Baseband-Free End-to-End Communication System Based on Diffractive Deep Neural Network

Shi Jin; Wankai Tang; Xiaokun Teng; Xiao Li

REVIEW 4 major objections 6 minor 41 references

Baseband-Free End-to-End Communication System Based on Diffractive Deep Neural Network

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that modulation, beamforming, and detection can all be performed by trained stacks of passive metasurfaces, eliminating the digital baseband.

desk verdict A plausible new E2E D2NN architecture, but the unvalidated ASM-RSF equivalence and an incomplete hardware comparison undercut the headline claims. read the letter →

arxiv 2506.02411 v1 pith:XTQIMT3C submitted 2025-06-03 eess.SP cs.ITmath.IT

Xiaokun Teng , Wankai Tang , Xiao Li , Shi Jin This is my paper

classification eess.SPcs.ITmath.IT

keywords diffractivedeepneuralnetworkstackedintelligentmetasurfacesbaseband-freecommunicationend-to-endlearningangularspectrummethodreconfigurablesurfacewave-domainsignalprocessingbeamforming

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

This paper proposes a wireless link in which modulation, beamforming, and detection are all done by electromagnetic wave propagation through stacks of transmissive metasurfaces, with no digital baseband processing. The transmitter maps each symbol to an antenna subarray; the radiated field passes through a transmitter-side diffractive neural network (D2NN), a wireless channel, and a receiver-side D2NN, whose phase patterns are trained as a complex-valued autoencoder. The paper reports that this single-RF-chain, 1024-element design achieves symbol error rates comparable to a conventional maximum-ratio-transmission system using 81 RF chains, and that it is reliable at low signal-to-noise ratios and under Rician fading. The claim matters because it suggests physical-layer functions traditionally done in digital circuits could be offloaded to passive wave computing at the speed of light, lowering cost, power, and latency.

What carries the argument

The central object is the diffractive deep neural network (D2NN): a stack of transmissive metasurface layers, each represented by a diagonal matrix of unit-modulus complex transmission coefficients, through which electromagnetic waves propagate. Each layer's propagation is computed with the angular spectrum method, whose transfer function $h(f_x,f_z)=\exp\!\left(jk\sqrt{1-(\lambda f_x)^2-(\lambda f_z)^2}\,y\right)$ is applied in the Fourier domain via FFT, giving an equivalent propagation matrix $\mathbf{W}_{\mathrm{equ}} = \mathbf{F}^H \bar{\mathbf{H}}_D \mathbf{F}$ that reduces the per-layer diffraction cost from $O(N^2)$ to $O(N \log N)$. The transmitter and receiver D2NNs together form the autoencoder, with the correlated Rician channel as the differentiable bottleneck, and the detector's sub-array power comparison as the decision rule.

What would settle it

Simulate the trained 4-layer, 16 by 16 design with a full Rayleigh-Sommerfeld or full-wave propagator at 1 mm inter-layer spacing and compare the optimized phase patterns and symbol error rate against the angular-spectrum-trained versions; if the error rate degrades materially, the angular spectrum equivalence assumption is the weak link. A hardware experiment at 28 GHz measuring the power focused at the intended detector subarray would settle the same point.

Watch

Extended reading notes

Core claim

The central claim is that a baseband-free end-to-end transceiver can learn to communicate purely by diffraction: the transmitter-side D2NN encodes symbols into a structured radiated field, the receiver-side D2NN decodes that field back into a focused energy pattern, and detection is simply an argmax over sub-array powers. The phase shifts of each metasurface layer are optimized with mini-batch stochastic gradient descent and Wirtinger calculus to minimize cross-entropy loss over a correlated Rician channel. The headline quantitative result is that with four transmitter and four receiver layers of 16 by 16 elements, the system matches a 16-QAM maximum-ratio-transmission baseline that uses 81 RF chains, while using only one RF chain and 1024 passive elements, and it outperforms that baseline at higher SNR.

Load-bearing premise

All training and testing assume that the angular spectrum method reproduces what physically happens between metasurface layers spaced 1 mm apart at 28 GHz, including how evanescent waves decay; if that equivalence fails for such sub-wavelength spacing, the learned phase patterns may not produce the claimed performance on real hardware.

Editorial extensions

If this is right

If the claim is correct, digital baseband modules for modulation and detection are not strictly necessary; those functions can be embedded in trained phase profiles of passive layers.
A single RF chain and passive metasurfaces can provide beamforming and diversity gains that would ordinarily require dozens of active RF chains, cutting hardware cost and power.
The receiver can operate non-coherently with simple power detectors, removing the need for strict carrier phase synchronization.
Increasing the number of D2NN layers or elements per layer improves symbol error rate, so performance can be traded against hardware complexity.
High-rank, rich-scattering channels improve the learned system, meaning the architecture exploits multipath rather than being degraded by it.

Reading between the lines

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

A direct experimental test at 28 GHz with 1 mm inter-layer spacing would be the real arbiter: the simulation's angular spectrum model may not capture evanescent-wave coupling between closely spaced layers, so trained phases might need hardware-in-the-loop recalibration.
The same autoencoder trick could extend to multi-user or integrated sensing and communication by letting multiple symbols share the field, with detection regions assigned per user or per sensing task.
A useful stress test would be to train the same architecture with the Rayleigh-Sommerfeld formula and compare the optimized phase patterns with those from the angular spectrum method; material disagreement would show that the diffraction model, not the architecture, drives the result.
If fabrication tolerances in phase-shift resolution are included during training, the design could become robust enough to deploy with lower-cost metasurface hardware.

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

Here's my read on 2506.02411. The core idea is genuinely interesting: train a pair of diffractive neural networks—one at TX, one at RX—as a complex-valued autoencoder, using one-hot subarray selection as modulation and power comparison as detection, so the entire modulation-beamforming-detection chain runs in the wave domain. That specific combination is new relative to the prior SIM/D2NN work, which handled one module at a time. The simulations are thorough in exploring depth, width, training SNR, channel rank, and Rician factor, and the visualizations show the system learning to concentrate energy in the correct detector region. Credit where due: the ASM-based forward pass is a sensible speedup, and the idea itself is worth thinking about.

That said, there are two load-bearing soft spots. First, the paper relies on ASM for all training and testing but never validates that the discrete ASM matches the Rayleigh-Sommerfeld model at the simulated spacing (d_L = 1 mm = 0.093 λ). At that spacing, evanescent coupling between metasurface layers is not negligible, and the discrete DFT implementation without zero-padding—despite Remark 4 claiming it—raises doubts that the learned phase masks would transfer to physical hardware. This is not a cosmetic issue; it directly challenges the 'baseband-free' claim.

Second, the hardware comparison is unfair. The '1024 passive elements' counts only the TX D2NN; the RX D2NN also has 1024, plus the detector array. And the MRT baseline presumably has a conventional receiver, so the comparison mixes transceiver architectures. The abstract and Section V-D need to be reworded or the comparison redone.

There's also a sign error in the gradient formulas (34) and (36): the cross-term should be added, not subtracted. It doesn't affect the trained results since they use auto-diff, but it needs to be fixed.

The paper has no code, data, or error bars, which makes independent verification harder, but the core idea is plausible and the flaws are addressable. I'd send it to peer review with a request for major revision, primarily to validate the ASM-RSF equivalence numerically at the physical spacing and to redo the cost comparison. If the simulation model is right, this is a useful contribution to the SIM/D2NN literature; if not, the headline claim collapses. Worth reading for anyone working on wave-domain processing.

Recommendation: engage with it, but treat the performance numbers with caution until the diffraction model is verified.

Referee Report

4 major / 6 minor

Summary. The paper proposes BBF-E2E, a wireless communication system in which a transmitter-side diffractive deep neural network (D2NN) and a receiver-side D2NN jointly perform modulation, beamforming, and detection entirely through electromagnetic wave propagation. The system is trained as a complex-valued autoencoder with cross-entropy loss, using mini-batch SGD and Wirtinger calculus, and the angular spectrum method (ASM) is used to accelerate diffraction calculations. Simulation results show low SER under Rician fading and low-SNR conditions, and the authors claim that the system matches the performance of a conventional 81-RF-chain MRT scheme while requiring only a single RF chain and 1024 passive metasurface elements.

Significance. If the physical modeling is sound, the paper presents a conceptually interesting step toward wave-domain, baseband-free transceivers, with potential hardware-cost and latency advantages. The manuscript has several strengths: it formulates an end-to-end trainable architecture with explicit forward and backward equations, it identifies the computational bottleneck of diffraction calculation and proposes an FFT-based ASM alternative, and it provides extensive simulations over layer count, element count, training SNR, channel rank, and Rician factor. However, the central performance claim rests on the equivalence between discrete ASM and Rayleigh-Sommerfeld diffraction at sub-wavelength layer spacing, and this equivalence is asserted rather than validated. The gradient derivation also contains a sign error, and the hardware comparison with MRT undercounts the receiver-side D2NN and detector array. These issues are load-bearing for the paper's headline claims and need to be addressed before the results can be accepted as reliable.

major comments (4)

[Section IV-A, Eq. (31), Remark 4] The training and testing of all BBF-E2E networks use the discrete ASM, but the manuscript never validates that this discrete ASM reproduces the Rayleigh-Sommerfeld model at the simulated spacing d_L=1 mm=0.093 lambda and d_x=d_z=0.125 lambda. The equivalence invoked in the text is a continuous-domain statement; the implemented version in Eq. (31), W_equ=F^H diag(vec(H_bar)) F, is an N-point DFT-based circular convolution, and Remark 4's proposed zero-padding does not appear in the equations. At sub-wavelength spacing, evanescent components are not negligible, and the RSF kernel in Eqs. (4) and (13) contains a 1/(2 pi r) term that is not identical to the ASM exponential decay e^{-gamma y}. The authors should provide a quantitative comparison between discrete ASM and discrete RSF for the exact array sizes, spacings, and layer distances used in Figs. 6-10, and clarify whether zero-padding was actually used in training. Without this validation, the learned phase masks may exploit ASM-specific artifacts, and the SER curves do not transfer to a physical implementation.
[Section IV-B, Eqs. (34) and (36)] The cross-entropy gradient formulas have a sign error. For L=-log p'_m with p'_m=softmax(p)_m, the derivative is dL/dpsi = (p'_m - 1) dp_m/dpsi + sum_{tilde m != m} p'_tilde m dp_tilde m/dpsi. The second sum should be added, not subtracted. The minus signs in Eqs. (34) and (36) are inconsistent with the stated loss in Eq. (19) and with the update rule in Eq. (38), which uses gradient descent. If these formulas are used for training, the cross-entropy is not minimized for the cross terms. If, as stated later in the text, gradients are instead computed by automatic differentiation, the manuscript should say so explicitly and correct or remove the erroneous closed-form expressions, since they are presented as the derivation of the training algorithm.
[Section V-D and Abstract] The comparison with the 81-RF-chain MRT scheme undercounts the hardware of BBF-E2E. The configuration used for the headline result, L_TX=L_RX=4 and N_x=N_z=16, contains 1024 passive elements in the TX-D2NN and another 1024 passive elements in the RX-D2NN, for a total of 2048 passive phase-shifting elements, in addition to the 16-element feed subarrays and the 256-element receiving detector array. The claim that the system requires 'only a single RF chain and 1024 passive elements' omits the entire receiver-side D2NN and detector. The MRT baseline is also not fully specified: the number of receive antennas, the RF-chain count at the receiver, and the SNR definition should be stated so that the comparison is a fair end-to-end comparison. The abstract and Section V-D should be revised to reflect the total hardware count and the precise baseline configuration.
[Section V-A and Fig. 3] The training procedure introduces batch normalization layers that are removed during validation and testing 'to match the hardware constraint.' This creates a train-test mismatch: the forward model used for gradient updates includes minibatch-dependent normalization, while the SER results are generated from a forward model without BN. The manuscript does not explain how the trained phase values are adapted for BN-free inference, nor does it quantify the effect of removing BN. Since the learned phase masks must work in the physical system, which has no BN, the authors should either train without BN, fold the BN parameters into the phase values using a principled procedure, or provide an ablation study showing that the reported SER is not an artifact of the training-only BN.

minor comments (6)

[Section V-B] The phrase 'computating-by-propagation' should be corrected to 'computing-by-propagation'.
[Section III] The notation N={1,2,...,N} uses the same symbol for the set and its cardinality; using a different symbol for the index set would improve readability.
[Section IV-A, Eq. (29)] The discrete frequency normalization in the transfer function h_bar(bar n_x, bar n_z) is not explicitly defined; the mapping from continuous spatial frequencies f_x, f_z to the DFT indices bar n_x, bar n_z should be stated.
[Section V-D] The statement that BBF-E2E 'outperforms MRT in higher SNR conditions' should be accompanied by the confidence intervals or the number of Monte Carlo runs, since the difference appears to be within one order of magnitude in Fig. 7.
[Section V-C] The training SNR of -20 dB is chosen as optimal in Fig. 8, but the main comparison in Fig. 7 uses a training SNR of -10 dB; the authors should justify this choice and report how sensitive the conclusions are to the training SNR.
[Figure 3] The figure caption spells 'Rayleigh-Sommerfield' with one 'f'; the standard spelling is 'Rayleigh-Sommerfeld'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the BBF-E2E phase masks are learned and evaluated on held-out data, the ASM/RSF equivalence is an external textbook claim, and author self-citations are background only.

full rationale

The paper's central claim—that the trained D2NN transceiver matches a conventional 81-RF-chain MRT system with one RF chain and 1024 passive elements—is an empirical simulation result, not a quantity forced by a fitted parameter or by construction. The phase shifts are optimized with SGD on the cross-entropy loss (P1) and tested on fresh samples over a range of SNRs; no equation in the paper defines the testing SER in terms of the trained phases in a way that makes the reported performance tautological. The closest candidate for circularity is the substitution of ASM for RSF in Section IV-A. The paper asserts 'Since both RSF and ASM are derived from the Helmholtz equation under the same physical assumptions, they are theoretically equivalent [38],' but this equivalence is cited to Goodman's external textbook, not to the authors' own work, and it is a modeling assumption about the forward diffraction simulation rather than a derivation that reduces the output to its input. Whether ASM faithfully reproduces RSF at d_L = 1 mm and lambda = 10.7 mm is a real model-fidelity question, but it is a correctness or validation concern, not circularity, because training and testing both use the same ASM operator. The self-citations in the paper—references [5], [7], and [12], which involve the present authors—are used only as background on RIS and transmissive RIS architectures; none of them supplies an essential premise, uniqueness theorem, or ansatz on which the central performance claim depends. The training-SNR study in Fig. 8 is a hyperparameter sensitivity analysis; the choice of a favorable training SNR is not a fitted parameter that is later renamed as a prediction. The modulator's antenna-subarray mapping resembles spatial modulation, but the paper's contribution is the jointly learned wave-domain encoder/decoder, and it does not present that known modulation pattern as a derived first-principles result. Overall, the derivation chain is self-contained and independently evaluable from the simulation, so no circular step is present.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the trainable phase parameters, a handful of hand-chosen geometric and channel parameters, and five idealizations about diffraction modeling, metasurface hardware, and channel statistics. No new physical entities or forces are introduced.

free parameters (6)

TX-D2NN phase shifts beta_l,n = trained via SGD
All trainable transmissive phase coefficients in TX D2NN; define the learned modulation and beamforming.
RX-D2NN phase shifts gamma_l,n = trained via SGD
All trainable phase coefficients in RX D2NN; define the learned detection.
Inter-layer spacing d_L = 1 mm
Hand-chosen; sub-wavelength relative to lambda=10.7 mm, affecting evanescent wave contributions.
Element spacing d_x, d_z = 0.125 lambda
Hand-chosen; determines aperture size and the ASM spatial filtering passband.
Training SNR = -10 dB in Fig. 7; optimal near -20 dB in Fig. 8
Hand-chosen or tuned post hoc based on testing SER; affects what the network learns.
Rician K-factor K_R = 0 dB (default)
Hand-chosen channel parameter; performance strongly depends on it per Fig. 10.

assumptions (5)

domain assumption Scalar diffraction theory (Helmholtz equation) models EM propagation in the D2NN.
Used in Sections III and IV to derive RSF and ASM; ignores polarization and vector diffraction.
ad hoc to paper Discrete ASM with zero-padding accurately reproduces the discrete RSF at sub-wavelength spacing.
All training and testing use ASM (Section IV-A). The paper provides no numerical comparison of the two at d_L=1 mm and lambda=10.7 mm.
domain assumption Metasurface elements are phase-only, unit-modulus, independently reconfigurable, lossless, and uncoupled.
Section III models phi=e^{j beta}; real transmissive metasurfaces have amplitude loss, quantization, and inter-element coupling.
domain assumption Correlated Rician fading with sinc-based Kronecker correlation models the wireless channel.
Section III Eqs. (7)-(12); standard but idealized. The sinc correlation function (12) is chosen, not measured.
domain assumption Training and testing on samples from the same channel distribution yields deployment-ready phases.
All evaluation uses the same statistical model; no out-of-distribution or time-varying channel tests are performed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Baseband-Free End-to-End Communication System Based on Diffractive Deep Neural Network." pith.science (2026). https://pith.science/paper/XTQIMT3C

@misc{pith2026250602411,
  author       = {Pith},
  title        = {Pith review of: Baseband-Free End-to-End Communication System Based on Diffractive Deep Neural Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTQIMT3C}},
  note         = {Machine review of arXiv:2506.02411}
}

read the original abstract

Diffractive deep neural network (D2NN), also referred to as reconfigurable intelligent metasurface based deep neural networks (Rb-DNNs) or stacked intelligent metasurfaces (SIMs) in the field of wireless communications, has emerged as a promising signal processing paradigm that enables computing-by-propagation. However, existing architectures are limited to implementing specific functions such as precoding and combining, while still relying on digital baseband modules for other essential tasks like modulation and detection. In this work, we propose a baseband-free end-to-end (BBF-E2E) wireless communication system where modulation, beamforming, and detection are jointly realized through the propagation of electromagnetic (EM) waves. The BBF-E2E system employs D2NNs at both the transmitter and the receiver, forming an autoencoder architecture optimized as a complex-valued neural network. The transmission coefficients of each metasurface layer are trained using the mini-batch stochastic gradient descent method to minimize the cross-entropy loss. To reduce computational complexity during diffraction calculation, the angular spectrum method (ASM) is adopted in place of the Rayleigh-Sommerfeld formula. Extensive simulations demonstrate that BBF-E2E achieves robust symbol transmission under challenging channel conditions with significantly reduced hardware requirements. In particular, the proposed system matches the performance of a conventional multi-antenna system with 81 RF chains while requiring only a single RF chain and 1024 passive elements of metasurfaces. These results highlight the potential of wave-domain neural computing to replace digital baseband modules in future wireless transceivers.

Figures

Figures reproduced from arXiv: 2506.02411 by the authors.

**Figure 2.** Detailed illustration of partial components of BBF-E2E. (a) The modulator and layer 1 of TX-D [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The learning framework of BBF-E2E system. The EM wave propagation in the upper half is abstracted as the neural [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: The process of diffracted field calculation with ASM, which exploits the equivalence between spatial diffraction and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Forward propagation process of the BBF-E2E system. The heatmaps illustrate the amplitude distributions of EM waves [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Impact of the number of layers and elements on the performance of BBF-E2E. (a) Training loss. (b) Testing SER. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: SER performance of the BBF-E2E system under differ [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

**Figure 8.** Figure 8: SER performance of the BBF-E2E system versus [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 10.** Figure 10: SER performance of the BBF-E2E system versus [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

**Figure 11.** Figure 11: Spatial filtering effect with ASM. The amplitude of [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 33 canonical work pages

[1]

On limits of wireless communications in a fading environment when using multiple antennas,

G. Foschini and M. Gans, “On limits of wireless communications in a fading environment when using multiple antennas,”Wirel. Pers. Commun., vol. 6, no. 3, pp. 311–335, Mar. 1998

work page 1998
[2]

Noncooperative cellular wireless with unlimited num- bers of base station antennas,

T. L. Marzetta, “Noncooperative cellular wireless with unlimited num- bers of base station antennas,”IEEE Trans. Wirel. Commun., vol. 9, no. 11, pp. 3590–3600, Nov. 2010

work page 2010
[3]

Smart radio environments empowered by reconfig- urable intelligent surfaces: How it works, state of research, and the road ahead,

M. Di Renzoet al., “Smart radio environments empowered by reconfig- urable intelligent surfaces: How it works, state of research, and the road ahead,”IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2450–2525, Nov. 2020

2020
[4]

Intelligent reflecting surface-aided wireless communications: A tutorial,

Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelligent reflecting surface-aided wireless communications: A tutorial,”IEEE Trans. Commun., vol. 69, no. 5, pp. 3313–3351, 2021

2021
[5]

Wireless communications with reconfigurable intelligent surface: Path loss modeling and experimental measurement,

W. Tanget al., “Wireless communications with reconfigurable intelligent surface: Path loss modeling and experimental measurement,”IEEE Trans. Wirel. Commun., vol. 20, no. 1, pp. 421–439, 2020

work page 2020
[6]

Reconfigurable intelligent surfaces for energy efficiency in wireless communication,

C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in wireless communication,”IEEE Trans. Wirel. Commun., vol. 18, no. 8, pp. 4157–4170, Aug. 2019

work page 2019
[7]

Large intelli- gent surface-assisted wireless communication exploiting statistical CSI,

Y . Han, W. Tang, S. Jin, C.-K. Wen, and X. Ma, “Large intelli- gent surface-assisted wireless communication exploiting statistical CSI,” IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 8238–8242, Aug. 2019

work page 2019
[8]

Simultaneously transmitting and reflecting (STAR) RIS aided wireless communications,

X. Mu, Y . Liu, L. Guo, J. Lin, and R. Schober, “Simultaneously transmitting and reflecting (STAR) RIS aided wireless communications,” IEEE Trans. Wirel. Commun., vol. 21, no. 5, pp. 3083–3098, May 2022

work page 2022

Show all 41 references

[9]

Active RIS vs. passive RIS: Which will prevail in 6G?

Z. Zhanget al., “Active RIS vs. passive RIS: Which will prevail in 6G?” IEEE Trans. Commun., vol. 71, no. 3, pp. 1707–1725, Mar. 2023

2023
[10]

A wideband amplifying and filtering reconfigurable intelligent surface for wireless relay,

L. Wuet al., “A wideband amplifying and filtering reconfigurable intelligent surface for wireless relay,” Dec. 2024. [Online]. Available: http://arxiv.org/abs/2501.09759

2024 arXiv
[11]

Intelligent adaptive metasurface in complex wireless environments,

H. Q. Yanget al., “Intelligent adaptive metasurface in complex wireless environments,” Nov. 2024. [Online]. Available: http://arxiv.org/ abs/2411.08538

2024 arXiv
[12]

Wireless communications with programmable meta- surface: New paradigms, opportunities, and challenges on transceiver design,

W. Tanget al., “Wireless communications with programmable meta- surface: New paradigms, opportunities, and challenges on transceiver design,”IEEE Wirel. Commun., vol. 27, no. 2, pp. 180–187, Apr. 2020

2020
[13]

Toward transmissive RIS transceiver enabled uplink com- munication systems: Design and optimization,

Z. Liet al., “Toward transmissive RIS transceiver enabled uplink com- munication systems: Design and optimization,”IEEE Internet Things J., vol. 11, no. 4, pp. 6788–6801, Feb. 2024

2024
[14]

All-optical machine learning using diffractive deep neural networks,

X. Linet al., “All-optical machine learning using diffractive deep neural networks,”Science, vol. 361, no. 6406, pp. 1004–1008, Sep. 2018

2018
[15]

Interplay between RIS and AI in wireless commu- nications: Fundamentals, architectures, applications, and open research problems,

J. Wanget al., “Interplay between RIS and AI in wireless commu- nications: Fundamentals, architectures, applications, and open research problems,”IEEE J. Sel. Areas Commun., vol. 39, no. 8, pp. 2271–2288, Aug. 2021

2021
[16]

Stacked intelligent metasurfaces for efficient holographic MIMO communications in 6G,

J. Anet al., “Stacked intelligent metasurfaces for efficient holographic MIMO communications in 6G,”IEEE J. Sel. Areas Commun., vol. 41, no. 8, pp. 2380–2396, Aug. 2023

2023
[17]

AI and memory wall,

A. Gholamiet al., “AI and memory wall,”IEEE Micro, vol. 44, no. 3, pp. 33–39, May 2024

2024
[18]

Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit,

T. Zhouet al., “Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit,”Nat. Photonics, vol. 15, no. 5, pp. 367–373, May 2021

2021
[19]

A programmable diffractive deep neural network based on a digital-coding metasurface array,

C. Liuet al., “A programmable diffractive deep neural network based on a digital-coding metasurface array,”Nat. Electron., vol. 5, no. 2, pp. 113–122, Feb. 2022

2022
[20]

All-analog photoelectronic chip for high-speed vision tasks,

Y . Chenet al., “All-analog photoelectronic chip for high-speed vision tasks,”Nature, vol. 623, no. 7985, pp. 48–57, Nov. 2023

2023
[21]

Super-resolution diffractive neural network for all-optical direction of arrival estimation beyond diffraction limits,

S. Gaoet al., “Super-resolution diffractive neural network for all-optical direction of arrival estimation beyond diffraction limits,”Light: Sci. Appl., vol. 13, p. 161, Jul. 2024

2024
[22]

Classification of metal handwritten digits based on microwave diffractive deep neural network,

Z. Gu, Q. Ma, X. Gao, J. W. You, and T. J. Cui, “Classification of metal handwritten digits based on microwave diffractive deep neural network,” Adv. Opt. Mater., vol. 12, no. 7, p. 2301938, 2024

2024
[23]

Diffractive deep neural network for microwave imaging based on a multi-layer transmitarray an- tenna,

D. Liu, T. Shan, M. Li, F. Yang, and S. Xu, “Diffractive deep neural network for microwave imaging based on a multi-layer transmitarray an- tenna,” in2023 International Applied Computational Electromagnetics Society Symposium (ACES-China), Hangzhou, China, 2023, pp. 1–3

2023
[24]

Inference in artificial intelligence with deep optics and photonics,

G. Wetzsteinet al., “Inference in artificial intelligence with deep optics and photonics,”Nature, vol. 588, no. 7836, pp. 39–47, Dec. 2020

2020
[25]

Diffractive deep neural networks: Theories, optimiza- tion, and applications,

H. Chenet al., “Diffractive deep neural networks: Theories, optimiza- tion, and applications,”Applied Physics Reviews, vol. 11, no. 2, p. 021332, Jun. 2024

2024
[26]

Efficient beamforming and radiation pattern control using stacked intelligent metasurfaces,

N. U. Hassan, J. An, M. Di Renzo, M. Debbah, and C. Yuen, “Efficient beamforming and radiation pattern control using stacked intelligent metasurfaces,”IEEE Open J. Commun. Soc., vol. 5, pp. 599–611, 2024

2024
[27]

Stacked intelligent metasurfaces for multiuser beamforming in the wave domain,

J. An, M. Di Renzo, M. Debbah, and C. Yuen, “Stacked intelligent metasurfaces for multiuser beamforming in the wave domain,” inICC 2023 - IEEE International Conference on Communications, Rome, Italy, 2023, pp. 2834–2839

2023
[28]

Two-dimensional direction-of-arrival estimation using stacked intelligent metasurfaces,

J. Anet al., “Two-dimensional direction-of-arrival estimation using stacked intelligent metasurfaces,”IEEE J. Sel. Areas Commun., vol. 42, no. 10, pp. 2786–2802, Oct. 2024

2024
[29]

Multi-user ISAC through stacked intelligent metasur- faces: New algorithms and experiments,

Z. Wanget al., “Multi-user ISAC through stacked intelligent metasur- faces: New algorithms and experiments,” inGLOBECOM 2024 - 2024 IEEE Glob. Commun. Conf., Cape Town, South Africa, 2024, pp. 4442– 4447

2024
[30]

Channel estimation for stacked intelligent metasurface-assisted wireless networks,

X. Yao, J. An, L. Gan, M. Di Renzo, and C. Yuen, “Channel estimation for stacked intelligent metasurface-assisted wireless networks,”IEEE Wirel. Commun. Lett., vol. 13, no. 5, pp. 1349–1353, May 2024

2024
[31]

Stacked intelligent metasurfaces for task-oriented semantic communications,

G. Huanget al., “Stacked intelligent metasurfaces for task-oriented semantic communications,”IEEE Wirel. Commun. Lett., vol. 14, no. 2, pp. 310–314, Feb. 2025

2025
[32]

An introduction to deep learning for the physical layer,

T. O’Shea and J. Hoydis, “An introduction to deep learning for the physical layer,”IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 4, pp. 563–575, Dec. 2017

2017
[33]

Deep learning based communication over the air,

S. D ¨orner, S. Cammerer, J. Hoydis, and S. ten Brink, “Deep learning based communication over the air,”IEEE J. Sel. Top. Signal Process., vol. 12, no. 1, pp. 132–143, Feb. 2018

2018
[34]

AI empowered wireless communications: From bits to semantics,

Z. Qinet al., “AI empowered wireless communications: From bits to semantics,”Proc. IEEE, vol. 112, no. 7, pp. 621–652, Jul. 2024

2024
[35]

Deep learning based end-to-end wireless communication systems without pilots,

H. Ye, G. Y . Li, and B.-H. Juang, “Deep learning based end-to-end wireless communication systems without pilots,”IEEE Trans. Cogn. Commun. Netw., vol. 7, no. 3, pp. 702–714, Sep. 2021

2021
[36]

End-to-end learning for OFDM: From neural receivers to pilotless communication,

F. Ait Aoudia and J. Hoydis, “End-to-end learning for OFDM: From neural receivers to pilotless communication,”IEEE Trans. Wirel. Com- mun., vol. 21, no. 2, pp. 1049–1063, Feb. 2022

2022
[37]

Deep learning-based end-to-end wireless communication systems with conditional GANs as unknown channels,

H. Ye, L. Liang, G. Y . Li, and B.-H. Juang, “Deep learning-based end-to-end wireless communication systems with conditional GANs as unknown channels,”IEEE Trans. Wirel. Commun., vol. 19, no. 5, pp. 3133–3143, May 2020

2020
[38]

J. W. Goodman,Introduction to Fourier Optics. Roberts and Company publishers, 2005

2005
[39]

R. W. Heath Jr and A. Lozano,Foundations of MIMO Communication, 1st ed. Cambridge University Press, Dec. 2018

2018
[40]

Complex-valued neural networks: A comprehensive survey,

C. Lee, H. Hasegawa, and S. Gao, “Complex-valued neural networks: A comprehensive survey,”IEEE/CAA J. Autom. Sin., vol. 9, no. 8, pp. 1406–1426, Aug. 2022

2022
[41]

Maximum ratio transmission,

T. Lo, “Maximum ratio transmission,”IEEE Trans. Commun., vol. 47, no. 10, pp. 1458–1461, Oct. 1999

1999

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

On limits of wireless communications in a fading environment when using multiple antennas,

G. Foschini and M. Gans, “On limits of wireless communications in a fading environment when using multiple antennas,”Wirel. Pers. Commun., vol. 6, no. 3, pp. 311–335, Mar. 1998

work page 1998

[2] [2]

Noncooperative cellular wireless with unlimited num- bers of base station antennas,

T. L. Marzetta, “Noncooperative cellular wireless with unlimited num- bers of base station antennas,”IEEE Trans. Wirel. Commun., vol. 9, no. 11, pp. 3590–3600, Nov. 2010

work page 2010

[3] [3]

Smart radio environments empowered by reconfig- urable intelligent surfaces: How it works, state of research, and the road ahead,

M. Di Renzoet al., “Smart radio environments empowered by reconfig- urable intelligent surfaces: How it works, state of research, and the road ahead,”IEEE J. Sel. Areas Commun., vol. 38, no. 11, pp. 2450–2525, Nov. 2020

2020

[4] [4]

Intelligent reflecting surface-aided wireless communications: A tutorial,

Q. Wu, S. Zhang, B. Zheng, C. You, and R. Zhang, “Intelligent reflecting surface-aided wireless communications: A tutorial,”IEEE Trans. Commun., vol. 69, no. 5, pp. 3313–3351, 2021

2021

[5] [5]

Wireless communications with reconfigurable intelligent surface: Path loss modeling and experimental measurement,

W. Tanget al., “Wireless communications with reconfigurable intelligent surface: Path loss modeling and experimental measurement,”IEEE Trans. Wirel. Commun., vol. 20, no. 1, pp. 421–439, 2020

work page 2020

[6] [6]

Reconfigurable intelligent surfaces for energy efficiency in wireless communication,

C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in wireless communication,”IEEE Trans. Wirel. Commun., vol. 18, no. 8, pp. 4157–4170, Aug. 2019

work page 2019

[7] [7]

Large intelli- gent surface-assisted wireless communication exploiting statistical CSI,

Y . Han, W. Tang, S. Jin, C.-K. Wen, and X. Ma, “Large intelli- gent surface-assisted wireless communication exploiting statistical CSI,” IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 8238–8242, Aug. 2019

work page 2019

[8] [8]

Simultaneously transmitting and reflecting (STAR) RIS aided wireless communications,

X. Mu, Y . Liu, L. Guo, J. Lin, and R. Schober, “Simultaneously transmitting and reflecting (STAR) RIS aided wireless communications,” IEEE Trans. Wirel. Commun., vol. 21, no. 5, pp. 3083–3098, May 2022

work page 2022