A Survey on Stacked Intelligent Metasurfaces: Fundamentals, Recent Advances, and Challenges

arXiv:2603.05633 · eess.SP, cs.IT, math.IT · Submitted 2026-03-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "A Survey on Stacked Intelligent Metasurfaces".

Dev: Here's the gist of the paper. Comprehensive Research Synthesis: Stacked Intelligent Metasurfaces (SIMs) This document synthesizes information from two distinct sources—a structured survey overview (Paper A) and a detailed technical review of challenges and future directions (Paper B)—to provide a comprehensive, high-fidelity summary of the research landscape surrounding Stacked Intelligent Metasurfaces (SIMs).

Rosa: As an AI researcher, I recognize that the core contribution lies in bridging the gap between rigorous electromagnetic theory, practical hardware constraints, and advanced system-level optimization.

Dev: Core Concept and Scope: Stacked Intelligent Metasurfaces (SIMs) represent an advanced architectural paradigm where multiple programmable metasurface layers interact through wave propagation to achieve richer and more flexible electromagnetic transformations than conventional single-layer designs.

Rosa: SIMs fundamentally extend the capabilities of traditional Reconfigurable Intelligent Surfaces (RIS) from planar, largely diagonal wave manipulation into a three-dimensional, strongly coupled wave-domain processing platform.

Title and authors: Rosa: So, let's talk about the title of this paper, "A Survey on Stacked Intelligent Metasurfaces: Fundamentals, Recent Advances, and Challenges," and who wrote it. It’s by Chandan Kumar Sheemar, Wali Ullah Khan, Sourabh Solanki, George C. Alexandropoulos, and Symeon Chatzinotas <ref:2603.05633#pg1>.

Dev: That’s a lot of authors for a survey paper; it tells you this is intended to be really comprehensive across the whole stack of SIM research right now.

Taro: A survey implies they’ve seen a lot, but does it mean they haven't found the real, hard problems yet?

Rosa: It means they are trying to unify everything—from the physics modeling to the different ways people are actually building these things and where they are trying to apply them <ref:2603.05633#pg3>.

Dev: They’re aiming for a unified view across electromagnetic modeling, hardware designs, communication uses, sensing, and computing applications <ref:2603.05633#pg1>.

Taro: Unifying that much stuff sounds ambitious; what’s the biggest gap they are trying to bridge with this paper?

Rosa: They are trying to connect the dots between the rigorous electromagnetic theory, like those S, Z, and T parameters we talk about in some of these older papers, and the actual practical hardware constraints <ref:2603.05633#pg1>.

Dev: They are trying to bridge that gap between perfect theoretical models and what's actually achievable when you’re dealing with physical loss mechanisms or non-idealities <ref:2603.05633#pg1>.

Taro: So the main takeaway is that they are trying to create a single language for understanding SIM research, not just a list of separate papers <ref:2603.05633#pg1>.

Rosa: Exactly; they want to identify the open challenges and future research directions in one place, which helps everyone know where the community needs to focus its efforts <ref:2603.05633#pg1>.

Dev: It sets a baseline for what’s been done so we can see where we need to push the envelope next, especially regarding those complex joint optimization problems <ref:2603.05633#pg1>.

The paper's summary: Rosa: So, walking through the summary section of this survey, it lays out the core concept of stacked intelligent metasurfaces as a system where multiple programmable layers interact through wave propagation <ref:2603.05633#pg1>.

Dev: It emphasizes that unlike a single surface that just does local phase or amplitude weighting, the inter-layer propagation inherently mixes those spatial field samples in a way that creates a much richer input-output mapping <ref:2603.05633#pg2>.

Taro: So it’s not just adding more layers; it’s about how the waves travel between them and mix with each other, which makes the processing inherently richer <ref:2603.05633#pg2>.

Rosa: That richness is what allows you to synthesize expressive linear transformations, and if you add nonlinear elements, you get more general mappings between what comes in and what goes out <ref:2603.05633#pg2>.

Dev: They frame this as a wave-domain signal processing paradigm where part of the transceiver function is implemented physically by the SIM instead of being purely digital <ref:2603.05633#pg2>.

Taro: So, implementing processing in the physical layer instead of just a software algorithm sounds like it could save a lot of energy, which I’m interested in hearing about <ref:2603.05633#pg2>.

Rosa: It does suggest that analog neural processing with reduced hardware complexity and improved energy efficiency is possible through this approach <ref:2603.05633#pg1>.

Dev: They also show how this structure allows for the synthesis of expressive linear transformations, which is a key concept for many signal processing tasks <ref:2603.05633#pg2>.

Taro: I wonder if that physical implementation actually makes the system more robust when things go wrong compared to a purely digital one <ref:2603.05633#pg1>?

Rosa: That’s what they are exploring, but they also point out the modeling difficulty—reconciling physically consistent models with models that are computationally tractable is a central challenge <ref:2603.05633#pg1>.

Dev: The paper also covers the various application domains systematically, covering communication, intelligence and control, sensing and computing <ref:2603.05633#pg1>.

Taro: I’m interested in the sensing and computing parts specifically; can this actually be used for something like over-the-air inference or localization?

Rosa: Yes, that’s one of the areas they highlight, suggesting SIMs could serve as wave-domain computing engines for those tasks <ref:2603.05633#pg1>.

Dev: But they also warn about the complexity of joint optimization; when you have all these layers and variables interacting, it gets numerically unstable very quickly <ref:2603.05633#pg1>.

The paper's improvements: Rosa: Now let’s talk about what this survey suggests should be the next steps for research, which is where they suggest a really integrated co-design approach <ref:2603.05633#pg1>.

Dev: It stresses that we need to develop structure-preserving reduced-order electromagnetic operators that explicitly keep physical constraints like passivity and reciprocity intact while making them easier to optimize <ref:2603.05633#pg1>.

Taro: So the goal is to create a model that’s physically accurate enough for real testing but simple enough for the optimization software to actually run without timing out <ref:2603.05633#pg1>.

Rosa: They also call for scalable optimization strategies, which means using layer-wise decomposition and Riemannian optimization when things get strongly coupled in those systems <ref:2603.05633#pg1>.

Dev: And they mention constraint-aware algorithms that specifically handle those discrete or quantized phase controls, which is a big deal for realizing the hardware <ref:2603.05633#pg1>.

Taro: That’s where my experience comes in; if you can handle those discrete phases reliably, it means we can move closer to implementing the actual programmable hardware designs <ref:2603.05633#pg1>.

Rosa: They also push hard on joint training and estimation, using tensor or structured Bayesian inference to exploit that multilayer coupling when acquiring channel information <ref:2603.05633#pg1>.

Dev: And they suggest co-designing dispersion engineering with signal processing, so you can design the physical stack to handle wide bandwidths while optimizing for subcarrier performance <ref:2603.05633#pg1>.

Taro: So it’s a lot of intertwined work: modeling, optimization, and hardware design all happening at the same time?

Rosa: Precisely; they argue that hardware-algorithm co-design is essential to incorporating things like measured loss and quantization directly into the design loop <ref:2603.05633#pg1>.

Dev: It sounds like the next big push is making sure we move from these theoretical concepts to a practical, stable platform that can handle real-world nonidealities <ref:2603.05633#pg1>.

Conclusion: Rosa: So wrapping this up on the survey of "A Survey on Stacked Intelligent Metasurfaces: Fundamentals, Recent Advances, and Challenges," the main implication is that we need a holistic approach <ref:2603.05633#pg1>.

Dev: We need to integrate physically consistent electromagnetic models with scalable wave- and baseband optimization techniques, all while being acutely aware of the hardware constraints <ref:2603.05633#pg1>.

Taro: So in short, the paper says that to make SIMs useful, we can’t just focus on one part of it—like just the communication or just the sensing—we have to look at how everything interacts <ref:2603.05633#pg1>.

Rosa: That means moving toward validated reduced-order models and physics-constrained learning methods to get these systems from theoretical ideas into deployable platforms <ref:2603.05633#pg1>.

Dev: It’s a call for a framework that handles the complexity of joint optimization in a way that is stable and scalable for real-time operation <ref:2603.05633#pg1>.

Taro: I think the biggest lesson here is that we need to prove it works outside the lab, and then figure out how to make it gracefully recover when the world misbehaves <ref:2603.05633#pg1>.

Rosa: It really seems like they’ve laid out a very clear roadmap for what needs to be built next in this area, pushing us toward practical deployment <ref:2603.05633#pg1>.

Chandan Kumar Sheemar, Wali Ullah Khan, Sourabh Solanki, George C. Alexandropoulos, Symeon Chatzinotas

SnT University of Luxembourg · National Institute of Technology Warangal Institute of Electrical and Computer Engineering University of Illinois Chicago National and Kapodistrian University of Athens

eess.SP, cs.IT, math.IT

Submitted: 2026-03-05

Updated: 2026-11-02

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 100/100

The gist: Stacked Intelligent Metasurfaces (SIMs) This document synthesizes information from two distinct sources—a structured survey overview (Paper A) and a detailed technical review of challenges and…

Key concepts

Stacked Intelligent Metasurfaces (SIMs)
SIMs are advanced structures where several programmable metasurface layers interact. They enable richer electromagnetic manipulation than single layers, allowing for 3D wave-domain processing and superior beam control by coupling the layers' wave propagation.
Multiport Network Theory
This framework uses mathematical models like Z-, S-, and T-parameters to describe how waves interact at multiple points. It is essential for accurately modeling complex phenomena such as power conservation, mutual coupling, and loss within dense or wideband stacks.
Joint Optimization Complexity
SIMs involve thousands of coupled variables subject to strict hardware limits (like discrete phases). Optimizing these variables alongside baseband processing creates a highly non-convex problem that is numerically unstable and difficult to solve in real-time.

Terminology

Summary

Stacked Intelligent Metasurfaces (SIMs)

This document synthesizes information from two distinct sources—a structured survey overview (Paper A) and a detailed technical review of challenges and future directions (Paper B)—to provide a comprehensive, high-fidelity summary of the research landscape surrounding Stacked Intelligent Metasurfaces (SIMs). As an AI researcher, I recognize that the core contribution lies in bridging the gap between rigorous electromagnetic theory, practical hardware constraints, and advanced system-level optimization. p. 2

Where the paper says it

In contrast to existing tutorial and magazine articles that focus on selected aspects or application domains, this survey develops a rigorous and physically grounded framework that connects electromagnetic propagation theory, multiport network modeling, and cascaded wave-domain processing with practical wireless system design. Specifically, the survey consolidates physically consistent SIM representations across impedance, scattering, and transfer-parameter formulations, clarifying their interrelationships, modeling tradeoffs, and implications for scalable optimization and practical implementation.

Core Concept and Scope:

Stacked Intelligent Metasurfaces (SIMs) represent an advanced architectural paradigm where multiple programmable metasurface layers interact through wave propagation to achieve richer and more flexible electromagnetic transformations than conventional single-layer designs. SIMs fundamentally extend the capabilities of traditional Reconfigurable Intelligent Surfaces (RIS) from planar, largely diagonal wave manipulation into a three-dimensional, strongly coupled wave-domain processing platform. p. 0

Where the paper says it

Alexandropoulos, Senior Member, IEEE, and Symeon Chatzinotas, Fellow, IEEE Abstract—Reconfigurable intelligent surfaces (RISs) have emerged as a promising technology for shaping wireless propagation environments. While most existing works focus on RIS deployments within the propagation environment, an increasingly important paradigm consists of integrating programmable metasurfaces directly at the antenna front end, where they enable direct control of the radiated electromagnetic field and introduce new opportunities for wave-domain signal processing.

This expanded controllability enables superior beam control, multi-functional operation, and enables true wave-domain computation. p. 1
Where the paper says it

, localization and imaging), as well as wave-domain analog computation through programmable transformations and feature extraction [46]–[50]. Furthermore, the high-dimensional and reconfigurable structure naturally enables learning-based optimization, whereby SIM coefficients can be trained to approximate electromagnetic operators or to perform task-oriented inference directly on the wavefront, analogous to physical neural networks [51].

Modeling and Foundational Frameworks:

The survey establishes a rigorous foundation by reviewing the electromagnetic modeling frameworks underpinning SIM operation. This includes describing the physical architecture at both transmitter/base station and base station/receiver ends, detailing the cascaded propagation-modulation mechanism. p. 2

Where the paper says it

FUNDAMENTALS OF SIM This section introduces the fundamental principles that underpin SIM operation as a wave-domain signal processing platform. We first describe the SIM architecture at the transmitter/base station and its cascaded propagation-modulation mechanism, and then present the main electromagnetic modeling frameworks.

Crucially, it presents physically consistent multiport network-theoretic frameworks that account for complex phenomena such as radiation, mutual coupling, power conservation, dispersion, polarization mixing, and loss mechanisms—effects that become non-negligible in dense or wideband stacks. p. 4
Where the paper says it

A physically consistent modeling of SIM requires an electromagnetic formulation that rigorously accounts for radiation, mutual coupling, and power conservation. To this end, a multiport network-theoretic framework can be adopted, which provides an exact and unified description of programmable metasurfaces and their interaction with transmitting and receiving arrays, without relying on far-field approximations or uncoupledelement assumptions.

The literature consolidates various representations (Z-, S-, and T-parameters) to provide a unified taxonomy across impedance, scattering, and transfer-parameter formulations. p. 2
Where the paper says it

In contrast to existing tutorial and magazine articles that focus on selected aspects or application domains, this survey develops a rigorous and physically grounded framework that connects electromagnetic propagation theory, multiport network modeling, and cascaded wave-domain processing with practical wireless system design. Specifically, the survey consolidates physically consistent SIM representations across impedance, scattering, and transfer-parameter formulations, clarifying their interrelationships, modeling tradeoffs, and implications for scalable optimization and practical implementation.

Key Areas of Advancement (Section III Review):

The review systematically covers recent developments across numerous application domains:

  • Communication: Channel estimation and dynamics, sum rate analysis, wideband communications (addressing frequency selectivity and delay-Doppler coupling), near-field processing, and physical-layer security.

  • Intelligence & Control: Learning-driven control (including reinforcement learning, meta-learning, and multiagent learning) for adaptive configuration under mobility or partial Channel State Information (CSI). End-to-end neural and differentiable optimization frameworks are emerging for direct parameter tuning. p. 0

    Where the paper says it

    We review existing modeling approaches based on cascaded operators, multiport impedance formulations, and network parameter representations, and discuss their implications for scalable optimization and system design. The survey further examines key communication functionalities enabled by front-end metasurface processing, including communication performance optimization, near-field and wideband transmission, learning-driven control, integrated sensing and communications, and emerging architectures such as cell-free and non-terrestrial networks.

  • Sensing & Computing: Integrated Sensing and Communications (ISAC), localization, imaging, and the potential of SIMs as wave-domain computing engines for over-the-air inference.

  • System Design: Studies on optimization and fairness, energy-efficient designs, and various hardware realizations (static, programmable-passive, programmable-active).

Major Challenges Identified (The Technical Hurdles):

Despite the potential, the literature identifies several critical bottlenecks that define the current research frontier:

  1. Modeling Fidelity vs. p. 6

    Where the paper says it

    SUBMITTED TO IEEE (FOR REVIEW) 7 Table III: Comparison of Z-, S-, and T-parameter approaches for SIM modeling. Aspect Z-parameters (impedance) [63] S-parameters (scattering) [64] T-parameters (transfer scattering) [65] Primary viewpoint Circuit-complete multiport description that preserves self/mutual coupling and inter-layer feedback Wave-based reflection/transmission description aligned with RF characterization; can be physically consistent and include coupling Wave-based transfer description tailored to cascaded stacks; reorganizes cascade for tractability Key strength in SIM context High modeling fidelity; avoids restrictive simplifying assumptions that may cause mismatch in optimization Measurement-native interpretation; convenient for describing interfaces and reflection/transmission effects Cascade-friendly: overall stack formed by multiplying per-layer/inter-layer transfer blocks Cascading across layers Not inherently multiplicative; global model assembled from structured blocks, where SIM neighbor coupling yields sparse/banded matrices Equivalent stack obtained through repeated multiport cascading; produces nested operations Equivalent stack obtained by linear matrix products, avoiding nested cascade operators Computational bottleneck (deep stacks) Naive dense inversion can be heavy; tractability relies on exploiting SIM band/sparsity and iterative evaluation strategies Nested block operations lead to a rapidly increasing number of matrix inversions with the number of layers, limiting scalability Forming the cascade is multiplicationdominated; channel evaluation typically requires only a single terminal inversion Optimization tractability Designed for optimization: preserves coupling/feedback and leverages structure for scalable iterative computations Challenging for optimization in deep stacks due to variable entanglement inside nested inverses and poor controllability Well suited for gradient-based optimization over many layers due to compact cascade structure Physical constraints (lossless/reciprocal layers) Physical consistency ensured by complete network modeling (coupling and feedback retained); constraint form is model dependent Constraints can be imposed on scattering matrices for ideal layers (e.

    Tractability: The central tension is reconciling physically consistent yet computationally tractable models. Simplified cascaded operators are often used for optimization but neglect inter-layer feedback and coupling, while full-wave models are too demanding for system-level evaluation. p. 6
    Where the paper says it

    SUBMITTED TO IEEE (FOR REVIEW) 7 Table III: Comparison of Z-, S-, and T-parameter approaches for SIM modeling. Aspect Z-parameters (impedance) [63] S-parameters (scattering) [64] T-parameters (transfer scattering) [65] Primary viewpoint Circuit-complete multiport description that preserves self/mutual coupling and inter-layer feedback Wave-based reflection/transmission description aligned with RF characterization; can be physically consistent and include coupling Wave-based transfer description tailored to cascaded stacks; reorganizes cascade for tractability Key strength in SIM context High modeling fidelity; avoids restrictive simplifying assumptions that may cause mismatch in optimization Measurement-native interpretation; convenient for describing interfaces and reflection/transmission effects Cascade-friendly: overall stack formed by multiplying per-layer/inter-layer transfer blocks Cascading across layers Not inherently multiplicative; global model assembled from structured blocks, where SIM neighbor coupling yields sparse/banded matrices Equivalent stack obtained through repeated multiport cascading; produces nested operations Equivalent stack obtained by linear matrix products, avoiding nested cascade operators Computational bottleneck (deep stacks) Naive dense inversion can be heavy; tractability relies on exploiting SIM band/sparsity and iterative evaluation strategies Nested block operations lead to a rapidly increasing number of matrix inversions with the number of layers, limiting scalability Forming the cascade is multiplicationdominated; channel evaluation typically requires only a single terminal inversion Optimization tractability Designed for optimization: preserves coupling/feedback and leverages structure for scalable iterative computations Challenging for optimization in deep stacks due to variable entanglement inside nested inverses and poor controllability Well suited for gradient-based optimization over many layers due to compact cascade structure Physical constraints (lossless/reciprocal layers) Physical consistency ensured by complete network modeling (coupling and feedback retained); constraint form is model dependent Constraints can be imposed on scattering matrices for ideal layers (e.

    Reconciling electromagnetic fidelity with algorithmic tractability across heterogeneous architectures (passive/active, dispersive) remains paramount. p. 6
    Where the paper says it

    SUBMITTED TO IEEE (FOR REVIEW) 7 Table III: Comparison of Z-, S-, and T-parameter approaches for SIM modeling. Aspect Z-parameters (impedance) [63] S-parameters (scattering) [64] T-parameters (transfer scattering) [65] Primary viewpoint Circuit-complete multiport description that preserves self/mutual coupling and inter-layer feedback Wave-based reflection/transmission description aligned with RF characterization; can be physically consistent and include coupling Wave-based transfer description tailored to cascaded stacks; reorganizes cascade for tractability Key strength in SIM context High modeling fidelity; avoids restrictive simplifying assumptions that may cause mismatch in optimization Measurement-native interpretation; convenient for describing interfaces and reflection/transmission effects Cascade-friendly: overall stack formed by multiplying per-layer/inter-layer transfer blocks Cascading across layers Not inherently multiplicative; global model assembled from structured blocks, where SIM neighbor coupling yields sparse/banded matrices Equivalent stack obtained through repeated multiport cascading; produces nested operations Equivalent stack obtained by linear matrix products, avoiding nested cascade operators Computational bottleneck (deep stacks) Naive dense inversion can be heavy; tractability relies on exploiting SIM band/sparsity and iterative evaluation strategies Nested block operations lead to a rapidly increasing number of matrix inversions with the number of layers, limiting scalability Forming the cascade is multiplicationdominated; channel evaluation typically requires only a single terminal inversion Optimization tractability Designed for optimization: preserves coupling/feedback and leverages structure for scalable iterative computations Challenging for optimization in deep stacks due to variable entanglement inside nested inverses and poor controllability Well suited for gradient-based optimization over many layers due to compact cascade structure Physical constraints (lossless/reciprocal layers) Physical consistency ensured by complete network modeling (coupling and feedback retained); constraint form is model dependent Constraints can be imposed on scattering matrices for ideal layers (e.

  2. Joint Optimization Complexity: SIMs introduce thousands of tightly coupled variables subject to severe nonconvexity (discrete phase, hardware bounds). Jointly optimizing SIM coefficients with digital precoding, power control, and multi-user constraints quickly becomes numerically unstable and difficult to scale for deep stacks or strongly coupled regimes. p. 9

    Where the paper says it

    Hybrid wave-domain and digital estimation strategies illustrate the importance of jointly optimizing SIM configurations and baseband processing to fully leverage available degrees of freedom. Moreover, recent studies on channel aging call for adaptive and scalable estimation frameworks tailored to dynamic and large-scale SIM deployments.

    Real-time configuration under practical latency budgets is a significant barrier. p. 7
    Where the paper says it

    Each meta-atom supports discrete phase tuning, commonly at 1-bit resolution, allowing effective wavefront shaping and fully analog beamforming. This hardware paradigm significantly reduces RF chain requirements and power consumption, making programmablepassive SIMs attractive for scalable MIMO transceivers and real-time reconfigurable wireless systems.

  3. CSI Acquisition and Tracking: High-dimensional SIM-assisted channels strain practical RF chains, leading to challenges in pilot overhead, estimator robustness, and tracking under mobility or channel aging. Designing SIM-aware acquisition pipelines that remain reliable with partial or outdated CSI is a major hurdle. p. 8

    Where the paper says it

    In [76], CSI acquisition is studied for SIM-assisted holographic MIMO systems under RF chain constraints. By collecting pilot observations across multiple SIM configurations, a subspace estimation framework is developed that leverages partial and statistical CSI, while SIM phase shifts are optimized during training to minimize estimation mean-squared error (MSE).

  4. Wideband and Doubly Dispersive Operation: Maintaining stability across wide bandwidths requires co-design between electromagnetic dispersion engineering and signal processing (e.g., delay augmentation). p. 2

    Where the paper says it

    Specifically, the survey consolidates physically consistent SIM representations across impedance, scattering, and transfer-parameter formulations, clarifying their interrelationships, modeling tradeoffs, and implications for scalable optimization and practical implementation. Furthermore, it presents a detailed review of SIM hardware platforms, including static, programmable-passive, and programmable-active realizations, as well as emerging novel architectural paradigms such as 1) delay-augmented, 2) dual-polarized, 3) flexible, 4) meta-fiber-assisted, and 5) nonlinear SIM structures.

    Understanding fundamental limits linking bandwidth, stacking depth, and focusing robustness is still an open problem.

  5. Hardware Nonidealities: Stacking increases expressiveness but exacerbates issues like accumulated insertion loss, sensitivity to tolerances/misalignment, quantization effects in phase control, and thermal drift. Translating algorithmic gains into repeatable hardware gains under these realistic constraints requires sophisticated co-design (e. p. 9

    Where the paper says it

    Extending toward fully programmable electromagnetic transceivers, [86] examined joint transmit–receive wave-domain processing in SIM-assisted holographic MIMO systems, where SIM phase shifts are tightly coupled with digital covariance and beamforming variables. Iterative gradient- and projection-based updates under unit-modulus constraints yielded notable gains over single-layer and weakly coupled architectures, highlighting SIM’s role as an integrated wave-domain transceiver.

    g., hybrid passive–active stacks). p. 2
    Where the paper says it

    Specifically, the survey consolidates physically consistent SIM representations across impedance, scattering, and transfer-parameter formulations, clarifying their interrelationships, modeling tradeoffs, and implications for scalable optimization and practical implementation. Furthermore, it presents a detailed review of SIM hardware platforms, including static, programmable-passive, and programmable-active realizations, as well as emerging novel architectural paradigms such as 1) delay-augmented, 2) dual-polarized, 3) flexible, 4) meta-fiber-assisted, and 5) nonlinear SIM structures.

  6. Calibration and Reproducibility: Characterizing multilayer stacks is difficult due to dependencies on inter-layer spacing, coupling, and environment. The lack of standardized calibration procedures leaves large residual mismatches between simulated models and physical prototypes. p. 9

    Where the paper says it

    Leveraging a multilayer structure, [81] introduced a tensor-based framework that models SIM observations using nested tensor representations capturing inter-layer coupling, enabling joint channel and layer-response estimation via structured decomposition. This approach achieved significantly lower NMSE with moderate complexity, demonstrating how SIM’s physical structure can be exploited for efficient estimation.

Promising Research Directions (The Roadmap):

Future research is strongly directed toward a tightly integrated co-design approach:

  1. Structure-Preserving Reduced-Order Models: Developing reduced-order EM operators that explicitly preserve physical constraints (passivity, reciprocity) while remaining optimization-friendly. This includes frequency-dependent learning and low-rank factorizations exploiting neighbor interactions. p. 6

    Where the paper says it

    SUBMITTED TO IEEE (FOR REVIEW) 7 Table III: Comparison of Z-, S-, and T-parameter approaches for SIM modeling. Aspect Z-parameters (impedance) [63] S-parameters (scattering) [64] T-parameters (transfer scattering) [65] Primary viewpoint Circuit-complete multiport description that preserves self/mutual coupling and inter-layer feedback Wave-based reflection/transmission description aligned with RF characterization; can be physically consistent and include coupling Wave-based transfer description tailored to cascaded stacks; reorganizes cascade for tractability Key strength in SIM context High modeling fidelity; avoids restrictive simplifying assumptions that may cause mismatch in optimization Measurement-native interpretation; convenient for describing interfaces and reflection/transmission effects Cascade-friendly: overall stack formed by multiplying per-layer/inter-layer transfer blocks Cascading across layers Not inherently multiplicative; global model assembled from structured blocks, where SIM neighbor coupling yields sparse/banded matrices Equivalent stack obtained through repeated multiport cascading; produces nested operations Equivalent stack obtained by linear matrix products, avoiding nested cascade operators Computational bottleneck (deep stacks) Naive dense inversion can be heavy; tractability relies on exploiting SIM band/sparsity and iterative evaluation strategies Nested block operations lead to a rapidly increasing number of matrix inversions with the number of layers, limiting scalability Forming the cascade is multiplicationdominated; channel evaluation typically requires only a single terminal inversion Optimization tractability Designed for optimization: preserves coupling/feedback and leverages structure for scalable iterative computations Challenging for optimization in deep stacks due to variable entanglement inside nested inverses and poor controllability Well suited for gradient-based optimization over many layers due to compact cascade structure Physical constraints (lossless/reciprocal layers) Physical consistency ensured by complete network modeling (coupling and feedback retained); constraint form is model dependent Constraints can be imposed on scattering matrices for ideal layers (e.

  2. Scalable Optimization Strategies: Future solvers must exploit SIM structure via layer-wise decomposition, Riemannian optimization for stable updates in strong coupling regimes, and constraint-aware algorithms that handle discrete/quantized control effectively. Online and distributed optimization are key for real-time performance in large systems. p. 1

    Where the paper says it

    It outlines the foundational principles of SIM operation and identifies their role in extending conventional metasurface capabilities toward multifunctional electromagnetic processing. Complementing this perspective, [53] positions SIM as a key enabling technology for future 6G wireless networks, with particular emphasis on distributed and large-scale deployments.

  3. Joint Training and Estimation: CSI acquisition must be co-designed with SIM states using tensor/structured Bayesian inference to exploit multilayer coupling. Predictive tracking models combining statistical CSI with mobility models will reduce overhead. p. 8

    Where the paper says it

    By collecting pilot observations across multiple SIM configurations, a subspace estimation framework is developed that leverages partial and statistical CSI, while SIM phase shifts are optimized during training to minimize estimation mean-squared error (MSE). Closed-form MSE expressions demonstrated that appropriately designed wave-domain training significantly improves estimation accuracy with manageable pilot overhead, highlighting the intrinsic coupling between SIM configuration and CSI acquisition.

  4. Dispersion Engineering: Co-designing delay-augmented/dispersive stacks for controllable group delay and developing multi-band designs that optimize across subcarriers simultaneously is essential for wideband robustness.

  5. Hardware-Algorithm Co-Design: Developing stability-aware control, low-power biasing architectures, and impairment-aware beamforming that explicitly incorporates measured hardware models (loss, quantization) into the design loop.

  6. Practical Calibration and Benchmarking: Establishing scalable calibration pipelines—including in-situ self-calibration and digital twin approaches—alongside standardized measurement protocols to ensure reproducibility across diverse prototypes.

Conclusion:

SIMs are positioned as a transformative wave-domain technology for 6G and beyond, fundamentally enhancing achievable rate performance by enabling wave-domain precoding and expanding spatial degrees of freedom. However, realizing this potential requires moving beyond isolated optimization techniques toward a holistic framework that tightly integrates physically consistent electromagnetic models, scalable wave- and baseband-domain optimization, hardware-aware architectures under realistic nonidealities, and robust network control protocols. p. 0

Where the paper says it

The survey further examines key communication functionalities enabled by front-end metasurface processing, including communication performance optimization, near-field and wideband transmission, learning-driven control, integrated sensing and communications, and emerging architectures such as cell-free and non-terrestrial networks. Finally, we identify open research problems related to physical modeling, scalability, hardware–algorithm co-design, and network integration, and outline promising directions toward realizing SIM-based antenna front ends as fully programmable electromagnetic processors for future sixth-generation (6G) wireless systems Index Terms—Stacked intelligent metasurfaces, wave-domain signal processing, modeling, optimization, wireless applications.

The path forward demands validated reduced-order models and physics-constrained learning methods to transition SIMs from theoretical concepts to scalable, deployable platforms.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI System can perform wave-domain signal processing directly at the antenna front end by implementing expressive linear transformations and, when augmented with nonlinear elements, more general nonlinear mappings between incident and transmitted/reflected wavefronts. This allows for realizing multi-layer analog processing directly at the speed of wave propagation.

  2. Improved AI System can achieve high-dimensional beamforming and spatial filtering directly in the electromagnetic domain by jointly optimizing SIM configurations and baseband processing to improve spectral efficiency and spatial multiplexing while reducing RF-chain count and computational burden.

  3. Improved AI System can perform learning-driven control of SIM architectures by training coefficients to approximate electromagnetic operators or to perform task-oriented inference directly on the wavefront, analogous to physical neural networks. This enables realizing analog neural processing with reduced hardware complexity and improved energy efficiency.

  4. Improved AI System can enable near-field and discrete constraint optimization for beamforming by utilizing a framework that allows it to jointly optimize discrete SIM phases and power allocation for ISAC systems, achieving near continuous-phase performance even with low-resolution control.

  5. Improved AI System can perform wideband communications by implementing fully analog SIM precoding/combining architectures that diagonalize the wideband channel across subcarriers, achieving near interference-free transmission within an effective bandwidth and over 300% capacity gain compared to center-frequency SIM designs.

Abstract

Stacked intelligent metasurfaces (SIMs) cascade several closely spaced programmable metasurface layers so that an electromagnetic wave undergoes repeated modulation-propagation interactions before leaving the stack. The stack thereby implements, at the speed of propagation and with few radio-frequency chains, high-dimensional analog transformations that would otherwise require digital precoding, combining, spatial filtering, focusing, or inference. Since 2023, SIM research has produced more than two hundred articles, a dedicated magazine special issue, and two survey-level overviews organized around applications and wave-domain computing. This article complements them with a model-to-hardware perspective. First, we present the electromagnetic modeling hierarchy of SIMs, making explicit the assumptions under which the models coincide, their validity regimes, and their computational complexity, with a common notation for hardware non-idealities and power consumption. Second, we present a methodology-first taxonomy (SIM-aware channel acquisition, model-based optimization, statistical-CSI design, learning-based control, and hardware-aware robust design), compare the families in terms of CSI requirements, complexity, convergence, and hardware constraints, and reconcile conflicting findings in the literature. Third, we synthesize applications by the physical role of the SIM. Fourth, we quantitatively compare experimentally demonstrated multilayer prototypes, separate validated from conceptual architectures, and discuss thermal, tolerance, and calibration issues. Fifth, we provide a tutorial-style performance and trade-off analysis anchored in reported results. Finally, we map evidence-based challenges to research directions with measurable targets. Lessons learned are summarized at the end of every section.

Sources

Related papers