Semantic-aware Digital Twin for AI-based CSI Acquisition

Jiajia Guo; Shi Jin; Yiming Cui

REVIEW 4 major objections 6 minor 15 references

Semantic-aware Digital Twin for AI-based CSI Acquisition

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper argues that a semantic-aware digital twin—a streamlined replica of the propagation environment—can supply the knowledge and training data that AI-based CSI acquisition lacks, cutting feedback overhead and data-collection cost…

desk verdict Useful taxonomy of semantic-DT-assisted CSI acquisition, honest about its limits, but the load-bearing distribution-preservation premise is asserted, not shown. read the letter →

arxiv 2506.21126 v1 pith:ZFDCK5MJ submitted 2025-06-26 cs.IT math.IT

Jiajia Guo , Yiming Cui , Shi Jin This is my paper · ORCID

classification cs.ITmath.IT

keywords semantic-awaredigitaltwinCSIacquisitionmassiveMIMOchannelstateinformationfeedbackhypernetworkraytracingtrainingdatageneration6Gairinterface

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

AI-based channel state information (CSI) acquisition in massive MIMO systems is held back by two problems: it leans only on radio-signal data, and it needs large real-world training datasets that are expensive to collect. This paper argues that a semantic-aware digital twin—a stripped-down virtual copy of the propagation environment that keeps only the features shaping the radio channel—can address both. In the first role, the twin is a fresh knowledge source: path parameters and environment indicators extracted from it improve channel estimation, feedback, and prediction. In the second role, the twin supports deployment by generating synthetic CSI samples for pre-training, or even by directly generating neural-network parameters through a hypernetwork, so that less real data and less retraining are needed. If the argument holds, 6G CSI acquisition can achieve lower feedback overhead and far lower data-collection cost.

What carries the argument

The load-bearing object is the semantic-aware digital twin replica: a discretized representation of the propagation environment—for instance a matrix in which walls, spatial boundaries, and open space take distinct values—built from scene graphs or 3D maps and paired with ray tracing to produce channel responses. It is 'semantic-aware' because it keeps only environmental features that affect channel distribution, trading full fidelity for efficiency. The second mechanism is the hypernetwork, a neural network that takes the twin's representation as input and outputs the weights and biases of a CSI reconstruction layer, which is how scene-specific adaptation happens without retraining the whole model. Together these carry the argument: the replica supplies channel statistics, and the hypernetwork converts those statistics into either training samples or network parameters.

What would settle it

Build a semantic-aware twin of one site, generate synthetic CSI by ray tracing at the same user positions where real CSI is measured, train two feedback networks—one on twin data only and one on real data—and compare reconstruction accuracy on a held-out real test set. If the twin-trained network does not beat a random-initialization baseline, or if closing the gap requires fine-tuning on nearly the full real dataset, the central transfer claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that what a CSI-acquisition neural network needs to learn—the channel distribution of a site—is essentially determined by the propagation environment and the user-position distribution. A semantic-aware digital twin is a compact replica that encodes exactly those semantics (wall layout, spatial boundaries, materials) and discards irrelevant detail, so it can stand in for the real environment in training. The paper sorts the resulting schemes into two classes: integration, where the twin supplies extra knowledge such as path parameters and environment indicators that feed into AI-based CSI acquisition; and deployment, where the twin generates training data (white-box) or neural-network parameters (black-box) to replace or reduce real data collection. A semi-black-box variant splits the acquisition network into a general pre-trained part and a scene-specific personalized part, with a hypernetwork producing only the personalized parameters from the twin's scene representation. The paper's claim is that this division addresses the two weaknesses it attributes to current AI-based CSI acquisition: reliance on single-modality information and the practical burden of dataset collection.

Load-bearing premise

The load-bearing premise is that a simplified, discretized replica of the environment preserves the real channel distribution closely enough that networks trained or parameterized in the digital world transfer to the physical world; the paper acknowledges that unpredictable propagation factors and unknown user-position distributions break this match, which is why real-data fine-tuning is still required.

Editorial extensions

If this is right

CSI feedback networks can be pre-trained on twin-generated channels, so real-world data collection is needed only to fine-tune on the samples the pre-trained model handles poorly.
A hypernetwork conditioned on the scene representation can generate a personalized reconstruction layer for each deployment, removing the need to retrain the whole network per environment.
Because the twin can expose slowly varying propagation conditions, pilot density and CSI feedback intervals can be enlarged, directly reducing acquisition overhead.
Multimodal information such as user position, vision, and path parameters can be fused with CSI in the acquisition network, improving estimation and prediction beyond radio-signal-only input.
Some real CSI remains necessary for fine-tuning since the simplified replica cannot capture every real-world factor; choosing which real samples to collect is posed as an open algorithmic problem.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct consequence of the paper's framing is that user-position distribution deserves modeling effort equal to geometry: if the distribution is wrong, the twin's channel distribution is wrong regardless of how accurate the map is.
The black-box idea becomes scalable only if the hypernetwork generates small adapters rather than full networks; the semi-black-box example already points in that direction.
The paper's efficiency argument implies a testable trade-off curve: how much of the replica's detail can be discarded before pre-training transfer degrades measurably.
The data-selection fine-tuning strategy would be sharper with a coverage metric over channel space, since the paper itself questions whether low-scoring samples are truly missing from the twin's dataset.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Read the semantic-aware DT for AI-based CSI acquisition paper. It is a coherent position/taxonomy article rather than a results paper. The useful part is the white-box/black-box/semi-black-box distinction for how a digital twin can feed an AI-based acquisition pipeline: generate training data, generate NN parameters, or generate only a small personalized part. That framing is genuinely helpful and could shape how people think about this space.

It does several things well. The paper is honest about its own limits — Section IV explicitly states that an imprecise replica would significantly compromise the proposed strategies, and Section III-B1 admits unavoidable mismatches from unpredictable propagation factors and unknown user distributions. The examples are drawn from the authors' prior work, which is fine as illustration. Nothing is presented as measured fact.

The soft spot is exactly the stress-test concern. The entire benefit story rests on the claim in Section III-B1 that channel distribution in the semantic-aware DT replica closely parallels that in the real environment. That sentence is doing all the load-bearing work for both the white-box pre-training/fine-tuning argument and the black-box hypernetwork parameter generation argument. In the white-box case, the claimed reduction in real data collection only materializes if a few fine-tuning samples close the gap; if the domain shift is large, fine-tuning approaches the cost of full real-data training. In the black-box case, there is no fine-tuning at all, so any systematic replica-reality discrepancy is baked straight into the deployed NN weights. The paper acknowledges the premise is shaky but offers no quantitative bound or small simulation to show the premise holds in some realistic setting.

Given the venue expectations, I would not reject it. A position paper can legitimately propose a research agenda without full validation. But the authors should be asked during review to label the transfer premise as an explicit conjecture rather than a stated fact, and ideally to include one modest experiment — even a ray-tracing vs. real-measured-channel comparison — to indicate the premise is not vacuous.

This paper is for researchers working on DT-assisted wireless or CSI feedback who want a structured map of options and open problems. It is not a breakthrough, but it is a fair and honest framework article. I'd send it to peer review with clear expectations that it is a vision paper, not an experimental one.

Referee Report

4 major / 6 minor

Summary. This position paper proposes using a semantic-aware digital twin (DT) to improve AI-based CSI acquisition in massive MIMO/6G systems. The authors first motivate the difficulty of CSI acquisition with large antenna arrays and the data-collection burden of AI-based approaches, then argue that a semantic-aware DT, which captures only environment semantics relevant to propagation, can serve both as an additional knowledge source for CSI acquisition and as a generator of training data or even neural network parameters. The proposed frameworks are organized into two classes: (i) integration of AI and semantic-aware DT for CSI acquisition, where DT-derived path parameters and environment indicators are extracted to aid channel estimation, feedback, or prediction; and (ii) deployment of AI-based CSI acquisition using DT, including white-box training-data generation with real-data fine-tuning and black-box/semi-black-box NN parameter generation via hypernetworks. The paper concludes with research challenges such as replica accuracy, user-position modeling, multimodal fusion, multi-DT collaboration, standardization, and privacy. The text is conceptual and contains no experiments, simulations, or derivations; its main illustrative architectures are drawn from the authors' prior work [10] and [12].

Significance. If the core premise is correct, the paper addresses a real and growing bottleneck: collecting enough real CSI samples to train and update CSI acquisition neural networks in practical 6G deployments. The proposed taxonomy, separating white-box data generation from black-box/semi-black-box parameter generation, is a useful organizational contribution, and the acknowledgement of replica mismatch and its consequences in Sections III-B1 and IV is honest. The paper also usefully highlights the distinction between building a maximally accurate DT and building an efficient semantic replica that only preserves channel-relevant information. However, the central quantitative claim—that a simplified semantic replica preserves the real-world channel distribution well enough to reduce data-collection overhead or to produce directly deployable NN weights—is not supported by any measurement, simulation, or bound. The manuscript is a plausible position statement rather than a validated proposal. The strengths are the clarity of the categorization, the constructive discussion of semi-black-box parameter generation, and the explicit enumeration of open problems.

major comments (4)

[III-B1] The load-bearing premise of the white-box deployment method is the sentence 'Given that the semantic-aware DT mirrors real-world CSI distribution, the channel distribution of the CSI within the semantic-aware DT replica closely parallels that in the real propagation environment.' This premise is asserted without evidence, and it directly implies that a small number of real fine-tuning samples suffices to close the gap. The paper itself immediately acknowledges in the same subsection that 'unavoidable mismatches' arise from unpredictable propagation factors and unknown user position distributions, and Section IV warns that 'an imprecise replica would significantly compromise the proposed DT-assisted strategies.' Please provide at least one quantitative validation: for example, a simulation study using a ray-tracing tool (e.g., Sionna RT) against a measured channel dataset, reporting a distribution-divergence metric (such as Wasserstein distance on channel gains or eigenvalue distributions) and the number of real samples needed to reach a target accuracy with and without DT-based pre-training. Without such evidence, the claimed reduction in data-collection overhead is not established.
[III-B2] In the black-box variant, the hypernetwork that generates NN parameters is itself trained on simulated or twin data. Consequently, any systematic discrepancy between the simplified semantic replica and the real environment is directly baked into the generated NN parameters, and in the pure black-box case there is no fine-tuning stage to correct this error. The paper's statement that 'the channel distribution has a mapping relationship with both the propagation environment and the user position distribution' is a plausible existence claim, but it does not justify the stronger claim that a hypernetwork trained on discretized scene graphs can approximate this mapping well enough for deployment in unseen real environments. The semi-black-box variant in Figure 6 has the same limitation for the personalized component. A concrete test would be to measure CSI feedback NMSE or channel estimation error on real indoor channels using parameters generated only from scene-graph inputs, and to compare against conventionally trained baselines. Absent such an experiment, the effectiveness of black-box parameter generation is an untested hypothesis rather than a demonstrated capability.
[III-A2] The claimed benefit of DT-aided knowledge extraction is stated only qualitatively. For example, 'the path information obtained from the semantic-aware replica can substantially enhance the performance of CSI acquisition' is not supported by any quantitative comparison to a system without side information. Since the paper is a position paper, this would be acceptable as a research direction, but the text asserts it as a likely outcome rather than as an open question. Please either present a small illustrative experiment (e.g., showing feedback accuracy as a function of auxiliary path information) or explicitly label such statements as hypotheses with the conditions under which they are expected to hold.
[Overall] The manuscript contains no experiments, simulations, or derivations. For a paper whose central contribution is a set of claims about reducing CSI feedback overhead and training-data costs in 6G, this is a load-bearing gap. Even for a position paper, the level of certainty in the text ('it makes logical sense', 'therefore, it is a natural idea') is higher than the evidence justifies. The authors should either add a validation study, or substantially reframe the paper as a taxonomy of open problems and clearly separate established results (e.g., from [10] and [12]) from speculative proposals. The current presentation risks leading readers to treat unverified architectures as ready-to-use solutions.

minor comments (6)

[Abstract] The sentence 'We categorizes the semantic-aware DT' contains a grammatical error; it should be 'We categorize'.
[III-B2] There is a typo in 'the hypernetwork generares the parameters'; it should be 'generates'.
[II-B] The definition of 'semantic-aware DT' would benefit from a more formal statement; currently it is described as a DT that 'captures environmental semantics critical to channel characteristics while maintaining high efficiency,' but the term 'semantics' is not defined.
[III-B1] The discussion of the unknown user position distribution cites [10] as assigning locations randomly in a 200m × 230m rectangle, but it does not report the quantitative findings of [10] regarding the resulting performance gap; a brief summary of the reported numbers would make the argument more concrete.
[Figure 1] Figure 1 is introduced in Section II-B and described as 'an AI-based CSI feedback framework enabled by a semantic-aware DT in [12]', but the full architecture involving the hypernetwork is only explained later in Section III-B2. Please add a forward reference or move the detailed figure description closer to its detailed discussion.
[References] Reference [12] is an arXiv preprint; if a peer-reviewed version has appeared by the time of publication, it should be cited instead of or in addition to the preprint.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's central transfer premise is an acknowledged assumption, and its illustrative claims come from prior work that is externally falsifiable rather than from an in-paper reduction.

full rationale

The paper is a speculative survey, not a derivation chain, so there is no equation-level reduction to exhibit. Section III-B1's sentence 'Given that the semantic-aware DT mirrors real-world CSI distribution, the channel distribution of the CSI within the semantic-aware DT replica closely parallels that in the real propagation environment' is exactly the transfer assumption on which the white-box benefit rests; it is not derived from the definition of semantic-aware DT (which is only said to capture 'environmental semantics critical to channel characteristics'), and the paper explicitly concedes in Section III-B1 'unavoidable mismatches' and in Section IV that 'an imprecise replica would significantly compromise the proposed DT-assisted strategies.' The Section III-B2 proposal to generate NN parameters directly from a DT world is conditional on the assumption that the determining information is integrated in the DT world ('which are assumed to be integrated within the semantic-aware DT world in this article'), so the conclusion is not presented as an empirical first-principles prediction. The self-cited [12] is used as an illustrative scene-graph example, but [12] is an externally falsifiable prior paper with its own experiments, and the DT-aided fine-tuning discussion also relies on [10] from an independent group; citing it does not make the present argument circular. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported. The appropriate finding is therefore no significant circularity.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or invented entities. Three domain assumptions carry the central argument; all are acknowledged as challenges in the paper's own conclusion section.

assumptions (3)

domain assumption The channel distribution of CSI within the semantic-aware digital twin replica closely parallels that in the real propagation environment.
This premise underlies the entire digital twin aided pre-training and fine-tuning approach; if replica and real distributions diverge, pre-training provides no benefit. Stated in Section III-B1.
domain assumption The channel distribution is determined by the propagation environment and the user position distribution, and the semantic-aware digital twin captures the relevant semantics of both.
Used to justify both white-box data generation and black-box parameter generation in Section III-B2. It assumes that simplification to key semantics preserves channel statistics.
domain assumption User position can be obtained accurately enough (via GNSS, vision-based localization, or other sensors).
Inputs to the knowledge extraction module and digital twin based data generation rely on accurate user positioning; the paper itself notes GNSS may be insufficient in Section III-A2.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semantic-aware Digital Twin for AI-based CSI Acquisition." pith.science (2026). https://pith.science/paper/ZFDCK5MJ

@misc{pith2026250621126,
  author       = {Pith},
  title        = {Pith review of: Semantic-aware Digital Twin for AI-based CSI Acquisition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZFDCK5MJ}},
  note         = {Machine review of arXiv:2506.21126}
}

read the original abstract

Artificial intelligence (AI) substantially enhances channel state information (CSI) acquisition performance but is limited by its reliance on single-modality information and deployment challenges, particularly in dataset collection. This paper investigates the use of semantic-aware digital twin (DT) to enhance AI-based CSI acquisition. We first briefly introduce the motivation and recent advancements in AI-driven CSI acquisition and semantic-aware DT employment for air interfaces. Then, we thoroughly explore how semantic-aware DT can bolster AI-based CSI acquisition. We categorizes the semantic-aware DT for AI-based CSI acquisition into two classes: enhancing AI-based CSI acquisition through integration with DT and using DT to aid AI-based CSI deployment. Potential integration frameworks are introduced in detail. Finally, we conclude by outlining potential research directions within the semantic-aware DT-assisted AI-based CSI acquisition.

Figures

Figures reproduced from arXiv: 2506.21126 by the authors.

**Figure 1.** A semi-black box approach for AI-based CSI feedback, as proposed in [12]. The hypernetwork processes the discretized scene graph to produce the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Illustration of integration of AI and semantic-aware DT for enhanced [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Illustration of deploying AI-based CSI acquisition through semantic-aware DT: semantic-aware-DT-based white-box training data generation. The [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Illustration of the refinement data selection proposed in [10]. CSI [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Illustration of semantic-aware-DT-based black-box training data [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: Semantic-aware-DT-based semi-black box training data generation, which entails segmenting the entire CSI acquisition NN into two components: [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 10 canonical work pages

[10]

Digital twin aided massive MIMO: CSI compression and feedback,

S. Jiang and A. Alkhateeb, “Digital twin aided massive MIMO: CSI compression and feedback,” in Proc. IEEE Int. Conf. Commun. (ICC) , 2024, pp. 1–6

work page 2024
[12]

AdapCsiNet: Environment-Adaptive CSI Feedback via Scene Graph-Aided Deep Learning

J. Liu, J. Guo, Y . Cui, C.-K. Wen, and S. Jin, “AdapCsiNet: Environment-adaptive CSI feedback via scene graph-aided deep learn- ing,” arXiv preprint arXiv:2504.10798 , 2025

work page Pith review arXiv 2025
[1]

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

ITU-R WP5D, “Framework and overall objectives of the future development of IMT for 2030 and beyond,” Tech. Rep., Jun. 2023, Accessed on Mar. 20, 2025. [Online]. Available: https: //www.itu.int/md/R19-WP5D-230612-TD-0905/en

work page 2023
[2]

Future technology trends of terrestrial international mobile telecommunications systems towards 2030 and beyond,

ITU-R M.2516-0, “Future technology trends of terrestrial international mobile telecommunications systems towards 2030 and beyond,” Tech. Rep., Nov. 2022, Accessed on Mar. 20, 2025. [Online]. Available: https: //www.itu.int/dms pub/itu-r/opb/rep/R-REP-M.2516-2022-PDF-E.pdf

work page 2022
[3]

Deep learning-based channel estimation,

M. Soltani, V . Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, “Deep learning-based channel estimation,” IEEE Commun. Lett. , vol. 23, no. 4, pp. 652–655, Apr. 2019

2019
[4]

Overview of deep learning- based CSI feedback in massive MIMO systems,

J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Overview of deep learning- based CSI feedback in massive MIMO systems,” IEEE Trans. Commun., vol. 70, no. 12, pp. 8017–8045, Dec. 2022

2022
[5]

Deep multimodal learning: Merging sensory data for massive MIMO channel prediction,

Y . Yang, F. Gao, C. Xing, J. An, and A. Alkhateeb, “Deep multimodal learning: Merging sensory data for massive MIMO channel prediction,” IEEE J. Sel. Areas Commun. , vol. 39, no. 7, pp. 1885–1898, Jul. 2021

2021
[6]

Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,

L. U. Khan, Z. Han, W. Saad, E. Hossain, M. Guizani, and C. S. Hong, “Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,” IEEE Commun. Surveys Tuts. , vol. 24, no. 4, pp. 2230– 2254, 4th Quart. 2022

work page 2022

Show all 15 references

[7]

6G digital twin networks: From theory to practice,

X. Lin, L. Kundu, C. Dick, E. Obiodu, T. Mostak, and M. Flaxman, “6G digital twin networks: From theory to practice,” IEEE Commun. Mag. , vol. 61, no. 11, pp. 72–78, Nov. 2023

2023
[8]

Digital twin based beam prediction: Can we train in the digital world and deploy in reality?

S. Jiang and A. Alkhateeb, “Digital twin based beam prediction: Can we train in the digital world and deploy in reality?” in Proc. IEEE Int. Conf. Commun. Workshops (ICC Workshops) , 2023, pp. 36–41

2023
[9]

Machine learning for CSI recreation in the digital twin based on prior knowledge,

B. Vilas Boas, W. Zirwas, and M. Haardt, “Machine learning for CSI recreation in the digital twin based on prior knowledge,” IEEE Open J. Commun. Soc. , vol. 3, pp. 1578–1591, 2022

2022
[11]

Prompt-enabled large AI models for CSI feedback,

J. Guo, Y . Cui, C.-K. Wen, and S. Jin, “Prompt-enabled large AI models for CSI feedback,” arXiv preprint arXiv:2501.10629 , 2025

2025 arXiv
[13]

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, Nov. 2023

2023
[14]

Sionna RT: Technical report,

F. A. Aoudia, J. Hoydis, M. Nimier-David, S. Cammerer, and A. Keller, “Sionna RT: Technical report,” arXiv preprint arXiv:2504.21719 , 2025

2025
[15]

Hypernetworks,

D. Ha, A. M. Dai, and Q. V . Le, “Hypernetworks,” in Proc. 5th Int. Conf. Learn. Representations (ICLR) , 2017, pp. 1–18

2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [10]

Digital twin aided massive MIMO: CSI compression and feedback,

S. Jiang and A. Alkhateeb, “Digital twin aided massive MIMO: CSI compression and feedback,” in Proc. IEEE Int. Conf. Commun. (ICC) , 2024, pp. 1–6

work page 2024

[2] [12]

AdapCsiNet: Environment-Adaptive CSI Feedback via Scene Graph-Aided Deep Learning

J. Liu, J. Guo, Y . Cui, C.-K. Wen, and S. Jin, “AdapCsiNet: Environment-adaptive CSI feedback via scene graph-aided deep learn- ing,” arXiv preprint arXiv:2504.10798 , 2025

work page Pith review arXiv 2025

[3] [1]

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

ITU-R WP5D, “Framework and overall objectives of the future development of IMT for 2030 and beyond,” Tech. Rep., Jun. 2023, Accessed on Mar. 20, 2025. [Online]. Available: https: //www.itu.int/md/R19-WP5D-230612-TD-0905/en

work page 2023

[4] [2]

Future technology trends of terrestrial international mobile telecommunications systems towards 2030 and beyond,

ITU-R M.2516-0, “Future technology trends of terrestrial international mobile telecommunications systems towards 2030 and beyond,” Tech. Rep., Nov. 2022, Accessed on Mar. 20, 2025. [Online]. Available: https: //www.itu.int/dms pub/itu-r/opb/rep/R-REP-M.2516-2022-PDF-E.pdf

work page 2022

[5] [3]

Deep learning-based channel estimation,

M. Soltani, V . Pourahmadi, A. Mirzaei, and H. Sheikhzadeh, “Deep learning-based channel estimation,” IEEE Commun. Lett. , vol. 23, no. 4, pp. 652–655, Apr. 2019

2019

[6] [4]

Overview of deep learning- based CSI feedback in massive MIMO systems,

J. Guo, C.-K. Wen, S. Jin, and G. Y . Li, “Overview of deep learning- based CSI feedback in massive MIMO systems,” IEEE Trans. Commun., vol. 70, no. 12, pp. 8017–8045, Dec. 2022

2022

[7] [5]

Deep multimodal learning: Merging sensory data for massive MIMO channel prediction,

Y . Yang, F. Gao, C. Xing, J. An, and A. Alkhateeb, “Deep multimodal learning: Merging sensory data for massive MIMO channel prediction,” IEEE J. Sel. Areas Commun. , vol. 39, no. 7, pp. 1885–1898, Jul. 2021

2021

[8] [6]

Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,

L. U. Khan, Z. Han, W. Saad, E. Hossain, M. Guizani, and C. S. Hong, “Digital twin of wireless systems: Overview, taxonomy, challenges, and opportunities,” IEEE Commun. Surveys Tuts. , vol. 24, no. 4, pp. 2230– 2254, 4th Quart. 2022

work page 2022