Reconfigurable Intelligent Metasurface Antenna (RIMSA)

• RIMSA is a programmable antenna array that leverages tunable metasurfaces to dynamically control electromagnetic fields and beam patterns.
• It integrates varied architectures—including reflective, transmissive, and multi-feed designs—to enable beam steering, holography, and index modulation.
• Advanced electromagnetic models combined with deep learning optimization allow rapid, precise reconfiguration for enhanced communication and sensing performance.

Reconfigurable Intelligent Metasurface Antenna (RIMSA) denotes a programmable metasurface-based antenna concept in which the metasurface is treated as an antenna array or acts as the primary radiating and/or receiving interface, rather than only as an environmental reflector. In the literature, the term is used in closely related but not identical ways: one system-level formulation states that the RIS metasurface is modeled as an antenna array, “this is called RIMSA,” while another proposes a wireless system “where the RIS acts as an antenna, which we call Reconfigurable Intelligent Metasurface Antennas (RIMSA)” (Gu et al., 2022, Wei et al., 23 Jun 2025). Related survey work places integrated metasurfaces at the antenna front end, where they directly manipulate the radiated electromagnetic field and enable wave-domain signal processing, in the same technological lineage (Sheemar et al., 5 Mar 2026).

1. Conceptual scope and relation to RIS and metasurface antennas

RIMSA sits at the intersection of reconfigurable intelligent surfaces, metasurface antennas, and programmable antenna front ends. In the RIS literature, the core idea is a passive metasurface composed of many discrete elements that act collectively to reflect and control electromagnetic waves, with tunable phase and, in some formulations, amplitude (Gu et al., 2022). In the metasurface-antenna literature, a waveguide-fed metasurface antenna consists of an array of subwavelength metamaterial elements distributed over an electrically large structure, each subwavelength in dimension and with subwavelength separation between elements; each element locally samples a reference wave and radiates into free space (Smith et al., 2017).

A central distinction from conventional phased arrays and electronically scanned antennas is repeatedly emphasized. A dynamic metasurface antenna does not require active phase shifters and amplifiers, but rather achieves reconfigurability by shifting the resonance frequency of each individual metamaterial element (Smith et al., 2017). This point matters because RIMSA is often misread as a simple synonym for phased-array beam steering. The published formulations instead describe a class of apertures whose control variables are surface states, polarizabilities, impedances, or metasurface phase responses, and whose electromagnetic behavior is constrained by the underlying meta-atom physics rather than by unconstrained complex weights.

The broader RIS tutorial literature adds a second axis of interpretation. Metasurface-based RIS implementations are described as thin 2D arrangements of subwavelength resonators whose tiles can be manipulated nearly independently, enabling anomalous reflection, reflection-angle control, and programmable shaping of the wireless propagation environment (ElMossallamy et al., 2020). This suggests that RIMSA is best understood as a convergence point: a metasurface no longer treated only as a passive environmental modifier, but as a reconfigurable antenna aperture in its own right.

2. Representative physical architectures and hardware realizations

One representative realization is the deep-learning-assisted holographic metasurface antenna of Hyunjun Ma et al. The metasurface is modeled as a two-dimensional array of $30 \times 30$ reconfigurable dipole elements, totaling 900 units, each spaced $0.2$ wavelengths apart. Each unit element is treated as a point dipole characterized by a variable polarizability $a_n$, and the encoded state satisfies $a_n = k_n a_{max}$ with $-1 \leq k_n \leq 1$; binary operation uses $k_n=\pm 1$ (Ma et al., 2024). This architecture is explicitly aimed at real-time holographic beam steering.

A full-wave reconfigurable intelligent surface realization uses a $10 \times 10$ array of planar square metallic patches, each $1\,\text{cm}^2$ in area, with 180 varactor diodes interconnecting the patches. The substrate has $\epsilon_r = 4.4$, conductivity $\sigma = 0.0025\,\text{S/m}$, and thickness $0.2$0, with a perfect electric conductor ground plane beneath the substrate and perfectly matched layer boundaries at the simulation bounds. In that model, each diode is represented as a single FDTD cell of $0.2$1 mm$0.2$2, with capacitance $0.2$3 pF or $0.2$4 pF used to represent “closed” and “open” conditions (Colella et al., 2023).

A distinct RIMSA-array architecture for multi-user downlink emphasizes parallel coaxial feeding. All elements are excited simultaneously through a power distribution network, and each element’s phase is continuously tunable via varactor diodes and reflective-type phase shifters (Wei et al., 23 Jun 2025). A related transmissive implementation replaces a conventional multi-RF-chain array by a single feed antenna, an RMS panel with $0.2$5 passive transmissive elements, and an intelligent controller that adjusts the transmissive coefficient $0.2$6 of each element (Li et al., 2021). These designs show that RIMSA is not restricted to one feed topology: reflective, transmissive, and directly radiating implementations all appear in the literature.

Hybrid and dual-functional unit-cell designs extend the hardware repertoire further. One hybrid RIS design uses a unit cell of size $0.2$7 at $0.2$8 GHz, combining a ring antenna and a central circular disc antenna that share the same phase center. The central disc is for sensing, the ring antenna is for reconfigurable reflection, and each unit cell includes an SP4T switch connected to tunable loads under microcontroller control (Birari et al., 23 Jan 2025). This is not merely a communications surface; it is a co-designed sensing-and-reflection aperture.

3. Electromagnetic models, aperture physics, and design constraints

Several electromagnetic models recur across RIMSA work. In the dipole-array formulation for holographic beam synthesis, the far-field response is governed by the scattering equation

$0.2$9

where $a_n$0 is the dyadic Green’s function and $a_n$1 is the induced polarization of the $a_n$2-th dipole. Training and inference use the Born approximation up to third order, while analytic Green’s function calculation is used to check the validity of the Born approximation (Ma et al., 2024).

In waveguide-fed metasurface antennas, each metamaterial element is modeled as an electrically small, non-interacting polarizable magnetic dipole with frequency-dependent magnetic polarizability

$a_n$3

If the reference wave is $a_n$4, the far-field array factor is

$a_n$5

The same analysis also states an important physical constraint: real elements cannot implement arbitrary phase and amplitude independently, because the response is constrained by the resonator’s Lorentzian nature, with phase range limited to $a_n$6 and amplitude approaching zero near those extremes (Smith et al., 2017). This directly counters a common oversimplification in high-level wireless models.

Surface-impedance formulations provide another widely used description. For an intelligent metasurface with continuously tunable local surface impedance,

$a_n$7

and the reflection coefficient is

$a_n$8

By adjusting $a_n$9 and $a_n = k_n a_{max}$0, one can tune reflection phase and absorption, and full absorption occurs when $a_n = k_n a_{max}$1 (Liu et al., 2018). This formulation is central when RIMSA is treated not just as a phase-only surface but as a locally programmable impedance boundary.

At the channel level, RIS-assisted communications literature distinguishes between a dyadic backscatter model and a metasurface-oriented spatial scattering model. The dyadic form is

$a_n = k_n a_{max}$2

with $a_n = k_n a_{max}$3. For metasurface-based implementations, the channel is also written as

$a_n = k_n a_{max}$4

The same source distinguishes a reflector regime, where path loss scales with the sum of distances, from a scattering regime, where it scales with the product of distances (ElMossallamy et al., 2020). For RIMSA, this distinction is consequential because whether the surface behaves as a reflector, scatterer, or radiating aperture determines both modeling fidelity and expected gain.

4. Control strategies, inverse design, and optimization algorithms

Control of RIMSA apertures spans iterative optimization, physics-informed neural networks, manifold methods, sparse recovery, and reinforcement learning. In the holographic beam-steering system of Ma et al., suitable states for desired far-field patterns can be identified using iteration, but this is very slow and needs to be done for each far-field pattern. The proposed alternative is an autoencoder in which the encoder is a ResNet-based neural network that maps the desired far-field pattern to the vector of dipole states, while the decoder is the analytic scattering equation using the Born approximation (Ma et al., 2024). The reported computing time for determining the 900-element state vector is under $a_n = k_n a_{max}$5, specifically $a_n = k_n a_{max}$6–$a_n = k_n a_{max}$7, whereas the Genetic Algorithm requires about $a_n = k_n a_{max}$8 seconds and the Gerchberg-Saxton algorithm about $a_n = k_n a_{max}$9 seconds; training takes about $-1 \leq k_n \leq 1$0–$-1 \leq k_n \leq 1$1 minutes on an Nvidia A6000 GPU (Ma et al., 2024).

For multi-user downlink with a RIMSA array, sum-rate maximization is formulated jointly over digital processing matrices and phase responses at both the base station and the users. In the MU-MISO setting, an alternating optimization algorithm uses fractional programming for the digital processing matrix and product manifold optimization for the phase responses of the RIMSA arrays. In the MU-MIMO setting, the problem is converted into a weighted sum of mean square errors minimization problem, where the digital precoder and digital combiner subproblems have closed-form solutions and the RIMSA configuration subproblem is again solved by product manifold optimization (Wei et al., 23 Jun 2025). The technical significance is that the unit-modulus and block-diagonal metasurface constraints are handled directly on the manifold rather than by loose relaxation.

A different extension introduces spatial reconfigurability in addition to electromagnetic control. The flexible intelligent metasurface framework compares EM-only, PBF-only, and EM-PBF modes in a SISO system, where element positions can be moved in the $-1 \leq k_n \leq 1$2-$-1 \leq k_n \leq 1$3 plane and reflection phases can also be adjusted. In a multi-element, multi-path scenario, the EM-only mode improves the received signal power by $-1 \leq k_n \leq 1$4 compared to the PBF-only mode, while EM-PBF further enhances performance. The same work formulates channel estimation as a compressive sensing problem and proposes clustering mean-field variational sparse Bayesian learning for joint cascaded and direct channel estimation (Yang et al., 14 Mar 2025). A plausible implication is that RIMSA control is increasingly moving beyond phase design into joint aperture-state and geometry-state design.

Further work extends the optimization toolkit to alternating optimization with successive convex approximation and the penalty convex-concave procedure for jointly optimizing beamforming, phase shifts, and the positions of fluid antennas and liquid metasurface elements (Shen, 22 Jul 2025), and to deep reinforcement learning for distributed anti-jamming sensing, where the policy network selects metasurface phase configurations and a sensing network maps received signals to object probability maps (Wang et al., 7 Aug 2025). These developments indicate that RIMSA control is now treated as a high-dimensional inverse problem rather than as a static beam codebook problem.

5. Communication functions: beam steering, holography, multiplexing, and system-level gains

Beam steering and wavefront shaping are the canonical RIMSA functions. The deep-learning-assisted metasurface antenna generates metasurface states that accurately reproduce arbitrary far-field patterns, including beams, multi-beams, letters, and MNIST digits, and it generalizes to untrained target images. Binary encoding by taking the sign of the output states still closely reproduces the target far-field pattern, with better fidelity and lower error than GA or GS in the reported tests (Ma et al., 2024). In this setting, RIMSA is explicitly a holographic antenna capable of real-time operation.

Analytical and full-wave work on waveguide-fed metasurface antennas shows that amplitude-only, binary amplitude, and Lorentzian-constrained phase holograms can all produce well-formed beams. Reported examples include a main beam at $-1 \leq k_n \leq 1$5, HPBW $-1 \leq k_n \leq 1$6 and first sidelobe $-1 \leq k_n \leq 1$7 dB for amplitude-only control, HPBW $-1 \leq k_n \leq 1$8 and first sidelobe $-1 \leq k_n \leq 1$9 dB for binary amplitude control, and HPBW $k_n=\pm 1$0 with first sidelobe $k_n=\pm 1$1 dB for Lorentzian-constrained phase control (Smith et al., 2017). These results are notable because they show useful beam quality even under non-ideal amplitude-only or binary constraints.

System-level studies connect these aperture-level controls to network metrics. In a far-field LOS-dominant RIS-assisted wireless system where the RIS metasurface is modeled as an antenna, simulations in a 7-cell, 21-sector layout show up to $k_n=\pm 1$2 dB increase in median received power and median SINR gains up to $k_n=\pm 1$3 dB, depending on RIS parameters. The same study reports that 2-bit phase control achieves near-optimal performance, that smaller element spacings provide better performance, and that RIS at cell edge outperforms mid-cell placement, especially when users are at the cell boundary (Gu et al., 2022). These are communications-system results rather than only field-pattern results.

RIMSA has also been proposed as a programmable aperture for index modulation in 6G wireless networks. One electromagnetics-compliant design reports a via-free meta-atom with continuously variable reflection phase, achieving $k_n=\pm 1$4 phase tunability with $k_n=\pm 1$5 dB loss at $k_n=\pm 1$6 GHz and $k_n=\pm 1$7 degree tunable phase across $k_n=\pm 1$8–$k_n=\pm 1$9 GHz, corresponding to $10 \times 10$0 bandwidth. The same framework supports phase modulation, spatial modulation, subcarrier or harmonic index modulation, time-slot indexing, and channel-domain radiation-pattern indexing (Hodge et al., 2023). This broadens the usual interpretation of RIMSA from beam steering to direct physical-layer modulation and multiplexing.

Transmissive reconfigurable meta-surface multi-antenna systems provide another communication mode. In the downlink, the feed antenna and users are on opposite sides of the metasurface, eliminating blockage and self-interference issues common in reflective systems; in the uplink, users transmit through the metasurface to a single receiving antenna. The reported optimization framework jointly designs transmission coefficients, power allocation, and, in the uplink, subcarrier allocation, with numerical results showing that sum-rate improves as the number of RMS elements increases (Li et al., 2021). This reinforces that RIMSA should not be reduced to reflective-only hardware.

6. Sensing, imaging, terminological ambiguities, and emerging directions

RIMSA increasingly appears in sensing and imaging contexts. A dual-functional hybrid RIS integrates simultaneous sensing and reconfigurable reflections by embedding two interleaved sensing arrays with half-wavelength spacing, orthogonal polarization, and quarter-wavelength offset inside a $10 \times 10$1 reflective metasurface centered around $10 \times 10$2 GHz (Birari et al., 23 Jan 2025). A separate distributed-RIMSA sensing framework deploys multiple RIMSA receivers at different places and formulates sensing as joint optimization of beamforming pattern and mapping of received signals to sensing outcomes; when jamming is present, the loss function includes a $10 \times 10$3 term to improve robustness (Wang et al., 7 Aug 2025).

Imaging work pushes the antenna interpretation further. A single transmitting-receiving antenna combined with a RIS can reconstruct the reflection matrix from far-field measurements, effectively transforming the RIS into a programmable synthetic antenna array. In that formulation, the RIS has $10 \times 10$4 elements switchable between two phase states $10 \times 10$5 or $10 \times 10$6, and the measured signal for configuration $10 \times 10$7 is

$10 \times 10$8

After calibration, the full scene reflection matrix is reconstructed from far-field measurements by solving a linear inverse problem (Goïcoechea et al., 13 Dec 2025). This shows that, in at least one strand of the literature, the metasurface is not just antenna-like; it is used as a programmable synthetic array.

The terminology itself remains non-uniform. Some works reserve RIMSA for systems where the RIS acts as an antenna or is modeled as an antenna array (Gu et al., 2022, Wei et al., 23 Jun 2025). Other works discuss dynamic metasurface antennas, transmissive RMS multi-antenna systems, flexible intelligent metasurfaces, or stacked intelligent metasurfaces without always using the exact acronym, even though the functional overlap is substantial (Smith et al., 2017, Li et al., 2021, Yang et al., 14 Mar 2025, Sheemar et al., 5 Mar 2026). A common misconception is therefore to treat RIMSA as a single standardized architecture. The published record instead shows a family of related architectures unified by programmable metasurface apertures and by direct electromagnetic-field control.

Emerging directions expand the design space beyond fixed planar apertures. Flexible intelligent metasurfaces support both element movement and passive beamforming (Yang et al., 14 Mar 2025). Liquid intelligent metasurfaces introduce joint electromagnetic and spatial reconfigurability by allowing the metasurface elements themselves to move within the metasurface area (Shen, 22 Jul 2025). Stacked intelligent metasurfaces place multiple programmable layers at the antenna front end, with the cascaded-operator model

$10 \times 10$9

thereby enabling richer electromagnetic transformations than conventional single-layer designs (Sheemar et al., 5 Mar 2026). Edge-deployed structures such as Reconfigurable Intelligent Surface & Edge extend reflection and refraction over surfaces to diffraction around obstacles’ edge (Xiang et al., 2023). This suggests that the future scope of RIMSA will be defined less by a single hardware archetype than by the degree of programmable control over radiated, reflected, transmitted, or diffracted fields.