On-board AI-based Channel Estimation for LEO NTNs

Alessandro Vanelli-Coralli; Bruno De Filippo; Carla Amatetti; Mahdi Abdollahpour

REVIEW 4 major objections 4 minor 18 references

This paper claims that a lightweight, convolution-based neural channel estimator called MDELAN can outperform LS and LMMSE channel estimation for LEO satellite uplinks while cutting FLOPs by about 29x, making real-time onboard AI estimation

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

2026-08-02 00:03 UTC pith:KW466F6K

load-bearing objection Incremental but useful neural receiver work for LEO NTN; the MSE story is plausible, but the headline 29x FLOP advantage over LMMSE is not reproducible until the baseline implementation is specified. the 4 major comments →

arxiv 2607.15127 v1 pith:KW466F6K submitted 2026-07-16 eess.SP

On-board AI-based Channel Estimation for LEO NTNs

Mahdi Abdollahpour , Bruno De Filippo , Carla Amatetti , Alessandro Vanelli-Coralli This is my paper

classification eess.SP

keywords channel estimationLEO satellitesnon-terrestrial networksneural receiverMDELANOFDMMSEcomputational complexity

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.

This paper argues that AI channel estimation can be made light enough to run onboard a LEO satellite under strict power budgets. Its central claim is that a convolutional neural block called MDELAN, inserted into a model-driven neural receiver, estimates the uplink OFDM channel more accurately than standard LS and LMMSE methods in realistic rural, suburban, and urban satellite scenarios, while needing roughly 29x fewer floating-point operations than LMMSE and 27% fewer parameters than the earlier MDX model. The practical payoff is direct-to-satellite 6G links: accurate channel estimation without large onboard compute, memory, or statistical covariance knowledge. The authors position this as evidence that efficiency and accuracy do not have to be traded off when the architecture exploits NTN channel structure.

Core claim

The paper claims that a model-driven neural receiver, which first forms pilot-aided and data-aided least-squares channel estimates and then refines them with a new convolutional block called MDELAN, achieves lower channel-estimation MSE than LS and LMMSE baselines across rural, suburban, and urban LEO NTN scenarios. It also claims roughly 29x fewer FLOPs than LMMSE and about 27% fewer learnable parameters than the prior MDX model, while maintaining superior MSE performance. The authors interpret this as evidence that efficient, lightweight AI can deliver practical channel estimation on power-constrained spaceborne nodes.

What carries the argument

The load-bearing object is the MDELAN block: a stack of depthwise-separable convolutions with multiple progressive dilations and residual connections, inserted between the pilot/data LS estimates and the final channel output. Dilations enlarge the receptive field over the OFDM time-frequency grid without larger kernels; depthwise separation cuts multiply-accumulate cost; residual connections and per-resource-block learnable multipliers stabilize gradient flow and allow fine-tuning.

Load-bearing premise

The performance and complexity claims rest on the simulated LEO channels being representative of real S-band satellite propagation, and on the LMMSE baseline being implemented and counted with a fair, not unnecessarily costly, covariance computation.

What would settle it

Run the estimator against measured or worst-case S-band LEO channel data with stronger multipath and Doppler than the paper's simulator uses; if MDELAN's MSE advantage over LMMSE disappears or inverts, the accuracy claim fails. Separately, re-count LMMSE with a reduced-rank or approximate covariance implementation; if its FLOP total drops below the reported 29x gap, the complexity claim fails.

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

If this is right

Onboard inference becomes plausible for LEO payloads: the proposed estimator achieves LMMSE-class accuracy with a compute budget closer to LS.
The neural approach does not rely on second-order channel statistics, so it avoids covariance estimation and storage at the satellite.
A 27% parameter reduction versus MDX helps fit the model into memory-limited satellite processors.
The same architecture generalizes across rural, suburban, and urban scenarios with only optional fine-tuning, so one model can serve diverse coverage areas.
Better MSE on the PUSCH link translates into more reliable 6G direct-to-satellite connectivity under QPSK and constrained power.

Where Pith is reading between the lines

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

The 29x FLOP advantage is measured on the paper's simulated test grid; a hardware-aware implementation with fixed-point arithmetic or quantization might shrink or enlarge the gap, so a deployment-oriented complexity benchmark would be a natural next test.
Because LEO channels change slowly and have limited multipath richness, MDELAN's dilated time-frequency receptive field suggests it could also act as a channel predictor, potentially reducing pilot overhead—an extension the paper does not test.
The same block could plausibly transfer to downlink channel estimation or to other non-terrestrial platforms, though those settings would need their own validation.
The strongest real-world test would be against measured in-orbit S-band channel data; simulation-based training may underplay multipath or Doppler effects that appear in actual LEO propagation.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Incremental but useful neural receiver work for LEO NTN; the MSE story is plausible, but the headline 29x FLOP advantage over LMMSE is not reproducible until the baseline implementation is specified.

read the letter

You should know this paper is a solid, incremental engineering contribution, not a breakthrough. The authors extend their MDX neural receiver with a new block, MDELAN, that borrows the ELAN gradient-path idea and adds dilated depthwise-separable convolutions. Evaluated on QuaDRiGa-generated LEO NTN channels, the model consistently beats LS and often LMMSE on MSE, with fewer parameters than MDX. That part is credible as far as it goes.

The new architecture is legitimate and the evaluation scenario is realistic for the field. The paper does what it claims on the MSE side: the curves show MDX:MDELAN below LS and mostly below LMMSE, and the neural complexity formula (Eq. 7) is transparent. I believe the central MSE claim for the simulated setting.

The soft spots are real and concentrate in the complexity analysis and the baseline comparisons. The stress-test note is right: the 29x FLOP reduction versus LMMSE is not reproducible because the manuscript never says how LMMSE complexity is computed. Full-grid inversion over 1848 REs is orders of magnitude more expensive than a sliding-window Wiener filter, and the authors don't say which one they used. That is a load-bearing omission because the headline is about efficiency. They also don't compare against any other learned NTN estimators, which matters because the novelty claim is architectural, not result-based. And there are no error bars, so I can't tell whether the MSE gaps are stable or noise. The fine-tuning protocol is mentioned but not described, and the code isn't out yet, though they promise it.

None of these are fatal. They are addressable reporting gaps. The paper would benefit from a clear LMMSE complexity derivation, at least one other learned baseline, and confidence intervals on the MSE curves. The deployment language about real-time onboard inference should be softened or supported by actual hardware numbers.

Who is this for? Researchers working on neural receivers for satellite links, especially anyone worried about compute and memory budgets on a LEO payload. It's worth reading if you work in that area. For peer review, I'd send it out. The architecture is sensible, the evaluation is real, and the gaps are fixable. An editor should ask for the missing details and then it could be a decent conference paper.

Referee Report

4 major / 4 minor

Summary. The paper proposes a lightweight AI-based channel estimator for LEO non-terrestrial networks (NTN) based on a new block called Multi-Dilated Efficient Layer Aggregation Networks (MDELAN), which extends the authors' prior MDX neural receiver. The estimator operates on pilot-aided and data-aided least-squares channel estimates fused with a residual CNN. The authors evaluate MSE performance in rural, suburban, and urban LEO scenarios generated with QuaDRiGa, comparing against LS and LMMSE baselines. They claim superior MSE for MDELAN-based models (especially after fine-tuning), a 27% parameter reduction versus MDX, and approximately 29× fewer FLOPs than LMMSE. The central claim is that efficient, domain-aware AI models can achieve strong channel estimation accuracy under the compute and memory constraints of onboard satellite processing.

Significance. If substantiated, the work would be a useful contribution to 6G NTN receiver design, demonstrating that a compact, convolution-based estimator can outperform conventional LMMSE while being far cheaper to compute. The paper is grounded in realistic simulation with QuaDRiGa and uses a standard PUSCH framework, which lends credibility to the MSE comparisons. The authors also provide a practical complexity analysis framework (MACs/FLOPs/parameters) and commit to open-sourcing the implementation, both of which are strengths. However, the headline complexity claim is not reproducible as written because the LMMSE implementation is unspecified, and the best MSE results depend on a fine-tuning protocol that is not described. The comparison to prior learned NTN estimators is also missing, so the claimed state-of-the-art parameter efficiency is premature.

major comments (4)

[§IV-D, Fig. 6, Eq. (7)] The 29× FLOP reduction versus LMMSE is not verifiable because the LMMSE complexity is never defined. Eq. (7) only counts MACs for depthwise separable convolutions; no analogous expression is provided for LMMSE. The manuscript must state the LMMSE implementation: the size of the channel covariance matrix, the number of pilot REs, whether a full-grid inversion (O((F·S)^3)) or a sliding-window/approximate inversion is used, and how the 132×14 resource grid and pilots enter the count. Without this, the complexity ratio spans orders of magnitude and the headline claim is not reproducible.
[§III-D, §IV-B, §IV-C] The fine-tuning procedure is not specified. The paper states that fine-tuned MDX:MDELAN attains the lowest MSE in the rural scenario and that fine-tuning improves both neural models, yet no details are given: fine-tuning data (which scenarios, what split), learning rate, number of iterations, whether any layers are frozen, and how the model is initialized. Since the central MSE claims rely on the fine-tuned models, this omission prevents reproduction of the reported results and must be addressed.
[§IV-A–§IV-C] MSE curves are reported as single realizations without error bars, confidence intervals, or seed variations. Given that user speeds are drawn uniformly in [0,100] km/h and SNR is sampled over a range (Table I), the observed differences between models, especially at high SNR where curves may converge, could be within run-to-run variability. The authors should state whether the reported curves are averaged over a fixed test set and provide either multiple-seed standard deviations or a statistical test for the key comparisons (e.g., MDX:MDELAN vs. LMMSE).
[Abstract, §IV-D] The claim of '27% parameter efficiency improvement compared with state-of-the-art AI models' is an overstatement: the comparison is made only against the authors' own MDX baseline, not against prior learned NTN estimators such as [7]–[9] or other recent works. To support the state-of-the-art claim, the authors should either include such comparisons or rephrase the claim to be explicitly relative to MDX.

minor comments (4)

[§II-A, Eq. (3)] The LMMSE symbol estimate in Eq. (3) uses an estimated noise variance σ̂ but the paper does not describe how σ̂ is computed. For reproducibility, this should be specified.
[§III, Table I] The OFDM frame has S OFDM symbols and F subcarriers; Table I gives TTI length as 14 OFDM symbols and PRBs as 5 (train) / 11 (test), but F and S are not explicitly defined in the text. Please state F = 132 (or the actual value) and S = 14 explicitly.
[Abstract, §II] The abstract mentions evaluation 'under stringent NGSO power budgets,' but no power or energy model is presented in the manuscript. Either quantify the power budget or rephrase this to 'under lightweight computational constraints' to avoid overclaiming.
[General] Typos and grammar: 'accomodate' in §III, 'theLMMSE' in §IV-D, and inconsistent spacing in equations. Reference [18] is a GitHub URL that currently does not contain the promised implementation; please ensure the repository is publicly accessible before publication.

Circularity Check

0 steps flagged

No significant circularity: the results are empirical comparisons against external channel data and conventional baselines, with only descriptive self-citation.

full rationale

The paper's claims are empirical: MSE performance is obtained by training and testing neural estimators on QuaDRiGa-generated NTN channels and comparing with LS, LMMSE, and the prior MDX architecture. The training loss (Eq. 6) uses ground-truth channel samples from an external simulator, and the reported improvements are measured on held-out test configurations. No fitted parameter is renamed as a prediction, and no derivation reduces to its own input. The only self-citation is [10], which describes the MDX architecture; this is not load-bearing because MDX serves as a baseline and is also described in the paper, and the proposed MDELAN block is independently compared against it. The 27% parameter reduction and 29x FLOP reduction are arithmetic complexity comparisons, not fitted results. The unspecified LMMSE implementation is a reproducibility/comparability concern, not a circularity: it does not make the claimed complexity ratio true by construction within the paper's stated equations. Overall, the derivation chain is self-contained with respect to circularity, though external validity and implementation fairness are correctness risks.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim imports two large external objects: a simulator (QuaDRiGa) trusted to stand in for LEO propagation, and an LMMSE baseline whose complexity is not transparent. The neural weights and hyperparameters are fitted, which is normal for ML but limits the claims to the simulated training distribution.

free parameters (4)

MDX/MDELAN trainable weights = not reported (order 10^3 parameters; see Fig. 6)
MSE results are produced by networks trained on QuaDRiGa-generated channels; the learned weights are the primary source of the reported gains.
Training hyperparameters = lr=1e-3, regularization=0.01, 10,000 iterations, batch=128
Chosen by hand; no sensitivity analysis is reported, so MSE could depend on these choices.
MDELAN architecture hyperparameters (channel widths, dilation rates, kernel counts) = not disclosed
The block is described qualitatively and via Fig. 3, but the exact configuration used in simulation is not given, making the 27% parameter reduction and 29x FLOPs claims under-specified.
Fine-tuning protocol = not described
Fine-tuning is credited with the best rural results, but the dataset, iterations, and learning rate used for fine-tuning are not specified.

axioms (5)

domain assumption QuaDRiGa NTN channel realizations are representative of real LEO S-band propagation for rural, suburban, and urban scenarios.
Used in Sec. IV-A; all MSE claims are generated in this simulator, so external validity depends on this assumption.
domain assumption The LMMSE baseline is implemented with a fair and realistic covariance estimate.
Sec. II-A defines LMMSE as needing covariance and Sec. IV-D attributes roughly 125M FLOPs to it, but the covariance source and inversion method are not described, so the 29x FLOPs comparison is not auditable.
domain assumption Residual Doppler is negligible relative to subcarrier spacing.
Sec. II assumes this; if realistic LEO residual Doppler is larger, pilot placement and DA-LS estimates would behave differently.
domain assumption The OFDM signal model with QPSK and known pilots (Eq. 1) captures the link.
System model in Sec. II excludes synchronization errors, nonlinearities, interference, and hardware impairments.
standard math Standard mathematical background of LMMSE and OFDM estimation.
LMMSE formulas and OFDM signal processing are taken as known background from the cited literature.

pith-pipeline@v1.3.0-alltime-deepseek · 6097 in / 12339 out tokens · 168336 ms · 2026-08-02T00:03:32.292807+00:00 · methodology

0 comments

read the original abstract

Artificial Intelligence(AI) methods have shown strong channel estimation performance in terrestrial networks, but they typically rely on substantial computational resources. As 6G moves toward a unified architecture that will include Non-Terrestrial Networks (NTN) from day 0, availability of large and power hungry computational resources shall not be taken for granted. At the same time, NTN propagation often exhibits high predictability, limited multipath richness and significant Doppler shifts, representing a specific channel estimation problem. In this work, we propose a lightweight convolution-based channel estimator designed specifically for NTN operation and real-time onboard inference. We evaluate its channel estimation accuracy under stringent NGSO power budgets and quantify the resulting end-to-end impact on link performance. We show the improvement in terms of Mean Squared Error (MSE) achieved by the proposed approach compared with established algorithms, demonstrating that efficient AI models can deliver robust performance even on power-constrained spaceborne nodes. In addition, the proposed design by exploiting the domain knowledge, improves parameter efficiency by $27\%$ compared with state-of-the-art AI models and requires approximately $29\times$ fewer floating-point operations than conventional methods while achieving superior MSE performance.

Figures

Figures reproduced from arXiv: 2607.15127 by Alessandro Vanelli-Coralli, Bruno De Filippo, Carla Amatetti, Mahdi Abdollahpour.

**Figure 2.** Figure 2: MDX channel estimation block diagram residual connections. The residual connections have a perPhysical Resource Block (PRB) learnable multiplier. For more details please refer to [10]. We use normalized horizontal and vertical location of every RE as Positional Encoding (PE), P ∈ R 𝐹×𝑆×2 , where for each resource element ( 𝑓 , 𝑠), the two channels correspond to the normalized vertical and horizontal locat… view at source ↗

**Figure 3.** Figure 3: MDELAN Block Diagram diverse features and achieve higher parameter utilization efficiency. We use multiple progressive dilations in our MDELAN structure to capture the spectral and temporal correlations of the OFDM channel structure, increasing the receptive field without enlarging the kernels. To lower the complexity we use depthwise separable convolutions [16] throughout the structure. C. Neural Model T… view at source ↗

**Figure 4.** Figure 4: MSE vs. SNR performance in NTN rural scenario. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Figure 5: MSE vs. SNR performance in NTN suburban scenario. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

**Figure 6.** Figure 6: FLOPs and number of parameters D. Complexity Analysis All proposed neural blocks employ depthwise separable convolutions to reduce computational complexity. As a result, the overall complexity of the models is dominated by these operations. A depthwise separable convolution consists of a depthwise convolution followed by a pointwise (1 × 1) convolution. Assuming the input tensor have spatial dimensions 𝐹 ×… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 2 linked inside Pith

[1]

Design of a standard-compliant real-time neural receiver for 5G NR,

R. Wiesmayr, S. Cammerer, F. A ¨ıt Aoudia, J. Hoydis, J. Zakrzewski, and A. Keller, “Design of a standard-compliant real-time neural receiver for 5G NR,” inIEEE ICMLCN, 2025, pp. 1–6

2025
[2]

Deeprx: Fully convolutional deep learning receiver,

M. Honkala, D. Korpi, and J. M. Huttunen, “Deeprx: Fully convolutional deep learning receiver,”IEEE J. WCOM, vol. 20, no. 6, pp. 3925–3940, 2021

2021
[3]

Eqdeeprx: Learning a scalable mimo receiver,

M. Honkala, D. Korpi, E. Raninen, and J. M. Huttunen, “Eqdeeprx: Learning a scalable mimo receiver,”arXiv preprint arXiv:2602.11834, 2026

arXiv 2026
[4]

3GPP highlights issue 7,

3GPP, “3GPP highlights issue 7,” https://www .3gpp.org/images/ newsletters/3gpp highlights issue 7 web version.pdf, 2023

2023
[5]

Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,

ITU-R, “Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,” International Telecommunication Union, Tech. Rep., 2023

2030
[6]

The path to 5g-advanced and 6g non-terrestrial network systems,

A. Guidottiet al., “The path to 5g-advanced and 6g non-terrestrial network systems,” inASMS/SPSC Workshop. IEEE, 2022, pp. 1–8

2022
[7]

Channel prediction with temporal convolutional networks: A new paradigm to enable non- orthogonal multiple access in 6g ntn,

C. Amatetti, B. D. Filippo, and A. Vanelli-Coralli, “Channel prediction with temporal convolutional networks: A new paradigm to enable non- orthogonal multiple access in 6g ntn,”IEEE Commun. Stand. Mag., pp. 1–9, 2026

2026
[8]

Revolutionizing future connectivity: A contemporary survey on ai-empowered satellite-based non-terrestrial networks in 6g,

S. Mahboob and L. Liu, “Revolutionizing future connectivity: A contemporary survey on ai-empowered satellite-based non-terrestrial networks in 6g,”IEEE Commun. Surv. Tutor., vol. 26, no. 2, pp. 1279– 1321, 2024

2024
[9]

Enhanced 6g non-terrestrial network link performance using deep learning-based channel estimation and doppler compensation techniques,

C. J. Pawase, A. U. Rehman, and K. Chang, “Enhanced 6g non-terrestrial network link performance using deep learning-based channel estimation and doppler compensation techniques,”J. Commun. Netw., vol. 28, no. 1, pp. 73–85, 2026

2026
[10]

A compute&memory efficient model-driven neural 5G receiver for edge AI-assisted RAN,

M. Abdollahpour, M. Bertuletti, Y. Zhang, Y. Li, L. Benini, and A. Vanelli- Coralli, “A compute&memory efficient model-driven neural 5G receiver for edge AI-assisted RAN,” inIEEE GLOBECOM, 2025, pp. 5248–5253

2025
[11]

QuaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,

S. Jaeckel, L. Raschkowski, K. B ¨orner, and L. Thiele, “QuaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,”IEEE Trans. Antennas Propag., vol. 62, no. 6, pp. 3242–3256, 2014

2014
[12]

Solutions for NR to support Non-Terrestrial Networks (NTN),

3GPP, “Solutions for NR to support Non-Terrestrial Networks (NTN),” TR 38.821, March 2023, version 16.2.0

2023
[13]

Lmmse channel estimation in ofdm context: a review,

V. Savaux and Y. Lou ¨et, “Lmmse channel estimation in ofdm context: a review,”IET Signal Processing, vol. 11, no. 2, pp. 123–134, 2017

2017
[14]

Designing network design strategies through gradient path analysis,

C.-Y. Wang, H.-Y. M. Liao, and I.-H. Yeh, “Designing network design strategies through gradient path analysis,”arXiv preprint arXiv:2211.04800, 2022

Pith/arXiv arXiv 2022
[15]

YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in Proceedings of the IEEE/CVF conference on CVPR, 2023, pp. 7464–7475

2023
[16]

Xception: Deep learning with depthwise separable convolu- tions,

F. Chollet, “Xception: Deep learning with depthwise separable convolu- tions,” inIEEE Conference on CVPR, July 2017, pp. 1251–1258

2017
[17]

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

J. Hoydiset al., “Sionna: An open-source library for next- generation physical layer research,” 2023. [Online]. Available: https://arxiv.org/abs/2203.11854

Pith/arXiv arXiv 2023
[18]

Available: https://github.com/Mahdi-Abdollahpour/mdx

[Online]. Available: https://github.com/Mahdi-Abdollahpour/mdx

[1] [1]

Design of a standard-compliant real-time neural receiver for 5G NR,

R. Wiesmayr, S. Cammerer, F. A ¨ıt Aoudia, J. Hoydis, J. Zakrzewski, and A. Keller, “Design of a standard-compliant real-time neural receiver for 5G NR,” inIEEE ICMLCN, 2025, pp. 1–6

2025

[2] [2]

Deeprx: Fully convolutional deep learning receiver,

M. Honkala, D. Korpi, and J. M. Huttunen, “Deeprx: Fully convolutional deep learning receiver,”IEEE J. WCOM, vol. 20, no. 6, pp. 3925–3940, 2021

2021

[3] [3]

Eqdeeprx: Learning a scalable mimo receiver,

M. Honkala, D. Korpi, E. Raninen, and J. M. Huttunen, “Eqdeeprx: Learning a scalable mimo receiver,”arXiv preprint arXiv:2602.11834, 2026

arXiv 2026

[4] [4]

3GPP highlights issue 7,

3GPP, “3GPP highlights issue 7,” https://www .3gpp.org/images/ newsletters/3gpp highlights issue 7 web version.pdf, 2023

2023

[5] [5]

Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,

ITU-R, “Recommendation itu-r m.2160-0: Framework and overall objectives of the future development of imt for 2030 and beyond,” International Telecommunication Union, Tech. Rep., 2023

2030

[6] [6]

The path to 5g-advanced and 6g non-terrestrial network systems,

A. Guidottiet al., “The path to 5g-advanced and 6g non-terrestrial network systems,” inASMS/SPSC Workshop. IEEE, 2022, pp. 1–8

2022

[7] [7]

Channel prediction with temporal convolutional networks: A new paradigm to enable non- orthogonal multiple access in 6g ntn,

C. Amatetti, B. D. Filippo, and A. Vanelli-Coralli, “Channel prediction with temporal convolutional networks: A new paradigm to enable non- orthogonal multiple access in 6g ntn,”IEEE Commun. Stand. Mag., pp. 1–9, 2026

2026

[8] [8]

Revolutionizing future connectivity: A contemporary survey on ai-empowered satellite-based non-terrestrial networks in 6g,

S. Mahboob and L. Liu, “Revolutionizing future connectivity: A contemporary survey on ai-empowered satellite-based non-terrestrial networks in 6g,”IEEE Commun. Surv. Tutor., vol. 26, no. 2, pp. 1279– 1321, 2024

2024

[9] [9]

Enhanced 6g non-terrestrial network link performance using deep learning-based channel estimation and doppler compensation techniques,

C. J. Pawase, A. U. Rehman, and K. Chang, “Enhanced 6g non-terrestrial network link performance using deep learning-based channel estimation and doppler compensation techniques,”J. Commun. Netw., vol. 28, no. 1, pp. 73–85, 2026

2026

[10] [10]

A compute&memory efficient model-driven neural 5G receiver for edge AI-assisted RAN,

M. Abdollahpour, M. Bertuletti, Y. Zhang, Y. Li, L. Benini, and A. Vanelli- Coralli, “A compute&memory efficient model-driven neural 5G receiver for edge AI-assisted RAN,” inIEEE GLOBECOM, 2025, pp. 5248–5253

2025

[11] [11]

QuaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,

S. Jaeckel, L. Raschkowski, K. B ¨orner, and L. Thiele, “QuaDRiGa: A 3- D multi-cell channel model with time evolution for enabling virtual field trials,”IEEE Trans. Antennas Propag., vol. 62, no. 6, pp. 3242–3256, 2014

2014

[12] [12]

Solutions for NR to support Non-Terrestrial Networks (NTN),

3GPP, “Solutions for NR to support Non-Terrestrial Networks (NTN),” TR 38.821, March 2023, version 16.2.0

2023

[13] [13]

Lmmse channel estimation in ofdm context: a review,

V. Savaux and Y. Lou ¨et, “Lmmse channel estimation in ofdm context: a review,”IET Signal Processing, vol. 11, no. 2, pp. 123–134, 2017

2017

[14] [14]

Designing network design strategies through gradient path analysis,

C.-Y. Wang, H.-Y. M. Liao, and I.-H. Yeh, “Designing network design strategies through gradient path analysis,”arXiv preprint arXiv:2211.04800, 2022

Pith/arXiv arXiv 2022

[15] [15]

YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” in Proceedings of the IEEE/CVF conference on CVPR, 2023, pp. 7464–7475

2023

[16] [16]

Xception: Deep learning with depthwise separable convolu- tions,

F. Chollet, “Xception: Deep learning with depthwise separable convolu- tions,” inIEEE Conference on CVPR, July 2017, pp. 1251–1258

2017

[17] [17]

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

J. Hoydiset al., “Sionna: An open-source library for next- generation physical layer research,” 2023. [Online]. Available: https://arxiv.org/abs/2203.11854

Pith/arXiv arXiv 2023

[18] [18]

Available: https://github.com/Mahdi-Abdollahpour/mdx

[Online]. Available: https://github.com/Mahdi-Abdollahpour/mdx