Dynamic Reconfigurable Intelligent Surface

• Dynamic RIS is a programmable metasurface that updates its electromagnetic response in real time to control signal propagation and improve communication performance.
• It employs methods like element-wise phase switching, delay control, and adaptive grouping to create artificial fading and dynamically optimize network throughput.
• Advanced control algorithms, ranging from model-based optimization to learning-driven approaches, enable robust, scalable, and CSI-independent performance in diverse scenarios.

Dynamic reconfigurable intelligent surface (RIS) denotes a programmable metasurface whose electromagnetic response is updated in real time, or multiple times within a transmission interval, in order to manipulate propagation conditions for a given communication objective. In the literature, this dynamicity is realized through element-wise phase switching, amplitude and delay control, adaptive grouping, mode switching between reflection and transmission, mobility of the surface itself, and even diffraction-oriented structures deployed on obstacle edges. Across these variants, the common principle is that the propagation environment is treated as a controllable system component rather than as a fixed channel impairment (Liu et al., 2024, Liu et al., 2020, An et al., 2021, Xiang et al., 2023).

1. Dynamicity as a Distinct RIS Operating Regime

A foundational distinction in the RIS literature is between static and dynamic configuration. In RIS-aided multi-user networks, a static RIS configuration is set once and remains unchanged throughout transmission, whereas a dynamic RIS configuration is reconfigured multiple times within a single transmission instance; in the formulation of RIS-aided NOMA networks, the reflection coefficients are updated $N$ times per transmission with equal time intervals, thereby creating artificial fading even in otherwise quasi-static channels (Liu et al., 2020). For large $N$, the dynamic capacity region becomes the convex hull over all feasible static configurations,

$$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$

which is reported as capacity-achieving for the RIS-aided broadcast channel with NOMA (Liu et al., 2020).

This dynamic interpretation extends beyond phase-only reconfiguration. In wideband OFDM, delay adjustable metasurfaces (DAMs) introduce an additional controllable delay per RIS element, so that reflection coefficients and delays are jointly optimized with transmit power allocation (An et al., 2021). In beyond-diagonal RIS (BD-RIS), dynamicity appears as CSI-dependent regrouping of ports into a permuted block-diagonal scattering matrix rather than as a fixed diagonal phase profile (Li et al., 2022). In mobile and aerial variants, the RIS itself becomes repositionable, either through joint optimization of installation position and phase shifts or by mounting the surface on a UAV for adaptive 3D placement (Ji et al., 2020, Abdalla et al., 2020).

A plausible implication is that “dynamic RIS” is better understood as a control regime than as a single hardware archetype. The literature uses the term to cover time-varying reflection states, hybrid transmitting-and-reflecting structures, mobile placements, and hardware that changes operational mode in response to channel or service conditions (Basar et al., 2021, Liu et al., 2024).

2. Hardware Architectures and Control Interfaces

Dynamic RIS implementations differ substantially in how control is delivered to the surface and in which physical degrees of freedom are exposed to the controller.

Architecture	Dynamic mechanism	Representative source
1-bit modular reflective RIS	Varactor-switched phase states with distributed infrared control	(Sayanskiy et al., 2022)
Delay adjustable metasurface	Joint phase and delay control via EIT properties	(An et al., 2021)
Wave-controlled RIS	Standing bias waves replace dense per-element wiring	(Ayanoglu et al., 2022)
Multi-mode “ultimate RIS”	Dynamic switching among active/passive reflect, active/passive transmit, absorb	(Basar et al., 2021)
Dynamic BD-RIS	CSI-adapted group-connected block structure	(Li et al., 2022)

A concrete modular implementation is a 1-bit RIS for the 5-GHz Wi-Fi band assembled from identical building blocks, each containing four separately phase-switchable patch antennas with varactor diodes and a common microcontroller extracting commands from modulated infrared light illuminating the entire RIS (Sayanskiy et al., 2022). Each patch uses two SMV2019-040LF SMD varactors, and the two reflection states are approximately $-90^\circ$ at $U=0$ V and $+90^\circ$ at $U\approx 3.2$ V. The remote controller broadcasts 38 kHz, 940 nm infrared signals; each block includes a TSOP34338 receiver and an ATTiny441 microcontroller, and the protocol supports 128 unique block addresses. An experimental demonstrator with 20×20 patches achieved full 2D phase control and robust IR control from several meters, while preserving the ability to scale the aperture without redesigning control circuitry (Sayanskiy et al., 2022).

The DAM architecture generalizes conventional reflective cells by allowing each element to store and retrieve impinging waves through dynamically controlled electromagnetically induced transparency. Its controllable state includes both $\phi_{k,m}$ and $\tau_{k,m}$, enabling alignment of dominant taps while keeping the total delay spread within the cyclic prefix (An et al., 2021). By contrast, wave-controlled RIS replaces dense bias wiring with a reduced set of spatial voltage standing waves,

$$v(x,t)=V_0+\sum_{p=1}^{P_x} V_p \sin(k_{b,p}x+\phi_{e,p})\cos(\omega_pt+\phi_{v,p}),$$

so that a small set of full-domain modes biases varactors across the surface (Ayanoglu et al., 2022).

At the most expansive architectural end, the “ultimate RIS” is described as a surface that can dynamically switch among active-reflect, active-transmit, passive-reflect, passive-transmit, and absorb modes, and can be partitioned into subsurfaces for PHY slicing (Basar et al., 2021). ETSI’s reference architecture places a micro-controller on the RIS panel and a RIS controller in the BS, the UE, or a dedicated control node, enabling network-controlled or UE-controlled operation in real time (Liu et al., 2024).

3. Optimization and Adaptive Control Algorithms

The algorithmic literature on dynamic RIS spans model-based, feedback-driven, and learning-based control.

A near-field approach based on full-wave electrodynamics uses the multidimensional minimization routines of the GNU Scientific Library (GSL) to optimize the capacitances of varactor diodes on a 5×5 RIS array within finite-difference time-domain simulations. The minimizer operates on $N$0, where the $N$1 are varactor capacitances; the maximization of received intensity is implemented by reversing the objective, and convergence is determined by the gradient norm criterion $N$2 (Colella et al., 2023). In a blocked link scenario with a PEC wall, the optimized RIS increased the time-domain received signal amplitude from approximately $N$3 V to approximately $N$4 V, while $N$5 was consistently higher over 2.5–5 GHz and especially pronounced from 2.5–3.25 GHz (Colella et al., 2023).

For moving users, a self-adaptive beamforming algorithm uses a low-rate feedback loop from UE to RIS controller without requiring UE position knowledge or CSI. In the reported implementation, the UE sends minimal feedback over UDP whenever received power falls below a predefined threshold; the controller then performs greedy subgroup toggling on a hexagonal 127-element RIS at 23.8 GHz (Radpour et al., 23 Apr 2025). The method produced more than 9 dB gain within 50 iterations for a static point and up to 24 dB gain over the baseline with inactive RIS elements across a 60 cm × 100 cm observation area in a semi-anechoic environment (Radpour et al., 23 Apr 2025).

Learning-based approaches become prominent when channel models are hard to obtain or when the control space is very large. In dynamic D2D networks, a CNN-based deep Q-network jointly optimizes RIS position and phase shifts to maximize network sum rate under QoS constraints, using user positions, RIS position, and RIS phases as the state and penalized rewards for QoS violations (Ji et al., 2020). In dynamic rich-scattering environments, a deep neural network is trained as a surrogate forward model for the second-order moments of the frequency response, and a genetic algorithm then searches the discrete RIS configuration space; with $N$6 binary-tunable elements and $N$7 frequency bins, the combined DNN+GA approach achieved a rate up to 15% higher than the average random configuration and within approximately 1 dB of exhaustive search (Stylianopoulos et al., 2022).

Scalable partial-CSI control is represented by RISnet, which combines a WMMSE precoder at the BS with a dedicated neural architecture for RIS configuration. RISnet is reported to optimize 1296 RIS elements while requiring partial CSI from only 16 anchor elements, and to configure the RIS in milliseconds, whereas the cited BCD baseline requires 10+ minutes per configuration (Peng et al., 2023). A separate evolutionary line is memory-enhanced dynamic RIS control, where a genetic search is augmented with memory pools that exploit time/space correlation in user trajectories and channel evolution. The objective is max–min fairness, expressed through the worst-case user throughput, and the algorithm requires only the cascaded BS-to-user channel rather than separate BS–RIS and RIS–user Green’s matrices (Zardi et al., 2023).

4. Electromagnetic and Communication-Theoretic Models

Dynamic RIS research is not limited to optimization routines; it also depends on models that describe which forms of reconfiguration are physically or information-theoretically meaningful.

In single-user semantic communication, the RIS is modeled by

$N$8

and the received signal is

$N$9

The phase optimization problem

$$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$0

admits the coherent-combining solution $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$1, with $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$2 for fair power distribution (Shi et al., 2023). Under Rayleigh fading, the RIS-assisted semantic communication system achieved higher BLEU scores than the point-to-point semantic communication system, especially at low SNR, and degraded more gracefully under channel estimation errors (Shi et al., 2023).

For wideband channels, DAM explicitly incorporates delay control. The implemented delay is quantized as $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$3, and the achievable OFDM rate is jointly optimized over $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$4, $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$5, and $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$6, subject to the cyclic-prefix constraint $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$7 (An et al., 2021). Simulation results in that work report up to 70% rate improvement over traditional RIS in ideal conditions and 40% improvement under practical constraints (An et al., 2021).

Electromagnetic planning tools also appear in dynamic RIS analysis. Dynamical Energy Analysis (DEA) models energy flow in phase space and incorporates local RIS phase gradients through the generalized law of reflection,

$$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$8

so that anomalous reflection can be studied within a meshed environment (Terranova et al., 2022). At the system level, far-field LoS-dominant simulations introduce RIS element patterns and optimal phase compensation for passive elements; under a 7-cell, 21-sector setup with 8 or 16 RIS panels per sector, deployment of RIS increased received power by about 4.6 dB to 6.3 dB and improved SINR by up to 5.4 dB, while 2-bit phase quantization produced performance close to continuous phases (Gu et al., 2022).

Theoretical work also clarifies that dynamicity can enlarge the achievable rate region rather than merely sharpen beams. In RIS-aided OMA/NOMA networks, dynamic reconfiguration yields a larger rate region than static reconfiguration, and the relative gain of dynamic over static operation is reported to be larger for OMA than for NOMA, even though NOMA with dynamic RIS achieves the full information-theoretic capacity region (Liu et al., 2020).

5. System Variants, Deployment Scenarios, and Use Cases

Dynamic RIS has been studied across a wide range of embodiments and operating scenarios.

In THz multiuser MIMO, a dynamic RIS subarray structure divides the RIS into $$\mathcal{C}_{\text{D-RIS}} = \mathrm{Conv}\left(\bigcup_{\boldsymbol{\Theta}\in\mathcal{F}} \mathcal{R}_{\text{NOMA}}(\boldsymbol{\Theta})\right),$$9 subarrays, pairing each subarray with one user so that each subarray reflects beams only to its corresponding user. On top of this structure, a weighted minimum mean square error–local search scheme handles limited phase shifts, and an adaptive BCD-aided algorithm addresses continuous-phase joint beamforming (Liu et al., 2022). Simulations showed higher weighted sum rate than a whole-RIS structure, especially as the number of users or RIS elements increased, with the joint beamforming algorithm often converging in approximately 50 iterations (Liu et al., 2022).

In BD-RIS with hybrid transmitting and reflecting mode, dynamic grouping partitions the RIS cells into non-overlapping subsets according to CSI, yielding permuted block-diagonal reflection and transmission matrices under the passivity constraint

$-90^\circ$0

The reported simulations show that dynamically group-connected BD-RIS outperforms fixed group-connected architectures and can approach fully-connected performance with lower complexity (Li et al., 2022).

Mobile embodiments extend dynamicity into geometry. UAV-mounted RIS is proposed as a mobile RIS paradigm providing adaptive 3D placement for extended coverage, increased capacity, massive access, spectrum sharing, physical layer security, and SWIPT, while raising new issues in channel modeling, channel estimation, control reliability, and payload constraints (Abdalla et al., 2020). Joint optimization of RIS position and phase shifts is also treated directly in D2D communications, where the RIS position $-90^\circ$1 and phase matrix $-90^\circ$2 are optimized for sum-rate maximization under QoS constraints (Ji et al., 2020).

Coverage enhancement near obstacles motivates another extension, Reconfigurable Intelligent Surface & Edge (RISE), which adds diffraction-oriented structures at obstacle edges. In this framework, a diffraction enhancement edge (DEE) combines an ultrathin perfect polarization rotator with a waveguide so that energy is bent into shadowed regions behind edges (Xiang et al., 2023). Simulations at 5.5 GHz report that ES/DEE outperforms conventional surface structures in edge-blocked scenarios, whereas conventional surface RIS is preferred in wall-blocked scenarios (Xiang et al., 2023).

A recurring pattern across these scenarios is scenario-specific specialization rather than architectural convergence. This suggests that future dynamic RIS deployments are likely to remain heterogeneous: near-field focusing, wideband delay control, THz subarraying, obstacle-edge diffraction, and mobile aerial placement solve different propagation problems and therefore privilege different controllable variables.

6. Challenges, Misconceptions, and Standardization Trajectory

The literature identifies CSI acquisition as a central obstacle. Because RIS elements are passive, direct estimation of RIS-assisted channels is difficult, and dynamic RIS compounds the problem by requiring estimation or tracking for multiple configurations (Liu et al., 2020, Abdalla et al., 2020). Several responses are already present in the record: hybrid RIS with a few active sensing elements, low-rate UE feedback without position knowledge, partial-CSI learning using anchor elements, and optimization based only on end-to-end cascaded channels (Radpour et al., 23 Apr 2025, Peng et al., 2023, Zardi et al., 2023).

A second challenge concerns control overhead and hardware realism. Dynamic RIS is often discussed through ideal continuous phase shifts, but reported implementations include 1-bit coding, 2-bit quantization, subgroup toggling, remote infrared control, and wave-controlled standing-bias architectures (Sayanskiy et al., 2022, Gu et al., 2022, Radpour et al., 23 Apr 2025, Ayanoglu et al., 2022). This suggests that dynamic RIS should not be conflated with ideal per-element continuous control. The realizable design space is structured by quantization, shared control resources, electromagnetic coupling, and the latency of the control link.

Power and complexity tradeoffs remain unresolved. Active RIS can mitigate multiplicative path loss and, under the cited parameters $-90^\circ$3 dBm, $-90^\circ$4 dBm, and $-90^\circ$5, can yield approximately 40 dB SNR gain over passive RIS, but it introduces amplified noise and higher power consumption (Basar et al., 2021). UAV-mounted RIS adds payload and endurance constraints, while standalone RIS architectures require embedded sensors or RF chains and more complex local processing (Abdalla et al., 2020, Basar et al., 2021).

Standardization has begun to frame these issues at system level. ETSI launched the Industry Specification Group on RIS in September 2021, with work covering deployment scenarios, use cases, requirements, hardware architectures, operating modes, and future directions (Liu et al., 2024). Two system-level indicators proposed there are the Area of Influence,

$-90^\circ$6

which characterizes the geographical region where a required QoS is met, and the Bandwidth of Influence, defined through the contrast

$-90^\circ$7

which quantifies frequency-domain reconfigurability (Liu et al., 2024).

Taken together, the current literature presents dynamic RIS as an increasingly heterogeneous field: theoretically grounded in capacity-region enlargement and anomalous-wave control, experimentally validated under mobility and near-field focusing, and architecturally expanding toward delay control, diffractive edges, mobile carriers, and multi-mode programmable surfaces. The remaining open problems—robust CSI acquisition, scalable control, realistic hardware abstraction, and cross-layer orchestration—are not peripheral implementation details; they define the research frontier of dynamic RIS itself (Liu et al., 2020, Liu et al., 2024).