1 Ris

Digital Communications and Networks 8 (2022) 87–104
Contents lists available at ScienceDirect
Digital Communications and Networks

journal homepage: www.keaipublishing.com/dcan
Reconfigurable intelligent surface: design the channel – a new opportunity

for future wireless networks
Miguel Dajer, Zhengxiang Ma, Leonard Piazzi, Narayan Prasad, Xiao-Feng Qi, Baoling Sheen,
Jin Yang, Guosen Yue *
Wireless Access Lab, Futurewei Technologies, Inc., Bridgewater, NJ, 08807, USA
A R T I C L E I N F O A B S T R A C T
Keywords: In this paper, we survey state-of-the-art research outcomes in the burgeoning field of Reconfigurable Intelligent
Reconfigurable intelligent surface Surface (RIS), given its potential for significant performance enhancement of next-generation wireless commu-
MIMO nication networks by means of adapting a propagation environment. Emphasis has been placed on several aspects
Channel estimation
gating the commercial viability of future network deployment. Comprehensive summaries are provided for
Beamforming
Reflectarray
practical hardware design considerations and broad implications of artificial intelligence techniques, as are in-
Transmitarray depth outlooks on the salient aspects of system models, use cases, and physical layer optimization techniques.
Machine learning
Deep learning
1. Introduction dielectric material comprised of an array of passive sub-wavelength

scattering elements with a specially designed physical structure. Its ele-
As the world witnesses growing deployments of the Fifth Generation ments can be controlled in a software-defined manner to change the
(5G) network, the wireless industry is casting its eyes on Beyond-5G (B5G), ElectroMagnetic (EM) properties (e.g., phase shift) of the reflection of the
or 6G, technological breakthroughs in anticipation of proliferating new incident Radio Frequency (RF) signals. By using a joint phase control of
applications and use cases that demand wireless communication-and- all scattering elements, the reflected radiation pattern of the incident RF
sensing networks with vastly higher agility, coverage, and throughput signals can be arbitrarily tuned in real time, thus creating new degrees of
[1–3]. Emerging B5G applications include autonomous driving, tactile freedom for the optimization of the overall wireless network perfor-
remote interaction, and augmented reality. New technological enablers mance. The phase-only version of the RIS is often referred to as Intelli-
capable of delivering a tenfold performance gain over their 5G counterparts gent Reflecting Surface (IRS) [4].
will be required for such applications in much the same way that massive The passive and conformal nature of an RIS allows for straight-
Multi-Input-Multi-Output (MIMO) emerged as a promising research topic a forward integration on existing surfaces (e.g., walls and furniture), non
decade ago, becoming a critical enabler for the evolution of commercial invasively augmenting the environment of existing wireless networks for
deployment from 4G to 5G. increased spectral efficiency and energy efficiency without altering the
Among candidate technologies at a nascent research stage, Recon- wireless standards and designs for transceivers already in deployment. In
figurable Intelligent Surfaces (RISs) have received much interest from addition to enhancing the performance of wireless communications
academia and industries [4,5]. A wireless mobile communication networks, RISs have been identified as aiding sensing, localization, and
network traditionally adapts transmission schemes of the transceiver wireless power transfer (see, e.g., Refs. [6–8]). In particular, it is
end-points to mitigate or leverage the dynamic multipath propagation potentially beneficial for networks operating in high-frequency bands
environment for optimal performance; however, such networks do not (e.g., millimeter-wave and terahertz) where blockage- and
have the means to control the environment. RISs offer a potential solu- absorption-induced dead zones and channel rank deficiency are perva-
tion to achieve a software-configurable smart radio environment. sive performance bottlenecks that would otherwise necessitate
Conceptually, an RIS is a large and thin metasurface of metallic or ultra-dense Base Station (BS) deployment with increased power
* Corresponding author.
E-mail addresses: [email protected] (M. Dajer), [email protected] (Z. Ma), [email protected] (L. Piazzi), [email protected] (N. Prasad), xiaofqi@
gmail.com (X.-F. Qi), [email protected] (B. Sheen), [email protected] (J. Yang), [email protected] (G. Yue).
https://doi.org/10.1016/j.dcan.2021.11.002
Received 12 October 2020; Received in revised form 25 October 2021; Accepted 10 November 2021
Available online 15 November 2021
2352-8648/© 2022 Chongqing University of Posts and Telecommunications. Publishing Services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an
open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Dajer et al. Digital Communications and Networks 8 (2022) 87–104
consumption and backhaul cost (e.g., Refs. [9–11]).

Several review articles on RIS have been published [12–15]. More
recent ones [4,16] provide comprehensive, state-of-the-art accounts of
surface and network models; ultimate limits on network performance;
optimal design at the EM, physical and network layers; comparisons
against legacy technologies (massive MIMO and active relay); and
outstanding challenges and future directions of this burgeoning field. In
particular [4], emphasized the importance of adopting physically
consistent surface interaction models based on synthesis and analysis
tools derived from Surface EM (SEM) theory. A common theme is that
system-level understanding of the ultimate performance limit of an RIS
network involving both active endpoints and distributed passive sur-
faces, and the network architecture responsible for such performance, is
still in its infancy [4,13,16]. It is apparent that the progress made so far
has been primarily achieved through means of analysis and simulation,
often on systems under ideal assumptions (e.g., perfectly known envi-
ronment state information, infinite-resolution phase control, and
simplified electromagnetic model of the surface) while also overlooking
protocol and computation overhead associated with real-world network
deployment. Furthermore, as the propagation environment has a much
larger physical footprint than the communications endpoints, power Fig. 1. Reconfigurable surface.
concentration on receivers, which was previously achieved by colocated
coherent endpoint beamforming (e.g., massive MIMO), must now be between the angle of incidence and angle of reflection and refraction is
achieved by multiple distributed passive surfaces over considerably governed by Snell's law. Particularly for the reflection, for regular sur-
larger surroundings to capture sufficient transmitted energy that can then faces of a medium, the angle of incidence is the same as the angle of
be manipulated by the surfaces. To justify a commercially viable reflection, as shown in Fig. 1(a). Recent research advances on reflectar-
deployment, unit-area cost and power consumption must be sufficiently rays with metasurfaces have enabled changing the impedance of the
low for the total cost of ownership to be on par with that of a conven- surface and achieving a certain phase shift between the incident and
tional co-located or distributed MIMO network with similar performance. scattered waves [19]. When the surface is divided into a large number of
Therefore, an inexpensive low-power surface is critical for its successful closely spaced elements and each metasurface element is made to have
commercial deployment, with tailored physical and network layer opti- appropriate phase shifts, the EM wave is tuned to some other angle
mization techniques being required to deliver better performance at a instead of the symmetric reflective wave based on Snell's law. Ideally, as
lower cost and power consumption. Therefore, a more comprehensive shown in Fig. 1(b), if the phase shift of each metasurface element can be
understanding of recent experimental studies to this end (see, e.g., Refs. configured to any value, a reflected beam can be formed at any angle
[17,18]) is warranted. On a separate note, Machine Learning (ML) and [20]. However, to form the reflected beam with an incident EM wave, the
Artificial Intelligence (AI) techniques have been covered in the afore- phase shift of the surface element has to be set appropriately or smartly;
mentioned review papers, albeit in a brief and somewhat ad hoc manner, here, signal processing design techniques or machine learning-based
as a promising augmentation to model-based approaches in RIS approaches can be applied, resulting in an RIS. In practice, precise
system-level optimization, especially for massive deployment of inex- continuous or analog phase shifts may be difficult or expensive to achieve
pensive surfaces that defy parametric modeling. due to calibration and control requirements. Recently, as reported in
This review paper aims to accentuate individual RIS research aspects Refs. [19,21,22], configurable RISs with quantized phase shifts have
deemed relevant to a system-level tradeoff between three prerequisites been designed.
for viable real-world network deployment: performance, cost, and power While most research interests in RISs focus on the IRS, as already
consumption. Thus, in addition to the basic concepts of RIS, channel described, other RIS functionalities are worthy of further research. As
model, and the PHYsical layer (PHY) design optimization, we provide an illustrated in Fig. 2, if situations are limited to either the input or output
in-depth review of hardware design and implementation along with a signal being a plane wave, the basic functionalities of RIS can be sum-
comprehensive summary of ML and AI techniques applied to an RIS marized as follows [4]: 1) reflection/refraction, where a plane wave is
system. The rest of the paper is organized as follows. Section 2 summa- diverted from its original direction of propagation to another direction;
rizes practical system-level RIS use cases and corresponding channel 2) absorption, where the amplitude of a plane wave is substantially
models. Section 3 highlights studies on PHY designs and algorithms by reduced; 3) focusing/collimation, where a plane wave is focused to a
considering design constraints resulting from system performance/cost/ single point, or a spherical wave from a point source is converted to a
power tradeoffs. Section 4 elucidates the hardware needs, reconfigurable single plane wave; and 4) polarization modification, where reflection,
nature, complexities, and future directions that link today's RIS efforts to refraction, focusing, or collimation involves changing the polarization of
the learned experiences from traditional transmitarray and reflectarray the incoming wave, such as from linear polarizations to circular polari-
antenna research. Section 5 highlights recent works on applying ML and zations. There is virtually no difference between RIS and the more
AI techniques to the optimization of real-world RIS networks that defy traditional reflectarray/transmitarray antennas [23,24] as far as these
simple analytical characterization. Section 6 articulates system-level basic functionalities are concerned. In addition, reconfigurability re-
open questions facing a viable real-world deployment of RIS- quires that the parameters of each functionality can be dynamically
augmented wireless networks. Section 7 concludes the article. changed, for example, the direction of reflection/refraction, the focal
point position, etc.
2. RIS systems and channel models
2.1. Reconfigurable surface and basic functionalities 2.2. RIS systems
Fundamentally, when a wireless signal, or in general, an EM wave As shown in Fig. 3, a basic RIS system for research studies comprises a
reaches the boundary between two isotropic mediums, the relationship transmitter, a receiver, and an RIS panel with programmable phase shifts
88
Currently, most research studies on RISs focus on only phase-shift vari-

ants, including system design [26–28] and hardware design [22,29].
Phase and Amplitude Control: As the heading suggests, in this RIS
category, both phase shift and reflection amplitude can be controlled.
Designed in Ref. [21], the so-called digital metasurface is a reflective
surface loaded with Positive-Intrinsic-Negative (PIN) diodes. By applying
Fig. 2. Illustration of basic functionalities of an RIS (Adapted from Ref. [4]). control voltages to a particular PIN diode, the reflection coefficients of
the surface elements can be dynamically controlled with discrete phase
or amplitude states. With more controlled variables, more design flexi-
bilities can be exploited. Consequently, superior performance can be
achieved. However, this category's disadvantage includes a considerable
increase in complexity. Another scenario in this category is that the
amplitude is controlled by the phase shift. As shown in Ref. [30], the
reflection amplitude coefficient of the surface element is a function of a
configured phase shift. In this case, although the amplitude cannot be
independently configured, it can be designed through phase consider-
ation by jointly considering the impacts on the amplitude and phase-shift
over the channel. If the model in Ref. [30] holds for most practical
hardware design even with the phase shift, the joint consideration on
both phase shift and amplitude is important for future studies.
A special case of the third category is a combination of the first two. If
the hardware design allows, the phase-shift RIS can have an off state
Fig. 3. A wireless transceiver system assisted with a reconfigurable intelli- introduced to each element. Rather than always reflecting the incident
gent surface.
signal with a phase shift, the element can be switched off to let the signal
pass through, as in Ref. [17]. In such a case, the amplitude is set to
and/or the reflective amplitude. The RIS panel reflects the incident signal 0 when the element is off and to a fixed scalar when the phase shift is
from the transmitter. From the perspective of a wireless communication configured. This design perspective for future research can have a large
system with a controllable phase shift on each element, the RIS changes impact on the resulting performance, particularly for current hardware
the channel environment from the transmitter to the receiver. The designs with very limited discrete phase shifts [22].
resulting channel between the transmitter and receiver now comprises
two components: the direct channel from the transmitter to the receiver 2.2.2. System settings
without involving the RIS and the channel with the RIS interaction. Using We now describe several system settings for RIS PHY research based
appropriate designs for the RIS phase shifts, amplitude, or both, a certain on a different antenna and user configurations. From the PHY design
metric objective, such as the system achievable rate or the coverage, can considerations, different settings in the following may lead to different
be optimized by changing the channel environment. This is fundamen- design objectives and philosophies.
tally different from conventional wireless communication research,
where the design and optimization opportunities are confined to the pair Single-antenna transceiver systems: In RIS-assisted wireless communi-
of transceivers. cation with a single antenna at the transmit and receiver ends, the
channel between the RIS and the transmitter as well as the reflected
2.2.1. RIS configurability channel after RIS interactions are both vectors. The effective channel
The key feature that makes the RIS attractive is its ability to change from the transmitter to the receiver for the transmitter-RIS-receiver path
the wireless environment (specifically, the wireless channel between the is scalar, i.e., a Single-Input Single-Output (SISO) channel. With a single
transmitter and receiver), and improve the system performance. We antenna at the transmitter, the transmission design does not affect the
present a brief summary of RIS configurability based on hardware de- spatial channel characteristics at the RIS but focuses on the RIS config-
signs as follows. urations. The SISO case (commonly used for indoor design) is a system
On-Off RIS: A simple RIS configuration is a one-bit on-off switch, setting for studying the RIS behavior and its impact on the overall
which has been designed and built in Ref. [17]. Each surface element can channel. Many RIS research works start with an emphasis on single-
be configured in one of two states, i.e., either the on state, when the antenna systems such as [31] deriving the error probabilities of an
signal is reflected, or the off state, when the signal passes through the RIS-assisted SISO transceiver system [32], providing the design of
element. There is no configurable phase shift on the incident signal for discrete phase values on an RIS to maximize the achievable rate [26,33],
reflection in the design. Using a properly designed method of on-off se- considering wideband OFDM transmissions and design phase shifts for
lection, the RIS can achieve power gain in the order of N2, where N is the the wideband transmission with different channels over different
number of elements in the RIS [17]. However, with an ideal on-off RIS subcarriers.
(e.g., 1-bit on-off RIS without any phase shift), there is limited potential
to reshape the reflected beam pattern for a far-field incident radio wave.
MIMO systems: With multiple antennas at the transmitter and the
The power of the reflected signal that reaches the user location may be
receiver, MIMO has been widely adopted in commercial wireless sys-
significantly lower than the total power of the reflected signal. Another
tems. Here, we include cases of a single receive antenna, i.e., Multiple-
design of on-off RIS can be found in Ref. [25], where the surface material
Input Single-Output (MISO), as the transmit beamforming exploiting
can be switched between nearly total reflection and total absorption of a
channel spatial diversity is a key feature of MIMO systems (other than
particular polarized incident wave through appropriate bias voltages that
spatial multiplexing). As transmit precoding or beamforming changes the
turn the diodes on and off in the RIS design.
spatial characteristics of the signal, the interaction with RIS is then
Phase-Shift RIS: In a phase-shift RIS, a phase shift is introduced to the
different. As such, the RIS design needs to be adapted to the particular
impinging radio wave so that the radio signal reflected over all the ele-
transmit precoding. The transmit precoding can be very dynamic (e.g.,
ments can be directed toward the desired location or coherent signal
changing over every time slot), making the joint design of RIS and
combining at the desired receiver [19]. The reflection amplitudes can
transmit precoding essential.
differ among the surface elements, but they cannot be actively controlled.
89
Single-user MIMO: For single-user MIMO, the spatial multiplexing the RIS is N. Denote the channel between the transmitter and RIS and
and/or spatial diversity of the MIMO channel can be exploited. These two between RIS and receiver as hTx,RIS of sizes N 1 and hRIS,Rx, respec-
aspects are also studied in RIS-assisted single-user MIMO systems, e.g., tively. Denote φi and ci, i ¼ 1, …, N, as the phase shift and scaling factor,
how to improve the beamforming gain with the channel path through a respectively, at the RIS element i. The overall channel from the trans-
large RIS or the spatial multiplexing by enriching the channel scattering. mitter to the receiver is given in Refs. [33,44] as
In Ref. [34] the joint design of active transmit beamforming, and passive
RIS beamforming was considered to exploit channel diversity with RIS in hRIS ¼ hTTx;RIS ΦhRIS;Rx (1)
a downlink MISO system. In Ref. [35], the authors considered the RIS
phase shift design to improve the rank of the channel matrix, i.e., the where Φ of size N N is a diagonal matrix with the entry Φi;i ¼ ci ejφi . In
spatial multiplexing gain and, consequently, the MIMO capacity. The most studies, it is assumed that ci ¼ 1 for all elements, i.e., full signal
study in Ref. [36] considered the joint transmit and RIS beamforming reflection [26,37,45–48]. The model can be easily extended to cases of
design for RIS-assisted Simultaneous Wireless Information and Power multiple antennas at the transmitter and the receiver. From Eq. (1), it is
Transfer (SWIPT) with the design objective of transmit power clear that with 1-bit on-off RIS, a reflected beamforming vector from RIS
minimization. is formed with selected elements of the Tx-RIS channel vector hTx,RIS.
Therefore, an on-off RIS has limited potential to change the beam pattern
Multi-user MIMO (MU-MIMO): In MU-MIMO, particularly massive of the incoming channel hTx,RIS. Compared with a phase-shift RIS, an
MIMO, the transmitter has an antenna array that can serve multiple users on-off RIS beamforming gain (after combining with the RIS-Rx channel
simultaneously by exploiting multiuser spatial diversity with multiuser hRIS,Rx) is limited and very opportunistic.
beamforming/precoding. For RIS-assisted MU-MIMO, it is clear that Recently, a novel RIS channel model was proposed in Ref. [30], where
multiple RISs allow spatial diversity among users to be improved. it is claimed that the proposed model is more practical than the simple
However, with a single RIS panel, designing an efficient RIS to improve phase-shifting channel model. Instead of a constant scaling factor ci for
MU-MIMO performance is challenging as the transmit-RIS path is the each surface element, it was modeled as a function of the phase shift φi,
same for all users, particularly for the rank-deficient Line-of-Sight (LoS) i.e., ci(φi), given as
channel between the transmitter and RIS. Therefore, the joint design of k
sinðφi φ0 Þ
RIS and transmit MU precoding is essential. Here are several studies on ci ðφi Þ ¼ ð1 cmin Þ þ cmin (2)
2
how an RIS improves MU-MIMO performance.
In [37], considering the Zero-Forcing (ZF) precoding at the trans- where cmin, φ0, and k denote the minimum amplitude, the horizontal
mitter for composite channels with an RIS phase shift, the RIS design for distance between π /2 and cmin, and the steepness factor of the function
MU-MISO was formulated as a joint optimization of transmit power ci(φi), respectively. All three parameters are nonnegative constants
allocation and phase shifts of the surface reflecting elements. The prob- determined by the specific design and implementation.
lem was extended to a wideband OFDM system in Ref. [38], where the
transmit beamformer and RIS phase shifts were jointly designed to 2.3.2. Channel model for transmitter-RIS-receiver link
maximize the average sum-rate over all subcarriers. [39] considered Path loss model: Assuming the far-field approximation, the large-scale
symbol-level precoding for RIS-aided MU-MISO to minimize symbol path loss for the transmitter-RIS-receiver is the product of path losses of
error probability with Quadrature Amplitude Modulation (QAM) and the transmitter-RIS and RIS-receiver links. The average received power
Phase-Shift Keying (PSK) inputs. [40,41] examined the robust beam- PRIS at the receiver for the transmitter-RIS-receiver link follows a pro-
forming for RIS-assisted MU-MIMO transmissions; while [41] considered duction model that is inversely proportional to the product of the
a cognitive radio system and optimized a secondary transmit beam- transmitter-RIS and RIS-receiver distances [44,49], which are given by
forming subject to limit the interference imposed on the primary user. In
Ref. [42], joint symbol level precoding and reflection coefficient design is 1
PRIS ∝ aTxRIS aRISRx (3)
considered for RIS-assisted MU-MISO systems to minimize the transmit dTxRIS d RISRx
power with guaranteed Quality of Service (QoS).
Investigations into multiple RIS-assisted single- and multi-user MIMO where dTxRIS and dRISRx denote the distances between the transmitter
communications have also been performed. In particular, theoretical and RIS and between the RIS and receiver, respectively, while aTxRIS and
analysis on the usage of two surfaces to establish a double reflection link aRISRx are the corresponding path loss exponents.
for single-user communications was pursued in Ref. [43]; the authors For more practical considerations, the path loss models as well as the
showed that under specific geometrical channel models and configura- shadowing in the specifications of the 3rd Generation Partnership Project
tions, a scaling in received Signal-to-Noise Ratio (SNR) of O(N4), where N (3GPP) can be considered. Depending on the RIS system deployments
is the total number of elements across the two RISs, is possible. In (e.g., indoor or outdoor, urban or rural, high-rise building or on the
contrast [11], considered a network with several blockers and employed ground), the path loss and shadowing models for the transmitter-RIS and
stochastic geometry-based tools to quantify the impact of coating a RIS-receiver links can be selected accordingly from Refs. [50,51].
fraction of blockers with configurable metasurfaces on several system Channel model for fading: Appropriate models for small-scale fading
metrics (such as the probability of being in a blind spot etc.). channels hTx,RIS and hRIS,Rx are desirable for RIS research efforts. For
Here, we aim to categorize pertinent system settings for future conventional MIMO, an i.i.d. Rayleigh fading model has been widely
research and highlight a few exemplary existing works for further considered in the literature, which is acceptable for a limited number of
reading. A recent and comprehensive summary of the existing research widely spaced antennas. Although in some research works, the study and
works for the various system settings described above can be found in design of RIS are still based on the i.i.d. fading channel [48], for the RIS
Ref. [16]. with closely spaced elements, such a fading model may not be appro-
priate. The design methodologies proposed in these papers may still be
valid. However, the results as well as the conclusions drawn from these
2.3. Channel models results should be scrutinized, particularly when designing experimental
prototypes and practical systems, as a channel can never be i.i.d. in a
2.3.1. RIS interaction real-world RIS environment.
The following simple channel model of RIS interaction has been For appropriate RIS channel modeling, it is necessary to consider the
considered in many research studies. Consider a single antenna at both Spatial Channel Model (SCM). A simple SCM is the multiscattering ring
the transmitter and the receiver. Assume that the number of elements at model for multiple channel clusters, as described in Ref. [52] for a
90
one-dimensional array. Given an angle spread [ Δl, Δl] and power the free-space path loss of RIS-assisted beamforming and broadcasting.
distribution ρl(θ) for scattering ring l, the spatial correlation for each Experiments were performed in a microwave anechoic chamber; the
scattering ring Rl of size N N can be obtained with each entry given as measurement results matched the modeling results well. Most recent
[53,54]: studies and analysis on near-field channels for MIMO and RIS-assisted
Z communications can be found in Refs. [59,60].
Δl
T ðαþθ
l Þðui uj Þ
½Rl i;j ¼ ρl ðαÞejk dα (4)
Δl 2.4. Challenges and future studies
where θl is center of Angle of Arrival (AoA) for cluster l, k ¼ A significant challenge for the modeling efforts described in this
2π
λ ðcosðαÞ; sinðαÞÞ
T section is channel modeling with RIS interaction. Almost all existing
is the wave number vector for a planar wave with an
AoA of α, λ is the carrier wavelength, and ui and uj are the position research on the PHY design of RIS assumes perfect phase shift modeling
vectors of elements i and j, respectively. The channel power distribution for the composite transmitter-RIS-receiver channel after RIS interaction.
ρl(θ) can be a uniform, truncated Laplace, or truncated Gaussian distri- Therefore, the accuracy of the phase-shift model for the impinging radio
bution. The overall spatial covariance matrix for multiscattering rings is a signal is critical as model error would greatly affect design performance.
P
linear combination of Rl, i.e., R ¼ l ρ~l Rl , where the weight ρ~l is the In most RISs designed using the phase shift, optimization problems are
P
power ratio for scattering ring l and l ρ~l ¼ 1 [52]. Replaced with 3D formulated and solved. With the modeling error, the solution may no
angles in Eq. (4), the multiring model can be easily extended to a longer be optimal. If it is difficult to model the precise phase shift, an
3D-channel model with a 2D array for RIS systems. error model on the phase shift needs to be studied to evaluate the design
Based on an eigen decomposition of covariance matrix R, the fading algorithms in terms of performance sensitivity of the solutions with
channel can be generated using the Karhunen-Loeve transform with respect to the modeling assumptions.
Gaussian complex random vectors. For correlated MIMO channels, we The modeling error issue became critical when the authors of [30]
obtain the covariance matrices for both ends, Rt and Rr, using the mul- showed that the amplitude of RIS reflection is a function of phase shift
tiring model and generate the MIMO channel as in Ref. [55]. and provided a model of the scaling factor. The proposed model was
For wideband OFDM systems, the channel delay profile can be verified by comparing the numerical results with the experimental design
introduced using an extended scattering ring model [56], where each parameters in Ref. [61]. However, the question of how well the model
scattering ring is assigned a path delay. One scattering ring is assigned matches practical hardware implementation remains, giving rise to the
zero delay to represent the channel cluster for the first arriving signal question of how an error might affect the system design when using the
path, while other path delays are uniformly distributed in a region of [0, exact mathematical expression in the model for design formulations.
τDS], where τDS is the delay spread. With multi-scattering rings and a Nonetheless, it may be safe to use the model presented in Ref. [30] for
delay profile, a spatial-frequency covariance is defined for the vectorized numerical evaluations.
channel in both spatial and frequency domains [56]. The wideband Another key aspect in IRS modeling is to incorporate coupling be-
channel can be generated based on the specified spatial-frequency tween closely spaced elements. While a model that captures coupling
covariance, again using the Karhunen-Loeve transform. with high fidelity is clearly beneficial, the model's tractability will
The aforementioned scattering ring model is a simple far-field sta- determine how conducive it is for optimizing different criteria, such as
tistical channel model sufficient for modeling the channels of the energy efficiency and spectral efficiency. In this context, tools developed
transmitter-RIS and RIS-receiver links. Based on the corresponding in Ref. [62] for modeling compact MIMO arrays can be a useful starting
discrete SCM with power angle profile, the MIMO fading channel for a 1- point for the far-field regime.
qffiffiP Better modeling of the RIS interactions is an important direction for
D antenna array can be generated as H ¼ 1L Ll¼1 hl br ðθr;l ÞbH
t ðθt;l Þ, where future research, particularly for various RIS hardware designs described
L is the number of paths, bt(θt,l) and br(θr,l) are the beam steering vectors in Section 4. That being said, if it is difficult to reduce the modeling error,
for the transmitting and receiving ends, respectively (with their entries it is worthy of investigating robust design in future studies.
T
ðθÞut;i T
ðθÞur;i
given as bt;i ðθÞ ¼ ejk and br;i ðθÞ ¼ ejk , respectively), and hl with As previously noted, near-field channel modeling with RIS in-
2
Efjhl j g ¼ pl is the complex channel gain for each path. The power profile teractions has received less attention. In particular, such modeling of RIS
{pl} can be generated according to a certain distribution (e.g., a log- deployed in indoor environments, homes, or enterprise settings is
normal distribution). For some deployment scenarios (in particular, the another important research area for future studies.
transmitter-RIS channel), the channel can be considered sparse, meaning
that the number of paths or clusters is small; moreover, for the multiring 3. PHY design and algorithms
model, the angle spread is also small. For an outdoor high-rise RIS sce-
nario, the channel can be modeled as single-path LoS. For more practical 3.1. Channel estimation and reconstruction
channel models, particularly for system-level evaluations, the 3D channel
model specified in 3GPP reports [51,57] can be used for RIS studies of It is evident that without the availability of reliable channel estimates,
different deployment scenarios. Given a specific 3D environment map, a phase patterns (and RIS response) cannot be optimized. However,
ray-tracing based channel model can also be considered [26,58]. channel estimation in RIS-aided communications is quite challenging in
Near-field Channel Modeling: The far-field assumption may not hold for practice and is elucidated in this section. As almost all studies on channel
scenarios in which the RIS is sufficiently large or close to either trans- estimation consider Time-Division Duplexing (TDD), we will begin by
mitter or receiver, particularly for indoor environments with a large RIS, considering RIS-assisted TDD systems and later briefly consider the
e.g., the settings in Ref. [17]. Therefore, in such cases, we need to extension to Frequency-Division Duplexing (FDD) systems. The key
consider near-field propagation properties in the channel modeling [14]. advantage in TDD systems is channel reciprocity. However, there is a
However, few studies have done so. In Ref. [49], a free-space path loss penalty to pay in terms of the latency and coverage in the uplink. The
model for RIS-assisted communications was developed, providing a latter arises from having to switch in time between uplink and downlink,
general path loss formula that can be applied to the near-field scenario. It forsaking the use of a lower carrier frequency in the uplink, thereby
showed that the free-space path loss in the near field is proportional to incurring a higher path loss. There is another classification for RIS-
ðdTxRIS þ dRISRx Þ2 instead of the product d2TxRIS d2RISRx of the far-field assisted communications. In particular, we can classify explicit
case. Based on the general path loss formula, the path loss formulas for channel-estimation schemes into two broad categories: (1) schemes that
near-field beamforming and broadcasting were derived to characterize require the use of a few active elements and have RF chains or baseband
processing capability at the RIS, and (2) schemes that assume purely
91
passive RIS elements with no such baseband processing capability. Let us Ref. [67], wherein a particular type of inherent sparsity was exploited.
consider these two categories in more detail. All the works discussed so far have considered TDD, wherein analog
observations are available as inputs to the reconstruction algorithm. One
3.1.1. Channel estimation with baseband processing capability at the RIS of the very few works directly applicable to an FDD scenario (where
Schemes in this category separately estimate each of the channel links instantaneous channel reciprocity cannot be exploited, or more broadly,
in Eq. (1), namely hTx,RIS and hRIS,Rx. As a result, the rich toolkit of applied to a scenario where quantized feedback information must be
techniques developed for massive MIMO channel estimation can be employed as input to the reconstruction algorithm) is [17], which
readily applied, especially those that can reconstruct the full channel by directly estimated an optimized RIS pattern by relying on the average
relying on a limited number of analog observations, each observation received signal strength feedback. However, while the authors assumed
being a projection of that channel corrupted by noise. Indeed, by no simplifying structure in the underlying cascaded channel, the
exploiting reciprocity in TDD, each of these channel vectors can be voting-based algorithm suggested applicability to only on-off control of
estimated by transmitting pilots from the transmitter and receiver, RIS elements. Indeed, directly determining an optimized RIS on-off
respectively, and by collecting the corresponding analog observations at pattern instead of reconstructing the underlying effective channel was
the RIS. An estimate (or reconstruction) of each channel vector is then justified as more sample-efficient by invoking the limited on-off control
obtained by processing these collected analog observations. In particular of possible RIS elements [17]. More recently, a multiple beam training
[26], reconstructed channel vectors by relying on compressive sensing scheme was proposed in Ref. [69], albeit after adopting a simplified
tools. The channel vectors, hTx,RIS and hRIS,Rx, are assumed to have a geometrical channel model. This scheme allows users to determine their
sparse representation in the beam domain, which is then exploited via respective optimized beams and can reduce the overhead considerably
sparse reconstruction formulations. By contrast [63], proposed an while offering similar performance compared to conventional beam
Alternating Direction Method of Multipliers (ADMM) based algorithm by training schemes that entail a full sweep over available beam directions.
adopting a matrix completion-based formulation. In addition to sparsity, On the contrary, in channel links having non-LoS components and where
the formulation in Ref. [63] also exploited a low-rank property of an finer control of RIS elements is possible, explicit channel reconstruction
underlying matrix and modeled the available observations as randomly becomes important while seeking the gains achievable via an optimized
sampled entries of the matrix that were further corrupted by noise. RIS pattern. For such reconstructions, an interesting avenue for research
would be to extend the nonconvex, quadratically constrained, quadratic
3.1.2. Channel estimation without baseband processing capability at the RIS programming based FDD massive MIMO channel reconstruction pro-
Schemes in this category assume the RIS to have only passive ele- posed in Ref. [70], which combined subspace side-information with
ments. An early representative work under this category proposed a two- limited instantaneous quantized feedback information. (see Ref. [71] for
stage scheme for (single-user) channel estimation [64]. Here, in the first related work) In this context, dictionary-based interpolation approaches
stage, approximate message-passing techniques were utilized for sparse for inferring downlink covariance (and hence downlink subspace infor-
bilinear matrix factorization to obtain the RIS-to-user channel along with mation) based on observed uplink estimates have been gainfully
another matrix term. In the second stage, the latter matrix term was used exploited over conventional massive MIMO communications [72].
as the input to a low-rank matrix completion algorithm to obtain the Another key aspect in training via RIS-assisted communications is the
RIS-to-transmitter channel. However, this two-stage scheme assumes that design of RIS patterns suitable for the training phase. While [66] pro-
each RIS element can be turned on and off and that in its on state, its posed certain formulation for designing RIS patterns based on an inco-
reflection coefficient can be configured to any arbitrary phase term. herence criterion that facilitates sparsity-based reconstruction, the
Another representative example is the three-stage scheme proposed in design of RIS patterns that can be effectively used during the training
Ref. [65] that derives dimensioning rules to decide the number of pilot phase is an important open problem. We recall that generalized Lloyd
resources for a given configuration, which comprises the number of RIS algorithms have traditionally been used for vector quantizer design and,
and BS antennas as well as the number of users and antennas per user. more recently, for massive MIMO analog codebook design [73]. Such
While the authors in Ref. [65] assumed a worst-case modeling in which algorithms can also be leveraged to construct a set of RIS patterns suit-
no sparsity was assumed for any channel links and all the channels were able for training, where finite alphabet constraints on these pattern
assumed to be full rank, a key observation that the RIS-transmitter link is vectors can be imposed via a proximal distance-based approach (see
common for all users was exploited. This allowed a much more benign Ref. [67] for one such provably convergent algorithm).
scaling of pilot overhead in the number of users. Yet another notable In summary, while channel reconstruction (estimation) for a single-
work proposed a multistage scheme [66], exploiting the aforementioned RIS-assisted communication has developed rapidly, an effective multi
fact that the RIS-transmitter link was common for all users. In addition, it time scale framework to estimate covariance (subspace) and instanta-
assumed a uniform linear array at the RIS and exploited a particular neous channel coefficients has not yet been developed. Such a framework
row-column block sparsity structure possessed by the effective channel in would entail effective provisioning of training resources, reconstruction
RIS-aided communications; such an exploit does not arise in conven- algorithms, as well as tracking and change detection algorithms. The
tional massive MIMO systems. design challenges and potential gains are substantially magnified as we
A key fact in Refs. [64–66] is that channel estimation was initiated move onto multi-surface scenarios. One of the very few practical works
from scratch, assuming that no a priori side-information, which can that have leveraged multiple reconfigurable surfaces is [74]. However,
reduce overhead, was available. In this context, exploiting subspace the scenario in Ref. [74] was single-user communications with multiple
side-information in combination with other reconstruction methods can surfaces providing a selection diversity where, at each instant, a single
greatly reduce the training overhead, as demonstrated by the results in surface was chosen (based on location and directional information) and
Ref. [67]. Moreover, akin to ML-based approaches, a location-indexed configured to assist communications.
dictionary can be maintained to provide such side information ([27])
or at least a good initial point (aka hot start) for subsequent refinement. 3.2. Optimizing RIS-assisted communications
Constructing and maintaining such a dictionary for RIS-assisted com-
munications is an interesting avenue for future research. Further, a key Multiple works have designed algorithms for throughput gains,
advantage of subspace estimation is that it needs to be done at a coarser coverage improvements, enhancing energy efficiency, and for optimizing
time scale compared to instantaneous channel estimation. Effective al- other metrics over RIS-assisted communication systems, given accurate
gorithms for subspace estimation from low dimensional projections over channel state information. A key motivation is the O(N2) rate of
massive MIMO systems have been proposed in Ref. [68]. An attempt to improvement in received signal power achieved by increasing the number
tailor these algorithms to RIS-assisted communications has been made in N of RIS elements [75]. However, this rate of growth does not hold for
92
arbitrarily large N and breaks down as we transition to a near-field regime; planer or conformal with curvatures. When we limit our consideration to
nevertheless, it can be accrued for large RIS configurations of practical a small area (a few wavelengths) of the RIS that is nearly flat and ignore
interest [59]. Notable examples of RIS optimization are [37], wherein the long-range interactions between different parts of a large RIS, there is
energy efficiency advantages of RIS over amplify-forward relaying were no difference between RIS and more traditional reflectarray/trans-
demonstrated, and [76], wherein a joint beamforming and RIS interaction mitarray antennas [23,24,80,81] in terms of the EM functionalities they
matrix optimization was considered under the restriction that RIS ele- can provide, as summarized in Section 2.1. However, as RIS is envisioned
ments could be assigned discrete phase shifts. to modify the radio environment and as wavefronts in the real world are
More recently, nonideal RIS elements that entail phase-dependent usually more complex than a single plane wave due to complex multipath
amplitude attenuation were modeled in Ref. [30] by leveraging previous propagation, the design of RIS is much more complex and challenging, as
modeling results and experimental validation provided in Ref. [61]. Design elaborated below.
algorithms matching this nonideal model were proposed, and it was shown Any complex wavefront can be decomposed into a linear combination
that even a single-user SNR (or effective channel power gain) maximization of plane waves of various directions and complex amplitudes. To deal
problem with a simple coherent combining and maximal phase with such complex wavefronts, the RIS should ideally support each of the
alignment-based optimal solution in the case of RIS with ideal elements can functionalities with multiple parameters simultaneously. For example, to
become nontrivial under a nonideal element model. The nonideality in- focus a complex wavefront to a single point requires the RIS to focus each
troduces a tradeoff between ensuring phase alignment and reducing of the component plane waves with different incoming directions
amplitude attenuation. Penalty-based algorithms were suggested in coherently to the same focal point. For a more sophisticated example, in a
Ref. [30] to address single and multiuser joint beamforming and pattern multiuser situation, an RIS may be required to simultaneously focus
optimization problems. The key message there was that nonidealities must multiple complex wavefronts to multiple focal points corresponding to
be explicitly incorporated in the problem formulation; moreover, naive the locations of the multiple users. Simple examples of such complex
approaches that directly apply derived solutions by assuming ideal RIS el- requirements have been implemented in reflectarray/transmitarray [82,
ements can incur substantial performance degradation. 83], without incorporating reconfigurability. In Ref. [82], a reflectarray
The design optimization considered in all previously mentioned antenna that can support four beams simultaneously from a single
works assumed a model with uncoupled RIS elements. This assumption is left-hand circular-polarized feed horn was designed and prototyped. In
particularly convenient for optimization as it allows controlling the RIS Ref. [83], a quadrifocal transmitarray antenna that can support four focal
response on a per-element basis. While this assumption is better justified points simultaneously was designed and prototyped,.
for an RIS with elements sufficiently spaced apart (a.k.a. Nyquist In the rest of this section, we begin with conceptual architectures to
spacing), its suitability is questionable in practical RIS designs realize the basic functionalities of RIS using nearly passive RF hardware;
comprising closely spaced unit cells [77]. In addition, optimizing RIS in addition, we provide examples of actual RIS implementations based on
configuration on a per-element (per-unit cell) basis entails a very large that architecture. We then provide an overview of reflectarray and
dimensional optimization problem over RISs comprising several thou- transmitarray antennas, which are the predecessors to RIS as far as RF
sand elements. These issues were recently examined in Ref. [77], hardware design is concerned, and typical technologies for realizing
wherein closely spaced unit cells of an RIS were partitioned into multiple reconfigurable reflectarray and transmitarray antennas. We conclude the
tiles, and the RIS configuration was optimized on a per-tile basis. The section with RIS hardware design challenges and potential future
study showed that even for a given assigned configuration, an RIS affects research directions.
the incident signal differently depending on the incident angle, polari-
zation state, and reflected angle. By jointly optimizing the transmit
precoder and the assignment of modes to tiles, it was shown that speci- 4.1. Conceptual architectures of RIS
fied QoS targets for all active users could be met with significantly
reduced transmit power. The first straightforward conceptual architecture of RIS that can
A relevant problem that arises for a given RIS aperture size is deter- support a good set of basic functionalities is simply a receive antenna
mining the quantum of additional gains achievable by closely packing a array with its tunable beam-forming network and a transmit antenna
larger number of elements as opposed to using fewer elements with array with its independent tunable beam-forming network connected
Nyquist spacing. To assess these gains, a proper modeling of compact RIS together, as shown in Fig. 4(a). The receive antenna array with proper
(along with potentially compact transmit and receive arrays) followed by configuration of its beam-forming network captures the energy of the
the design of well-matched optimization algorithms are imperative. A incoming plane or spherical wave and then passes it to the transmit an-
related useful development was the careful modeling of an active and tenna array, which, in turn, through a proper setting of its beam-forming
compact large intelligent surface by deriving its impedance matrix [78]. network, sends the energy toward the desired direction, as either a plane
Such a surface is capable of transmitting and receiving EM waves. or spherical wave. An example of such an RIS at a small scale can be
Incorporating mutual coupling among elements is not only appropriate found in Ref. [10], where a simple programmable mmWave mirror was
modeling, but it also changes transceiver optimization and permits used to enable a high-quality untethered virtual reality application. The
considerable performance gains. In the context of IRS aided communi-
cations, an initial step has been taken in Ref. [79], wherein impedance
matrices were numerically computed via the method of moment tools,
and the proposed models were validated via antenna simulations.
Further, an effective proximal distance-based algorithm was proposed to
maximize spectral efficiency. Moreover, the study showed that
coupling-aware optimization of RIS loading could be considerably su-
perior in terms of achieved spectral efficiency when compared to its
naive counterpart that disregards coupling.
Fig. 4. Conceptual architectures of RIS with an array of antenna elements: (a) a
4. RIS hardware design receive antenna array and a transmit antenna array connected together with
independent tunable beam-forming networks; (b) a receive antenna array and a
In this section, we discuss the hardware design aspects of RIS. The RIS transmit antenna array connected with element-wise tunable phase shifts; (c) an
envisioned here is an electrically thin (thickness ≪ wavelength) 2D antenna array with tunable reflecting load; (d) a reflectarray with tunable an-
structure with a large footprint (width/length ≫ wavelength). It can be tenna elements to change its response.
93
reflection [4].
A further refinement of the second conceptual implementation is to
reuse the antenna elements for receive and transmit: this reduces the
implementation to simply an antenna array with a tunable reflecting
load. The signals received at each element are essentially reflected back
through the same element with a variable phase controlled by its cor-
responding tunable reactive load, as illustrated in Fig. 4(c). There are
more ways to modify the response of an antenna element than just
Fig. 5. A programmable mmWave mirror prototype and block diagram.
changing its load impedance. The element itself can be tuned directly to
(Reprinted with permissions from Ref. [10]).
modify its properties, such as its resonant frequency, resonant mode, or
current distribution, to effect a change in its response. Thus, we arrive at
mirror prototype is shown in Fig. 5. The advantage of such an architec- the architecture illustrated in Fig. 4(d). This is essentially the architecture
ture is its conceptual simplicity and flexibility. However, the analog of a reflectarray antenna. An example of such RIS implementation can be
beam-forming network can quickly become very complex and lossy, found in Ref. [22], in which 256-element RIS prototypes for 2.3 and 28.5
except for simple plane-wave excitations, especially when the array size GHz were developed, with each element supporting a 2-bit phase shift
is scaled up. In addition, for large arrays and low frequencies, the room implemented using PIN diodes. One of the prototype designs is shown in
needed for the two separate arrays could be problematic. Fig. 7. The PIN state configurations modify the RF current distribution,
The second equally straightforward conceptual architecture is similar resulting in a change in the phase of the reflected waves [22].
to the first, except that it does not employ the analog beam-forming Antenna elements are still traditional in the sense that the element
networks and instead opts for a one-to-one connection between the ele- size and spacing is about a half of the wavelength of the incident waves,
ments of the receive antenna array and that of the transmit antenna array making coupling between elements sufficiently weak and ignorable, such
with tunable phase shifters, as illustrated in Fig. 4(b). This is essentially that each element's effect can be modeled independently from its
the architecture of a transmitarray antenna. An example of such RIS neighbors. However, as the element size and spacing shrink below one
implementation can be found in Ref. [84], where a prototype of a half of the wavelength, the implementation becomes more of a meta-
36-element array of the inexpensive antenna was built using program- surface [85], where an electrically thin surface with a large footprint is
mable phase shifters to demonstrate the concept of programming the composed of unit cells much smaller than a wavelength closely packed
radio environment. The prototype and architecture of this design are together with a spacing smaller than the wavelength. Examples of such
shown in Fig. 6. Besides the conceptual simplicity and flexibility, this RIS implementation can be found in Refs. [17,86]. In Ref. [17], a very
architecture is also simple to implement and the easiest to model math- large low-cost RIS was prototyped with 3200 elements, each with a size
ematically. In addition, as the phase response of each individual pair of of λ/4 and a spacing of λ/10, as shown in Fig. 8. Each element had only 1
receive-transmit elements can be independently controlled, it can deal RF switch to control its on-off state. In Ref. [86], a metasurface RIS design
with complex wavefronts more easily, as discussed in the previous sub- with multibit encoding was proposed and studied theoretically. The
section. However, its major limitation is that it cannot support the schematic of the design is shown in Fig. 9. The elements were claimed to
functionalities that require nonlocal interactions well, such as anomalous be much smaller than a wavelength; however, detailed dimension in-
formation was not provided.
Due to the close spacing of the elements or unit cells in a metasurface,
there are strong interactions among them, and they can no longer be
thought of or modeled as independently controlled scattering/reflecting
elements. A more vigorous treatment would be to abstract the surface as
an infinitely thin boundary that creates a discontinuity in the EM fields
[87,88]. A conceptual generic model of such metasurfaces was proposed
in Ref. [4] and is reproduced here in Fig. 10.
In this conceptual model, the RIS comprises multiple layers of peri-
odic structures called super-cells. Each super-cell comprises several unit
cells of various shapes and sizes that act as passive scatters. In general,
the unit cell is much smaller than a wavelength (~5 to 10 times smaller),
and the spacing between unit cells is of the same order. Strong interaction
exists between unit cells, and their responses must be modeled together
instead of individually. A configuration network is necessary to control
and program the tunable electronic circuits (e.g., PIN diodes or varactors)
to achieve the desired response from the surface. The model also includes
a means to communicate with the external world so that it can be
controlled and programmed remotely or potentially provide sensing in-
Fig. 6. Prototype and architecture of a large array of inexpensive antennas formation collected by optional on-board sensors [4]. contains a lengthy
(Reprinted with permissions from Ref. [84]). section detailing design methodology for specific functionalities based on
this conceptual model, which interested readers are encouraged to read
further.
4.2. Brief overview of transmitarray and reflectarray antennas
The concept of using a planar or conformal surface with properties

that can be manipulated to dynamically influence scattered fields can be
traced to transmitarray and reflectarray antennas. Traditional reflector
Fig. 7. Photo of a fabricated 2.3 GHz RIS prototype and the design of the patch antennas utilize a shaped reflector surface with a fixed transmit feed
element with phase control (©2020, IEEE. Reprinted with permissions location to produce antenna systems with large gains and achieve beam
from Ref. [22]). steering by physically orienting the shaped reflector in the desired
94
Fig. 8. Photo and schematic of an RFocus prototype (Adapted from Ref. [17]).
Fig. 11. Illustration of (a) reflectarray antenna and (b) transmitarray an-
tenna geometries.
element, and k0 is the free-space wave number.

From Eq. (5), there are two methods to induce a dynamic phase shift
over the aperture. One method involves modifying the first term in Eq.
(5) using feed-tuning techniques. As an example, this can be achieved by
switching through multiple feeds or by moving the feed relative to the
reflect array. The second method involves modifying the second term in
Eq. (5) using aperture phase tuning techniques. The two methods are not
mutually exclusive and can be simultaneously employed. In this paper,
Fig. 9. Schematic structure of b-bit encoded RIS (©2020, IEEE. Reprinted with more details about the second method are provided as more research
permissions from Ref. [86]). efforts are focused on this area with novel enabling technologies.
Generally, reflectarray antennas are constructed from lumped ele-
direction. Reflectarray antennas maintained the concept of a feed an- ments (not unlike the driven radiating elements found in some traditional
tenna but introduced various methods to synthesize a phase response on antennas). For instance, patch- or microstrip-type antennas on Printed-
a typically planar reflector surface such that the scattered fields can be Circuit-Board (PCB) material are commonly used. To enable these ele-
directed in the desired direction. In a similar vein, transmitarray an- ments to dynamically change the phase over the aperture, the charac-
tennas (sometimes referred to as array lens) utilize a synthesized phase teristics of element resonance must be changed, or a variable phase shift
response to focus the impinging fields on the other side of the source must be added per element [19]. As an example, three common elements
through the surface. These reflectarray and transmitarray concepts are are shown in Fig. 12.
illustrated in Fig. 11(a) and Fig. 11(b), respectively [80]. The element rotation method shown in Fig. 12(c) is used for circularly
This section of the paper attempts to elucidate the hardware re- polarized fields and can be realized by rotating the element via micro-
quirements, reconfigurable nature, complexities, and directions of RIS actuators or electronically using a spiral phase shifter.
through the learned experiences of transmitarray and reflectarray an- While Fig. 12 shows some examples of reflectarray unit cells, many
tenna research community. Of course, RIS requires progress beyond other designs are possible. One area of growing interest is to use engi-
current designs to dramatically upscale the hardware to create very large, neered materials, or metamaterials, with unit features that are signifi-
low-profile, invisible, and conformal surfaces that must be controlled by cantly smaller than the wavelength of operation. An example of this is
an intelligent autonomous agent in the communication system. given in Fig. 13, where the meta particle trace is approximately λ0/8 in
The phase distribution over the reflectarray aperture must be size.
dynamically changed to realize beam scanning functionality. For the case
shown in Fig. 11(a), the phase distribution on each element of the 4.3. Typical technologies to achieve reconfigurable transmitarray and
reflectarray can be given by reflectarray
!
φm;n ¼ k0 R m;n Δφm;n (5) The previous section provided examples of elements that can be used
to construct a reflectarray. In this section, some of the technology needed
! to implement the reconfiguration of a reflectarray or transmitarray is
where R m;n represents the phase due to the spatial phase difference be-
tween the (m, n)th element and feed location, Δφm,n represents the phase given in Table 1 which was recreated from Ref. [19]; the table provides
shift introduced between the incident and scattered fields on the (m, n)th some examples of the technologies that can be used with particular de-
signs to tune the phase across the aperture of the array. An assessment of
the level of the technology parameters is given with associated references
for further study.
Fig. 12. Reconfigurable phase tuning for reflectarray elements, (a) tunable
Fig. 10. Conceptual structure of an RIS based on metasurface. (©2020, IEEE. resonator (or size), (b) tunable delay line, (c) element rotation. (©2014, IEEE.
Reprinted with permissions from Ref. [4]). Reprinted with permissions from Ref. [19]).
95
Fig. 13. Unit cell for a metamaterial-based reflectarray elements (©2010, IEEE.
Reprinted, with permissions, from Ref. [89]).
While a number of technologies are available for providing phase-

tuning in the transmitarrays or reflectarrays, the best choice must be
part of the overall antenna and system design. As frequencies increase
into the THz band, as envisioned for 6G, some solutions (such as var-
actors) become too large and require reconfigurability by controlling the
material properties.
An example of a typical reconfigurable reflectarray prototype is
shown in Fig. 14 and is worth reviewing in more detail. The top layer
(Fig. 14(a)) comprises patch elements on PCB, followed by the ground
plane reflector with feed slots (Fig. 14(b)), followed by the PCB board
with varactor loaded lines and corresponding bias lines (Fig. 14(c)), and Fig. 14. Prototype of a 3 15-element x-band reconfigurable reflectarray with
finally the Digital-to-Analog Converter (DAC) board (Fig. 14(d)) that four major components: (a) patch elements, (b) a ground plane with slots feeds,
(c) varactor-loaded lines with biasing traces, and (d) a DAC board that sets bias
interfaces between the microcontroller and varactor bias line.
control line voltages (©2012, IEEE. Reprinted with permissions from Ref. [96]).
Fig. 15(a) shows a reflectarray comprising Frequency-Selective Sur-
face (FSS)-backed dual-band unit cell design, which allows simultaneous
operation in the X band (8–12 GHz) and Ku band (12–18 GHz). From transmission amplitude of the field to remain as close to unity as possible
Fig. 15(b), it can be seen that the reflectarray comprises two distinct while simultaneously allowing phase shifts through 360 . In this way, the
surfaces and that the use of the FSS for these surfaces allows for the losses are minimized while the lensing capabilities are maximized.
isolation to be maintained between the bands. Novel designs like this An example of a transmitarray unit cell is shown in Fig. 16(a), illus-
further enhance the frequency range and bandwidth of reflectarray an- trating the stack-up of a multilayer patch element designed to operate in
tennas and are a research area for future explorations. the Ku band [98]. Fig. 16(b) shows the prototype of the transmitarray
Transmitarrays or array lenses are similar in nature to reflectarrays. utilizing 36 patch elements. However, what is not seen is the distribution
Transmitarrays combine the principles of phased arrays and lenses to network and varactors that provide a reconfigurable phase shifter design
allow impinging waves to travel through surfaces. As such, the source is to control the transmit array response.
always on the opposite side of the desired user location. One way to Finally, a complete transmitarray prototype is shown in Fig. 17. The
conceptualize the transmit array is illustrated in Fig. 11(b), where the measurements from the prototype compare very well to the simulation
surface can be thought of as two antennas: one receiving the field from results [98]. A couple of scenarios were measured. In the first case, the
the feed location and the other transmitting the signal after a phase lens array allowed a field to pass through and did not change the radi-
adjustment. This design approach is known as the guided wave method. ation pattern; however, the system corrected for any phase mismatches.
Another method is the layered-scattered approach, where periodic In the second case, the lens array changed the radiation pattern along
structures (such as slots, rings, or patches) in single or multiple layers with one of the main axes by applying a 9-degree tilt. The last scenario
change the electrical behavior of a surface to allow transmission. In any demonstrates the ability of the transmitarray to steer beams dynamically
transmitarray design, it is critical that the surface features allow the [98].
Table 1
Selected technologies for the implementation of reconfigurable reflectarray and transmitarray and qualitative assessment of a few related parameters (‘þ’,‘O’,‘-’ refer to
good, neutral, and poor, respectively. Reproduced from Ref. [19]).
Type Technology Reliability - Integration D/A Complexity Loss Bias Power Linearity Switching References
Maturity Control (Cost) (Microwave/ Consumption Time
THz)
Lumped PIN Diodes þ – D þ / – O þ [90]

Elements Varactor Diodes þ – D þ / þ – þ [91]
RF - MEMS O þ D þ þ/O þ þ O [92]
Hybrid Ferro-electric O þ A O O/- þ O þ [92]
Thin Films
Tunable Liquid Crystal O O A O /þ O O [93]
Materials Graphene – þ A O /þ þ – þ [94]
Photo- O – A O / – – þ [95]
Conductive
Mechanical Fluidic O – A O O/þ þ O – [92]
Micromotors – – A – þ O þ – [92]
96
In general, the synthesis of metasurfaces from desired EM properties

is still an open problem, especially due to the complex requirements
likely to arise for RIS. Thus, it is a rich field for innovation.
The trade-off between size, cost, and complexity is a challenge that
needs to be tackled before RIS application on any meaningful scale. For
example, trying to cram many features into a smaller size will likely in-
crease the hardware design complexity and cost. By keeping it large and
simple, it can be divided into multiple subarrays, with each supporting
only limited functionality, thus allowing lower material cost; however,
Fig. 15. (a) FSS-backed reflectarray unit cell for a frequency-selective surface- this may increase the deployment cost considerably. Innovative ways to
based dual-band design and (b) fabricated prototype of the FSS-backed reflec- fundamentally change such trade-offs and limitations are needed for
tarray (Reprinted with permissions From Ref. [97]). successful RIS application in the real world. For example, the integration
of RIS with building materials (such as window glasses) is a very inter-
esting direction [99].
Finally, with large-scale RIS, the endpoints of communication will
likely be in the near-field of the RIS, especially in indoor applications. As
most models and our own intuitions about antennas are based on far-field
assumptions, there is still much to explore and understand. Potentially,
some phenomena unique to near-field RIS can be leveraged to develop
new applications or considerably enhance the performance of existing
applications.
Fig. 16. (a) Transmit array unit cell and (b) transmit array complete assembly 5. AI/ML for RIS solution design
without feed (©2010, IEEE. Reprinted with permissions from Ref. [98]).
RISs have the potential to revolutionize the design of wireless net-
There is a growing interest in the transmitarray and reflectarray works, particularly when combined and integrated with other 6G
technology. Some research areas being explored are bandwidth candidate technologies, such as AI-empowered wireless networks and
enhancement, improved linearity, active reflectarrays that integrate terahertz communications.
amplifiers in the design, and THz operation. Further, electrically large Recent years have seen an overwhelming interest in using AI for the
arrays (100s λ) have already been built in the Ka band, which gives hope design and optimization of wireless networks. It has been gradually
that RIS-scale surfaces with transmitarray and reflectarray designs are introduced into 5G cellular networks (though more as isolated applica-
feasible. tions) and has been envisioned as a key enabling technology for 6G
[100–102].
4.4. Hardware design challenges and potential research directions In RIS-aided communication applications, there are more appealing
reasons and motivations to apply AI/ML as part of the total solution,
As discussed earlier, the basic RF functionalities expected from an RIS especially to control and optimize the reconfigurable panels and ele-
differ little from a reflectarray or transmitarray. Design concepts from ments. The massive use of meta-surfaces, reconfigurable reflectarrays,
more mature reflectarrays and transmitarrays are readily applicable to and reconfigurable large-intelligent surfaces provides a large number of
many RIS designs. However, RIS does have additional challenges as it is degrees of freedom, whose optimization presents a huge challenge to
envisioned to be much larger in size electrically and be used in much traditional analytical and numerical approaches and entails a large
more complex and dynamic environments. First, the wavefronts an RIS computational complexity. We would like to discuss the motivations
must handle are much less likely to be simple plane waves or spherical behind AI/ML-based approaches in Section 5.1, followed by a detailed
waves due to multipath propagation, unlike the case for reflectarray/ discussion of recent research works and experiments in Section 5.2. We
transmitarray design. Second, an RIS may be expected to implement conclude the section with a summary of the strengths and challenges of
multiple basic functionalities with a larger number of goals simulta- AI/ML based-approaches for RIS-aided applications in Section 5.3.
neously (for example, beam forming simultaneously to multiple users).
Third, due to the large size envisioned, an RIS must support wideband/ 5.1. AI/ML for RIS solution - motivations
multiband operation, complicating the design requirements further. For
example, as the size increases, the bandwidth of an implementation using The growing interests and research on applying AI/ML to RIS-aided
a simple phase shifter decreases considerably. The true-time delay may applications are motivated by two forces: “push” and “pull.”
have to be introduced to achieve wideband operation. “Push” indicates growing problem complexity and demanding
requirements. RIS-assisted networks/systems, such as communication,
sensing, and wireless charging, are complex and challenging to design.
This originates from a large number of parameters requiring optimization
based on the contextual information and real-time decisions necessary
every time the network conditions change (e.g., channel conditions and
positions of the users).
Traditional and analytical solutions build upon communication the-
ories, mathematical models, and optimization algorithms, which have
been illustrated in earlier sections. While solid and successful, they face
new challenges in supporting RIS-assisted systems and applications.
To realize the optimal control of RIS panels, traditional approaches
take the first and fundamental steps of “sensing” and acquiring the
Channel State Information (CSI). CSI acquisition problems in an RIS-
Fig. 17. Complete transmit array prototype with feed antenna and reconfig- assisted system have their own unique challenges [14]. The high
urable surface (©2010, IEEE. Reprinted with permissions from Ref. [98]). dimensionality of the channel space, introduced by a large number of RIS
97
elements, implies longer delays and higher pilot training overhead due to
the number of pilot symbols required and the associated processing
power. The passiveness of an RIS poses additional challenges for signal
processing algorithms. Furthermore, as in any model-based approach, the
quality of an estimated or reconstructed channel depends heavily on the
accuracy of the underlying model. At present, the channel model of an
RIS-assisted MIMO system has not yet been well understood. Even
allowing for a well-suited model, the required accuracy of channel esti-
mation is very much application dependent. Current analytical solutions
have no well-established mechanism to address such a tradeoff between
required estimation accuracy and incurred algorithm processing
overhead.
The telecomminucations field is entering a completely new space
with “extreme” performance requirements for future communication
applications in terms of delay, reliability, and throughput accompanied
with higher frequency spectrum deployment [101]. In traditional
methods, as pointed out in Ref. [103], the “sensed” channel status is
mostly after the fact, and the system can only react to what has already
happened, which makes it extremely difficult to combat unfavorable or Fig. 18. Digital Twin Framework for RIS-embedded environment.
sudden changes in the propagation conditions, especially at high fre-
quency bands.
3) ML for other applications in RIS-assisted wireless communication
At the system level, with or without RIS components, we will see a
network.
growing mixture of solutions, including various cell layouts (macro cell,
small cells, terrestrial/nonterrestrial networks), spectrum ranges
(mmWave, terahertz, and visible light communication), and a whole new
5.2.1. ML to assist channel estimation or reconstruction
dimension of channel dynamics when RIS is employed. It will become
To handle the challenges in channel estimation or reconstructions,
very difficult, if not impossible, to completely rely on traditional model-
studies have applied a new ML-based approach or tried to learn the
based approaches when orchestrating and choosing the optimal combi-
structural features of the channels to reduce the overhead of CSI
nations and configurations of these tools and/or options at various
acquisition.
granularities.
Generally, ML-based approaches require data to train a model for
“Pull” indicates the success of data-driven knowledge learning.
supervised learning. For channel estimation, the input usually includes
We have witnessed the enormous success of AI and ML in computer
received pilot signals, and the output is the predicted channel state.
vision, Natural Language Processing (NLP), network planning and con-
Typically, the training is done in a centralized manner as proposed in
trol, and other fields via models that capture extremely complicated high-
Ref. [47], where the authors proposed ChannelNet, a twin Convolutional
dimension relationships that were not available or feasible from existing
Neural Network (CNN) architecture fed with the received pilot signals for
theories and domain knowledge.
estimating both direct and cascaded channels. In a multi-user scenario,
Traditional communication design relies on theories and models that
each user has access to the CNN model to estimate its own channel. La-
assume minimum or zero knowledge of the surrounding environment. As
bels for direct channel estimation are generated with all the elements
observed in Ref. [104], the extreme performance requirements of many
from the intelligent surface, and the cascaded channel is estimated by
new applications that are anticipated demand a close fitness between the
transmitting pilot symbols when each RIS element is turned on sequen-
radio control and resource management decisions and the underlying
tially. Then, the least square method is used to estimate the cascaded
physical environment (e.g., surroundings and user distribution-
channel. However, this centralized approach poses significant trans-
s/mobility). Conventional analytical approaches, while having shown
mission burdens on wireless links. To overcome this, the authors in
great strength in leading to this point, now face faces growing challenges
Ref. [105] developed a Federated Learning (FL)-based approach for
and limitations to find close fitness to underlying environments and meet
channel estimation, in which a CNN was trained on local datasets of the
higher performance demands.
User Equipment (UE) without sending the data to the BS. The same CNN
This does not have to be the case anymore as argued in Ref. [104]. We
model can be used to estimate both direct and cascaded channels in an
are at the beginning of a new era of industrial transformation enabled by
RIS-assisted massive-MIMO network. Simulation results showed that
sensing, digitization, and connectivity. Combined with the growing
such an approach could reduce transmission overhead by 16 times
power of AI/ML algorithms, digital twins can be constructed for
compared to a centralized approach.
real-world entities (e.g., organizations, buildings, roads, cities, products,
In [106], the authors leveraged ML to design a deep denoising
people, and processes). Such digital twins allow communication systems
NN-assisted compressive channel estimation for mmWave RIS systems to
to self-learn the relationships among high dimensional factors and
reduce the training overhead. The proposal used a hybrid passive/active
leverage the discovered knowledge about underlying environments and
RIS architecture, where a small number of RIS elements were activated to
network behaviors to make decisions according to various granularity
estimate the uplink UE-to-RIS channels. The complete channel matrix
levels. The generalized digital-twin framework for an RIS-embedded
was first estimated based on compressive sensing, whereby the common
environment is depicted in Fig. 18.
sparsity of angular domain mmWave MIMO channels across different
subcarriers was leveraged. After preliminary estimation, a
5.2. Research and experiments on AI/ML in RIS solution design Complex-Valued Denoising CNN (CV-DnCNN) was used to jointly process
the real and imaginary parts of the estimated channel and output the
Recently, many research publications have studied and experimented residual noise. The output from the CV-DnCNN was then used together
with AI/ML-based approaches for RIS-aided application solutions. In this with the preliminary estimation to improve the channel estimation
paper, we organize them in the following three groups: accuracy.
The methods discussed above still apply communication theories to
1) ML to assist channel estimation or reconstruction. the RIS solution framework in general while leveraging ML to facilitate
2) ML for phase shift configuration at RIS and beamforming at BS.
98
the channel-estimation step; then, the optimal RIS configuration and that the proposed method can approach the semi-definite relaxation
beamforming vector at the BS can be determined. An alternative is to performance. In Ref. [48], the authors leveraged the actor-critic DRL
apply ML to predict directly the optimal RIS phase shifts and the beam- approach to study the joint design of a transmit beamforming matrix at
forming vector from available pilots or environment/receiver attributes the BS and the RIS phase shift configuration for a multiuser MISO system.
while bypassing the explicit channel estimation step, which will be dis- The sum rate was utilized as an instant reward to train the DRL-based
cussed next. algorithm. The transmit beamforming matrix and the phase shifts were
jointly obtained (as the output of the DRL NN) by gradually maximizing
5.2.2. ML for phase shift configuration at RIS and beamforming at BS the sum rate through observing the reward and iteratively adjusting the
To determine optimal phase shifts on RIS and beamforming vectors at parameters. In Ref. [46], the authors designed a potential standalone RIS
a BS, an analytics-based solution typically relies on channel estimation, solution also based on DRL. States used in the DRL solution were the
which can be costly given a large number of elements on the RIS. Recent normalized concatenated sampled channel (by the active elements) of
studies showed promising results in leveraging ML techniques to directly each transmitter-receiver pair, with actions being represented by each
predict the optimal RIS phase shifts and beamforming vector at the BS candidate's interaction vector and the achievable rate at the receiver
while either completely bypassing the intermediate step of channel being used as the reward function to train a fully connected NN. Simu-
estimation or with limited knowledge of channel information. lation results based on accurate ray-tracing channels showed the pro-
In [45], the authors developed an RIS architecture that introduced posed DRL approach can converge to near-optimal rates, close to the
only a small number of active RIS elements connected to the baseband, performance of the baseline approach, which utilized supervised
thus granting them sensing capabilities. The solution exploited Deep learning.
Learning (DL) techniques to predict optimal RIS reflection matrices using Unsupervised learning and reinforcement learning do not require
knowledge of the channels seen only by those active RIS elements. collecting a large amount of training data in advance. However, their
Through simulation validation, the proposed approach achieved performance remains a challenge compared to the supervised learning-
near-optimal data rates with relatively low training overhead and based approach. Another challenge (and candidate research topic) is
without any knowledge of the RIS geometry. In Refs. [107,108], the studying the mechanism to control potential performance swings during
authors presented methods that use received pilots as input to train the exploration in a reinforcement learning setting.
proposed DNNs to predict the optimal RIS phase shifts and beamforming
vector at the BS without the channel estimation step. 5.2.3. ML for other applications in RIS-assisted wireless communication
In [27], the authors proposed a DL-based approach to learn the network
mapping between the measured coordinate information at a user location While RIS offers promising performance improvement and more
and the configuration of each RIS element that would maximize the user's efficient wireless communication, there are also many potential RIS-
received signal strength in an offline phase utilizing a preconstructed aided applications. With the advances in ML technology, recent
fingerprint database when training the DNN. In the online phase, the research works studied the feasibility and performance of integrating ML
trained DNN model was fed with the measured coordinates of the target into the solutions for these RIS-aided applications as well. In this sub-
user location to output the optimal phase configuration with a focus on section, we cover a few examples that leverage the strength of ML.
the intended location. To reduce BS power consumption [113], studied the energy efficiency
In [109], the authors introduced a different paradigm in predicting optimization problem in an RIS-assisted wireless network. The authors
the achievable rate at the receiver for any RIS phase shift and then proposed a framework to maximize the average energy efficiency by
derived the optimal phase shift configuration. The proposal does not enabling the BS to jointly determine the transmit power and RIS
require explicit channel estimation effort, thus reducing the costly configuration. The proposal leveraged DRL to enable the BS to derive the
overhead associated with acquiring such information. The proposed so- optimal action for energy efficiency, which comprised the transmit
lution learned the mapping function between the receiver location at- power allocation and RIS phase shift configuration.
tributes with any intended RIS configuration, and the corresponding With the presence of RIS, the communication topology between a BS
achievable rate directly via a domain knowledge-inspired DNN archi- and a device becomes a two-hop (BS-RIS-UE) environment, which makes
tecture. Simulation results showed that such an approach is very Beam Management (BM) more challenging. In Ref. [114], the authors
sample-efficient, which can achieve near-optimal data rates on testing proposed an ML-empowered BM framework for a RIS-assisted mmWave
locations while using less than 2% of the total number of receiver loca- network. The framework included a DL-based environmental awareness
tions during the training phase. module that was trained offline to encode the communication environ-
Simulation results showed good performance for the approaches ment geometries and an online learning model for mobility awareness to
discussed above. However, they all leverage supervised learning tech- predict the future motion state information and short-term trajectory of
nique that comes with the cost of collecting a large number of labeled the device. The BS then utilized the output from the modules to assist the
training samples in advance. Such training samples are usually difficult BM.
and time-consuming to obtain. In [115], the authors proposed a DRL-based secure beamforming
To avoid the overhead of collecting a large amount of labeled data, approach in an RIS-embedded environment where the system needs to
unsupervised learning and reinforcement learning are appealing options. secure the communications of multiple legitimate users with different
The authors in Ref. [110] leveraged an unsupervised learning technique QoS requirements in the presence of eavesdroppers. The objective for the
and designed an RIS BeamForming Neural Network (RISBFNN) archi- DRL was to jointly optimize the BS's transmit beamforming matrix and
tecture to predict the optimal phase shift configuration, using estimated the reflection beamforming matrix at the RIS to maximize the worst-case
channels at the BS as input and the negated transmission rate as the loss secrecy rate. To help improve learning efficiency, post-decision state and
function. Simulation results showed the proposed method could achieve prioritized experience replay were also utilized in the learning phase.
decent performance compared to the baseline semi-definite relaxation In [116], the authors developed models for RIS-enabled wireless
approach while greatly reducing the system running time. Another ML fingerprinting localization to support various use cases requiring the
technique, Deep Reinforcement Learning (DRL), which uses the data capability of localizing target devices accurately within the propagation
collected online to train the model, has gained momentum in various environment (e.g., Internet of Things (IoT), communication, and
wireless network scenarios, especially for optimization problems [111]. sensing). The proposed approach leveraged ML for feature selection to
For RIS-assisted wireless communication, the authors in Ref. [112] prune the large state space of the RIS.
designed an efficient DRL-based framework to optimize RIS phase shifts As RIS-assisted wireless communication is still in its initial research
using the received SNR as the reward function; simulation results showed stage, we believe use cases are also evolving, and ML technology is likely
99
to be explored for them. However, given the dynamic channel conditions (i.e., training-free) ([104,120]), thus reducing the overhead and la-
and data collection challenge in a real-world environment, the perfor- tency from pilot-based training. This is important as we move toward
mance benefit and associated overhead need to be carefully studied to more Ultra-Reliable Low-Latency Communication (URLLC)-driven
assess the practicality of applying ML to solutions for RIS-aided applications, in conjunction with other types of traffic.
applications. 2) Predictive control of RIS elements. As illustrated by several pub-
We summarize the above-discussed AI/ML-based RIS solutions and lished works [117,118], with the help of ML, the channel character-
experiments in Table 2. istics become predictable based on device locations and/or other
fine-granularity context information. It is also anticipated that the
predictions can be combined with other ML-learned models for
5.3. Strength of AI/ML and challenges of their applications to RIS-aided
user/device trajectory/mobility and/or traffic distribution pre-
systems
dictions. [117] illustrated an approach where user distribution pre-
diction was used to control the panels' ON/OFF settings. For
AI/ML provides powerful tools to gain additional knowledge and
high-frequency-band communication, the presence of a LoS path is
intelligence of the surrounding environment that conventional systems
an important assurance for high reliability. RIS can help create such
and solutions fall short of achieving. For the first time, as pointed out in
LoS channel d but only if we can predict the channel dynamics and
Ref. [104], a communication system can learn the relationships among
configure RIS elements proactively. An interesting study in this re-
high dimensional factors, which is hard to capture using a traditional
gard can be seen in Ref. [118], where a split-learning architecture was
analytical approach, and leverage the discovered knowledge of the un-
proposed to enable vision-aided mmWave received power prediction.
derlying environment and network behaviors for performance and other
3) Broad deployment and application scenarios. With an ML-based
optimizations.
approach, the signal processing burden or requirements for commu-
For AI/ML in RIS-aided applications, we summarize the key advan-
nication devices and nodes are considerably reduced. For example,
tages of a learning-based approach as follows:
the RIS panel can be completely passive, allowing standalone
deployment [46] without deep integration with existing communi-
1) Low overhead and real-time control of RIS elements. When suffi-
cation systems. [46] leverages a few active elements on the RIS to
cient prediction accuracy can be achieved by the learned models, it is
perform sensing and channel estimation. It is also possible for RIS to
possible not to rely on explicit channel estimation/reconstruction
Table 2
Summary of existing ML-based RIS research works: (1) ML used to assist channel estimation; (2) ML used for phase shift configuration at RIS and beamforming at BS; and
(3) ML used for other applications in RIS-assisted communication networks.
Cate- Scenario/objective Approach ML Model Refer- Key contribution
gory ence
(1) Channel estimation Supervised learning Twin CNN [47] Estimate both direct channel and cascaded channel at the same
(DL) time
(1) Channel estimation with lower Supervised learning CNN [105] Reduce transmission overhead by 16X compared to centralized
transmission overhead (FL) leaning approach
(1) Channel estimation accuracy Supervised learning CNN [106] Proposed approach uses ML to improve channel estimation
enhancement (DL) accuracy in real-world deployment scenario where ideal channel
information is challenging to obtain
(2) RIS-assisted network with few active Supervised learning FCNN [45] Approach requires only a few active elements on the RIS, thus
elements (DL) reduces beam training overhead
(2) Configuring RIS phase shift and DL Supervised learning FCNN [107] Predict RIS phases and DL beamforming vector using available
beamforming vector without explicit (DL) pilot signals while bypassing explicit channel estimation. Two DNN
channel estimation models are proposed, and the second model reduces pilot overhead
while still achieving decent performance
(2) Configuring RIS phase shift and DL Supervised learning FCNN [108] Proposed method learns to configure the RIS and beamforming at
beamforming vector without explicit (DL) the BS to maximize the system sum rate directly based on received
channel estimation pilots
(2) RIS-assisted indoor environment Supervised learning FCNN [27] Learn the mapping between a user's position and the optimal
(DL) configuration of each RIS element with a fingerprinting database
(2) Configuring RIS phase shift without Supervised learning CNN [109] Predict achievable rate for any receiver and intended RIS phase
using channel information or pilots (DL) shift combination without requiring channel information while
achieving close to optimal rate and using less than 2% data in
training
(2) RIS-assisted network with no labeled Unsupervised FCNN [110] Introduce unsupervised learning approach with no need of
data learning (DL) collecting labeled training data in advance
(2) RIS-assisted multi-user MISO system Model-free actor- FCNN [48] Jointly optimize the transmit beamforming vector and the phase
optimization critic DRL shifts together without the need to collect large amount of training
data in advance
(2) Standalone RIS operation Reinforcement FCNN [46] Approach does not require any control from communication
learning (DL) infrastructure. Validation showed achieving comparable
performance as supervised learning
(3) Energy efficient RIS-assisted Reinforcement FCNN [117] Proposed approach enables the BS to jointly optimize the transmit
communication network learning (DRL) power, RIS reflector's on/off state, and RIS phase shifts
(3) RIS-assisted communication network Supervised learning Gaussian Process [118] Proposed framework takes into account environment and mobility
with efficient BM þ online learning (mobility awareness to predict mobile's future motion state and efficiently
awareness) perform BM/switching
(3) RIS-assisted secure communication Reinforcement Deep Q Learning [115] Proposed approach utilizes PDS-PER to improve the learning
system learning (DRL) (DQN) efficiency and enhances the system's secrecy rate while also
satisfies different device's QoS requirement
(3) RF fingerprint localization in RIS- Supervised learning Genetic algorithm [119] Proposed framework leverages ML to select optimal set of features
assisted communication KNN that minimize the localization error. The approach reduces the
complexity and enhances position acquisition time
100
work with devices that have limited or no signal processing capa- deployment of RIS-embedded wireless networks.
bility. This will broaden applications of the RIS technology, for
example, in IoT applications or in the wireless power transfer use 1. An RIS-integrated B5G network must offer superior combined cost,
case. power, performance, and ease-of-deployment advantages over
4) Beyond simple scenarios. Most of the research so far was limited to incumbent network technologies to be commercially competitive,
simple problem settings (e.g., single-panel-single-user) and applying especially for realistic and complex multi-panel and/or multiuser
AI/ML to panel element control. Yet we can anticipate that ML scenarios. Factors such as surface design, manufacturing cost, overall
models, which capture the interaction patterns of panel/elements computation complexity, power consumption (see Section 4 for a
with device locations etc., can be used for more complicated and comprehensive summary), and network control protocol overhead
challenging settings, e.g., multi-panel-multi-user scenarios. Further, become variables in the overall system design tradeoff.
they can help other stages of RIS solution design, e.g., optimizing ● For instance, there is a fundamental tradeoff between the protocol
multi-panel placements during the planning phase. overhead associated with environment state information acquisi-
tion and network capacity, especially for a system that employs a
It is only natural that while promising and with many potential ad- large number of nearly passive distributed surfaces to ensure
vantages, the AI/ML approach for RIS solution design faces many chal- robust network coverage.
lenges. These challenges are also opportunities for future research ● Similarly, such massive deployments compound the computa-
directions. tional complexity required for the optimal configuration of each
First, the effectiveness of the learned models depends on the avail- surface, with the associated power consumption offsetting the
ability, granularity, and quality of the observation data for both sur- power saving promised by passive surfaces.
rounding environment and the channel characteristics, with performance ● Cost-effective nearly passive hardware architectures, such as
being observed at the device side. The richer, more granular, and higher severely quantized phase control, and the absence of sensing ele-
quality of the measurements, the better the quality of the trained models, ments on the surface may necessitate additional capabilities in
enabling a greater number of more powerful applications. digital signal processing and optimization to compensate for per-
Second, domain knowledge plays a critical role in model input/output formance degradation.
design, offering a clear understanding of the underlying causal rela-
tionship and predictability while considering the coherence time of the Therefore, in addition to analytical scaling laws governing the
wireless environment. In some cases, due to the limited observation of fundamental system capacity limits, those that govern the performance-
state information, there is a high degree of uncertainty regarding the cost-power tradeoff in a real-world system require careful investigation,
output. Reducing such uncertainty would require enrichment of the which in turn relies on realistic, physically consistent surface interaction
environment state representation and/or further online exploration/ and propagation models to allow simulation-based comparative evalua-
exploitation as indicated in Ref. [119]. Domain knowledge may also tion against legacy systems. This is especially the case for compact
accelerate the convergence of the ML algorithms within the coherence metasurfaces employing sub-wavelength spaced unit cells and for the
time of the wireless channel. newly emerging near-field application scenarios enabled by large
Third, the tradeoff between the site-specific model and generalizable surfaces.
knowledge learning needs to be better understood. Machine-learned
models are likely site-specific. However, it would be much desirable for 2. The network architecture that is optimized for an RIS-embedded
such models to capture a more general knowledge of the underlying network may look considerably different from the ones for its RIS-
physics law (e.g., from the relative locations, distance, and objects in be- free counterpart. For instance, nearly-passive RIS surfaces require
tween). An interesting experiment described in Ref. [104] shows that it is frequent endpoint transmission to probe the channels between the
possible to learn such knowledge from a small amount of labeled data. surfaces and endpoints, and between different surfaces, in dynamic
However, it remains a challenge to learn more transferable knowledge scenarios (such as user mobility). In the case that the number of de-
beyond simple problem settings, e.g., environments with multiple users or grees of freedom of the surface-enabled environment exceeds the
devices. capability of the endpoint pairs to probe (which is conceivable given
Lastly, we face a suite of challenges to make the new AI-based that extra and perhaps redundant number of degrees of freedom from
approach work in real-world experiments. In almost all the works we artificial environment is sought after in the RIS paradigm), protocol
surveyed, the studies and experimental results are based on numerical overhead may dominate the system performance. Either surfaces with
simulations, which are tractable first steps. However, it is generally sensing and transmission capability or additional probing endpoints
recognized that the data from the simulation are clean, and the simulated may be required. Both involve the careful design of low-power pro-
behaviors are more deterministic. The simulation also provides the tocol to transport sensed channel states from distributed measure-
convenience to generate the large quantities of data required for machine ment points to distributed or centralized processing entities.
learning. In a real-world environment, these benefits would not be 3. In contrast to current conventional cellular deployments, where
available. It would be critical to go beyond simulation-based experiments endpoint transmitter power spectral density masks suffice to ensure
to study the practical feasibility and effectiveness of the AI/ML-based RIS tolerable out-of-band emissions between neighboring licensed bands
solutions. It is encouraging to see a large-scale real-world design and belonging to different network operators in an overlapping environ-
implementation being introduced using a nonconventional approach ment, an RIS surface configuration optimized for one operator's band
without explicit channel estimation for its control (the majority voting allocation may incur performance degradation to another operator's
algorithm employed for RIS configuration exhibits an AI/ML flavor) [17]. network that shares the same surface as part of its propagation
Many other practical issues and challenges remain to be addressed for environment, unless the configuration of the surface response can be
AI/ML-based RIS solutions, e.g., the distributed model training, updates, made band-selective, or different networks deploy dedicated band-
and transfers. limited surfaces specific to their own band allocation. Such multi
operator system-level compatibility requirements require further
6. System-level challenges and future directions investigation.
4. The optimization along the new environmental dimension enabled by
The challenges presented in the preceding sections give rise to RIS, and the latter's potential for joint sensing, learning, and control of
exciting new research directions. Collectively, they also contribute to the environment of the dynamic ambient surroundings, requires
system-level research opportunities to ensure a commercially viable scalable solutions in an extremely high-dimensional space.
101
Traditional numerical optimization techniques may fall short. [13] Q. Wu, R. Zhang, Towards smart and reconfigurable environment: intelligent
reflecting surface aided wireless network, IEEE Commun. Mag. 58 (1) (2020)
Furthermore, low-cost and low-power surface architectures may
106–112.
render them and the associated propagation characteristics less [14] X. Yuan, Y.-J.A. Zhang, Y. Shi, W. Yan, H. Liu, Reconfigurable-intelligent-surface
amiable for parametric modeling. Model-free techniques, such as AI, Empowered Wireless Communications: Challenges and Opportunities, 2020 arXiv:
offer unique advantage here. Effective means of capturing the un- 2001.00364.
[15] M.A. ElMossallamy, H. Zhang, L. Song, K.G. Seddik, Z. Han, G.Y. Li,
derlying structural features of the environment remains a challenge Reconfigurable intelligent surfaces for wireless communications: principles,
for any AI technique. See Section 5.3 for AI-specific system-level challenges, and opportunities, IEEE Trans. Cognit. Commun. Netw. 6 (3) (2020)
challenges. 990–1002.
[16] S. Gong, X. Lu, D.T. Hoang, D. Niyato, L. Shu, D.I. Kim, Y. Liang, Towards smart
5. The question of “killer” applications for an RIS-embedded system wireless communications via intelligent reflecting surfaces: a contemporary
remains open. One pressing research direction would be to identify survey, IEEE Commun. Surv. Tutorials (2020) 1–31.
which industrial verticals show the most significant benefit in terms [17] V. Arun, H. Balakrishnan, Rfocus: beamforming using thousands of passive
antennas, in: Proc. 17th USENIX Symposium on Networked Systems Design and
of return on investment and total cost of ownership. It appears that Implementation (NSDI), 2020, pp. 1047–1061.
applications that require ubiquitous spatial coverage with a modest [18] Z. Li, Y. Xie, L. Shangguan, R.I. Zelaya, J. Gummeson, W. Hu, K. Jamieson,
degree of mobility (and/or those that derive added value from a Programmable radio environments with large arrays of inexpensive antennas,
GetMobile Mob. Comput. Commun. 23 (3) (2019) 23–27.
synergistic combination of communication and near-field sensing, [19] S.V. Hum, J. Perruisseau-Carrier, Reconfigurable reflectarrays and array lenses for
localization and power transfer functionalities) are likely to benefit dynamic antenna beam control: a review, IEEE Trans. Antenn. Propag. 62 (1)
the most, provided that the overall value to customers and network (2014) 183–198.
€
[20] O. Ozdogan, E. Bj€
ornson, E.G. Larsson, Intelligent reflecting surfaces: physics,
operators justifies the added cost associated with mass fabrication
propagation, and pathloss modeling, IEEE Wireless Commun. Lett. 9 (5) (2020)
and protocol overhead. 581–585.
[21] L. Zhang, X.Q. Chen, S. Liu, Q. Zhang, J. Zhao, J.Y. Dai, G.D. Bai, X. Wan,
7. Conclusions Q. Cheng, G. Castaldi, V. Galdi, T.J. Cui, Space-time-coding digital metasurfaces,
Nat. Commun. 9 (1) (2018) 1–11.
[22] L. Dai, B. Wang, M. Wang, X. Yang, J. Tan, S. Bi, S. Xu, F. Yang, Z. Chen,
This paper provides a literature survey and an outlook on the practical M.D. Renzo, C. Chae, L. Hanzo, Reconfigurable intelligent surface-based wireless
design aspects of RIS, as well as system-level challenges facing a communications: antenna design, prototyping, and experimental results, IEEE
Access 8 (2020) 45913–45923.
commercially viable deployment of an RIS-integrated wireless network. [23] D. Berry, R. Malech, W. Kennedy, The reflectarray antenna, IEEE Trans. Antenn.
In addition to summaries on salient recent works on system modeling and Propag. 11 (6) (1963) 645–651.
PHY design, the paper fully introduces both hardware architecture de- [24] King-Wai Lam, Sze-Wai Kwok, Yeongming Hwang, T.K. Lo, Implementation of
transmitarray antenna concept by using aperture-coupled microstrip patches, in:
signs and challenges, and unique opportunities such large-scale multi- Proc. Asia-Pacific Microwave Conference, vol. 1, 1997, pp. 433–436.
purpose intelligent networks facing AI/ML techniques. [25] B. Zhu, Y. Feng, J. Zhao, C. Huang, T. Jiang, Switchable metamaterial reflector/
absorber for different polarized electromagnetic waves, Appl. Phys. Lett. 97
(051906) https://doi:10.1063/1.3477960.
Declaration of competing interest [26] A. Taha, M. Alrabeiah, A. Alkhateeb, Enabling Large Intelligent Surfaces with
Compressive Sensing and Deep Learning, 2019 arXiv:1904.10136.
None. [27] C. Huang, G.C. Alexandropoulos, C. Yuen, M. Debbah, Indoor signal focusing with
deep learning designed reconfigurable intelligent surfaces, in: Proc. IEEE
Workshop on Sig. Proc. Advances in Wireless Commun. (SPAWC), 2019, pp. 1–5.
References Cannes, France.
[28] C. Huang, R. Mo, C. Yuen, Reconfigurable Intelligent Surface Assisted Multiuser
[1] T. Nakamura, 5G evolution and 6G[C]// ICDCN '21: International Conference on MISO Systems Exploiting Deep Reinforcement Learning, 2020 arXiv:2002.10072.
Distributed Computing and Networking 2021, 2021. [29] Y. Zhou, G. Zhang, H. Chen, P. Zhoua, X. Wang, L. Zhang, L. Zhang, J. Xie, L. Deng,
[2] W. Saad, M. Bennis, M. Chen, A vision of 6G wireless systems: applications, trends, Design of phase gradient coding metasurfaces for broadband wave modulating,
technologies, and open research problems, IEEE Network 34 (3) (2020) 134–142. Sci. Rep. 8 (8672) (2020) 1–9.
[3] U. of Oulu, Key Drivers and Research Challenges for 6G Ubiquitous Wireless [30] S. Abeywickrama, R. Zhang, Q. Wu, C. Yuen, Intelligent reflecting surface:
Intelligence, 2019. http://jultika.oulu.fi/files/isbn9789526223544.pdf (accessed practical phase shift model and beamforming optimization, IEEE Trans. Commun.
10 March 2020). 68 (9) (2020) 1–13.
[4] M.D. Renzo, A. Zappone, M. Debbah, M. Alouini, C. Yuen, J.D. Rosny, [31] E. Basar, Transmission through large intelligent surfaces: a new frontier in
S. Tretyakov, Smart radio environments empowered by reconfigurable intelligent wireless communications, in: Proc. European Conf. Networks and Commun.
surfaces: how it works, state of research, and road ahead, IEEE J. Sel. Area. (EuCNC), 2019, pp. 112–117. Valencia, Spain.
Commun. 38 (11) (2020) 2450–2525. [32] C. You, B. Zheng, R. Zhang, Channel estimation and passive beamforming for
[5] N. Docomo, DOCOMO and metawave announce successful demonstration of 28 intelligent reflecting surface: discrete phase shift and progressive refinement, IEEE
GHz band, [Online]. Available: https://www.businesswire.com/news/home/20 J. Sel. Area. Commun. 38 (11) (2020) 2604–2620.
181204005253/en/NTTDOCOMO-Metawave-Announce-Successful-Demonstra [33] Y. Yang, B. Zheng, S. Zhang, R. Zhang, Intelligent reflecting surface meets ofdm:
tion-28GHz-Band. protocol design and rate maximization, IEEE Trans. Commun. 68 (7) (2020)
[6] C.J. Vaca-Rubio, P. Ramirez-Espinosa, R.J. Williams, K. Kansanen, Z.-H. Tan, E. de 4522–4535.
Carvalho, P. Popovski, A Primer on Large Intelligent Surface (LIS) for Wireless [34] Q. Wu, R. Zhang, Intelligent reflecting surface enhanced wireless network: joint
Sensing in an Industrial Setting, 2020 arXiv:2006.06563. active and passive beamforming design, in: Proc. IEEE Global Telecom. Conf.
[7] S. Hu, F. Rusek, O. Edfors, Beyond massive MIMO: the potential of positioning (GLOBECOM), UAE, Abu Dhabi, 2018, pp. 1–6.
with large intelligent surfaces, IEEE Trans. Signal Process. 66 (7) (2018) €
[35] O. Ozdogan, E. Bj€
ornson, E.G. Larsson, Using intelligent reflecting surfaces for
1761–1774. rank improvement in MIMO communications, in: Proc. IEEE Int. Conf. Acoustics,
[8] D. Mishra, H. Johansson, Channel estimation and low-complexity beamforming Speech, & Sigal Processing, ICASSP, 2020, pp. 1–6.
design for passive intelligent surface assisted MISO wireless energy transfer, in: [36] Q. Wu, R. Zhang, Joint active and passive beamforming optimization for
Proc. IEEE Int. Conf. Acoustics, Speech, & Sigal Processing, ICASSP, Brighton, UK, intelligent reflecting surface assisted SWIPT under QoS constraints, IEEE J. Sel.
2019, pp. 4659–4663. Area. Commun. 38 (8) (2020) 1735–1748.
[9] X. Tan, Z. Sun, D. Koutsonikolas, J.M. Jornet, Enabling indoor mobile millimeter- [37] C. Huang, A. Zappone, G.C. Alexandropoulos, M. Debbah, C. Yuen, Reconfigurable
wave networks based on smart reflect-arrays, in: Proc. IEEE Conf. on Computer intelligent surfaces for energy efficiency in wireless communication, IEEE Trans.
Communications, Honolulu, HI, 2018, pp. 270–278. Wireless Commun. 18 (8) (2019) 4157–4170.
[10] O. Abari, Enabling high-quality untethered virtual reality, in: Proc. ACM [38] H. Li, R. Liu, M. Li, Q. Liu, X. Li, Irs-enhanced wideband mu-miso-ofdm
Workshop on Millimeter-Wave Networks and Sensing Systems, mmNets, communication systems, in: Proc. IEEE Wireless Commun. Network Conf.
Snowbird, UT, 2017, 49. (WCNC), 2020, pp. 1–6.
[11] M. A. Kishk, M.-S. Alouini, Exploiting Randomly-Located Blockages for Large-Scale [39] M. Shao, Q. Li, W. Ma, Minimum symbol-error probability symbol-level precoding
Deployment of Intelligent Surfaces, https://doi:10.1109/JSAC.2020.3018808. with intelligent reflecting surface, IEEE Wireless Commun. Lett. 9 (10) (2020)
(accessed 10 March 2020). 1601–1605.
[12] E. Basar, M. Di Renzo, J. de Rosny, M. Debbah, M.-S. Alouini, R. Zhang, Wireless [40] G. Zhou, C. Pan, H. Ren, K. Wang, M. Di Renzo, A. Nallanathan, Robust
communications through reconfigurable intelligent surfaces, IEEE Access 7 (2019) beamforming design for intelligent reflecting surface aided miso communication
116753–116773. systems, IEEE Wireless Commun. Lett. 9 (10) (2020) 1658–1662.
102
[41] L. Zhang, C. Pan, Y. Wang, H. Ren, K. Wang, A. Nallanathan, Robust Beamforming [75] Q. Wu, R. Zhang, Intelligent reflecting surface enhanced wireless network via joint
Design for Intelligent Reflecting Surface Aided Cognitive Radio Systems with active and passive beamforming, IEEE Trans. Wireless Commun. 18 (11) (2019)
Imperfect Cascaded CSI, 2020 arXiv:2004.04595. 5394–5409.
[42] R. Liu, H. Li, M. Li, Q. Liu, Symbol-level precoding design for irs-assisted mu-miso [76] Q. Wu, R. Zhang, Beamforming optimization for wireless network aided by
systems, in: Proc. IEEE Wireless Commun. Network Conf. (WCNC), 2020, pp. 1–6. intelligent reflecting surface with discrete phase shifts, IEEE Trans. Commun. 68
[43] Y. Han, S. Zhang, L. Duan, R. Zhang, Cooperative Double-Irs Aided (3) (2020) 1838–1851.
Communication: Beamforming Design and Power Scaling, https://doi:10.1109/ [77] M. Najafi, V. Jamali, R. Schober, H.V. Poor, Physics-based Modeling and Scalable
LWC.2020.2986290. Optimization of Large Intelligent Reflecting Surfaces, Sept. 2020, https://doi.org/
[44] Q. Wu, S. Zhang, B. Zheng, C. You, R. Zhang, Intelligent Reflecting Surface Aided 10.1109/TCOMM.2020.3047098.
Wireless Communications: A Tutorial, 2020 arXiv:2007.02759. [78] R.J. Williams, E. De Carvalho, T.L. Marzetta, A Communication Model for Large
[45] A. Taha, M. Alrabeiah, A. Alkhateeb, Deep learning for large intelligent surfaces in Intelligent Surfaces, May 2020, https://doi.org/10.1109/
millimeter wave and massive MIMO systems, in: Proc. IEEE Global Telecom. Conf. ICCWorkshops49005.2020.9145091.
(GLOBECOM), 2019, pp. 1–6. [79] D. Kumar, N. Prasad, Z. Ma, L. Piazzi, X.-F. Qi, in: IRS Aided Communication
[46] A. Taha, Y. Zhang, F.B. Mismar, A. Alkhateeb, Deep reinforcement learning for Model for Compact MIMO Systems, IEEE ICC, 2021.
intelligent reflecting surfaces: towards standalone operation, in: Proc. IEEE [80] P. Nayeri, F. Yang, A.Z. Elsherbeni, Reflectarray Antennas: Theory, Designs, and
Workshop on Sig. Proc. Advances in Wireless Commun, 2020, pp. 1–5. Applications, Wiley-IEEE Press, 2018.
[47] A.M. Elbir, A. Papazafeiropoulos, P. Kourtessis, S. Chatzinotas, Deep channel [81] A.H. Abdelrahman, F. Yang, A.Z. Elsherbeni, P. Nayeri, Analysis and design of
learning for large intelligent surfaces aided mm-wave massive MIMO systems, transmitarray antennas, Synth. Lect. Antenn. 6 (1) (2017) 1–175.
IEEE Wireless Commun. Lett. 9 (9) (2020) 1447–1451. [82] P. Nayeri, F. Yang, A.Z. Elsherben, Design and experiment of a single-feed quad-
[48] C. Huang, R. Mo, C. Yuen, Reconfigurable intelligent surface assisted multiuser beam reflectarray antenna, IEEE Trans. Antenn. Propag. 60 (2) (2012) 1166–1171.
MISO systems exploiting deep reinforcement learning, IEEE J. Sel. Area. Commun. [83] P. Nayeri, F. Yang, A.Z. Elsherbeni, Design of multifocal transmitarray antennas
38 (8) (2020) 1839–1850. for beamforming applications, in: Proc. IEEE Antenna & Propagation Society
[49] W. Tang, M.Z. Chen, X. Chen, J.Y. Dai, Y. Han, M.D. Renzo, Y. Zeng, S. Jin, International Symposium, APSURSI, Orlando, FL, 2013, pp. 1672–1673.
Q. Cheng, T.J. Cui, Wireless Communications with Reconfigurable Intelligent [84] Z. Li, Y. Xie, L. Shangguan, R.I. Zelaya, J. Gummeson, W. Hu, K. Jamieson,
Surface: Path Loss Modeling and Experimental Measurement, 2019 arXiv: Towards programming the radio environment with large arrays of inexpensive
1911.05326. antennas, in: Proc. of 16th USENIX Symp. on Networked SystemsDesign and
[50] 3GPP, Evolved Universal Terrestrial Radio Access (E-UTRA); Further Implementation (NSDI), 2019, pp. 285–299.
Advancements for (E-UTRA) Physical Layer Aspects, 2010. TR36.814 V9.0.0 [85] F. Yang, Y. Rahmat-Samii, Surface Electromagnetics: with Applications in
(accessed 10 March 2020). Antenna, Microwave, and Optical Engineering, Cambridge University Press, 2019.
[51] 3GPP, Study on Channel Model for Frequencies from 0.5 to 100 GHz, 2020. [86] B. Di, H. Zhang, L. Song, Y. Li, Z. Han, H.V. Poor, Hybrid beamforming for
TR38.901 V16.1.0 (accessed 10 March 2020). reconfigurable intelligent surface based multi-user communications: achievable
[52] G. Yue, B. Zhang, X.-F. Qi, L.J. Cimini, Generalized JSDM with enhanced rates with limited discrete phase shifts, IEEE J. Sel. Area. Commun. 38 (8) (2020)
interference management for massive MIMO systems, in: Proc. IEEE Wireless 1809–1822.
Commun. Network Conf. (WCNC) Workshops, 2019, pp. 1–7. Marrakech, [87] E.F. Kuester, M.A. Mohamed, M. Piket-May, C.L. Holloway, Averaged transition
Morocco. conditions for electromagnetic fields at a metafilm, IEEE Trans. Antenn. Propag.
[53] Y. Xu, G. Yue, S. Mao, User grouping for massive MIMO in FDD systems: new 51 (10) (2003) 2641–2651.
design methods and analysis, IEEE Access 2 (2014) 947–959. [88] C.L. Holloway, M.A. Mohamed, E.F. Kuester, A. Dienstfrey, Reflection and
[54] B. Zhang, G. Yue, L.J. Cimini Jr., Efficient beamforming training and limited transmission properties of a metafilm: with an application to a controllable surface
feedback precoding for massive MIMO systems, IEEE J. Sel. Area. Commun. 38 (9) composed of resonant particles, IEEE Trans. Electromagn C. 47 (4) (2005)
(2020) 2197–2214. 853–865.
[55] A. Goldsmith, S.A. Jafar, N. Jindal, S. Vishwanath, Capacity limits of MIMO [89] T.H. Hand, S.A. Cummer, Reconfigurable reflectarray using addressable
channels, IEEE J. Sel. Area. Commun. 21 (5) (2003) 684–702. metamaterials, IEEE Antenn. Wireless Propag. Lett. 9 (2010) 70–74.
[56] G. Yue, X.-F. Qi, Multi-domain conversions of high dimensional channel [90] S.V. Hum, M. Okoniewski, R.J. Davies, Modeling and design of electronically
characteristics for massive MIMO-OFDM, in: Proc. IEEE Global Telecom. Conf., tunable reflectarrays, IEEE Trans. Antenn. Propag. 55 (8) (2007) 2200–2210.
GLOBECOM, Taibei, Taiwan, 2020, pp. 1–6. [91] A. Petosa, An overview of tuning techniques for frequency-agile antennas, IEEE
[57] 3GPP, Study on 3D Channel Model for LTE, 2018. TR36.873 V12.7.0. Antenn. Propag. Mag. 54 (5) (2012) 271–296.
[58] Remcom, Wireless insite, [Online]. Available: http://www.remcom.com/wirele [92] H. Rajagopalan, Y. Rahmat-Samii, W.A. Imbriale, RF MEMS actuated
ss-insite. (accessed 10 March 2020). reconfigurable reflectarray patch-slot element, IEEE Trans. Antenn. Propag. 56
[59] E. Bjornson, L. Sanguinetti, Power scaling laws and near-field behaviors of massive (12) (2008) 3689–3699.
MIMO and intelligent reflecting surfaces, IEEE Open J. Commun. Soc. (2020) [93] S.A. Long, G.H. Huff, A fluidic loading mechanism for phase reconfigurable
1–18. reflectarray elements, IEEE Antenn. Wireless Propag. Lett. 10 (2011) 876–879.
[60] A. Pizzo, T.L. Marzetta, L. Sanguinetti, Spatially-stationary model for holographic [94] E. Carrasco, J. Perruisseau-Carrier, Reflectarray antenna at terahertz using
MIMO small-scale fading, IEEE J. Sel. Area. Commun. 38 (9) (2020) 1964–1979. graphene, IEEE Antenn. Wireless Propag. Lett. 12 (2013) 253–256.
[61] B.O. Zhu, J. Zhao, Y. Feng, Active impedance metasurface with full 360 reflection [95] M.R. Chaharmir, J. Shaker, M. Cuhaci, A.-R. Sebak, Novel photonically-controlled
phase tuning, Sci. Rep. 3 (2013) 1–6. reflectarray antenna, IEEE Trans. Antenn. Propag. 54 (4) (2006) 1134–1141.
[62] L. Kundu, Information-Theoretic Limits on MIMO Antennas, Ph.D. dissertation, [96] F. Venneri, S. Costanzo, G. Di Massa, A. Borgia, P. Corsonello, M. Salzano, Design
Dept. Elec. Eng., NCSU, Raleigh, NC, 2016. of a reconfigurable reflectarray based on a varactor tuned element, in: Proc. 6th
[63] G.C. Alexandropoulos, E. Vlachos, A hardware architecture for reconfigurable European Conference on Antennas and Propagation (EUCAP), 2012,
intelligent surfaces with minimal active elements for explicit channel estimation, pp. 2628–2631.
in: Proc. IEEE Int. Conf. Acoustics, Speech, & Sigal Processing (ICASSP), 2020. [97] I. Derafshi, N. Komjani, E. Ghasemi-Mizuji, M. Mohammadirad, Dual-band X/Ku
[64] Z. He, X. Yuan, Cascaded channel estimation for large intelligent metasurface reflectarray antenna using a novel FSS-backed unit-cell with quasi-spiral phase
assisted massive MIMO, IEEE Wireless Commun. Lett. 9 (2) (2020) 210–214. delay line, J. Microw. Optoelectron. Electromagn. Appl. 15 (3) (2016) 225–236.
[65] Z. Wang, L. Liu, S. Cui, Channel estimation for intelligent reflecting surface [98] P. Padilla, A. Mu~ noz-Acevedo, M. Sierra-Casta~ner, M. Sierra-Perez, Electronically
assisted multiuser communications: framework, algorithms, and analysis, IEEE reconfigurable transmitarray at ku band for microwave applications, IEEE Trans.
Trans. Wireless Commun. 19 (10) (2020) 6607–6620. Antenn. Propag. 58 (8) (2010) 2571–2579.
[66] J. Chen, Y.-C. Liang, H.V. Cheng, W. Yu, Channel Estimation for Reconfigurable [99] NTT DoCoMo Inc., DOCOMO Conducts World's First Successful Trial of
Intelligent Surface Aided Multi-User MIMO Systems, 2019 arXiv:1912.03619. Transparent Dynamic Metasurface, Jan. 2020 [Online]. Available: https://www.
[67] N. Prasad, M.M. Chowdhury, X.-F. Qi, in: Channel Reconstruction in Intelligent nttdocomo.co.jp/english/info/media_center/pr/2020/0117_00.html.
Surface Aided Communications, IEEE Comsnets, 2021 (accessed 10 March 2020). [100] S. Ali, et al., 6G White Paper on Machine Learning in Wireless Communication
[68] S. Haghighatshoar, G. Caire, Low-complexity massive MIMO subspace estimation Networks, 2020 arXiv:2004.13875.
and tracking from low-dimensional projections, IEEE Trans. Signal Process. 66 (7) [101] J. Park, S. Samarakoon, H. Shiri, M.K. Abdel-Aziz, T. Nishio, A. Elgabli, M. Bennis,
(2018) 1832–1844. Extreme Urllc: Vision, Challenges, and Key Enablers, 2020 arXiv:2001.09683.
[69] C. You, B. Zheng, R. Zhang, Fast Beam Training for Irs-Assisted Multiuser [102] F. Tariq, M.R.A. Khandaker, K. Wong, M.A. Imran, M. Bennis, M. Debbah,
Communications, Jun. 2020 arXiv:2005.11652. A speculative study on 6G, IEEE Wireless Commun. 27 (4) (2020) 118–125.
[70] N. Prasad, Q. Zhang, X.-F. Qi, Channel reconstruction via quadratic programming [103] G. Jornod, R. Alieiev, A. Kwoczek, T. Kürner, Environment-aware communications
in massive MIMO networks, in: IEEE WiOpt, 2019, pp. 1–9. Avignon, France. for cooperative collision avoidance applications, in: Proc. IEEE Int. Symp. on a
[71] N. Prasad, M.M. Chowdhury, X.-F. Qi. Channel Reconstruction with Limited World of Wireless, Mobile and Multimedia Networks (WoWMoM), 2018,
Feedback in Intelligent Surface aided Communications, IEEE VTC-Fall, 2021. pp. 588–599. Chania, Greece.
[72] A. Decurninge, M. Guillaud, D. Slock, Channel covariance estimation in massive [104] B. Sheen, J. Yang, X. Feng, M.M.U. Chowdhury, A Digital Twin for Reconfigurable
MIMO frequency division duplex systems, in: , 2015, https://doi.org/10.1109/ Intelligent Surface Assisted Wireless Communication, Sept. 2020 arXiv:
GLOCOMW.2015.7413982. 2009.00454.
[73] M. Ganji, H. Cheng, Q. Zhan, K. Song, A New Codebook Design for Analog [105] A.M. Elbir, S. Coleri, Federated Learning for Channel Estimation in Conventional
Beamforming in Millimeter-Wave Communication, Feb. 2020 arXiv:1902.00838. and IRS-Assisted Massive MIMO, Aug. 2020 arXiv:2008.10846.
[74] O. Abari, D. Bharadia, A. Duffield, D. Katabi, Enabling high-quality untethered [106] S. Liu, Z. Gao, J. Zhang, M.D. Renzo, M.-S. Alouini, Deep denoising neural network
virtual reality, in: Proc. 14th USENIX Symp. on Networked SystemsDesign and assisted compressive channel estimation for mmWave intelligent reflecting
Implementation (NSDI), 2017, pp. 531–544. surfaces, IEEE Trans. Veh. Technol. 69 (8) (2020) 9223–9228.
103
€
[107] Ozgecan €
Ozdogan, E. Bj€
ornson, Deep learning-based phase reconfiguration for [114] C. Jia, H. Gao, N. Chen, Y. He, Machine learning empowered beam management
intelligent reflecting surfaces, in: Proc. Asilomar Conf. Signals, Systems and for intelligent reflecting surface assisted MmWave networks, China Commun. 17
Computers, 2020. https://doi.org/10.1109/IEEECONF51394.2020.9443516. (10) (2020) 100–114.
[108] T. Jiang, H.V. Cheng, W. Yu, Learning to beam form for intelligent reflecting [115] H. Yang, Z. Xiong, J. Zhao, D. Niyato, L. Xiao, Q. Wu, Deep reinforcement learning-
surface with implicit channel estimate, in: Proc. IEEE Global Telecom. Conf. based intelligent reflecting surface for secure wireless communications, IEEE
(GLOBECOM), 2020, pp. 1–6. Trans. Wireless Commun. 20 (1) (2021) 375–388.
[109] B. Sheen, J. Yang, X. Feng, M.M.U. Chowdhury, A deep learning based modeling of [116] C.L. Nguyen, O. Georgiou, G. Gradoni, Reconfigurable Intelligent Surfaces and
reconfigurable intelligent surface assisted wireless communications for phase shift Machine Learning for Wireless Fingerprinting Localization, 2020 arXiv:
configuration, IEEE Open J. Commun. Soc. 2 (2021) 262–272. 2010.03251.
[110] J. Gao, C. Zhong, X. Chen, H. Lin, Z. Zhang, Unsupervised learning for passive [117] L. Subrt, P. Pechac, Intelligent walls as autonomous parts of smart indoor
beamforming, IEEE Commun. Lett. 24 (5) (2020) 1052–1056. environments, IET Commun. 6 (8) (2012) 1004–1010.
[111] M. Chen, U. Challita, W. Saad, C. Yin, M. Debbah, Artificial neural networks-based [118] Y. Koda, J. Park, M. Bennis, K. Yamamoto, T. Nishio, M. Morikura, Distributed
machine learning for wireless networks: a tutorial, IEEE Commun. Surv. Tutorials Heteromodal Split Learning for Vision Aided mmWave Received Power Prediction,
21 (4) (2019) 3039–3071. 2020 arXiv:2007.08208.
[112] K. Feng, Q. Wang, X. Li, C. Wen, Deep reinforcement learning based intelligent [119] M.D. Renzo, M. Debbah, D.-T. Phan-Huy, A. Zappone, M.-S. Alouini, C. Yuen,
reflecting surface optimization for MISO communication systems, IEEE Wireless V. Sciancalepore, G.C. Alexandropoulos, J. Hoydis, H. Gacanin, J. de Rosny,
Commun. Lett. 9 (5) (2020) 745–749. A. Bounceu, G. Lerosey, M. Fink, Smart radio environments empowered by
[113] G. Lee, M. Jung, A.T.Z. Kasgari, W. Saad, M. Bennis, Deep reinforcement learning reconfigurable AI meta-surfaces: an idea whose time has come, EURASIP J. Wirel.
for energy-efficient networking with reconfigurable intelligent surfaces, in: Proc. Commun. Netw. 129 (2019) 1–20.
of Int. Conf. Commun., ICC, 2020, pp. 1–6. [120] Y. Zeng, X. Xu, Towards Environment-Aware 6G Communications via Channel
Knowledge Map, Jul. 2020 arXiv:2007.09332.
104

1 Ris

Uploaded by

Copyright:

Available Formats

1 Ris

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 Ris

Uploaded by

Copyright:

Available Formats

Digital Communications and Networks 8 (2022) 87–104

Contents lists available at ScienceDirect

Digital Communications and Networks

Reconﬁgurable intelligent surface: design the channel – a new opportunity

1. Introduction dielectric material comprised of an array of passive sub-wavelength

consumption and backhaul cost (e.g., Refs. [9–11]).

2.1. Reconﬁgurable surface and basic functionalities 2.2. RIS systems

Currently, most research studies on RISs focus on only phase-shift vari-

4.2. Brief overview of transmitarray and reﬂectarray antennas

The concept of using a planar or conformal surface with properties

element, and k0 is the free-space wave number.

While a number of technologies are available for providing phase-

Lumped PIN Diodes þ – D þ / – O þ [90]

In general, the synthesis of metasurfaces from desired EM properties

You might also like