, “Ray tracing for radio propagation modeling: Principles and applications,”IEEE Access, vol. 3, pp. 1089–1100, 2015

    2015

  10. [10]

    Sionna: An open-source library for next-generation physical-layer research,

    J. Hoydis, S. Cammerer, F. Ait Aoudiaet al., “Sionna: An open-source library for next-generation physical-layer research,” arXiv:2203.11854, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Sionna RT: Differentiable ray tracing for radio propagation modeling,

    J. Hoydis, F. Ait Aoudia, S. Cammereret al., “Sionna RT: Differentiable ray tracing for radio propagation modeling,” arXiv:2303.11103, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    Wireless InSite: 3D wireless propagation software,

    Remcom, Inc., “Wireless InSite: 3D wireless propagation software,” https://www.remcom.com/wireless-insite-em-propagation-software, 2024, state College, PA

    2024

  13. [13]

    Toward real-time digital twins of EM environments: Computational benchmark for ray launching software,

    M. Zhu, L. Cazzella, F. Linsalata, M. Magarini, M. Matteucci, and U. Spagnolini, “Toward real-time digital twins of EM environments: Computational benchmark for ray launching software,”IEEE Open J. Commun. Soc., vol. 5, pp. 6291–6302, 2024

    2024

  14. [14]

    Real-time digital twins: Vision and research directions for 6G and beyond,

    A. Alkhateeb, S. Jiang, and G. Charan, “Real-time digital twins: Vision and research directions for 6G and beyond,”IEEE Commun. Mag., vol. 61, no. 11, pp. 128–134, 2023

    2023

  15. [15]

    Framework and overall objectives of the future development of IMT for 2030 and beyond,

    ITU-R, “Framework and overall objectives of the future development of IMT for 2030 and beyond,” ITU-R, Recommendation M.2160-0, Nov. 2023

    2030

  16. [16]

    Study on supporting NR from 52.6 GHz to 71 GHz,

    3GPP, “Study on supporting NR from 52.6 GHz to 71 GHz,” 3GPP, Tech. Rep. TR 38.808, V17.0.0, Mar. 2021

    2021

  17. [17]

    Study on channel model for frequencies from 0.5 to 100 GHz,

    3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” 3GPP, Tech. Rep. TR 38.901, V17.0.0, Mar. 2022

    2022

  18. [18]

    6G wireless channel measure- ments and models: Trends and challenges,

    C.-X. Wang, J. Huang, H. Wanget al., “6G wireless channel measure- ments and models: Trends and challenges,”IEEE Veh. Technol. Mag., vol. 15, no. 4, pp. 22–32, Dec. 2020

    2020

  19. [19]

    Terahertz wireless channels: A holistic sur- vey on measurement, modeling, and analysis,

    C. Han, Y . Wu, Y . Liet al., “Terahertz wireless channels: A holistic sur- vey on measurement, modeling, and analysis,”IEEE Commun. Surveys Tuts., vol. 24, no. 3, pp. 1670–1707, 2022

    2022

  20. [20]

    Wireless communications and applications above 100 GHz: Opportunities and challenges for 6G and beyond,

    T. S. Rappaport, Y . Xing, O. Kanhereet al., “Wireless communications and applications above 100 GHz: Opportunities and challenges for 6G and beyond,”IEEE Access, vol. 7, pp. 78 729–78 757, Jun. 2019

    2019

  21. [21]

    Millimeter wave and sub-terahertz spatial statistical channel model for an indoor office building,

    S. Ju, Y . Xing, O. Kanhere, and T. S. Rappaport, “Millimeter wave and sub-terahertz spatial statistical channel model for an indoor office building,”IEEE J. Sel. Areas Commun., vol. 39, no. 6, pp. 1561–1575, Jun. 2021

    2021

  22. [22]

    A novel site-specific inference model for urban canyon channels: From measurements to modeling,

    J. Song, R. He, M. Yang, Z. Zhang, X. Chen, X. Zhang, and B. Ai, “A novel site-specific inference model for urban canyon channels: From measurements to modeling,” arXiv:2509.19275, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Measurement-based prediction of mmWave channel parameters using deep learning and point cloud,

    H. Mi, B. Ai, R. Heet al., “Measurement-based prediction of mmWave channel parameters using deep learning and point cloud,”IEEE Open J. Veh. Technol., vol. 5, pp. 1059–1072, 2024

    2024

  24. [24]

    Deep learning-based channel prediction in realistic vehicular communications,

    J. Joo, M.-C. Park, D. S. Han, and V . Pejovi ´c, “Deep learning-based channel prediction in realistic vehicular communications,”IEEE Access, vol. 7, pp. 27 846–27 858, 2019

    2019

  25. [25]

    ML-based massive MIMO channel prediction: Does it work on real-world data?

    M. K. Shehzad, L. Rose, S. Wesemann, and M. Assaad, “ML-based massive MIMO channel prediction: Does it work on real-world data?” IEEE Wireless Commun. Lett., vol. 11, no. 4, pp. 811–815, Apr. 2022

    2022

  26. [26]

    GAN-based massive MIMO channel model trained on measured data,

    F. Euchner, J. Sanzi, M. Henninger, and S. ten Brink, “GAN-based massive MIMO channel model trained on measured data,” inProc. 27th Int. ITG Workshop Smart Antennas (WSA), 2024, pp. 109–116

    2024

  27. [27]

    Generative network-based channel modeling and generation for air-to- ground communication scenarios,

    Y . Tian, H. Li, Q. Zhu, K. Mao, F. Ali, X. Chen, and W. Zhong, “Generative network-based channel modeling and generation for air-to- ground communication scenarios,”IEEE Commun. Lett., vol. 28, no. 4, pp. 892–896, Apr. 2024

    2024

  28. [28]

    Physics-informed generative modeling of wireless channels,

    B. B ¨ock, A. Oeldemann, T. Mayer, F. Rossetto, and W. Utschick, “Physics-informed generative modeling of wireless channels,” inProc. Int. Conf. Mach. Learn. (ICML), ser. PMLR, vol. 267, 2025, pp. 4602– 4626

    2025

  29. [29]

    How to bridge the sim-to-real gap in digital twin-aided telecommunication networks,

    C. Ruah, H. Sifaou, O. Simeone, and B. M. Al-Hashimi, “How to bridge the sim-to-real gap in digital twin-aided telecommunication networks,” arXiv preprint arXiv:2507.07067, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Learnable wireless digital twins: Reconstructing electromagnetic field with neural representations,

    S. Jiang, Q. Qu, X. Pan, A. K. Agrawal, R. Newcombe, and A. Alkha- teeb, “Learnable wireless digital twins: Reconstructing electromagnetic field with neural representations,”IEEE Open Journal of the Communi- cations Society, vol. 6, pp. 1568–1590, 2025

    2025

  31. [31]

    Calibrating wireless ray tracing for digital twinning using local phase error esti- mates,

    C. Ruah, O. Simeone, J. Hoydis, and B. M. Al-Hashimi, “Calibrating wireless ray tracing for digital twinning using local phase error esti- mates,”IEEE Transactions on Machine Learning in Communications and Networking, vol. 2, pp. 1193–1215, 2024

    2024

  32. [32]

    Site-specific location calibration and validation of ray-tracing simulator NYURay at upper mid-band frequencies,

    M. Ying, D. Shakya, P. Ma, G. Qian, and T. S. Rappaport, “Site-specific location calibration and validation of ray-tracing simulator NYURay at upper mid-band frequencies,”npj Wireless Technology, vol. 2, p. 8, 2026

    2026

  33. [33]

    Propagation-consistent wireless environment digital twin con- struction under sparse measurements,

    J. Ai, S. Xu, Y . Ren, Z. Liu, W. Chen, W. Tang, X. Li, C.-K. Wen, and S. Jin, “Propagation-consistent wireless environment digital twin con- struction under sparse measurements,”arXiv preprint arXiv:2605.22361, 2026

    Pith/arXiv arXiv 2026

  34. [34]

    Digital twin enabled site specific channel precoding: Over the air CIR inference,

    M. Haider, I. Ahmed, Z. Hassan, T. J. O’Shea, L. Liu, and D. B. Rawat, “Digital twin enabled site specific channel precoding: Over the air CIR inference,”IEEE Commun. Lett., vol. 29, no. 7, pp. 1559–1563, Jul. 2025

    2025

  35. [35]

    Wireless digital twin calibration: Refining DFT-domain channel information,

    H. Luo, S. R. Khosravirad, and A. Alkhateeb, “Wireless digital twin calibration: Refining DFT-domain channel information,”arXiv preprint arXiv:2603.16126, 2026

    Pith/arXiv arXiv 2026

  36. [36]

    Transfer learning enabled transformer-based generative adversarial networks for modeling and generating terahertz channels,

    Z. Hu, Y . Li, and C. Han, “Transfer learning enabled transformer-based generative adversarial networks for modeling and generating terahertz channels,”Commun. Eng., vol. 3, p. 153, 2024

    2024

  37. [37]

    Site-specific finetuning of neural receivers with real-world 5G NR measurements,

    N. B. Baytekin, R. Wiesmayr, S. Cammerer, C. Dick, and C. Studer, “Site-specific finetuning of neural receivers with real-world 5G NR measurements,” arXiv:2603.09644, 2026

    arXiv 2026

  38. [38]

    Ambi- guity in RMS delay spread of millimeter-wave channel measurements,

    R. Schulpen, U. Johannsen, A. Smolders, and L. A. Bronckers, “Ambi- guity in RMS delay spread of millimeter-wave channel measurements,” inProc. Eur. Conf. Antennas Propag. (EuCAP), Mar. 2023

    2023

  39. [39]

    Eravant Millimeter Wave Products & Components,

    Eravant, “Eravant Millimeter Wave Products & Components,” https:// www.eravant.com/, accessed: 2026-07-02

    2026

  40. [40]

    Digital- twin empowered site-specific radio resource management in 5G aerial corridor,

    P. Tarafder, I. Ahmed, D. B. Rawat, Z. Hassan, and K. Hasan, “Digital- twin empowered site-specific radio resource management in 5G aerial corridor,” inMILCOM 2025-2025 IEEE Military Communications Con- ference (MILCOM). IEEE, 2025, pp. 1–6

    2025

  41. [41]

    Villani,Optimal Transport: Old and New

    C. Villani,Optimal Transport: Old and New. Berlin, Germany: Springer, 2009

    2009

  42. [42]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProc. IEEE Int. Conf. Comput. Vis. (ICCV), 2017, pp. 2242–2251

    2017

  43. [43]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2016, pp. 770–778

    2016

  44. [44]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2016, pp. 694–711

    2016

  45. [45]

    Instance normalization: The missing ingredient for fast stylization,

    D. Ulyanov, A. Vedaldi, and V . Lempitsky, “Instance normalization: The missing ingredient for fast stylization,”arXiv preprint arXiv:1607.08022, 2016

    Pith/arXiv arXiv 2016

  46. [46]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inProc. Int. Conf. Learn. Represent. (ICLR), 2015

    2015

  47. [47]

    U-Net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- works for biomedical image segmentation,” inProc. Med. Image Com- put. Comput.-Assist. Intervent. (MICCAI), 2015, pp. 234–241

    2015

  48. [48]

    Robust estimation of a location parameter,

    P. J. Huber, “Robust estimation of a location parameter,”The Annals of Mathematical Statistics, vol. 35, no. 1, pp. 73–101, 1964

    1964