Smart Electromagnetic Entities Overview

• Smart Electromagnetic Entities (SEEs) are engineered systems that deliberately control electromagnetic wave propagation, scattering, sensing, and communications in designed environments.
• They encompass both passive and active variants, including static metasurfaces, one-time programmable skins, smart repeaters, and IAB nodes, balancing cost, power, and performance.
• SEE research integrates advanced hardware design, optimization algorithms, and network inference to enhance coverage, quality-of-service, and intelligent sensing applications.

Smart Electromagnetic Entities (SEEs) are engineered electromagnetic devices, structures, or assemblies that make wave propagation, scattering, sensing, and communication deliberately controllable within a Smart Electromagnetic Environment (SEME). In this view, the environment is not a passive obstacle but a designable asset: SEEs can be passive or active, static or reconfigurable, and may appear as electromagnetic skins, reconfigurable intelligent surfaces, smart repeaters, integrated access-and-backhaul nodes, programmable metasurface apertures, or broader embodied-intelligence systems that perceive, predict, and act on electromagnetic space (Benoni et al., 10 Sep 2025, Flamini et al., 2022, Li et al., 2019, Liu et al., 2 Jul 2026).

1. Conceptual scope and relation to SEME

Within the SEME framework, SEEs are the physical enablers that locally or globally manipulate electromagnetic fields so that coverage, Quality-of-Service (QoS), sensing utility, or scattering behavior can be improved in Regions-of-Interest (RoIs). A basic distinction is between Passive SEEs (PSEs), which reflect the incoming field without amplification, and Active SEEs (ASEs), which amplify and retransmit the signal. In heterogeneous SEMEs, different SEE technologies are jointly planned rather than selecting only one type, and the design objective is to recover or enhance electromagnetic coverage under installation-cost and energy-consumption requirements (Benoni et al., 10 Sep 2025).

An industrial mmWave perspective broadens this taxonomy by treating IAB nodes, Smart Repeaters, RISs, and passive surfaces (Smart Skins) as a layered ecosystem. The central argument is that not every coverage hole requires a base-station-like node: some situations can be served by cheaper and less power-hungry entities, while others require regenerative relays or more capable infrastructure. In that formulation, heterogeneous Smart EM deployment is driven by tradeoffs among coverage, capacity, installation cost, power consumption, deployment complexity, and EMF levels (Flamini et al., 2022).

Not all SEE-relevant work inserts new entities into the environment. In Opportunistic Sources Synthesis (OSS), the base transceiver station (BTS) itself is synthesized so that existing urban scatterers behave like opportunistic sources, reradiating energy constructively toward a target RoI. Conversely, in electromagnetic illusion in smart environments, the control is moved from the object coating to the surrounding environment, so that combined object–environment scattering yields the desired illusion without altering the object itself (Rù et al., 2024, Taghvaee et al., 2022). This suggests that SEE is best understood as a system-level role—an engineered electromagnetic degree of freedom—rather than a single hardware archetype.

2. Physical realizations and hardware taxonomy

SEEs span passive-static surfaces, reconfigurable surfaces, active relays, and intelligent apertures. Their hardware diversity is one reason the SEE concept is broader than a single RIS-style device.

SEE family	Defining mechanism	Indicative deployment profile
SP-EMS	Static passive EM skin; reflects and steers waves	\$500, 0 W
RP-EMS / RIS	Reconfigurable passive wave control	RP-EMS: \$750, <2 W; mmWave RIS target: <2 W
OTP-EMS	One-time configurable reflection via expendable fuses	No DC bias, no control circuitry, no power supply, no runtime updates, no maintenance
Smart Repeater	Network-controlled amplify-and-forward relay with two phased-array sides	\$3000, ~20 W
IAB node	Regenerative relay or micro-BTS-like node with wireless backhaul	\$7500, ≤350 W
Programmable metasurface aperture	Digitally controlled meta-atoms for task-adaptive illumination	$32 \times 24$ array around 2.4 GHz, controlled via PIN diodes and shift registers

The passive branch includes SP-EMSs, RP-EMSs, and OTP-EMSs. OTP-EMSs occupy a specific middle ground between reconfigurable passive EMSs / RISs and static passive EMSs: they are mass-producible and passive in operation, but configurable once at installation. Their basic unit is a meta-atom with a binary state $s_{pq}\in\{0,1\}$, where $s_{pq}=1$ denotes an intact fuse and $s_{pq}=0$ a burnt fuse. The local response is described by a reflection tensor,

$$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$

and the design objective is to make the phase difference between the two states close to $\pi$ while keeping reflection magnitude high in both states. The proof-of-concept meta-atom uses a square patch on a grounded substrate with two surface-mounted fuses, two vias, and a central excitation path; for the chosen LittleFuse R451, the measured average values are $0.6\,\Omega$ and $3.0\,\text{nH}$ when intact, and $<0.1\,\Omega$ and $<0.1\,\text{nH}$ when broken (Oliveri et al., 19 Jul 2025).

The active branch includes Smart Repeaters and IABs. Smart Repeaters are non-regenerative amplify-and-forward relays, whereas IAB nodes are Layer-2 regenerative relays or micro-BTS-like nodes. In the industrial viewpoint, Smart Repeaters are positioned as lower-cost alternatives to IAB when a full regenerative node is unnecessary, while IAB is appropriate when new infrastructure cells and wireless backhaul are required. RISs and Smart Skins complement these active devices by reshaping propagation with very low power or zero-power operation (Flamini et al., 2022, Benoni et al., 10 Sep 2025).

A separate SEE-like realization appears in intelligent sensing: a programmable metasurface functioning as a reconfigurable electromagnetic interface. In the microwave proof-of-principle sensing system, the metasurface comprises a $s_{pq}\in\{0,1\}$0 array of meta-atoms and is digitally controlled via PIN diodes and shift registers. Rather than probing with a fixed field, the aperture chooses coding patterns that adapt the illumination to the subject of interest and the downstream task (Li et al., 2019).

3. Mathematical and computational design paradigms

A defining trait of SEE research is that the electromagnetic front end is optimized jointly with inference, deployment, or source excitation. In intelligent sensing, the measurement process is modeled by a three-port deep ANN, with one port associated with the scene $s_{pq}\in\{0,1\}$1, one with the metasurface coding pattern $s_{pq}\in\{0,1\}$2, and one producing the raw measurement $s_{pq}\in\{0,1\}$3: $s_{pq}\in\{0,1\}$4 The full sensing chain is interpreted as a variational autoencoder with controllable conditional distribution $s_{pq}\in\{0,1\}$5 and inverse model $s_{pq}\in\{0,1\}$6, optimized through

$s_{pq}\in\{0,1\}$7

The coding patterns $s_{pq}\in\{0,1\}$8 and the reconstruction parameters $s_{pq}\in\{0,1\}$9 are found by alternating iterative optimization, using gradient-based backpropagation for continuous weights and randomized simultaneous perturbation stochastic approximation (r-SPSA) for binary metasurface codes (Li et al., 2019).

At the environment-planning level, the heterogeneous planning problem is cast as a multi-objective optimization over discrete deployment choices. For candidate sites $s_{pq}=1$0, the decision variable is

$s_{pq}=1$1

where $s_{pq}=1$2 means no SEE installed and $s_{pq}=1$3 means deploy SEE type $s_{pq}=1$4. The objectives minimize coverage mismatch $s_{pq}=1$5, installation cost $s_{pq}=1$6, and energy consumption $s_{pq}=1$7, subject to the QoS condition $s_{pq}=1$8 over all RoIs and times. The computational workflow is split into five blocks: Site Definition (SD), SEEs Design (SEED), Problem Definition (PD), Database Computation (DC), and Solution Space Exploration (SSE), with the search performed by an integer NSGA-II (Benoni et al., 10 Sep 2025).

In source-centric SEME design, OSS formulates the BTS as a phase-only synthesized source. The received field is reconstructed by superposing Embedded-plus-Environment Patterns (EPEPs): $s_{pq}=1$9 where each $s_{pq}=0$0 is computed once with only the $s_{pq}=0$1-th element excited and the others matched. This allows Particle Swarm Optimization (PSO) to search excitation phases $s_{pq}=0$2 without repeated ray-tracing of the full array-environment interaction, turning a prohibitive repeated-simulation loop into an offline database plus algebraic superposition (Rù et al., 2024).

A different formalism appears in programmable electromagnetic space via metasurface clusters, where the environment is represented as a virtual nodal network. Excitations and meta-atoms are treated as fully interconnected nodes, and the global far field is synthesized as a superposition of intrinsic radiation patterns weighted by nodal voltages: $s_{pq}=0$3 This shifts the design from local phase control toward deterministic programming of global scattering through cooperative multi-body assemblies (Li et al., 23 Mar 2026).

4. Sensing, inference, and embodied electromagnetic intelligence

One SEE-relevant direction is task-aware intelligent sensing. In the microwave proof-of-principle system, a transmitting horn antenna, a receiving horn antenna, a large-aperture programmable metasurface, and a vector network analyzer are combined into an end-to-end sensing chain. The raw data are transmission coefficients $s_{pq}=0$4. Two tasks are demonstrated: in-situ microwave imaging of the human body and in-situ recognition of human body gestures. For imaging, joint optimization of the measurement ANN and reconstruction ANN improves cross-validation loss and the structural similarity index metric (SSIM), especially in the highly compressed regime with $s_{pq}=0$5. For gesture recognition, the learned pipeline reports nearly 100% average recognition accuracy, with performance saturating around $s_{pq}=0$6, and gains of 5–35% in classification accuracy at low measurement counts relative to conventional approaches (Li et al., 2019).

A stronger autonomy-oriented formulation is metaEI-WM, which is highly relevant to SEE even though it does not use the exact term. The system combines a robotic perception layer, a semantic and geometric digital twin, an electromagnetic world model, and a decision/execution loop that outputs metasurface coding patterns. Its world model is represented as

$s_{pq}=0$7

where $s_{pq}=0$8 is reconstructed geometry, $s_{pq}=0$9 the semantic object set, $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$0 room topology, $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$1 material assignment, IMS metasurface poses, and $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$2 signal-source positions. The architecture includes perception and representation, EM dynamics prediction, and decision and execution, driven by a vision-LLM that interprets natural-language user instructions. Metasurface poses are estimated by homography decomposition, radio sources are localized with Wi-Fi CSI and SpotFi, and semantic reconstruction uses TARE, R3LIVE, SpatialLM, and RoomFormer. The control objective is

$$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$3

with prediction-inversion implemented through time-reversal symmetry, back-projection, and modified Gerchberg–Saxton inversion. The reported applications include zero-latency non-line-of-sight signal enhancement, symbiotic communications, and contactless physiological sensing, with average measured gains of 6.47 dB, 7.91 dB, and 5.10 dB, and BER reductions of 0.84 orders, 0.80 orders, and 1.11 orders across workplace, corridor, and apartment settings (Liu et al., 2 Jul 2026).

At the network inference layer, distributed SEE systems also require recognition of multiple simultaneous emitters. Specific multi-emitter identification (SMEI) reformulates identification as a multi-label problem with label vector

$$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$4

rather than a multiclass problem over $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$5 emitter subsets. Theoretical limits are derived with Fano’s inequality, mutual information is estimated using MINE, and computational complexity is reduced from $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$6 to $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$7. The improved I-SMEI adds multi-head attention to exploit correlations in overlapping signal mixtures; under 100% overlap at SNR = 12 dB, I-SMEI improves subset accuracy over SMEI by 0.33% at $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$8, 1.17% at $$\overline{\overline{\Gamma}}\!\left(\underline{g};s_{pq};\mathbf{k}_{inc}\right)= \begin{bmatrix} \Gamma^{TE} & \Gamma^{TE-TM}\ \Gamma^{TM-TE} & \Gamma^{TM} \end{bmatrix},$$9, and 4.63% at $\pi$0 (Chen et al., 22 Dec 2025).

5. Deployment, validation, and application domains

A major milestone for SEE/SEME validation is the first large-scale indoor experimental assessment of SEME using static-passive EM skins (SP-EMSs) in a real building with commodity Wi‑Fi devices. The test-bed is the Mesiano building of the University of Trento, about 13,000 m², using Aruba AP-304 access points at 5 GHz and smartphone-based measurements averaged over 10 s at each point. For hallway A, the reference low-coverage region is $\pi$1, corresponding to about 34% of hallway A’s area, with threshold $\pi$2. A chosen design with $\pi$3 yields predicted RoI reduction 70.00%, predicted average gain 2.64 dB, measured average gain $\pi$4, measured low-coverage area $\pi$5, and measured RoI reduction $\pi$6. User-level metrics also improve: average download throughput rises from 95.79 Mbps to 127.98 Mbps and average download latency falls from 122.87 ms to 52.13 ms. The reported total cost of ownership is \$\pi$71350 over 5 years for adding another access point (Benoni et al., 2023).

At urban scale, heterogeneous planning quantitatively demonstrates the complementarity of passive and active SEEs. In Trento Nord at $\pi$8 GHz, with $\pi$9 time instants, $0.6\,\Omega$0 RoIs, and $0.6\,\Omega$1 candidate sites, the Pareto front grows from 1 trivial solution to 40 non-dominated solutions after convergence. The best compromise solution (BCS) deploys 6 SP-EMSs, 1 RP-EMS, and 3 SRs, reducing the blind-spot area by 86.1% at $0.6\,\Omega$2 and 88.9% at $0.6\,\Omega$3, with cost <strong>\$6750 and energy 62 W. In San Martino, with $0.6\,\Omega$4, $0.6\,\Omega$5, and a search space around $0.6\,\Omega$6 configurations, the best compromise solution reaches below-threshold probabilities around 4.9% at $0.6\,\Omega$7 and 5.3% at $0.6\,\Omega$8. The paper also shows that heterogeneous deployment can outperform single-technology strategies: in one RoI, joint active-plus-passive deployment yields $0.6\,\Omega$9 and $3.0\,\text{nH}$0, whereas using only ASEs gives 43.5% and 28.6% (Benoni et al., 10 Sep 2025).

The industrial mmWave viewpoint reaches a related conclusion through a different setting. In a small area of Hong Kong at 28 GHz, the initial coverage is about 60% with omnidirectional UEs; adding three $3.0\,\text{nH}$1 RISs at strategic intersections yields about 20% coverage increase, more than 10 dB received-power improvement for cell-edge users, and almost 2× capacity improvement. The same study emphasizes that RISs do not amplify signals, do not introduce amplification noise, do not increase latency, and do not increase EMF levels, whereas IAB nodes and Smart Repeaters provide stronger but more power-intensive intervention (Flamini et al., 2022).

SEE applications are not limited to coverage enhancement. In electromagnetic illusion in smart environments, a metasurface is placed behind or in front of a dielectric slab so that the total reflection seen by an observer becomes identical to that of a bare PEC wall. The 1D proof-of-principle uses an FR4 slab with permittivity $3.0\,\text{nH}$2 and thickness $3.0\,\text{nH}$3, surrounded by an air region of thickness $3.0\,\text{nH}$4; after metasurface insertion, the reflection amplitude becomes unity, resonances disappear, and the reflection phase matches the PEC-only case. The required parameters are non-passive, so active elements are needed (Taghvaee et al., 2022). In programmable electromagnetic space via metasurface clusters, the same broad application class is expanded from 1D illusion to angle-resolved illusion spaces and deeply coupled meta-emitters capable of single-beam, dual-beam, and triple-beam collective radiation, with space itself treated as a functional scattering entity (Li et al., 23 Mar 2026).

6. Limitations, trade-offs, and open directions

Several limitations recur across the literature. The intelligent sensing architecture based on learnable acquisition and processing is still supervised and offline-trained, requiring labeled triplets and explicit optimization rather than autonomous lifelong adaptation; its metasurface codes are optimized for specific tasks and scenes, and generalization to substantially different environments or rapidly changing conditions is not fully addressed. The implementation is also a proof-of-concept microwave platform, and the use of r-SPSA for binary control patterns indicates that the physical design space remains challenging (Li et al., 2019).

Hardware simplicity and deployability come with electromagnetic trade-offs. OTP-EMSs are single-bit surfaces, so phase control is coarse and produces quantization lobes; the paper explicitly notes that these are a known limitation of 1-bit surfaces. In heterogeneous planning, full coverage often requires costly and energy-intensive active devices, passive-only solutions may be insufficient in every RoI, and optimization is computationally demanding enough to require precomputed coverage databases. In the industrial mmWave view, practical deployment is further constrained by quantized phase shifts, finite-resolution control circuitry, difficult CSI acquisition for transparent devices, site-specific codebooks, non-convex scheduling, and materials tradeoffs spanning PIN diodes, MEMS, SOI CMOS, GaN / PHEMT, liquid crystals, phase change materials, and ferroelectric materials (Oliveri et al., 19 Jul 2025, Benoni et al., 10 Sep 2025, Flamini et al., 2022).

A second recurring issue is the boundary between passive control and active functionality. Some illusion targets require negative real surface impedance or non-passive susceptibilities, so a purely passive metasurface cannot realize the specified camouflage. Conversely, active devices improve QoS strongly but increase cost and power draw. The surveyed works therefore resist any simple equivalence between SEE and a low-power reflecting surface: the category spans passive-static skins, one-time programmable skins, reconfigurable surfaces, active relays, smart sources, cooperative metasurface clusters, and autonomous world-model agents (Taghvaee et al., 2022, Rù et al., 2024, Li et al., 23 Mar 2026, Liu et al., 2 Jul 2026).

A plausible implication is that mature SEE systems will require simultaneous co-design across electromagnetic materials, array or metasurface hardware, world or scene models, network planning, and task-level inference. The literature already contains each of these ingredients separately—heterogeneous deployment, learnable sensing chains, deterministic programming of cooperative clusters, source-centric synthesis, and embodied world-model control—but not yet a single universal architecture that unifies them all.