{"id":"c24c6c09-c35c-42f2-ad4e-493e88e7fcc8","arxiv_id":"2501.05780","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Proposes a multi-layer reconfigurable intelligent surface as a single passive edge platform for MIMO communication, wave-based computation, and wireless power transfer to IoT devices.","lead":"A team proposes using stacks of programmable surfaces, called multi-layer RIS, as one cheap, low-power edge platform for wireless communication, computation, and power delivery to IoT devices. A smart generalist should read this to see a vision for replacing power-hungry digital hardware with wave-based, nearly passive processing at the network edge.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The universal paradigm rests on a fixed, known Rayleigh-Sommerfeld inter-layer channel; if that channel is state-dependent or uncalibrated, MIMO beamforming, computation, and WPT all fail.","rationale":"The reader's weakest assumption identifies the same gating issue: the paper treats the Rayleigh-Sommerfeld inter-layer channel as a known quantity when designing all three functions. I agree, and I extend it to a more specific physical risk: even if fabrication were perfect, the inter-layer coupling may depend on the tunable states of adjacent layers due to multiple scattering, so the channel is not a fixed matrix. This matters because every headline benefit—wave-domain beamforming, D2NN computation, and WPT focusing—is a phase-sensitive operation trained or optimized under this model. The paper is a vision article and should not be held to the standard of a full system demonstration, but the universal claim should be qualified until at least one experimental or full-wave validation of the inter-layer channel model is provided. The reader's CONDITIONAL verdict remains appropriate, so I recommend no change.","tokens_in":12586,"tokens_out":5294,"duration_ms":54857,"concrete_test":"Full-wave numerical check: simulate a two-layer transmissive RIS stack (e.g., 8x8 meta-atoms, layer spacing 2 wavelengths, meta-atom pitch lambda/4) in CST or HFSS. Impress a unit source at each layer-1 meta-atom and record the complex field at each layer-2 meta-atom location for two different phase configurations of layer 1 (e.g., all 0 degrees and a random 0/180-degree pattern) while layer 2 is fixed. Compare the resulting transmission matrix with the Rayleigh-Sommerfeld H matrix from Section II-A. If the complex transmission coefficients change by more than about 10% of the target phase step (or if max |Delta H| > 0.1 * mean |H|) when only the adjacent layer's tuning changes, then H is not a known constant and the optimized/trained designs in Sections IV-A through IV-C are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—one multi-layer RIS structure performs MIMO beamforming, wave-domain computation, and WPT—depends on an accurate inter-layer channel model at both design time and run time. Section II-A models inter-layer propagation as a fixed Rayleigh-Sommerfeld diffraction matrix: each meta-atom output is the product of its EM response and a weighted sum of previous-layer outputs. Sections IV-A, IV-B, and IV-C optimize or train layer responses assuming these weights are known and state-independent. Two unvalidated assumptions hide here. First, as the authors acknowledge in Section V-A, fabrication inconsistencies and assembly deformations make actual coefficients deviate from the ideal formula, misaligning beams and corrupting computation; the proposed error-back-propagation calibration is only sketched, with no demonstration. Second, and more fundamentally, the fixed H matrix ignores multiple scattering between adjacent layers. Because the layers are separated by only a few wavelengths and are densely populated with tunable meta-atoms, the field re-scattered by layer k+1 back into layer k changes with the tuning state of layer k+1. The effective inter-layer coupling is therefore state-dependent, not a known constant. If so, a single calibration or theoretical H cannot support the optimized phase profiles claimed in Section IV. This is load-bearing because MIMO capacity (Fig. 4), D2NN classification (Section IV-B), and WPT energy focusing (Fig. 6) all inherit the same channel-model error. The paper frames this as a future research direction, but the universal paradigm must be conditioned on resolving it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-layer RIS architecture as a universal edge platform for IoT, claiming that a single stacked-RIS structure can provide MIMO communication, wave-domain computation, and wireless power transfer (WPT) with minimal hardware changes, low cost, and low power consumption. It describes the multi-layer RIS structure, presents a system architecture with various communication/energy links, discusses the three functions in separate sections, reports two illustrative simulation results (system capacity versus number of layers, and SWIPT energy efficiency versus number of layers), and concludes with future research directions such as inter-layer channel modeling, resource allocation, channel estimation, and edge training.","tokens_in":12855,"tokens_out":5151,"duration_ms":53296,"significance":"If the proposed universal paradigm were validated, it could have substantial impact by replacing digital beamforming hardware, AI accelerators, and dedicated power transmitters with a single nearly passive structure at the edge. The paper usefully synthesizes prior prototypes (e.g., RIS-based MIMO, D2NN classification, RIS-based WPT) and clearly articulates a vision, while also being honest about open problems in Section V. However, the current manuscript provides no detailed derivations, no reproducible simulation models, and no experimental validation; the quantitative performance claims rest on underspecified figures. The central architectural claim is also sensitive to a physical assumption about the inter-layer channel that the paper itself flags as uncertain. As a vision/position paper the article has merit, but the strength of the claims currently exceeds the evidence provided.","major_comments":[{"comment":"The designs in Sections IV-A through IV-C assume the inter-layer channel is a fixed, known Rayleigh-Sommerfeld diffraction matrix. Because adjacent layers are separated by only a few wavelengths and each layer contains densely packed tunable meta-atoms, the field re-scattered by layer k+1 back into layer k depends on the tuning state of layer k+1. The effective inter-layer coupling is therefore state-dependent, not a known constant. The calibration procedure sketched in Section V-A (error back-propagation) would only capture a fixed channel and is not demonstrated. This concern is load-bearing because MIMO beamforming, D2NN computation, and WPT energy focusing all inherit the same channel assumption.","section":"II-A and V-A"},{"comment":"The simulation results are under-specified. No signal model, channel model, or optimization problem is stated, and no parameter values are given (e.g., number of meta-atoms per layer, inter-layer spacing, wavelength, transmit power budget, noise level, QoS thresholds). The figures do not clearly show single-layer baseline curves despite the text claiming comparison with single-layer counterparts. Without these details, the central quantitative claims—capacity increasing up to an optimal three layers and energy efficiency improving with layer count—cannot be evaluated or reproduced.","section":"IV-A, IV-C, Figs. 4 and 6"},{"comment":"The phrase 'power-free computation' is overstated. The paper itself states in Section IV-B that power consumption is primarily due to the FPGA used for system control. The FPGA and per-meta-atom bias circuitry consume power, so the computation is not literally power-free. The claim should be qualified as 'power-efficient' or 'computation whose power consumption is limited to control circuitry.'","section":"III and IV-B"}],"minor_comments":[{"comment":"The statement that 'only the transmission process of RF signals is altered' understates the need for per-meta-atom bias control lines and an FPGA controller; please clarify the intended meaning of 'minimal hardware changes.'","section":"III"},{"comment":"The claim that D2NN is a 'universal function approximator' extrapolates from the universal linear transformation result in [11] and the potential for nonlinear activation functions; please cite a specific result or temper the claim.","section":"IV-B"},{"comment":"There is a typo: 'Beside, as computation' should be 'Besides, as computation.'","section":"IV-B"},{"comment":"Calling the proposal 'an unprecedented attempt' is too strong given the existing prototypes and the cited works [10] and [12]; please qualify the novelty more carefully.","section":"I"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is a magazine-style overview rather than a full research paper. For IEEE Internet of Things Magazine, a vision paper with this level of detail may be acceptable if the authors substantially temper the performance claims and explicitly frame the results as illustrative under an idealized inter-layer channel assumption. However, if the venue expects archival technical depth, the lack of simulation details and the unaddressed state-dependent coupling problem are serious. The authors should be encouraged to either add a more realistic channel discussion or reframe the central claims as conditional on accurate inter-layer channel knowledge, which is currently listed as a future direction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful vision paper, not a rigorous research paper. It argues that a multi-layer RIS can be configured as a single nearly-passive platform for MIMO beamforming, wave-domain computation (D2NN-style), and WPT, and it maps out applications and open problems. The combination is new even though the components all come from existing prototypes. I think the central idea is plausible and the paper is honest about what is missing.\n\nWhat it does well: it connects three previously separate literatures (stacked intelligent metasurfaces for beamforming, diffractive neural networks for computation, and metasurface WPT) under one architecture. The system model in Section III is a reasonable way to organize the use cases. The paper gives credit to existing prototypes and explicitly admits that the inter-layer channel is treated as known. That is more honest than many papers in this space. The self-citations are to the authors' own prior work on semantic communication and AirJSCC; that is normal practice and not a red flag.\n\nThe soft spots, in proportion: First, the load-bearing assumption is the fixed Rayleigh-Sommerfeld inter-layer channel. The stress-test concern is right to push further than the authors do: adjacent layers are only a few wavelengths apart and densely populated with tunable meta-atoms, so the field re-scattered from layer k+1 back into layer k depends on the tuning state of layer k+1. That means the effective inter-layer coupling may be state-dependent, not a known constant. The paper flags calibration but not this multiple-scattering effect. For a vision paper this is acceptable as an open problem, but it does mean the universal paradigm is conditional on solving a hard physics problem. Second, the two simulations are described in prose with no equations, no parameter values, and no single-layer baselines in the figures. They read as illustrations, not evidence. Third, 'power-free computation' is overstated; the FPGA controller consumes power, and the meta-atoms themselves need biasing. The paper does qualify this later, but the abstract and Section IV-B use the stronger phrase.\n\nWho this is for: someone new to multi-layer RIS who wants a map of possible use cases, or a researcher looking for open problems in inter-layer channel modeling and calibration. It is not for someone who wants a demonstrated system. A serious referee could engage with it, but the right venue is a magazine, and the revisions should be about qualifiers and clearly labeling the simulations as illustrative.","headline":"A plausible and honestly written vision paper that integrates existing multi-layer RIS results, but its universal paradigm rests on an inter-layer channel model that is not yet shown to hold in practice.","tokens_in":13434,"tokens_out":2323,"would_cite":true,"duration_ms":23767,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multi-layer reconfigurable intelligent surface can act as one edge platform for MIMO communication, wave-domain computation, and wireless power transfer.","keywords":["multi-layer RIS","reconfigurable intelligent surface","wave-domain computation","MIMO beamforming","wireless power transfer","SWIPT","diffractive deep neural network","edge computing"],"falsifier":"A direct test is to fabricate a multi-layer RIS, illuminate it with a known input field, and compare the measured output field to the field predicted from the ideal Rayleigh-Sommerfeld inter-layer model; if the residual cannot be made small by the proposed back-propagation calibration of the inter-layer coefficients, the precise-beamforming and wave-computation claims fail. A second check would be to measure the capacity of a real 3-layer stack serving four users and see whether it exceeds a single layer by the amount the simulations report.","tokens_in":12385,"feed_emoji":"⚡","tokens_out":7302,"duration_ms":62024,"temperature":0.7,"pith_summary":"This paper proposes a universal edge platform built from a multi-layer reconfigurable intelligent surface (RIS), a sealed stack of programmable metasurface layers whose meta-atoms are tuned by a single FPGA controller. The claim is that the same passive structure can simultaneously perform MIMO communication, wave-domain computation, and wireless power transfer, replacing digital beamforming arrays, AI accelerators, and dedicated power transmitters. The motivation is that hardware-based scaling of communication and computing is too costly and power-hungry for massive IoT, while wave-based processing is nearly passive, low-latency, and scalable. Simulations reported in the paper show that MIMO capacity and SWIPT energy efficiency improve with the number of RIS layers up to a saturation point, with three layers optimal in the studied four-device scenarios.","feed_headline":"Stacked smart surfaces deliver MIMO, computing, and wireless power","feed_subtitle":"A single nearly passive stack could replace beamforming, AI accelerators, and power transmitters at the edge.","key_machinery":"The load-bearing object is the multi-layer RIS itself: a compact, sealed stack of transmissive RIS layers, each a dense lattice of passive meta-atoms (about a tenth to a half wavelength in size) with FPGA-controlled phase and amplitude response, separated by several wavelengths. The mechanism that carries the argument is the inter-layer diffraction channel: each meta-atom output on one layer is the weighted sum of all outputs from the previous layer, with weights given by Rayleigh-Sommerfeld diffraction factors. This makes the whole stack a physical deep neural network whose adjustable weights are the meta-atom responses and whose fixed connections are the propagation factors, so beamforming matrices, computation kernels, and power-focusing phases are all realized in the wave domain during propagation. The number of layers is the central tuning parameter: more layers add degrees of freedom and beam resolution but also penetration loss, producing the saturation behavior in the simulations.","core_discovery":"At the core is the idea that a multi-layer RIS, because its inter-layer propagation follows Huygens-Fresnel and Rayleigh-Sommerfeld diffraction, behaves mathematically like a fully connected neural network with the propagation factors as fixed weights and the meta-atom EM responses as adjustable parameters. By configuring those responses, the same stack can act as a MIMO beamformer, a diffractive deep neural network that executes computation tasks (classification, DOA estimation, encryption, over-the-air computation, joint source-channel coding), and an energy beamformer for wireless power transfer. The paper's simulations indicate that adding layers increases beam resolution and degrees of freedom, raising system capacity in MIMO and energy efficiency in SWIPT until penetration losses dominate, which is why the optimal layer count in the studied setups is three.","pith_inferences":["If the inter-layer channel model and calibration hold, the same wave-domain machinery could naturally extend to integrated sensing and communication, since DOA estimation and MIMO beamforming already share the diffraction structure.","The reported optimality of around three layers depends on the specific loss model and user count, so a design rule implied by the paper is to choose layer count per task and per hardware loss budget rather than maximize it universally.","The sketched error back-propagation calibration of inter-layer coefficients is directly testable: iteratively updating theoretical coefficients against measured output fields either validates the diffraction model or exposes its limits in practice.","A longer-term consequence is an energy-autonomous IoT loop, where a standalone multi-layer RIS powers devices by WPT and those devices use the harvested energy for uplink transmissions that the same RIS structure helps decode."],"forward_implications":["A base station equipped with multi-layer RIS could serve many IoT devices with MIMO beamforming performed in the wave domain, removing the need for excessive RF chains and high-precision DACs.","The same stack can execute light-speed computation tasks such as data-class-specific encryption, DOA estimation, over-the-air function computation, and joint source-channel coding, with latency independent of the number of meta-atoms and power cost dominated by the FPGA.","A standalone multi-layer RIS can act as a power beacon for WPT and SWIPT, using its higher beam resolution to focus energy on IoT devices and satisfy energy-harvest and SINR targets more easily as layers are added.","Because only the RF passage through the RIS is altered, the paradigm is backward-compatible with existing IoT devices and base stations, needing only the RIS stack and its FPGA controller as added hardware.","Adding RIS layers improves capacity and energy efficiency only up to a saturation point set by penetration loss, so layer count is a design parameter to optimize per scenario rather than something to maximize."],"supporting_citations":[{"why":"Supplies the alternating optimization algorithm, combining water-filling power allocation and projected gradient ascent, used in the MIMO capacity simulations.","marker":"[10]"},{"why":"Demonstrates the three-layer diffractive D2NN prototype that classifies handwritten digits in the microwave band, grounding wave-domain computation.","marker":"[5]"},{"why":"Provides the RIS-based WPT and SWIPT prototype that powers devices while transmitting data, the basis for the energy-transfer functions.","marker":"[9]"},{"why":"Presents the multi-user stacked-intelligent-metasurface prototype, cited as evidence that multi-layer RIS outperforms a single layer when inter-layer spacing is small.","marker":"[3]"},{"why":"The single-layer dual-polarized RIS MIMO prototype that serves as the baseline which multi-layer RIS is claimed to enhance.","marker":"[2]"},{"why":"Provides the analysis that single-layer RIS concentrates most power in central meta-atoms, motivating multi-layer stacking for higher beam resolution.","marker":"[15]"},{"why":"Frames the diffractional deep neural network approach that positions multi-layer RIS as a physical neural network for computation.","marker":"[4]"},{"why":"Shows stacked intelligent metasurfaces performing two-dimensional DOA estimation, used as an example wave-domain computation scheme.","marker":"[12]"}],"fun_headline_variants":["Multi-layer RIS does MIMO, computing, and wireless power","Stacked RIS: one passive surface for comms, computation, and energy","Three-layer RIS beams data, computes, and transfers power","Passive RIS stack handles MIMO, AI tasks, and WPT","Edge RIS stack delivers MIMO, computing, and wireless charging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The design assumes the inter-layer propagation between adjacent RIS layers is precisely described by Rayleigh-Sommerfeld diffraction and that those inter-layer coefficients are known when the surface is optimized or trained; if fabrication defects or assembly deformations make the real coefficients deviate, beam pointing misaligns and computation results become inaccurate, and the paper only sketches an error back-propagation calibration without showing it works.","fun_headline_variants_meta":{"raw":{"variants":["Multi-layer RIS does MIMO, computing, and wireless power","Stacked RIS: one passive surface for comms, computation, and energy","Three-layer RIS beams data, computes, and transfers power","Passive RIS stack handles MIMO, AI tasks, and WPT","Edge RIS stack delivers MIMO, computing, and wireless charging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1254,"prompt_tokens":981,"completion_tokens":273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":183}},"tokens_in":597,"tokens_out":273,"duration_ms":3069,"temperature":1.0,"reasoning_tokens":183,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:06:12.727475+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to fabricate a multi-layer RIS, illuminate it with a known input field, and compare the measured output field to the field predicted from the ideal Rayleigh-Sommerfeld inter-layer model; if the residual cannot be made small by the proposed back-propagation calibration of the inter-layer coefficients, the precise-beamforming and wave-computation claims fail. A second check would be to measure the capacity of a real 3-layer stack serving four users and see whether it exceeds a single layer by the amount the simulations report.","supporting_citations":[{"cited_title":"Stacked intelligent metasurfaces for multiuser beamforming in the wave domain","cited_arxiv_id":null,"evidence_quote":"Supplies the alternating optimization algorithm, combining water-filling power allocation and projected gradient ascent, used in the MIMO capacity simulations."},{"cited_title":"Classification of metal handwritten digits based on microwave diffractive deep neural network","cited_arxiv_id":null,"evidence_quote":"Demonstrates the three-layer diffractive D2NN prototype that classifies handwritten digits in the microwave band, grounding wave-domain computation."},{"cited_title":"Intelligent wireless power transfer via a 2-bit compact reconfigurable transmissive-metasurface-based router","cited_arxiv_id":null,"evidence_quote":"Provides the RIS-based WPT and SWIPT prototype that powers devices while transmitting data, the basis for the energy-transfer functions."},{"cited_title":"Multi-user ISAC through Stacked Intelligent Metasurfaces: New Algorithms and Experiments","cited_arxiv_id":"2405.01104","evidence_quote":"Presents the multi-user stacked-intelligent-metasurface prototype, cited as evidence that multi-layer RIS outperforms a single layer when inter-layer spacing is small."},{"cited_title":"Design and implementation of MIMO transmission based on dual-polarized reconfigurable intelligent surface","cited_arxiv_id":null,"evidence_quote":"The single-layer dual-polarized RIS MIMO prototype that serves as the baseline which multi-layer RIS is claimed to enhance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the analysis that single-layer RIS concentrates most power in central meta-atoms, motivating multi-layer stacking for higher beam resolution."},{"cited_title":"RIS-based on-the-air semantic communications — A diffractional deep neural network approach","cited_arxiv_id":null,"evidence_quote":"Frames the diffractional deep neural network approach that positions multi-layer RIS as a physical neural network for computation."},{"cited_title":"Two-dimensional direction-of-arrival estimation using stacked intelligent metasurfaces","cited_arxiv_id":null,"evidence_quote":"Shows stacked intelligent metasurfaces performing two-dimensional DOA estimation, used as an example wave-domain computation scheme."}],"review_version":1}