AUTO MS: Automated Service for mmWave Coverage

Optimization using Low-cost Metasurfaces

Ruichun Ma* Shicheng Zheng* Hao Pan†
Yale University University of Science and Microsoft Research Asia
ruichun.ma@yale.edu Technology of China panhao@microsoft.com
sc7543@mail.ustc.edu.cn

Lili Qiu Xingyu Chen‡ Liangyu Liu‡

Microsoft Research Asia University of California San Diego University of Science and
UT Austin xic063@ucsd.edu Technology of China
liliqiu@microsoft.com lly@mail.ustc.edu.cn

Yihong Liu‡ Wenjun Hu Ju Ren

University of Glasgow Yale University Tsinghua University
y.liu.6@research.gla.ac.uk wenjun.hu@yale.edu renju@tsinghua.edu.cn
ABSTRACT show that AUTO MS can increase the median received signal
mmWave networks offer wide bandwidth for high-speed strength by 11 dB in target rooms and over 20 dB at previous
wireless communication but suffer from limited range and blind spots, and improve the median throughput by over 3×
susceptibility to blockage. Existing coverage provisioning in real-world scenarios.
solutions not only incur high costs but also require signifi-
cant expert knowledge and manual efforts. In this paper, we CCS CONCEPTS
present AUTO MS, an automated service framework to opti- • Hardware → Wireless devices; • Networks → Wireless
mize mmWave coverage by strategically designing and plac- access points, base stations and infrastructure.
ing low-cost passive metasurfaces. Our approach consists of
three key components: (1) joint optimization of metasurface KEYWORDS
phase configurations and placement as well as access point Millimeter-wave, Coverage optimization, Cloud service, Meta-
beamforming codebooks. (2) a fast 3D ray-tracing simulator surfaces, Ray-tracing simulator
for accelerated large-scale metasurface channel modeling. (3)
ACM Reference Format:
a metasurface design amenable to ultra-low-cost hot stamping
Ruichun Ma, Shicheng Zheng, Hao Pan, Lili Qiu, Xingyu Chen,
fabrication, featuring high reflectivity, near 2𝜋 phase control, Liangyu Liu, Yihong Liu, Wenjun Hu, and Ju Ren. 2024. AUTO MS:
and wideband support. Simulation and testbed experiments Automated Service for mmWave Coverage Optimization using Low-
cost Metasurfaces. In International Conference On Mobile Comput-
* Ruichun Ma and Shicheng Zheng contribute equally to this research
† Hao Pan is the corresponding author ing And Networking (ACM MobiCom ’24), September 30–October 4,
‡ Ruichun 2024, Washington, DC, USA. ACM, New York, NY, USA, 15 pages.
Ma, Shicheng Zheng, Xingyu Chen, Liangyu Liu, and Yihong Liu
did this work as interns at Microsoft Research Asia https://doi.org/10.1145/3636534.3649347

Permission to make digital or hard copies of all or part of this work for 1 INTRODUCTION
personal or classroom use is granted without fee provided that copies are not
The millimeter wave (mmWave) frequency band is critical
made or distributed for profit or commercial advantage and that copies bear
this notice and the full citation on the first page. Copyrights for components
for high-speed wireless communication in 5G and beyond
of this work owned by others than the author(s) must be honored. Abstracting due to its large bandwidth [1, 19]. However, mmWave signals
with credit is permitted. To copy otherwise, or republish, to post on servers or suffer from limited range and blockage due to their high
to redistribute to lists, requires prior specific permission and/or a fee. Request frequency [52], which hinders wide adoption. One potential
permissions from permissions@acm.org. solution is to utilize massive phased antenna arrays (e.g.,
ACM MobiCom ’24, November 18–22, 2024, Washington, DC, USA
$3000+ [46]) and create narrow beams to increase the received
© 2024 Copyright held by the owner/author(s). Publication rights licensed to
ACM.
signal power and extend the line-of-sight (LOS) range [6]. To
ACM ISBN 979-8-4007-0489-5/24/09. . . $15.00 cover multiple non-line-of-sight (NLOS) areas in a complex
https://doi.org/10.1145/3636534.3649347 deployment environment, a common solution is to deploy

62
ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA Ruichun Ma, Shicheng Zheng, Hao Pan, et al.

multiple mmWave access points (APs) or base station [29]. Method Type Coverage Cost Hardware count
AP [4, 54] Active Manual ≈ $300 per AP
These solutions are not only costly (e.g., around $300 for
Relay [2] Active Manual ≥ $200 per relay
each AP [4]), but also lack scalability due to the considerable mmWall [11] Varactors Manual ≈ $10000 28 × 76
domain knowledge and manual labor required for deployment. milliMirror [41] Passive Manual ≈ $15 80 × 80
This is because the APs must be strategically positioned and AUTO MS Passive Auto ≈ $1 160 × 200
configured to prevent interference and guarantee coverage. Table 1: Comparison between mmWave coverage solu-
Smart surfaces or metasurfaces, composed of specially de- tions. We show cost details in section 4
signed sub-wavelength elements, have powerful capability
to manipulate electromagnetic (EM) wave propagation and employ passive metasurfaces that are fixed post-fabrication.
enhance mmWave coverage beyond what simple mirror re- Moreover, we leverage the AP’s dynamic beam steering ca-
flectors [8, 62] offer. Unbound by the law of reflection, they pability together with passive metasurfaces to serve different
enable flexible wavefront control, such as reflecting at vari- areas and handle user mobility.
ous angles in 3D space, and maintain high performance with Our system consists of following three major components:
minimal surface area, as demonstrated in Figure 13. Previous Automated optimization: To automate metasurface-assisted
works [5, 9, 13, 15, 26, 31, 59] design programmable surfaces mmWave coverage planning, we develop a novel joint op-
to enhance coverage in blind spots and notably improve chan- timization framework. It determines optimal network de-
nel conditions. However, these solutions often rely on tunable ployment configurations, including the phase settings and
circuit components and computing devices to enable rapid placement of one or more metasurfaces, AP(s)’ codebook
beam switching, resulting in costly hardware, considerable (phased array configurations), and, if necessary, the position-
computational overhead, and increased power consumption. ing of the APs. We iteratively identify and optimize hyper-
MilliMirror [41] uses passive metasurfaces to lower cost as a configurations, such as the quantity, sizes, and placement
proof of concept, but handles only limited scenarios – beam of metasurfaces. Each iteration refines the phase settings of
reflection towards a specific direction by manually identi- metasurfaces and APs through gradient descent. While our
fying the placement of metasurfaces. It remains open how framework primarily aims to maximize the aggregate channel
to unleash the full power of metasurfaces, balancing trade- capacity, it is also flexible enough to support alternative objec-
offs between effectiveness, complexity, cost, and deployment tives and may extend to active or programmable metasurfaces.
burden on users. Fast simulator: We develop a fast 3D ray tracing simulator
In order to realize the potential of metasurfaces, there capable of computing channel response matrices for meta-
are several significant challenges. First, existing solutions surfaces with tens of thousands of elements, utilizing both
require substantial expertise in wireless signal propagation software and hardware acceleration techniques. Our simula-
for guiding their design and deployment. Manual configura- tor processes approximately 1.6 billion wireless propagation
tion is error-prone and often results in less-than-ideal perfor- paths in 3 minutes using an NVIDIA A100 GPU, supporting
mance [32]. Second, automating optimization hinges on a large metasurfaces by element count. To our knowledge, no
reliable wireless channel model that accurately captures EM current ray-tracing simulators can handle large metasurfaces
wave propagation in complex scenarios. Existing simulators, with arbitrary phase configurations due to the prohibitive
like Wireless Insite [44], fall short by being either slow or computational costs.
offering limited metasurface support, such as only enabling Low-cost metasurface: We design a passive metasurface to
linear phase configurations [43]. Developing a fast and accu- achieve several important properties: high reflectivity (≥ 90%)
rate simulator that supports general metasurface operations to minimize energy loss, near 2𝜋 phase range for wavefront
is a challenging problem. Third, metasurface-based methods control, wideband support (over 8 GHz) on mmWave bands,
must be considerably more cost-effective than alternatives to and compatibility with low-precision but cost-effective hot
be appealing. Current strategies do not demonstrate a substan- stamping. By employing surface substrates, i.e., paper and
tial economic advantage over multi-AP solutions (Table 1). PVC plastic, we induce specific phase delays, forming an
This necessitates the creation of an affordable and efficient impedance transformer and thus “amplifying” phase shifts of
passive metasurface design. different metallic patterns. With this approach, we achieve
We propose AUTO MS, an automated service framework the desired properties using single-layer split-ring patterns
that optimizes mmWave coverage for given areas, such as that can be hot stamped onto paper, allowing for virtually
apartments or office floors, using strategically designed and cost-free metasurface production.
placed low-cost passive reflective metasurfaces, as depicted We implement AUTO MS as a cloud-based service featuring
in Figure 1. These metasurfaces alter mmWave propagation a user-friendly, multi-step workflow illustrated in Figure 2.
to overcome blockages and concentrate power toward de- Users can employ commercial applications, such as Poly-
sired regions. To minimize costs and simplify deployment, we cam [39], to capture their deployment environment using a

63
AUTO MS: Automating Metasurface-aided mmWave Coverage ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA

mobile device, resulting in a 3D layout model. Concurrently, 2 RELATED WORK

we design a straightforward yet effective technique to infer re- 2.1 mmWave Coverage Extension
flection ratios of walls and objects in the environment. This is
achieved by collecting Received Signal Strength (RSS) mea- Conventional Methods Passive mirror reflectors are used to
surements during the scanning process and correlating them extend mmWave coverage in data centers [62], indoor [23],
with RSS predictions from ray tracing, enabling automatic and outdoor settings [37], by reflecting signals around obsta-
estimation of both the 3D model and the wall reflectivity. cles into NLOS areas. However, they are limited by the law
Subsequently, AUTO MS leverages this data to compute opti- of reflection, where the reflection angle is tied to the incident
mized deployment configurations, metasurface designs, and angle and the orientation of the reflector. As illustrated in
an updated codebook for the AP to work together with the Figure 13, even optimized placement of traditional reflectors
metasurfaces. often fails to achieve complete mmWave coverage in complex
When a cable/DSL company installs Internet service, they environments. Some studies [22, 56] attempt to improve link
can offer a value-added service by capturing the 3D layout and reliability by utilizing pre-existing large reflectors, but this is
uploading it to the AUTO MS platform to generate metasur- generally only practical in certain scenarios. Others [2, 53]
faces and provide placement instructions. Customers facing introduce active mmWave relays that re-transmit signals to
mmWave Wi-Fi coverage issues can also submit their 3D widen coverage and bolster link resilience. Despite their effec-
layout to the AUTO MS online service for assistance. We in- tiveness, such active solutions are costly and power-intensive,
troduce a novel concept in network deployment – surfaces as akin to adding extra APs.
a service. This approach transfers the complexity of hardware
Reconfigurable/Active surfaces. Several existing works fo-
design to the cloud service, aligning with trends seen in 5G
cus on developing reconfigurable surfaces with active ele-
network development.
ments that can dynamically modulate wireless signals. A
We evaluate AUTO MS using a public 3D layout dataset [18]
recent series of research presents end-to-end systems aimed
and through real-world testing. Simulations show that our sys-
at manipulating signal paths to enhance receiver channel con-
tem markedly enhances mmWave coverage across various
ditions [5, 9, 13, 15, 26, 31, 57–59]. However, these systems,
scenarios, achieving a median RSS gain of 12.1 dB com-
tailored for sub-6 GHz Wi-Fi frequencies, are not directly
pared to a single-AP setup. Field tests further confirm that
transferable to mmWave communications due to differences
our method boosts median RSS by 11 dB in target areas
in channel modeling, endpoint architecture, and application
and over 20 dB in blind spots, elevating median throughput
scenarios. Compared to lower frequencies, mmWave has a
from 77 Mbps to 373 Mbps. Our metasurfaces surpass mir-
limited range and is more susceptible to blockage. Metasur-
ror reflectors four times their size by a median of 7 dB in
faces have emerged as a promising solution [10, 11, 20, 25].
RSS improvement. As discussed in subsection 5.4, AUTO MS
However, programmable metasurfaces face several challenges.
proves resilient to environmental changes. By restricting the
They require costly tunable circuits and printed circuit board
vertical search space for placements, we prioritize above-head
(PCB) substrates. While individual elements are inexpensive,
locations for APs and metasurfaces, reducing the likelihood
a metasurface with thousands of them can become more ex-
of human movement disrupting the links. Combined with
pensive than mmWave APs, diminishing the appeal of using
dynamic AP beam steering, AUTO MS ensures dependable
such surfaces. Additionally, they rely on external controllers
coverage regardless of environmental dynamics.
for wiring and power, adding complexity to the control system
To summarize, we make the following contributions:
and installation. The interaction and feedback loop between
an active surface and the endpoints remain unresolved issues.
• We present AUTO MS, the first automated service frame-
work for metasurface-aided mmWave network deployment. Passive metasurfaces. MilliMirror [41] introduces a cost-
• We build an efficient 3D ray-tracing wireless channel sim- effective, 3D-printed reflective metasurface with beam syn-
ulator with a series of acceleration techniques. We open- thesis capabilities, although its production process is not yet
source the code to the research community at https://github. mainstream. While showcasing several deployment case stud-
com/microsoft/AutoMS. ies, MilliMirror lacks a systematic method for automating
• We design novel hot-stamped passive metasurfaces for cov- metasurface design and placement and requires significant
erage enhancement. It achieves high reflectivity, near 2𝜋 expertise. It, along with other reconfigurable intelligent sur-
phase control, and wideband operation, with orders of mag- face (RIS) research, aims to enhance single links, whereas
nitude lower cost than prior solutions (Table 1). our work seeks to optimize network-wide performance. PM-
• We evaluate AUTO MS extensively in 5 simulated envi- Sat [35] offers a unique passive metasurface design targeting
ronments and 3 real-world environments, achieving robust the optimization of individual terrestrial and satellite links.
mmWave coverage in diverse deployment scenarios. Additionally, recent studies [47, 48] have utilized hot-stamped

64
ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA Ruichun Ma, Shicheng Zheng, Hao Pan, et al.

Cloud Service

2 Deployment Hyper-config
1
Network
Configurations tuner Coverage Deployment Configurations

3 3D model Wireless
AP Environment Surface & AP Metasurface Network
M ache

channel
att

Scan optimizer Phase fabrication deployment

eta d

h Reflection
itc modeling Maps
su on

sw ized Coeff.
rfa w

ly
ical stom
ce alls

m cu
na
s

Dy ween ,2,3É
AP

t
be am 1
be

Figure 1: Metasurfaces create paths Figure 2: AUTO MS overview. Based on an environmental scan, our service
to circumvent blockages (beam 1) framework outputs the optimized deployment configurations for passive meta-
and extend coverage (beams 2&3) in surfaces and AP(s) tailored to the target 3D model. After surface fabrication,
the 3D environment we deploy the metasurfaces and AP accordingly

passive metasurfaces for security applications, but these de- metasurface support and necessitates considerable training
signs suffer from a power loss of over 7 dB. In contrast, for each scenario. Considering that a metasurface can have
AUTO MS pioneers the use of hot stamping in metasurfaces tens of thousands of elements, creating a multitude of signal
to boost SNR with minimal power loss. paths, scalability becomes a significant challenge. To address
this, we have developed a 3D mmWave ray tracing simulator
Our approach. We develop a comprehensive approach that
capable of rapidly and accurately modeling wireless channels
inputs a 3D scene model and determines the best placement
for general metasurfaces.
and settings for metasurfaces and AP codebooks. We employ
high-performance passive metasurface designs to ease pro-
duction, deployment, and to lower costs. Despite its passive 3 AUTOMS DESIGN
nature, our metasurface can accommodate endpoint mobil- 3.1 Overview
ity and environmental shifts by collaborating with the AP’s
Considering the symmetry between upstream and downstream
phased array, which selects suitable beamforming codewords.
links in mmWave communications, our study focuses on a
downstream scenario specifically, transmission from the AP
2.2 Ray Tracing-based Channel Modeling to the receiver in an indoor setting. We presume the mmWave
Accurate simulation of wireless signal propagation and chan- AP has a phased array capable of beam steering and that re-
nel modeling is crucial for designing and deploying wireless flective metasurfaces are installed on the walls. In line with
networks. Ray tracing (RT) methods, which approximate the the IEEE 802.11ad standard, we utilize a beam scanning pro-
Maxwell equations at high frequencies (e.g., as optical rays), cess where the AP sequentially transmits using all codewords.
represent the field as a set of rays that reflect, diffract, and The receiver then selects and reports back the codeword that
scatter through the environment. While several RT simulators yields the highest SNR. Therefore, the achievable link for the
like WinProp [3], Aster [16], CrossWave [17], and Wireless receiver is the best link among all codewords.
Insite [44] are available for commercial or academic use, only Figure 2 illustrates the workflow of our system. Using
the latest Wireless Insite [43] version includes metasurface RoomPlan [12] on a mobile device (section 4), we first create
support. However, based on our discussions with their techni- a 3D model of the environment for network deployment, serv-
cal team, it only allows for reflection towards a fixed angular ing as the input to AUTO MS. Subsequently, our framework
range and lacks the capability to model individual metasurface iteratively executes the following steps: (1) Our proposed ray-
elements, precluding optimization of the elements themselves. tracing simulator generates wireless channel response matri-
Prior PIS research [27, 28] relied on a basic ray-tracing ces between the AP (TX), metasurfaces, and potential receiver
simulator that only offered control at the surface level and (RX) locations based on the 3D model, influenced by the de-
assumed environments without obstacles. Recently, some re- ployment’s hyper-configurations such as metasurface and AP
searchers proposed open-source RT simulators, like Opal [14], placements. (2) A phase map optimizer employs gradient de-
SimRIS [7], and NeRF2 [61]. Opal, fast due to its OptiX API scent to fine-tune the metasurfaces’ phase settings and APs’
implementation, unfortunately lacks metasurface modeling codebooks to maximize a network-wide metric, like sum ca-
capabilities [14]. SimRIS, a MATLAB-based simulator, can pacity, and provides feedback on the optimized performance
model channels for RIS-assisted MIMO systems but is not to the hyper-configuration tuner. (3) The hyper-configuration
equipped for handling large metasurfaces [7]. NeRF2 uses tuner, typically using simulated annealing, updates the de-
a neural network for channel modeling but does not include ployment hyper-configurations, prompting a return to Step 1

65
AUTO MS: Automating Metasurface-aided mmWave Coverage ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA

for updates. (4) The process concludes when the optimized gradient descent with the Adam optimizer to enhance sum
performance stabilizes, yielding the final phase maps and de- capacity. Conceptually, this optimization can be viewed as
ployment configurations for the actual metasurface fabrication the strategic distribution of power from the metasurfaces to
and installation. all RX locations.
We define the 𝑗th codeword among all 𝑁𝑐 code words as
3.2 Optimization Framework ( 𝑗)
W𝑎𝑝 ∈ C1×𝑁𝑡 and the phase configuration of 𝑘th surface
(𝑘)
Our optimization framework comprises a hyper-configuration among all 𝑀 surfaces as W𝑚𝑠 ∈ C𝑁𝑚𝑠 ×1 , where 𝑁𝑚𝑠 and 𝑁𝑡
tuner and a phase map optimizer for mmWave network de- are the numbers of metasurface elements and transmitting an-
ployment. Hyper-configurations are high-level deployment tennas, respectively. Each value in the phase configuration and
parameters, such as the quantity, dimensions (element count), codebook matrix is a phase shift without amplitude change,
and locations of metasurfaces, as well as the orientation(s) denoted as 𝑒 𝑗𝜃 , where 𝜃 is the phase we aim to determine. The
and location(s) of AP(s). The phase map optimizer operates sum of channel capacity between the TX and RX locations is
as the inner loop, determining the phase configurations for used as the optimization objective and derived as follows:
the given hyper-configurations. 𝑁𝑟
Õ 𝑆𝑖 (W𝑚𝑠 , W𝑎𝑝 )
Hyper-configuration tuner. AUTO MS allows for the opti- max log2 (1 + ) (1)
mization of certain deployment configurations based on the W𝑚𝑠 ,W𝑎𝑝
𝑖=1
𝑁𝑜𝑖𝑠𝑒
specific scenario and user preferences, while users can man- where 𝑁𝑟 denotes the number of RX locations and 𝑆𝑖 denotes
ually set the remaining according to their requirements. For the received signal at the 𝑖th receiving antenna as a function
instance, if the AP’s position is predetermined, AUTO MS of metasurface phase maps and AP codebook. We set a noise
focuses on optimizing the other parameters. Various algo- floor of −70 dBm to match our experiment hardware (i.e., 0
rithms can optimize these configurations; our implementation capacity if RSS is below −70 dBm). Next, we describe the
employs simulated annealing, which has proven effective for calculation for received signal strength 𝑆𝑖 . Let 𝑃𝑡 denote the
our purposes. The hyper-configuration tuner, informed by transmission power, 𝐺𝑡 denote the transmission antenna gain,
the phase map optimizer’s performance metrics, iteratively 𝐺𝑟 denote the receiver antenna gain. Since the AP can switch
explores and updates configurations until performance stabi- between multiple codewords, we take the maximal achievable
lizes (i.e., no improvements are observed in the last iterations). signal strength from all codewords as 𝑆𝑖 :
The resulting final configurations guide the fabrication and
( 𝑗)
p
deployment of metasurfaces. 𝑆𝑖 (W𝑚𝑠 , W𝑎𝑝 ) = 𝑃𝑡 𝐺𝑡 𝐺𝑟 max W𝑎𝑝 (H𝑚𝑠,𝑖 +H𝑇 −𝑅𝑖 ) (2)
𝑗 ∈ [1, 𝑁𝑐 ]
Phase map optimizer. The phase map optimizer identifies the
optimal phase configurations for metasurfaces and mmWave where H𝑚𝑠,𝑖 ∈ C𝑁𝑡 ×1 is the channel matrix characterizing the
AP phased arrays to maximize the coverage objective func- propagation paths affected by the metasurfaces between the
tion in a target environment. This optimization is based on TX and the 𝑖th RX, H𝑇 −𝑅𝑖 ∈ C𝑁𝑡 ×1 is the channel matrix be-
the channel matrices provided by the simulator and the cur- tween the TX and the 𝑖th RX, characterizing the propagation
rent hyper-configurations. It determines the phase shift values paths not traversing any metasurfaces. Due to a small 𝑁𝑐 , 𝑆𝑖 ’s
for each metasurface element, denoted as 𝑊𝑚𝑠 (𝑘) with 𝑘 optimal value can be obtained easily without an optimization
representing the metasurface index, and selects a set of beam- solver. The metasurface-affected channel matrix, H𝑚𝑠,𝑖 is a
(𝑘) 𝑀
forming codewords for the APs, indicated by 𝑊𝑎𝑝 ( 𝑗) where function of {W𝑚𝑠 }𝑘=1 . If we only consider the first-order
𝑗 is the codeword index. Upon achieving convergence, these reflections from metasurfaces, then we have
phase configurations are discretized for practical deployment. 𝑀
Õ
(𝑘)
Objective function. There are many possible metrics that can H𝑚𝑠,𝑖 = (H𝑇 −𝑚𝑠 W𝑚𝑠 H𝑚𝑠−𝑅𝑖 ) (3)
𝑘=1
quantify the coverage of an area. Our design is intentionally
decoupled from physical layer specifics, such as modulation where H𝑇 −𝑚𝑠 ∈ C𝑁𝑡 ×𝑁𝑚𝑠 is the channel matrix between the
schemes or hardware implementations, enabling our system’s transmitting antennas and the metasurface elements, H𝑚𝑠−𝑅𝑖 ∈
compatibility with diverse mmWave devices and protocols. C𝑁𝑚𝑠 ×1 is the channel matrix between the metasurface ele-
Consequently, we select the sum of channel capacities be- ments and the receiver. We further extend the equation above
tween the transmitter (TX) and all potential receiver (RX) to support reflections between multiple metasurfaces, but we
locations as our optimization goal, though our framework is omit the details here.
versatile enough to accommodate alternative objectives. As Thus, we need to know the following channel matrices to
outlined in subsection 3.3, channel matrices involving TX, compute the objective function: H𝑇 −𝑅 , H𝑇 −𝑚𝑠 , H𝑚𝑠𝑖 −𝑚𝑠 𝑗 , and
RX, and metasurfaces are computed using ray tracing. Our H𝑚𝑠−𝑅 . These channel matrices are calculated using the ray
optimization problem is differentiable, allowing us to employ tracing simulator, as described in subsection 3.3.

66
ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA Ruichun Ma, Shicheng Zheng, Hao Pan, et al.

3.3 Efficient Channel Simulator where 𝜏 denotes the reflection coefficient of the reflector or
To address the optimization problem, we simulate the channel surface element. We use the equations above iteratively for
response matrices between the mmWave APs, metasurfaces, multi-order reflections. Next, we sum the LOS and NLOS
and receivers in the environment. Each metasurface element is channel responses to calculate the overall channel response
modeled as a reflective antenna with a phase shift, allowing us for one source-destination pair as follow:
to treat different channels similarly to those between source 𝑃
and destination antenna(s). This modeling is independent of
Õ
(𝑖)
ℎ𝑠−𝑡 = ℎ𝑟𝑒 + ℎ𝐿𝑜𝑆 (6)
the metasurface hardware design and assumes no coupling 𝑖=1
between surface elements.
Ray tracing-based modeling. We employ ray tracing to ac- where 𝑃 is the number of NLOS propagation paths between
curately and efficiently simulate wireless signal propagation, the source and destination. By replacing the source and desti-
tracing rays from the source and modeling their paths and nation with TX, RX, or metasurfaces, we have all the chan-
directions to compute the channel response. Due to the high nel response matrices needed for the objective function, i.e.,
attenuation of mmWave, we primarily focus on modeling H𝑇 −𝑅 , H𝑇 −𝑚𝑠 and H𝑚𝑠−𝑅 . We plug into these values into Equa-
reflections and direct LOS propagation paths. tion 3 and solve the optimization problem.
We calculate the channel response in two stages. Initially, Acceleration. We develop several techniques to boost simu-
we launch rays to identify valid propagation paths, emitting lation efficiency. Initially, we optimize the propagation path
numerous rays from the sources. Some rays reach the destina- discovery, the most resource-intensive step in ray tracing, by
tion antennas, while others are discarded due to missing the sharing path information among nearby destinations. While
destination or low signal strength. We record the propagation casting a large number of rays reduces the likelihood of over-
paths together with any intersecting reflectors for rays that looked paths and heightens accuracy, it also incurs substantial
reach the destination. Next, we compute the channel response memory and computation costs. To mitigate these costs, we
for each valid path and aggregate these responses to determine divide the area of interest into a grid system where each grid
the overall channel response at the destination, encompass- can exchange propagation path details with its neighbors, al-
ing both amplitude and phase information. To validate the lowing for the discovery of paths using significantly fewer
accuracy of our channel modeling, we compare our simu- rays. Moreover, we refine data storage by cataloging path in-
lation results with real-world measurements. As shown in formation through the set of reflectors traversed from source
Figure 4, our simulations correspond closely with real-world to destination, calculating the cascaded coordinate transfor-
measurements. Further validation is conducted in section 5. mation matrices for mirroring. This approach allows multi-
Channel response calculation. Given the propagation paths ple TX-RX pairs in close proximity to reuse these matrices,
between a source and a destination, we need to calculate the conserving memory and computation time when calculating
channel response. Here, we describe the channel model used channel responses. Finally, we implement our ray tracing
in our simulation. We first analyze the LOS path. The channel using OptiX [36], which capitalizes on RT cores for GPU
response can be derived as ℎ𝐿𝑜𝑆 = 4𝜋𝑟 𝜆 2𝜋𝑟
𝑒 𝜆 , where 𝜆 is the acceleration. With this, we can simulate approximately 1.6
wavelength corresponding to the carrier frequency, and 𝑟 is billion wireless channel responses in just 3 minutes on an
the distance from the AP to the user. For NLOS paths that NVIDIA A100 GPU.
include multi-order reflections from objects, we analyze each Material reflection coefficients. The reflection coefficient of
reflection separately and multiply them together. When a path materials significantly influences the wireless channel. In our
goes from the previous point to the next point, two types of experiments, we use values from existing literature [24, 45]
reflection can happen as shown in Figure 3. When the reflector and enhance our database by measuring mmWave proper-
is sufficiently large to be considered as a mirror, e.g., walls ties of uncharted materials. To improve simulation precision,
and floors, such that the previous point is within the near-field we develop a method that utilizes RSS measurements to de-
distance of it, we have 𝜃 𝑖𝑛𝑐 = 𝜃𝑜𝑢𝑡 and the channel response duce material reflection coefficients. During environmental
of this reflection is scanning for 3D modeling, a mobile device with mmWave
𝜆 2𝜋 connectivity also gathers RSS data throughout the area. At
ℎ𝑟𝑒 = 𝜏 𝑒 𝜆 (𝑟𝑖𝑛𝑐 +𝑟𝑜𝑢𝑡 ) (4) each location where RSS is collected, our ray tracing sim-
4𝜋 (𝑟𝑖𝑛𝑐 + 𝑟𝑜𝑢𝑡 )
ulator identifies the propagation paths of all rays reaching
When the reflector is small enough to be considered as a far- that point. We then treat the reflection coefficients as vari-
field scatterer, e.g., a metasurface element, we have scattering: ables to be optimized, calculating RSS values based on these
𝜆 2𝜋 𝜆 2𝜋 propagation paths. Starting with coefficients from the litera-
ℎ𝑟𝑒 = 𝜏 𝑒 𝜆 𝑟𝑖𝑛𝑐 𝑒 𝜆 𝑟𝑜𝑢𝑡 (5) ture, we employ gradient descent to find values that minimize
4𝜋𝑟𝑖𝑛𝑐 4𝜋𝑟𝑜𝑢𝑡

67
AUTO MS: Automating Metasurface-aided mmWave Coverage ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA

AP

.5m
14

18
.2m

(a) 3D model of a conference room (b) Simulation result (c) Measurement result
Figure 3: Illustration of path loss when re- Figure 4: Comparison between simulation results of our ray-tracing
flected by a reflector vs. a metasurface unit based channel modeling and real-world RSS measurements
the discrepancy between the calculated and actual RSS mea- as a quarter-wave impedance transformer, as detailed in Mi-
surements. As shown in Figure 8, using optimized material crowave Engineering [40], which effectively transforms the
properties results in simulations that are more accurate than zero impedance of a metal sheet into infinite impedance. We
those based solely on literature values. control the impedance mismatch between air and this infinite
impedance using a metallic pattern, drawing inspiration from
3.4 Low-cost Metasurface Design prior work [30]. Therefore, a set of split-ring patterns, span-
ning a limited range of surface pattern admittance, results in
In this section, we present the development of a reflective and
a broad reflection phase range.
passive mmWave metasurface design that is employed in our
proposed AUTO MS. Equivalent circuit analysis. To enhance the understanding
of phase control, we refer to the equivalent circuit model
Design goals. We set the following design goals: (i) high depicted in Figure 5b. Here, air and the surface substrate are
reflectivity (e.g., ≥ 90%), to minimize energy loss due to represented as transmission lines with impedances denoted
metasurface.; (ii) full 2𝜋 phase control, which is crucial for by 𝑍 0 = 376.7 Ω and 𝑍 1 , respectively. The substrate-based
beamforming; (iii) wideband operation, covering a 60 GHz transmission line possesses a thickness 𝑙 and a phase constant
channel bandwidth; and (iv) ultra-low cost, i.e., considerably 𝛽. EM waves arriving at the surface induce current along
cheaper than other solutions (see Table 1), achieved through the pattern, enabling us to model the metallic pattern as a
hot-stamping fabrication on common materials. circuit component with an admittance 𝑌𝑝 . This admittance is
purely imaginary, as the resistance of the metallic patterns
Challenges and our approach. Achieving 2𝜋 phase control is negligible. The metal sheet provides a termination for the
and high reflectivity for mmWave with a low-complexity and substrate transmission line with an impedance 𝑍𝑚 = 0 (short-
low-cost design is a challenging task. Previous works have circuited). Consequently, the substrate and the metal sheet
used either meta-atoms with rib-like arrangements [11] or 3D together form a short-circuited transmission line, as described
printed materials with varying thicknesses [41], which poses in Microwave Engineering [40], resulting in an impedance:
challenges in terms of surface thickness, assembly, and fabri-
cation. In contrast, we adopt 2D metallic patterns on a thin 𝑍𝑚 + 𝑗𝑍 1 tan(𝛽𝑙)
𝑍𝐿 = 𝑍 1 = 𝑗𝑍 1 tan(𝛽𝑙) (7)
(∼0.5 mm) dielectric substrate and a metallic sheet to induce 𝑍 1 + 𝑗𝑍𝑚 tan(𝛽𝑙)
reflection phase shifts (Figure 5d), achieving simplicity and which describes the total impedance of the substrate and the
sub-mm thickness. Figure 5a shows a split ring-based pattern, metal sheet (the whole surface except metallic patterns). The
which is the template of different patterns for phase control. phase delay caused by the substrate is around 𝜋2 as a quarter
By varying the pattern dimensions, we can control the phase wave impedance transformer, so 𝑍 𝐿 ≈ 𝑗𝑍 1 tan( 𝜋2 ) = ∞𝑗,
shift. However, a single layer has a limited control on the which indicates the substrate transforms the impedance of
phase shift [38]. One solution is to cascade more layers of pat- metal sheet from 0 to ∞𝑗.
terns [35], but this significantly increases design complexity Subsequently, the surface can be modeled as two parallel
and power loss due to dissipation in the substrate. circuit components: (i) a metallic pattern with admittance
𝑌𝑝 , and (ii) a combination of the substrate and metal sheet,
Substrate as an impedance transformer. We propose a which we represent with an impedance 𝑍 𝐿 . The equivalent
novel design that leverages the surface substrate between impedance of the entire surface is determined as follows:
metallic patterns and a metal sheet as a microwave compo-
1 1
nent. As incident waves propagate through the substrate, it 𝑍𝑆 = ≈ (8)
functions like a transmission line, inducing a phase delay. 𝑌𝑝 + 1/𝑍 𝐿 𝑌𝑝
Consequently, we consider substrate thickness to be a critical We observe that the 𝑍𝑆 undergoes significant variations when
design parameter. With a phase delay of 𝜋2 , the substrate acts the admittance 𝑌𝑝 fluctuates near zero. Following this, the

68
ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA Ruichun Ma, Shicheng Zheng, Hao Pan, et al.

Reflection Coeff (dB)

0.04
2.5 mm 180
Air Surface (!! ) Admittance (S)
0.03 Phase (degree) -0.5
R !! !" " #" $ 90 Reflection Power Loss < 0.5 dB
W
""# Substrate 0.02 Paper+Plastic Only 0.3 dB higher than Rogers PCB
G Rogers PCB
0 -1
!"! Metal 0.01
50 55 60 65 70
200 Frequency (GHz)
"$%& Sheet

Phase (degree)
0.5 mm Patterns !! " # -90
Top-Down layers: 0
Pattern- Papers- PVC - Metal 0
-180
Substrate -0.01
-200
50 55 60 65 70

(a) Split Ring Pattern (b) Circuit Model Frequency (GHz)

(c) Phase vs. admittance (d) Wideband S11 performance.
Figure 5: AUTO MS metasurface element design. The reflection phase of the surface varies with different pattern
admittance in (c); It provides high reflection power and near 2𝜋 phase coverage over wideband (e.g., 57-65 GHz)

reflection from the surface can be modeled as the interaction metasurface’s beam pattern, our focus is on maintaining a
between the air transmission line with impedance 𝑍 0 and a steady phase differential between different patterns. The di-
load represented by 𝑍𝑆 . Consequently, we can express the mensions of the patterns, particularly the ring width, affect
reflection coefficient (S11), i.e., the ratio of the incoming both admittance and wideband performance. We chose mul-
signal 𝐸𝑖𝑛 to the reflected signal 𝐸𝑜𝑢𝑡 [40], as follows: tiple patterns that maintain high reflection and a consistent
𝐸𝑖𝑛 𝑍𝑆 − 𝑍 0 1/𝑌𝑝 − 𝑍 0 phase across 57 GHz to 65 GHz, adequately covering IEEE
Γ= = = (9) 802.11ad/ay channels. For precise phase control, we employ
𝐸𝑜𝑢𝑡 𝑍𝑆 + 𝑍 0 1/𝑌𝑝 + 𝑍 0
18 patterns in our design. Figure 5d illustrates the perfor-
Note that 𝑌𝑝 is an imaginary number and 𝑍 0 is a real number. mance of six such patterns. We anticipate that our design
∥1/𝑌 −𝑍 ∥
Thus, we have ∥Γ∥ = ∥1/𝑌𝑝𝑝 +𝑍 00 ∥ = 1, which means almost all could be adapted to other frequency bands as well.
incoming signal power is reflected. The phase of Γ, i.e., phase Metasurface pattern array. To create the metasurface de-
shift between incoming and reflected signals, becomes: sign, we integrate a set of patterns that facilitate phase control.
1 The patterns are spaced 2.5 mm apart, approximately half
∠Γ = 2 arctan +𝜋 (10)
Im(𝑌𝑝 )𝑍 0 of the wavelength, to minimize undesired coupling between
By modifying the dimensions of the split-ring pattern, we ef- different patterns.. The optimization framework yields the op-
fectively adjust the pattern’s admittance 𝑌𝑝 , for example, from timized phase configuration for each element, which we then
-0.005j S to 0.035j S in our design. Upon substituting these match with the corresponding patterns to construct the meta-
values, we achieve a nearly 2𝜋 phase shift range, provided surface. The efficacy of our design is demonstrated through
that 𝑌𝑝 offers an adequate tuning range. beam steering tests, as shown in Figure 10.
Selecting patterns. We select a set of split-ring patterns with Enabling low-cost fabrication. Our ultra-thin single-layer
varying admittances 𝑌𝑝 to span the entire 2𝜋 phase range. design enables the hot stamping of metallic patterns onto
Through HFSS simulations, we adjusted three ring pattern affordable materials without compromising performance. Al-
parameters—radius (R), width (W), and gap (G), as shown though standard PCB production offers high precision (up
in Figure 5a—to obtain a broad admittance spectrum. Fig- to 0.03 mm), it becomes costly for large metasurfaces—for
ure 5c displays the admittance and corresponding phase shifts instance, a 0.5 × 0.5 𝑚 2 RF PCB may exceed $1000. Metasur-
for six exemplary patterns. By varying the ring radius from faces, however, operate differently as RF signals propagate
0.6 mm to 1.1 mm, widths between 0.3 mm and 0.5 mm, orthogonally through the thin substrate, not along PCB traces.
and gaps from 0.3 mm to 0.7 mm, we achieved admittance At just 0.5 mm thick, excluding the metal layer, our design
values ranging from -0.005j S to 0.035j S. Given our hot- permits the use of less expensive substrates such as paper
stamping fabrication precision of approximately 0.1 mm, we and PVC, which, despite higher loss tangents, dramatically
adjusted the dimensions in increments of 0.1 mm, resulting in reduce costs. To confirm feasibility, we conducted simula-
a maximum phase error of 30 deg. This error is within accept- tions considering material loss tangents, precise substrate
able limits for wavefront manipulation and is more precise thickness, and small fabrication-induced discrepancies. The
than many commercial devices; for example, phased arrays results, depicted in Figure 5d, reveal that our design’s power
in commercial routers typically have only 𝜋4 granularity [60]. loss is minimal, only 0.3 dB higher than that of a Rogers
Wideband performance. Selecting patterns with wideband PCB substrate. The detailed fabrication process is outlined in
performance is crucial. The metasurface must ensure a high re- section 4. While recent studies [21, 48] have also employed
flection coefficient and consistent phase control across mmWave hot stamping, their designs solely affect cross-polarization
channels, such as 58-70 GHz for 802.11ad networks. Since phase and suffer from over 7 dB power loss, rendering them
a uniform phase offset across all patterns does not alter the unsuitable for our goal of coverage enhancement.

69
AUTO MS: Automating Metasurface-aided mmWave Coverage ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA

3.5 Practical Considerations

Robustness to changes. Our flexible optimization frame-
work employs several strategies to handle changes: (i) The
AP can pick the optimal beam steering codeword to dynami- Step1: Print patterns on Step2: Hot stamp aluminum
cally counter minor location variations, e.g., within < 20 cm. paper with a laser printer foil on paper with a laminator

(ii) For substantial changes in AP location or furniture re-

arrangement, we can adjust the AP codebook or metasurface
placement while keeping the metasurface configuration static.
(iii) To minimize disruptions from human activity, we po- Step3: Tear the aluminum foil
off to get the metallic patterns
sition the AP and metasurface above typical human height Step4: Paste patterns on the plastic sheet and aluminum board
during optimization (see Figure 18 and Figure 19). (iv) We Figure 6: Fabrication process of our proposed ultra-low-
account for various situations, optimizing for either average cost passive metasurfaces
or worst-case performance, e.g., different door states (refer to
Figure 20). Software implementation. We implement AUTO MS as a
software running on an Azure cloud server with one NVIDIA
3D modeling accuracy. We employ 3D Lidar mapping and A100 GPU. The computation hardware cost is amortized
image classification to scan environments, identifying objects and very low as a cloud instance. We leverage an existing
like wood tables, displays, walls, and doors, and determining toolkit [33] for the algorithm used in hyper-config tuner. We
their material types to assign appropriate reflection coeffi- use Optix library [36] to implement wireless channel simu-
cients (detailed in section 4). To improve accuracy, we refine lator and performs phase map optimization using gradient
these coefficients with RSS measurements (see subsection 3.3 descent in pyTorch.
and Figure 8). Our optimization is resilient to minor inaccura- Metasurface fabrication. Our metasurfaces are constructed
cies, such as wall shifts at the centimeter level, maintaining using common low-cost materials, detailed in Figure 6. Ini-
coverage integrity. Continuous RSS monitoring can also de- tially, a laser printer creates the pattern. Then, an aluminum
tect and adjust to environmental changes over time. hot-stamping foil is placed over the print and fed through a
Optimization customization and scalability. Our framework laminator at approximately 170◦ C, causing the metallic pow-
supports multiple optimization objectives, defaulting to sum der to bond to the ink. Once the foil is removed, the metallic
capacity. Utility weights can be assigned to different regions, patterns remain on the paper. To complete the assembly, we
allowing customization of the objective function for varied layer the materials: office printer paper (0.1 mm) holds the
use cases. It also supports optimization for single or multiple metallic pattern, and a plastic sheet (0.3 mm) augments the
AP placements with metasurfaces by including additional substrate thickness. These are combined with a metal ground
variables. We concentrate on single AP configurations with plane using double-sided tape, as illustrated in Figure 5a.
metasurfaces to effectively address coverage in blind spots. Cost comparison. Existing mmWave coverage enhancement
The scalability of our method depends on the GPU memory solutions are priced similarly to or above that of a mmWave
required for channel matrix calculations in the ray-tracing AP, diminishing their appeal (Table 1). For instance, mmWall
simulator; enhancing this calculation’s efficiency will boost features 76 × 28 elements with two MAVR-000120-14110P
our approach’s scalability. varactors each, and at $2.4 per varactor, the material cost for
varactors alone reaches $10000. Substituting varactors with
4 IMPLEMENTATION less expensive PIN diodes might reduce costs below $1000,
yet considerable engineering effort is required for comparable
Modeling 3D environment. Mobile devices, e.g., iPhones performance. Passive metasurfaces present a more affordable
and iPads, are equipped with a Lidar sensor to provide 3D alternative. A 10 × 10 cm2 MilliMirror sample with 80 × 80
depth information. RoomPlan [12] is an API by Apple that elements is priced at about $15, making a 40 × 50cm2 sur-
runs a machine learning algorithm on both Lidar and RGB face approximately $300, requiring specialized 3D printing
camera information to generate 3D room plan information. resources. In contrast, our AUTO MS metasurface of the same
Based on this feature, many mobile apps have been devel- size with 160 × 200 elements costs merely $3, utilizing readily
oped to create 3D models of real-world environments, e.g., available materials and tools. This cost accounts for hot stamp-
polycam [39]. In this paper, we use off-the-shelf solutions to ing foils ($0.1), printer paper ($0.1), a PVC sheet ($1), and a
implement 3D model and mainly focus on fast simulation of 1 mm thick metal plate ($2). Mass production could further
EM wave propagation in a given environment and effective reduce costs. Additionally, while we use a thick metal plate
optimization of metasurface design and placement. for stability and ease of manual fabrication, a thin 0.05 mm

70
ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA Ruichun Ma, Shicheng Zheng, Hao Pan, et al.

MS3
AP

9.8m
MS2

MS1
MS1 MS2 MS3 One AP Three APs One AP + Metasurfaces Gain vs. One AP
10.6m
i) Layout and deployment locations ii) Optimized metasurface designs (phase maps) iii) Coverage performance and gain from metasurface

(a) Simulation scenario I with an indoor AP

Base station

MS3
9.2m

MS2
MS1

13.8m Base Station Base Station + Metasurfaces

MS1 MS2 MS3 Gain vs. base station
i) Layout and deployment locations ii) Optimized metasurface designs (phase maps) iii) Coverage performance and gain from metasurface

(b) Simulation scenario II with an outdoor base station

Figure 7: Coverage performance results of our proposed AUTO MS on the simulation models

metal sheet suffices for functionality at a significantly lower protocol. For the receiver, we use a single antenna with an
cost, bringing the total metasurface cost to around $1. Our omni-directional pattern, which matches the default setting
ultra-low-cost design offers a viable option for enhancing of the router. The router driver reports the RSS values of the
mmWave coverage. best codeword to the Linux kernel, which we can access by
executing the command ‘iw dev $interface scan’
5 EVALUATION at the RX. This RSS value represents the received signal
strength at RX for the beacon from TX. We also set up a
5.1 Experiment Setup separate 2.4 GHz network for router control and remote data
We use simulation to study diverse environments and settings, collection.
and use testbed to evaluate coverage improvement in realistic Testbed experiments. We conduct experiments in three in-
scenarios. Additionally, we also evaluate AUTO MS’s ability door scenarios (Figure 14): (1) the AP is in an open office
to handle channel fluctuations and compare its performance area and AUTO MS extends the coverage to nearby conference
with alternatives like mirror reflectors and manual placement. rooms, (2) the AP is in a conference room and AUTO MS ex-
Simulation models. We obtain 3D models of various rooms tends the coverage to the hallway and neighboring room, (3)
from a public dataset [18] for AUTO MS analysis. Five models the AP is in the living room of an apartment and AUTO MS
are tested, with the initial two depicted in Figure 7, represent- extends the coverage to the bedroom and restroom. We use the
ing common mmWave coverage solutions: indoor APs and ploycam app [39] to obtain the 3D models of the experiment
outdoor base stations. In simulations, we use a 15 × 15 phased environment. Based on the 3D models, AUTO MS generates
array, with a fixed transmitter location and AUTO MS generat- the optimal metasurface phase map and deployment locations.
ing the metasurface design, placement, and AP’s codebook. We fabricate and deploy the metasurfaces according to our
Testbed setup. For our testbed experiments, we employ TP- optimization results. Since our commercial APs can only re-
Link Talon AD7200 commercial routers [54] equipped with port RSS above -70 dBm, we use -70 dBm as the baseline
QCA6335 BM [42] as mmWave endpoints (AP and client). value for RSS gain calculation when the client fails to report
By default, we designate the client as the receiver (RX), as RSS, i.e., blind spots.
metasurface coverage enhancement applies to both uplink
and downlink due to channel reciprocity. We flash a modi-
5.2 Micro benchmarks
fied image of OpenWrt [34, 50, 51] to control routers. For Channel modeling accuracy. We evaluate the simulator accu-
the transmitter, we replace the original antenna array with a racy based on the RSS error between real-world measurement
6 × 6 phased array [60] for better beam steering. By updat- and simulator output. In the real-world scenario III, we test
ing the codebook file loaded to the firmware, we change the with measurements from 133 locations. Initially, our simulator
code words used at AP [49, 55]. Each codeword alters the yields a median error of 4.10 dB using the reflection coeffi-
phase of signals from the phased array antennas to generate cients from the literature. After estimating these coefficients
a certain beam pattern. We use 30 codewords for each exper- via gradient descent based optimization (subsection 3.3), we
iment. The AP selects the best code according to 802.11ad achieve a reduced median error of 2.37 dB in the testing data

71
AUTO MS: Automating Metasurface-aided mmWave Coverage ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA

1 1 -30

Capacity improvement
0.8 -32

RSS (dB)
0.6 -34
CDF

0.5
0.4 -36
Scenario#1 Metal plate
0.2 Optimized
Optimized Scenario#2 -38 45deg Beam
Literature
Estimated Scenario#3 30~60deg Diffuser
0 0 -40
0 2.5 5 7.5 10 12.5 15 0 2 4 6 8 0 10 20 30 40 50 60 70 80 90
RSSI Error (dB) Optimization time (h) Angle (degrees)
Figure 8: Channel simula- Figure 9: Capacity opti- Figure 10: Metasurface Figure 11: Impact of AP an-
tor accuracy mization convergence beam steering validation tenna array and surface sizes

50x50cm Reflectors 1mx1m Reflectors 2mx2m Reflectors 25cmx25cm Metasurfaces

MS3 MS3 MS3

MS2 MS2 MS2

MS1 MS1 MS1

(a) Optimized placement (b) Manual placement 1 (c) Manual placement 2 Coverage: -57.69 dBm -55.61 dBm -53.63 dBm -50.89 dBm

Figure 12: Coverage gain comparison between optimized

and manually selected placement of metasurfaces

as shown in Figure 8. Our accuracy aligns with state-of-the-art Gain: 1.98 dB 4.06 dB 6.04 dB 8.78 dB

approach (e.g., Nerf2 [61]), which reports 2.6 dB error. Figure 13: Coverage and gain comparison between large
reflectors and metasurfaces, utilizing median room RSS
Optimization convergence time. Figure 9 shows the opti-
as a performance metric
mization results over time for three real-world scenarios. We
normalize the optimization objective metric, i.e., capacity, metasurface measuring 150 × 150 nearly maximizes perfor-
and observe that all three scenarios converge within 8 hours. mance, with marginal benefits from further size increases.
The scatter points show the results after each iteration of the Conversely, when the metasurface size is relatively large, the
hyper-configuration tuner, which runs the simulated annealing performance of coverage can be improved by increasing the
algorithm. Each iteration takes about 1 min to simulate the phased array size.
channel response matrices and compute Equation 1. The total
time increases with the number of annealing iterations, but Comparison to large mirror reflectors. We compare meta-
can potentially be reduced significantly through computation surfaces with large mirror reflectors for coverage performance
reuse across iterations. using simulation scenario I, accounting for a vertical offset
between the AP and the target area. The comparison results
Metasurface design validation. To validate our metasurface are shown in Figure 13, we find that metasurfaces provide
design, we conduct beam steering experiments. We use two superior coverage, with RSS gains 7 dB higher than mirror
metasurfaces: one that reflects the beam from the AP at 45 reflectors four times their size. Even large reflectors (2×2 m2 )
degrees and another that produces a wide reflection beam are unable to address all blind spots. In contrast to mirror
from 30 to 60 degrees. Figure 10 shows that the experimental reflectors which are limited to signal reflection at angles dic-
angles match the target angles well. We compare them with a tated by the law of reflection, metasurfaces offer the flexibility
metal plate that reflects the beam at 0 degrees. Since we mea- to tailor 3D beam directions.
sure the results in a conference room instead of an anechoic
Comparison to manual placement. To assess the impact of
chamber, we do not report the exact gain of the metasurface
metasurface placement on coverage and validate our system’s
beam steering. We only demonstrate the beam steering direc-
efficacy, we conduct experiments in simulation scenario II.
tions. Our results confirm that our metasurface design can
We select two distinct placement settings for each of the three
apply the desired phase shifts to the reflected signals.
metasurfaces based on expert knowledge. These placements
Impact of phased array and metasurface sizes. In simu- are then benchmarked against those optimized by our sys-
lation scenario I, we analyze coverage performance gains, tem, as depicted in Figure 12. The findings suggest that the
i.e., RSS gains, with three metasurfaces at optimal locations, choice of placement caps the metasurface’s potential gains.
varying the sizes of the phased array and the metasurfaces. With our system’s optimization of metasurface and AP phase
Assuming the optimal locations remain consistent across dif- configurations, we attain median gains of 6.7 dB and 8.9 dB
ferent setups, our results, presented in Figure 11, indicate that for manual placements 1 and 2, respectively, and 11.6 dB for
gains improve with the enlargement of both the AP phased the optimized placement, all with identical metasurface sizes,
array and metasurface sizes. With a smaller phased array, a highlighting our system’s advantage over manual methods.

72
ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA Ruichun Ma, Shicheng Zheng, Hao Pan, et al.

1. 3D model 2. Layout and deployment locations 3. Real-world scene and deployment One AP One AP + 1 Metasurface AP codeword#1 w/o Metasurface

AP
Transmitter
.2m
14

8.5 4 5 6
m 1 2 3
MS1
One AP + Reflectors Gain from Metasurfaces w/ Metasurfaces
4. Phase distribution maps 5. Fabricated metasurface details AP codeword#2

MS1

Receiver

MS1 MS1
i) 3D model, layout, and optimal metasurface design ii) Simulation results: coverage performance and gain iii) Measurement results

(a) Real-world scenario I (median gain in the hallway: 15 dB)

1. 3D model 2. Layout and deployment locations 3. Real-world scene and deployment One AP One AP + Metasurfaces w/o Metasurface

MS2

MS1
AP Receiver
m

MS2
14

10 Transmitter
.4m

MS1 One AP + Reflectors Gain from Metasurfaces w/ Metasurfaces

4. Phase distribution maps 5. Fabricated metasurface details

MS1 MS2 MS1 MS2

i) 3D model, layout, and optimal metasurface design ii) Simulation results: coverage performance and gain iii) Measurement results

(b) Real-world scenario II (median gain in rooms: 9.5 dB)

1. Layout and deployment locations 2. Real-world scene and deployment One AP One AP + 2 Metasurfaces

AP MS2
Transmitter
MS1
MS1
Receiver
14.2 m
ii) Simulation results: coverage performance
3. Phase distribution maps 4. Fabricated metasurface details
One AP One AP + Metasurfaces

MS1 MS1 MS2

MS2
i) 3D model, layout, and optimal metasurface design iii) Measurement results

(c) Real-world scenario III (median gain in rooms: 7 dB)

Figure 14: Coverage performance results of our proposed AUTO MS in the real-world scenarios.

5.3 Enhancing mmWave Coverage

Testbed coverage improvement. In Figure 14, we present
Simulated coverage improvement. Figure 7 displays the our experimental setups, fabricated metasurfaces, and both
coverage performance, i.e., RSS distribution heatmaps, with simulated and measured RSS data with and without metasur-
and without metasurfaces for two simulation scenarios. Meta- faces for three real-world scenarios. The baseline RSS values
surfaces, when optimized and deployed via AUTO MS, create with the AP alone align closely with our simulation results for
a coverage comparable to multiple APs but at a significantly all scenarios. In areas with blocked LOS paths, baseline RSS
reduced cost, virtually eliminating blind spots. Compared to is low, leading to connectivity issues, so we focus on such
a single AP deployment, AUTO MS achieves substantial RSS areas. Deployment of metasurfaces, however, significantly
gains in target rooms, i.e., rooms not fully covered by LOS improves mmWave coverage in such target areas-hallways
paths from a single AP. As evidenced in Figure 15, across (real-world scenario I), adjacent conference rooms (real-world
five simulation scenarios, we enhance the median RSS by scenario II), and neighboring rooms (real-world scenario III),
21.6 dB in target rooms, with a median RSS gain of 12.1 dB, corroborating our simulations. As Figure 15 illustrates, AU -
by deploying metasurfaces. TO MS increases the median RSS by 11 dB in these target

73
AUTO MS: Automating Metasurface-aided mmWave Coverage
CDF over RX locations ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA

1 1 1

CDF over Links

CDF over Links

0.5 0.5 0.5
Baseline (simu)
AutoMS (simu) channel 1 Baseline
Baseline (testbed) channel 2 AutoMS
AutoMS (testbed)
0 0 0
-70 -60 -50 -40 -30 -10 0 10 20 30 40 0 200 400 600 800 1000
RSS (dB) RSS Gain (dB) Throughput (Mbps)
Figure 15: RSS gain for 5 simulation Figure 16: Gain for two separate Figure 17: TCP Throughput improve-
models and 3 real-world scenarios mmWave channels ment
-40 -40 -30
Baseline
AutoMS
RSS (dBm)

RSS (dBm)
-50 -40

RSS (dB)
-50
-50
-60
-60
Baseline Baseline -60
AutoMS -70 AutoMS
-70 -70
0 20 40 60 1 2 3 4 5 6 0 15 30 45 60 75 90
Time (s) Position ID Door Opening Angle (degree)
Figure 18: RSS measurements over Figure 19: RSS stability at multiple lo- Figure 20: RSS values when the door
time when people walk around cations when people walk around opens at different angles

areas, and achieves gains exceeding 20 dB at previously blind higher and more consistent RSS levels over time than the
spots compared to the baseline without metasurfaces. baseline. In a static environment, these curves would appear
Multi-channel support. We collect RSS measurements on flat. Figure 19 confirms AUTO MS’s consistent performance
both channel 1 (57.24-59.40 GHz) and channel 2 (59.40- enhancement across multiple locations with minimal RSS
61.56 GHz) of the 802.11ad protocol in real-world scenario I. fluctuation, even with human movement.
Our surface design’s wideband phase control yields compara- Door Movements. We evaluate the performance of AUTO MS
ble RSS gains for both channels, with a median increase of in simulation scenario I, featuring 3 doors at varying angles.
9.5 dB and peaks up to 30 dB, as displayed in Figure 16. Simulations afford exact control over door angles, enabling
Throughput improvement. Next, we show the throughput extensive testing. Findings depicted in Figure 20 reveal that
improvement brought by AUTO MS. We run iperf3 TCP baseline RSS is affected by door positions, whereas AUTO MS
measurements for RX positions inside conference rooms from delivers stable and consistently superior RSS compared to the
scenario II. We notice that the commercial mmWave devices baseline.
fail to establish a network or TCP connection when RSS is
lower than -55 dBm. Our system benefits most to the sce- 6 CONCLUSION
narios where there are low or no throughput, and it can im- Optimizing mmWave deployment for comprehensive cover-
prove the RSS to provide a reliable TCP connection with high age is a significant hurdle for the broad implementation of 5G
throughput. As shown in Figure 17, AUTO MS improves me- networks. These challenges stem from the inherently complex
dian TCP throughput from 77.2 Mbps without metasurfaces nature of mmWave signal propagation as well as the special-
to 373 Mbps with metasurfaces. ized expertise required for effective network management.
This paper presents AUTO MS, an innovative framework that
5.4 Robustness to Dynamics automates optimal network deployment using ultra-low-cost
In this part, we evaluate the robustness of AUTO MSwhen passive metasurfaces to enhance mmWave coverage. AU -
facing environmental changes. TO MS significantly improves network performance, elimi-
nates coverage dead zones, provides high throughput, and
Human movements. In testbed scenario I, following opti-
adapts to diverse channel conditions in various settings. Thus,
mization outputs, we install the AP and metasurface over 2 m
AUTO MS emerges as a promising solution to accelerate the
high, above areas of human traffic, enabling signal reach to
deployment and adoption of mmWave networks.
the metasurface. The metasurface then redirects the beam
downward to cover RX locations. We perform 100 RSS mea-
surements at 6 hallway positions, as marked in Figure 14a, at ACKNOWLEDGMENTS
a rate of one per second amidst pedestrian traffic. Figure 18 We are grateful to anonymous reviewers for their constructive
illustrates the temporal RSS stability at location 1 with the comments. We also thank Dr. Yifei Shen for his enlightening
metasurface in place versus without. AUTO MS maintains discussions.

74
ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA Ruichun Ma, Shicheng Zheng, Hao Pan, et al.

REFERENCES 55(12):186–192, 2017.

[1] Next G Alliance Report: Roadmap to 6G. https://nextgalliance.org/wp- [20] Jean-Baptiste Gros, Vladislav Popov, Mikhail A Odit, Vladimir Lenets,
content/uploads/2022/02/NextGA-Roadmap.pdf,. and Geoffroy Lerosey. A reconfigurable intelligent surface at mmwave
[2] Omid Abari, Dinesh Bharadia, Austin Duffield, and Dina Katabi. En- based on a binary phase tunable metasurface. IEEE Open Journal of
abling High-Quality Untethered Virtual Reality. In NSDI, pages 531– the Communications Society, 2:1055–1064, 2021.
544, 2017. [21] Hichem Guerboukha, Yasith Amarasinghe, Rabi Shrestha, Angela Piz-
[3] Altair. Altair’s 2022.2 Simulation Software. https://altair.com/ zuto, and Daniel M. Mittleman. High-volume rapid prototyping tech-
simulation-2022-2, 2023. nique for terahertz metallic metasurfaces. Optics Express, 29(9):13806–
[4] Amazon. NETGEAR Nighthawk X10 Smart WiFi Router 13814, April 2021.
(R9000). https://www.amazon.com/NETGEAR-Nighthawk-Quad- [22] Ish Kumar Jain, Raghav Subbaraman, and Dinesh Bharadia. Two beams
Stream-Quad-core-Compatible/dp/B01M12RE4A, 2023. are better than one: Towards reliable and high throughput mmWave
[5] Venkat Arun and Hari Balakrishnan. RFocus: Beamforming Using links. In Proceedings of the 2021 ACM SIGCOMM 2021 Conference,
Thousands of Passive Antennas. In NSDI, pages 1047–1061, 2020. pages 488–502, 2021.
[6] GSM Association. 5G mmWave Coverage Solutions. https://www. [23] Wahab Khawaja, Ozgur Ozdemir, Yavuz Yapici, Fatih Erden, and Is-
gsma.com/futurenetworks/wp-content/uploads/2022/12/5gmmWave- mail Guvenc. Coverage enhancement for NLOS mmWave links using
coverage-extension-solutions-whitepaper_FINAL-v1.pdf, 2023. passive reflectors. IEEE Open Journal of the communications Society,
[7] Ertugrul Basar and Ibrahim Yildirim. Reconfigurable intelligent sur- 1:263–281, 2020.
faces for future wireless networks: A channel modeling perspective. [24] B Langen, G Lober, and W+ Herzig. Reflection and transmission
IEEE Wireless Communications, 28(3):108–114, 2021. behaviour of building materials at 60 ghz. In 5th IEEE International
[8] Justin Chan, Changxi Zheng, and Xia Zhou. 3D Printing Your Wireless Symposium on Personal, Indoor and Mobile Radio Communications,
Coverage. In Proc. of HotWireless, 2015. Wireless Networks-Catching the Mobile Future., volume 2, pages 505–
[9] Lili Chen, Wenjun Hu, Kyle Jamieson, Xiaojiang Chen, Dingyi Fang, 509. IEEE, 1994.
and Jeremy Gummeson. Pushing the physical limits of IoT devices [25] Wonwoo Lee, Semin Jo, Kanghyeok Lee, Hong Soo Park, Junhyuk
with programmable metasurfaces. In Symposium on Networked Systems Yang, Ha Young Hong, Changkun Park, Sun K Hong, and Hojin Lee.
Design and Implementation (NSDI), pages 425–438. USENIX, 2021. Single-layer phase gradient mmwave metasurface for incident angle
[10] Lili Chen, Bozhong Yu, Ju Ren, Jeremy Gummeson, and Yaoxue Zhang. independent focusing. Scientific reports, 11(1):12671, 2021.
Towards Seamless Wireless Link Connection. In Proceedings of the [26] Zhuqi Li, Yaxiong Xie, Longfei Shangguan, Rotman Ivan Zelaya,
21st Annual International Conference on Mobile Systems, Applications Jeremy Gummeson, Wenjun Hu, and Kyle Jamieson. Towards pro-
and Services, pages 137–149, 2023. gramming the radio environment with large arrays of inexpensive an-
[11] Kun Woo Cho, Mohammad H Mazaheri, Jeremy Gummeson, Omid tennas. In 16th USENIX Symposium on Networked Systems Design and
Abari, and Kyle Jamieson. mmWall: A steerable, transflective metamate- Implementation (NSDI 19), pages 285–300, 2019.
rial surface for NextG mmWave networks. In 20th USENIX Symposium [27] Christos Liaskos, Shuai Nie, Ageliki Tsioliaridou, Andreas Pitsillides,
on Networked Systems Design and Implementation (NSDI 23), pages Sotiris Ioannidis, and Ian Akyildiz. End-to-end wireless path deploy-
1647–1665, 2023. ment with intelligent surfaces using interpretable neural networks. IEEE
[12] Apple Developer. RoomPlan API. https://developer.apple.com/ Transactions on Communications, 68(11):6792–6806, 2020.
augmented-reality/roomplan/, 2023. [28] Christos Liaskos, Ageliki Tsioliaridou, Shuai Nie, Andreas Pitsillides,
[13] Manideep Dunna, Chi Zhang, Daniel Sievenpiper, and Dinesh Bhara- Sotiris Ioannidis, and Ian F Akyildiz. On the network-layer modeling
dia. ScatterMIMO: Enabling virtual MIMO with smart surfaces. In and configuration of programmable wireless environments. IEEE/ACM
Proceedings of the 26th Annual International Conference on Mobile transactions on networking, 27(4):1696–1713, 2019.
Computing and Networking, pages 1–14, 2020. [29] Yuchen Liu, Yubing Jian, Raghupathy Sivakumar, and Douglas M
[14] Esteban Egea-Lopez, Jose Maria Molina-Garcia-Pardo, Martine Lien- Blough. Maximizing line-of-sight coverage for mmWave wireless
ard, and Pierre Degauque. Opal: An open source ray-tracing prop- LANs with multiple access points. IEEE/ACM Transactions on Net-
agation simulator for electromagnetic characterization. Plos one, working, 30(2):698–716, 2021.
16(11):e0260060, 2021. [30] Ruichun Ma and Wenjun Hu. RF-Mediator: Tuning Medium Inter-
[15] Chao Feng, Xinyi Li, Yangfan Zhang, Xiaojing Wang, Liqiong Chang, faces with Flexible Metasurfaces. In Proceedings of the 30th Annual
Fuwei Wang, Xinyu Zhang, and Xiaojiang Chen. RFlens: metasurface- International Conference on Mobile Computing and Networking, 2024.
enabled beamforming for IoT communication and sensing. In Proceed- [31] Ruichun Ma, R. Ivan Zelaya, and Wenjun Hu. Softly, Deftly, Scrolls
ings of the 27th Annual International Conference on Mobile Computing Unfurl Their Splendor: Rolling Flexible Surfaces for Wideband Wire-
and Networking. ACM, 2021. less. In Proceedings of the 29th Annual International Conference on
[16] Forsk. Aster Propagation Model. https://www.forsk.com/aster- Mobile Computing and Networking, 2023.
propagation-model, 2023. [32] Ajay Mahimkar, Ashiwan Sivakumar, Zihui Ge, Shomik Pathak, and
[17] Forsk. CrossWave Propagation Model. https://www.forsk.com/ Karunasish Biswas. Auric: using data-driven recommendation to auto-
crosswave-propagation-model, 2023. matically generate cellular configuration. In Proceedings of the 2021
[18] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao ACM SIGCOMM 2021 Conference, pages 807–820, 2021.
Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d- [33] Microsoft Research. Neural Network Intelligence. https://github.com/
front: 3d furnished rooms with layouts and semantics. In Proceedings microsoft/nni,.
of the IEEE/CVF International Conference on Computer Vision, pages [34] Joan Palacios, Daniel Steinmetzer, Adrian Loch, Matthias Hollick, and
10933–10942, 2021. Joerg Widmer. Adaptive codebook optimization for beam training on
[19] Yasaman Ghasempour, Claudio RCM Da Silva, Carlos Cordeiro, and off-the-shelf ieee 802.11 ad devices. In Proceedings of the 24th Annual
Edward W Knightly. IEEE 802.11 ay: Next-generation 60 GHz com- International Conference on Mobile Computing and Networking, pages
munication for 100 Gb/s Wi-Fi. IEEE Communications Magazine, 241–255, 2018.

75
AUTO MS: Automating Metasurface-aided mmWave Coverage ACM MobiCom ’24, September 30–October 4, 2024, Washington, DC, USA

[35] Hao Pan, Lili Qiu, Bei Ouyang, Shicheng Zheng, Yongzhao Zhang, Channel Estimation. In Proceedings of the Twenty-fourth International
Yi-Chao Chen, and Guangtao Xue. PMSat: Optimizing Passive Meta- Symposium on Theory, Algorithmic Foundations, and Protocol Design
surface for Low Earth Orbit Satellite Communication. In Proceedings for Mobile Networks and Mobile Computing, pages 41–50, 2023.
of the 29th Annual International Conference on Mobile Computing and [50] Daniel Steinmetzer, Daniel Wegemer, and Matthias Hollick. Talon
Networking, pages 1–15, 2023. tools: The framework for practical ieee 802.11ad research, 2018.
[36] Steven G Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared [51] Daniel Steinmetzer, Daniel Wegemer, Matthias Schulz, Joerg Widmer,
Hoberock, David Luebke, David McAllister, Morgan McGuire, Keith and Matthias Hollick. Compressive Millimeter-Wave Sector Selection
Morley, Austin Robison, et al. Optix: a general purpose ray tracing in Off-the-Shelf IEEE 802.11ad Devices. In Proceedings of the 13th
engine. Acm transactions on graphics (tog), 29(4):1–13, 2010. International Conference on Emerging Networking EXperiments and
[37] Zhangyou Peng, Linxiao Li, Miao Wang, Zhonghao Zhang, Qi Liu, Technologies, CoNEXT ’17, pages 414–425, New York, NY, USA,
Yang Liu, and Ruoran Liu. An effective coverage scheme with passive- November 2017. Association for Computing Machinery.
reflectors for urban millimeter-wave communication. IEEE Antennas [52] Mushtaq Talib, Norazizah Binti Mohd Aripin, Noor Shamsiah Othman,
and Wireless Propagation Letters, 15:398–401, 2015. and Adheed Hasan Sallomi. Comprehensive overview on millimeter
[38] Carl Pfeiffer and Anthony Grbic. Metamaterial Huygens’ Surfaces: wave communications for 5g networks concentrating on propagation
Tailoring Wave Fronts with Reflectionless Sheets. Physical Review models for different urban environments. In Journal of Physics: Con-
Letters, 110(19):197401, May 2013. ference Series, volume 2322, page 012095. IOP Publishing, 2022.
[39] polycam. Polycam - LiDAR & 3D Scanner for iPhone & Android. [53] Xin Tan, Zhi Sun, Dimitrios Koutsonikolas, and Josep M Jornet. En-
https://poly.cam/, 2023. abling indoor mobile millimeter-wave networks based on smart reflect-
[40] David M Pozar. Microwave engineering. John wiley & sons, 2011. arrays. In IEEE INFOCOM 2018-IEEE Conference on Computer
[41] Kun Qian, Lulu Yao, Xinyu Zhang, and Tse Nga Ng. MilliMirror: 3D Communications, pages 270–278. IEEE, 2018.
printed reflecting surface for millimeter-wave coverage expansion. In [54] Tp-link. Talon AD7200 Multi-Band Wi-Fi Router. https://www.tp-
Proceedings of the 28th Annual International Conference on Mobile link.com/sg/home-networking/wifi-router/ad7200/, 2023.
Computing And Networking, pages 15–28, 2022. [55] Teng Wei and Xinyu Zhang. Pose Information Assisted 60 GHz Net-
[42] Qualcomm. 802.11ad MAC baseband transceiver. https://www. works: Towards Seamless Coverage and Mobility Support. In Proceed-
qualcomm.com/products/technology/wi-fi/qca6335, 2023. ings of the 23rd Annual International Conference on Mobile Computing
[43] Remcom. Engineered Electromagnetic Surfaces (EES). https: and Networking, pages 42–55, 2017.
//www.remcom.com/wireless-insite-engineered-electromagnetic- [56] Teng Wei, Anfu Zhou, and Xinyu Zhang. Facilitating Robust 60 GHz
surfaces-ees, 2023. Network Deployment By Sensing Ambient Reflectors. In NSDI, pages
[44] Remcom. Wireless InSite 3D Wireless Prediction Software. https: 213–226, 2017.
//www.remcom.com/wireless-insite-em-propagation-software/, 2023. [57] Allen Welkie, Longfei Shangguan, Jeremy Gummeson, Wenjun Hu,
[45] Katsuyoshi Sato, Takeshi Manabe, Toshio Ihara, Hiroshi Saito, Shigeru and Kyle Jamieson. Programmable Radio Environments for Smart
Ito, Tetsu Tanaka, Kazuyoshi Sugai, Norichika Ohmi, Yasushi Mu- Spaces. In Workshop on Hot Topics in Networks (HotNets). ACM,
rakami, Masanori Shibayama, et al. Measurements of reflection and 2017.
transmission characteristics of interior structures of office building in [58] R Ivan Zelaya, Ruichun Ma, and Wenjun Hu. Towards 6G and Beyond:
the 60-ghz band. IEEE transactions on antennas and propagation, Smarten Everything with Metamorphic Surfaces. In Proceedings of the
45(12):1783–1792, 1997. Twentieth ACM Workshop on Hot Topics in Networks, pages 155–162,
[46] Sivers Semiconductors. Sivers Evaluation Kits (EVK) and Evalua- 2021.
tion Boards (EVB). https://www.sivers-semiconductors.com/sivers- [59] R Ivan Zelaya, William Sussman, Jeremy Gummeson, Kyle Jamieson,
wireless/evaluation-kits/, 2023. and Wenjun Hu. LAVA: fine-grained 3D indoor wireless coverage for
[47] Zhambyl Shaikhanov, Sherif Badran, Hichem Guerboukha, Josep Jor- small IoT devices. In Proceedings of the 2021 ACM SIGCOMM 2021
net, Daniel Mittleman, and Edward Knightly. MetaFly: Wireless Back- Conference, pages 123–136, 2021.
haul Interception via Aerial Wavefront Manipulation. In 2024 IEEE [60] Renjie Zhao, Timothy Woodford, Teng Wei, Kun Qian, and Xinyu
Symposium on Security and Privacy (SP), pages 151–151. IEEE Com- Zhang. M-cube: A millimeter-wave massive MIMO software radio. In
puter Society, 2024. Proceedings of the 26th Annual International Conference on Mobile
[48] Zhambyl Shaikhanov, Fahid Hassan, Hichem Guerboukha, Daniel Mit- Computing and Networking, pages 1–14, 2020.
tleman, and Edward Knightly. Metasurface-in-the-Middle Attack: From [61] Xiaopeng Zhao, Zhenlin An, Qingrui Pan, and Lei Yang. NeRF2:
Theory to Experiment. In Proceedings of the 15th ACM Conference Neural Radio-Frequency Radiance Fields. In Proceedings of the 29th
on Security and Privacy in Wireless and Mobile Networks, WiSec ’22, Annual International Conference on Mobile Computing and Network-
pages 257–267, New York, NY, USA, May 2022. Association for Com- ing, 2023.
puting Machinery. [62] Xia Zhou, Zengbin Zhang, Yibo Zhu, Yubo Li, Saipriya Kumar, Amin
[49] Yiwen Song, Changhan Ge, Lili Qiu, and Yin Zhang. 2ACE: Spectral Vahdat, Ben Y Zhao, and Haitao Zheng. Mirror mirror on the ceiling:
Profile-driven Multi-resolutional Compressive Sensing for mmWave Flexible wireless links for data centers. ACM SIGCOMM Computer
Communication Review, 42(4):443–454, 2012.

76