114 IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS, VOL. 10, NO.

1, MARCH 2020

Coding Programmable Metasurfaces Based on

Deep Learning Techniques
Tao Shan, Xiaotian Pan, Maokun Li , Member, IEEE, Shenheng Xu , Member, IEEE,
and Fan Yang , Fellow, IEEE

Abstract— Programmable metasurfaces have recently been [15], [16], by manipulating the amplitudes, phases and polar-
proposed to dynamically manipulate electromagnetic (EM) waves izations of electromagnetic (EM) waves. With the introduc-
in both temporal and spatial dimensions. With active components tion of the generalized Snell’s laws [17], metasurfaces have
integrated into unit cells of the metasurface, states of the unit
cells can be adjusted by digital codes. The metasurface can experienced explorative development and various interesting
then construct complex spatial and temporal electromagnetic devices have been developed in microwave and optical regime
beams. Given the main parameters of the beam, the optimal [18]–[23]. Furthermore, the concept of space-time-modulated
codes can be computed by nonlinear optimization algorithms, programmable metasurfaces has been proposed by modulating
such as genetic algorithm, particle swarm optimization, etc. EM waves in both spatial and temporal domain simultane-
The high computational complexity of these algorithms makes
it very challenging to compute the codes in real time. In this ously [24], [25].
study, we applied deep learning techniques to compute the Programmable metasurfaces, also known as coding meta-
codes. A deep convolutional neural network is designed and surfaces, have recently been proposed, which are capable
trained to compute the required element codes in milliseconds, of dynamically manipulating EM waves impinging on them
given the requirement of the waveform. The average accuracy [8], [26]–[28]. This is achieved by adding active components
of the prediction reaches more than 94 percent. This scheme
is validated on a 1-bit programmable metasurface and both into elements of the metasurface, such as positive-intrinsic-
experimental and numerical results agree with each other well. negative (PIN) diodes, varactors, etc. Independently switching
This study shows that machines may “learn” the physics of the state of these components can adjust the phase of reflected
modulating electromagnetic waves with the help of the good wave from each element on the metasurface and form com-
generalization ability in deep convolutional neural networks. The plex spatial waveforms. This provides an additional degree
proposed scheme may provide us with a possible solution for
real-time complex beamforming in antenna arrays, such as the of freedom to control wave propagation in real time. The
programmable metasurface. element reflection phase can be quantized into binary codes.
For example, the 1-bit code “0/1” represents the element can
Index Terms— Programmable metasurface, space-time-
modulation, deep learning, deep convolutional neural network, have two different reflection coefficients, and the 2-bit code
complex beamforming. “00/01/10/11” indicates four different states in the element
reflection coefficients.
I. I NTRODUCTION Space-time-modulated metasurface can be implemented on

M ETAMATERIALS, as artificially engineered structures,

demonstrate many novel electromagnetic phenomena
[1], [2], such as negative permittivity [3], perfect lensing
a programmable metasurface by dynamically coding the meta-
surfaces in real time [14]. The control codes for certain
beam patterns are usually pre-computed and stored as coding
[4], invisibility cloaking [5]–[7], and etc. Metasurfaces are matrices in a field-programmable gate array (FPGA) in order
two-dimensional equivalence of metamaterials by distribut- to have a real-time response. However, it is difficult to store
ing artificial periodic or quasi-periodic sub-wavelength struc- all the codes for all the scenarios especially for complex
tures on two-dimensional surfaces [1], [8]–[10], which can beams. Although interpolation scheme may help to derive the
realize multiple functionalities, such as broadband diffusion element phase from stored values, the accuracy of this scheme
[11], [12], beam steering [13], [14], polarization conversion may deteriorate in discrete domain where the element codes
are described. Therefore, for a programmable metasurface
Manuscript received December 1, 2019; revised January 22, 2020; accepted
January 30, 2020. Date of publication February 10, 2020; date of current containing thousands of the active components, an efficient
version March 12, 2020. The work was supported by National Science and accurate coding scheme is crucial in shaping the radiated
Foundation of China (61571264) and (61971263), National Key R&D Pro- electromagnetic waves.
gram of China (2018YFC0603604), Guangzhou Science and Technology Plan
(201804010266), Beijing Innovation Center for Future Chip, and Research Traditional coding schemes include back-projection and
Institute of Tsinghua, Pearl River Delta. This article was recommended by nonlinear optimization. The former one is based on the fact
Guest Editor S. Abadal. (Corresponding author: Maokun Li.) that the far field radiation pattern can be considered as a
The authors are with the State Key Laboratory on Microwave and Digital
Communications, Beijing National Research Center for Information Science spatial-domain Fourier transform of the current distribution
and Technology (BNRist), Department of Electronic Engineering, Tsinghua on the metasurface [29]. Given the radiation pattern, inverse
University, Beijing 100084, China (e-mail: [email protected]). Fourier transform can compute the current distribution on the
Color versions of one or more of the figures in this article are available
online at https://ieeexplore.ieee.org. metasurface hence the element reflection phase [8], [30]. With
Digital Object Identifier 10.1109/JETCAS.2020.2972764 the help of fast Fourier transform (FFT), this scheme can be
2156-3357 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
SHAN et al.: CODING PROGRAMMABLE METASURFACES BASED ON DEEP LEARNING TECHNIQUES 115

very efficient. It works well for coding a metasurface to gener- deep learning can correctly generate the waveform according
ate a single beam. However, its accuracy cannot be guaranteed to the requirement.
for coding multiple beams with complex requirement on the This paper is organized as follows: Section II presents the
beam, such as multiple beam directions with specific sidelobe deep learning framework of the coding scheme; Section III
levels, and etc. This is mainly because these requirements from shows both numerical and experimental results to benchmark
real applications cast the computation of element codes into the performance of this scheme; Discussion and Conclusions
a nonlinear ill-posed optimization in discrete domain, while are presented in Section IV.
FFT itself cannot solve.
The above challenge can be solved with iterative nonlinear II. M ETHODOLOGIES
optimization schemes [31], [32], such as genetic algorithm A. Back-Projection Approach
(GA) [33], [34] or particle swarm optimization (PSO) [35],
Back-projection (BP) approach is based on the array-theory
etc. GA has been widely applied in array synthesis and
method that is an important approach to calculate the scattering
optimization [36]–[39]. These algorithms iteratively minimize
pattern of a programmable metasurface with M × N unit cells.
the difference between the required and predicted wave pattern
Assuming the metasurface is illuminated by a horn antenna,
and can compute the optimal code given specific requirements
the radiated field of the metasurface can be written as [30]:
of wave patterns scattered by metasurfaces. However, these
methods usually have large computational complexity and 
M 
N
cosq f θ f (m, n)
cannot provide real-time responses. Therefore, real-time and E(θ, ϕ) = |mn |e j (φmn −k|rmn −r f |)
|rmn − r f |
accurate computation of element codes for programmable m=1 n=1
ˆ
metasurfaces is still a challenging problem due to its nonlinear · cosqe θ e j k(rmn ·ob) , (1)
nature and large computational cost.
where ob ˆ = x̂ sin θ cos ϕ + ŷ sin θ sin ϕ + ẑ cos θ is the
Recently, deep learning (DL) has attracted growing atten-
tion [40] due to its remarkable successes in speech [41], observation direction, rmn is the position vector of the mn t h
[42], image [43], [44] and video [42], [45] processing. Deep unit cell. Here,we use a cosine q f model to approximate feed
learning can abstract multi-level representations of data by horn pattern and a cosine qe model to approximate element
computing multiple layers of neural networks, which demon- receive mode pattern. In this paper, q f = 5 and qe = 1. In
strates outstanding learning capacity and powerful approx- Eq. (1),
imation capability [40]. This may even allow deep neural cos q f θ f (m, n)
|mn |e j (φmn −k|rmn −r f |) , (2)
networks to “learn” physical principles from data with a good |rmn − r f |
generalization ability. Recent works have reported deep neural
ˆ
networks can learn and predict physical phenomena, such as represents the element excitation and cosqe θ e j k(rmn ·ob) denotes
approximation of smoke diffusion process [46], prediction of the element radiation pattern. |mn |e j φmn is the element reflec-
electric potential field [47], inversion of electromagnetic data tion coefficient. In a programmable metasurface, φmn can be
[48]–[50]. These works design and train a deep convolutional set to discrete values. The code on every element forms a
neural network to discover an intricate representation of phys- matrix that can shape the beam radiated by the programmable
ical principle from a big dataset in the offline stage [40]. Then metasurface. For example, the 1-bit code “1” and “0” can
in the online stage, the trained network can make reliable represent φmn to be π and 0, respectively. In Eq. (1), −k|rmn −
predictions from the input based on the learned representation. r f | denotes the phase introduced by the illuminating waves.
ˆ
With the help of massive parallelization, the prediction process With the summation of weighted e j k(rmn ·ob) , Eq. (1) can be
can be computed in a fast speed. considered as an inverse Fourier transform. Hence we can
In this study, we apply the deep learning techniques to code rewrite it as [8]:
a programmable metasurface which is a new data-driven rou-
E(u, v)
tine by learning inner dominating law from a large amount of
data. We design and train a deep convolutional neural network = (1 − u 2 − v 2 )qe /2 · M · N
 
to “learn” the coding schemes from traditional coding schemes cos q f θ f (m, n) j (φmn −k|rmn −r f |)
· IDFT |mn |e , (3)
including both back-projection and nonlinear optimization |rmn − r f |
methods. We can code the metasuface to achieve single-beam where IDFT stands for the abbreviation of 2D inverse discrete
and dual-beam steering based on the same deep neural network Fourier transform, u and v are defined by
simultaneously. Single-beam and dual-beam steering mean that
the scattered farfield pattern of the programmable metasur- u = sin θ cos ϕ, v = sin θ sin ϕ, (4)
face is towards one or two main outgoing directions. Given
Therefore, if a desired scattering pattern is known, the reflec-
requirement of the waveform as input, the deep neural network
tion phase of the mn-th element, i.e., φmn , can be obtained
can compute the required element codes in milliseconds. The
from the Fourier transform of the scattering pattern which can
output element codes are in good agreement with the opti-
be written as:
mized codes generated by nonlinear optimization approaches    
E(u, v)
with an average accuracy of more than 94 percent. This φmn =  DFT +k|
r mn − 
r f |
scheme is validated on a 1-bit programmable metasurface. The (1 − u 2 − v 2 )qe /2 · M · N
experimental results show that the coding scheme based on (5)

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
116 IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS, VOL. 10, NO. 1, MARCH 2020

where  (·) denotes the calculation of the angle of a complex

number. The element codes can be derived by quantizing the
computed element phase values.
The BP method can predict the element phase given a
single beam direction. However, its accuracy deteriorates for
more complex requirement on the radiation pattern, such as
dual-beam directions, limits on sidelobe levels, etc. It is not
straightforward for BP method to compute a coding matrix
that satisfies all these constraints simultaneously. In these
scenarios, we can use optimization techniques to compute the
coding matrix, i.e., the discrete element phase [8].

B. Nonlinear Optimization Approach

The computation of the coding matrix can be formulated
as an optimization problem, in which we minimize an objec-
Fig. 1. A 1D example of radiation pattern, pattern lower and upper mask.
tive function measuring the difference between the com- The pattern is plot from −30◦ to 30◦ for detailed illustration.
puted radiation pattern and the requirement. This optimization
problem is nonlinear and non-convex. Therefore, given the
requirement on the scattering pattern of the metasurface,
we usually use stochastic nonlinear optimization algorithms,
such as genetic algorithm (GA) and particle swarm optimiza-
tion (PSO), to compute the optimal coding matrices. The
stochastic optimization algorithms iteratively search for the
optimal solution. The randomness in the algorithms can help
to find a global minimum with which the coding matrix gives
metasurfaces the optimal performance.
In this paper, GA is used to compute the element codes of
the programmable metasurface based on the requirement of
the scattering waveform. GA is a search algorithm based on
the mechanics of natural genetics [51], which has been widely
used in electromagnetic and optical engineering. The objective
of GA is to minimize the difference between the computed and
required beam pattern. The requirement, such as the lower and Fig. 2. Flowchart of the multiple-population genetic algorithm.
upper bounds of the beam pattern can be converted into the
lower and upper pattern mask in the optimization process. With
E, E Lower Mask and EU pper Mask denoting the computed beam
azimuth angles. E Lower Mask and EU pper Mask are two arrays
pattern, desired lower and upper pattern mask, the objective
representing lower and upper boundary of E. With careful
function of GA can be written as:
design of E Lower Mask and EU pper Mask , we can determine the
FG A = W 1 · F1G A + W 2 · F2G A (6) beam direction, 3dB beamwidth, sidelobe levels, beam shape
and etc. Figure 1 illustrates a 1D example of beam pattern,
where W 1, W 2 are the weighting coefficients. In Eq. (6),
lower mask and upper mask. From Figure 1, we can see that
F1G A only considers the error where E is smaller than
the lower mask plays an important role in 3dB beamwidth and
E Lower Mask :
 beam direction. Beam width and direction are key parameters
F1G A = (E(i, j ) − E Lower Mask (i, j ))2 in dual-beam steering. Therefore, the optimization objective
i, j Eq. (6) of GA should focus more on the Eq. (7) to make beam
i f E(i, j ) < E Lower Mask (i, j ) (7) pattern satisfy lower mask. In this paper, the values of W 1 and
W 2 are set to 2 and 1 respectively in Eq. (6).
and F2G A only considers error where E is larger than The flowchart of GA used in this paper for dual-beam
EU pper Mask : optimization with the objective function Eq. (6) is illustrated
 in Figure 2. In this algorithm, the population is initialized
F2G A = (E(i, j ) − EU pper Mask (i, j ))2
by transforming the coding matrices into chromosomes, and
i, j
then they are evolved in every iteration to find the most
i f E(i, j ) > EU pper Mask (i, j ) (8)
optimal chromosomes in the population that will be converted
In both Eq. (7) and Eq. (8), i, j represents the index of the to the desired coding matrix. The process of evolution consists
discrete samples on the computed beam pattern E. Specifi- of selection based on fitness, recombination (crossover) to
cally, E is an array of 2D radiation pattern in the u, v plane. generate offspring, mutation of offspring and reinsertion based
Here u = si nθ cosϕ, v = si nθ si nϕ, θ and ϕ are elevation and on the fitness of offspring, all of which are inspired by

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
SHAN et al.: CODING PROGRAMMABLE METASURFACES BASED ON DEEP LEARNING TECHNIQUES 117

Fig. 3. Schematic of coding scheme for programmable metasurface based on deep learning techniques. a is the input of the ConvNet including (i) upper
mask, (ii) lower mask and (iii) direction information of the desired beam. The first row shows the input of single-beam steering and the second shows the
input of dual-beam steering. b shows the schematic of the proposed ConvNet. c shows one example of the ConvNet output that codes the metasurface to
obtain multi-functionalities. d is the conceptual illustration of the programmable metasurface.

natural biological evolution. In order to improve the perfor- kernels are suitable for our problem of which the output
mance of GA optimization, multiple-population genetic algo- is a large binary array. Shortcut connections [54], [55] are
rithm (MPGA) is applied by separating the whole populations introduced in the last several convolutional layers to add
into multiple subpopulations [52]. Each subpopulation evolves outputs of different layers together, which introduces neither
independently and simultaneously, and individuals of each extra parameter nor computation complexity [56].
subpopulation are selected to migrate between subpopulations The scattering beam pattern generated by the coding matrix
based on the fitness. of metasurface computed by the ConvNet needs to meet the
requirement. Therefore, we take the requirement parameters
of the scattering beam pattern as the input of ConvNet.
C. Deep Learning Approach The input consists of two types of requirement parameters,
To predict the coding matrix based on the requirement as illustrated in Figure 3a. One describes the lower and upper
of the beam pattern, we design a fully convolutional neural mask of the desired pattern. The lower and upper mask
network (ConvNet) based on the VGG net [53], as shown in embody the parameters of the desired beams including the 3dB
Figure 3b. beamwidth, beam direction, beam shape and etc. The other
The proposed ConvNet is composed of eleven convolutional type of parameters is the information of beam direction for
layers and five pooling layers. Pooling layers introduce spatial emphasizing the fine-grained scattering beam direction in the
pooling following the first, second, fourth, fifth and sixth u, v plane, which is important in the functionalities of beam
convolutional layers. Small-sized receptive fields are used steering. For scattering patterns with a single-beam, the array
throughout the ConvNet that includes 1×1, 3×3, 4×4 and 5×5 for beam direction information can be defined as [57]:
convolutional kernels. 5 × 5 convolutional kernels are used in 
the first layer, 3×3 kernels are in layer 2 to 6, 4×4 kernels are dirsingle (i, j ) = (u i − u beam )2 + (v i − v beam )2 , (9)
in the seventh layer and 1×1 convolutional kernels are in layer
8 to 11. In layer 1 to 7, using small convolutional kernels to where (u beam , v beam ) represents the beam direction in the u, v
replace relatively large kernels can introduce more nonlinearity coordinate and (u i , v i ) denotes the u, v coordinates of each
and increase the depth of ConvNet, which makes the proposed element in the array of beam information. For multiple beams,
ConvNet more discriminative [53]. In layer 8 to 11, the size the array of beam direction information can be defined as the
of each feature map is 1 × 1 that represents the state of each minimum value of each individual beam direction information
metasurface unit cell. Besides, 1 × 1 convolutional kernels can array, i.e.,
introduce nonlinearity, change channels of output but don’t
change the size of output. Therefore, 1 × 1 convolutional dirmult iple = mi n{dir1 , dir2 , · · · , dirn , · · · , dir N }, (10)

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
118 IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS, VOL. 10, NO. 1, MARCH 2020

where N is the number of beams, and dirn represents the

n-th beam direction information array. Figure 3a shows one
example of input in single-beam and dual-beam steering case.
The beam lower mask, upper mask and direction information
arrays are very large because their discrete grids are very small
in order for the high resolution of beam scanning angle in
beam steering. Therefore, these arrays are truncated to retain
main features by dashed boxes, and then the truncated arrays
are used as the input of ConvNet, as shown in Figure 3a.
The output of ConvNet is the coding matrix of the pro-
grammable metasurface that can generate scattering beam
patterns fulfilling the input requirement. Figure 3c shows an
example of output. The output is composed of binary number
“1/0” corresponding to the on/off state of PIN diodes in the
metasurface elements, which can tune the element reflection
coefficient on metasurfaces and manipulate the impinging EM
Fig. 4. The average loss and accuracy of the training and testing process.
waves as shown in Figure 3d.
The ConvNet can be finely trained to make correct pre-
dictions of coding matrix via the optimization process. The we keep 38,560 samples for single-beam steering, of which
training dataset contains the coding matrix computed by BP 30,840 samples are used for training and 7,720 samples for
method for single beam and by GA method for multiple testing. Moreover, 83,840 samples are generated for dual-beam
beams. The optimization process in training the ConvNet aims steering, of which 67,040 samples are used for training and
to minimize the difference between the ConvNet output and 16,800 samples for testing. Both the training and testing
the desired output by adjusting the parameters in the ConvNet. data are selected randomly. Overall, we generate a dataset
The objective function of the optimization process is defined of 122,400 samples, of which 97,880 samples are used for
as the normalized Euclidean distance between the predicted training the ConvNet and 24,520 samples for testing the
and computed element coding matrices that can be written as: performance of the ConvNet.
The adaptive moment estimation (Adam) method is used
ˆ 2
(w, I ) − 
Fobj = (11) to minimize the objective function Eq. (11) in the training
ˆ 2
 stage. An Nvidia P100 GPU is used as the computing platform.
where (w, I ) and  ˆ denote the coding matrices predicted by The convergence curve is shown in Figure 4. It shows that a
ConvNet and computed by BP or GA method, respectively. w little over fitting exists. In order to evaluate the performance
represents the trainable parameter set in the ConvNet and the in detail, we calculate the average accuracy. For one sample,
proposed ConvNet has 37,870,304 parameters. I represents the accuracy is defined as:

the input with size of 3 × 200 × 200. It consists of three abs(M D L − M B P−G A )
acc = 1 − (12)
concatenated 200 × 200 2D arrays that represent upper mask, Element Number
lower mask and direction information of the desired beam where M D L is the coding matrix obtained by deep learn-
respectively. ing approach and M B P−G A represents the coding matrix by
back-projection or genetic algorithm approach. The average
III. N UMERICAL AND E XPERIMENTAL R ESULTS accuracy can be written as:
A. Numerical Simulation
1 
M
To validate the proposed approach, a 1-bit programmable accaver = accm (13)
M
metasurface consisting of 48×48 elements is taken as m=1
an example. By building one ConvNet as shown in where M is the number of samples.
Figure 3b, we can successfully program the metasurface Figure 4 demonstrates that both the average training and
to obtain single-beam and dual-beam steering in real time testing accuracy are more than 94%. Figure 5 demonstrates
simultaneously. three randomly-chosen results of single-beam steering. The
For single-beam steering, we define the beam direction coding matrices obtained by DL approach agree well with
using the elevation and azimuth angles θ and ϕ. The ranges the coding matrices obtained by BP approach and the cor-
of θ and ϕ are [0◦ , 60◦ ] and [0◦ , 179.5◦], respectively. In the responding beam patterns are in good agreement with each
construction of the dataset, we sample the beam direction at other. The theoretical beam directions of Figure 5 a, b, c are
an interval of 0.5◦ . In dual-beam steering, the beam direction (θ = 5◦ , ϕ = 148◦ ), (θ = 14.5◦ , ϕ = 133.5◦), (θ = 14◦ , ϕ =
θ and ϕ of beam A vary from 15◦ to 20◦ and from 10◦ to 24.5◦) respectively. The beam directions coded by BP and DL
59◦ , respectively; and the θ and ϕ of beam B vary from 15◦ approach are the same with theoretical directions which also
to 20◦ and from −90◦ to −139◦, respectively. The dataset validates DL approach. Figure 6 shows three randomly-chosen
is constructed by sampling θ and ϕ of both A and B beam results of dual-beam steering. Coding matrices optimized by
at an interval of 1◦ . After checking quality of the data, GA are not regular but deep learning approach can still capture

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
SHAN et al.: CODING PROGRAMMABLE METASURFACES BASED ON DEEP LEARNING TECHNIQUES 119

Fig. 5. Single-beam coding matrices and their 3D and 2D patterns calculated by back-projection and deep learning method (a, b, c). I (i, ii, iii) show the
coding matrix obtained by back-projection and the corresponding 3D and 2D beam pattern. II (i, ii, iii) show the coding matrix obtained by deep learning
and the corresponding 3D and 2D beam pattern. The view angle of 3D beam pattern is (azimuth angle = −37.5◦ , elevation angle = 30◦ ). I (iv) shows the
absolute error between the coding matrices obtained by back-projection I (i) and deep learning II (i). II (iv) compares the beam pattern on the principle plane
generated by these two approaches.

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
120 IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS, VOL. 10, NO. 1, MARCH 2020

Fig. 6. Dual-beam coding matrices and their 3D and 2D patterns calculated by genetic algorithm and deep learning method (a, b, c). I (i, ii, iii) are coding
matrix obtained by genetic algorithm approach and its corresponding 3D, 2D beam pattern. II (i, ii, iii) are coding matrix obtained by deep learning approach
and its corresponding 3D, 2D beam pattern. The view angle of 3D beam pattern is (azimuth angle = 0◦ , elevation angle = 30◦ ). I (iv) is the absolute error
between coding matrix obtained by genetic algorithm I (i) and deep learning approach II (i). II (iv) shows the absolute error between the normalized radiation
patterns (on a linear scale) coded by genetic algorithm and deep learning approach.

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
SHAN et al.: CODING PROGRAMMABLE METASURFACES BASED ON DEEP LEARNING TECHNIQUES 121

Fig. 7. Metasurface Prototype. a Conceptual illustration of the fabricated

programmable metasurface. b Measurement setup in an near-field anechoic
chamber.

the inner law and output coding matrices directly. As shown Fig. 8. Structure of the metasurface unit cell. a Whole view of a unit
in Figure 6, the coding matrices and beam patterns generated cell in the metasurface. b Top view of a unit cell in the metasurface. The
dimensions are (in mm): h 1 = 1.58, h 2 = 0.5, h 3 = 1.5, w = 16, px =
by GA and DL approach agree well with each other. For 2.5, p y = 1.75, s x = 7.5, s y = 6, gx = 3.5, g y = 4.3.
more detailed comparison, we show the absolute error of the
normalized beam patterns (on a linear scale) coded by GA and
DL approach in Figure 6 iv. The absolute error is defined as:
EG A EDL
absolute err or = abs( − ), (14)
E max E max
where E G A , E D L represent radiation patterns coded by GA
and DL approach that are on a linear scale, and E max =
max(E G A , E D L ) is the maximum value of them.

B. Experimental Verification
A 1-bit programmable metasurface is used to further verify
the numerical simulation, as shown in Figure 7a. The meta-
surface works at a central frequency of 9.37GHz and consists
of 2304 unit cells. The dimension of the unit cell is 16 mm ×
16 mm and its structure is illustrated in Figure 8. The unit
cell has three layers of different materials, of which the first Fig. 9. Measured reflection phase and phase difference of programmable
is made of Taconic TLX-8 (dielectric constant 2.55, loss of metasurface unit cell.
tangent 0.0019) and the other two are FR4 (dielectric constant
4.4, loss of tangent 0.02). In each unit cell, a rectangular
patch is printed on the first layer with a PIN diode (Skyworks aperture of the metasurface as the excitation source. Its
SMP1340), the second layer serves as the ground of the patch linear polarization direction is along the y axis, spa-
and the bias system for the PIN diode is printed on the third tial position is (0m, −0.15782m, 0.589m) and half-power
layer. The state of the PIN diode in each unit cell is controlled beamwidth (HPBW) is 36◦ . The wavefront impinging on
by its bias voltages (+0.8v for the ON state and -12v for the programmable metasurface is spherical wave. A probe is
the OFF state), which further provides two opposite phases used to measure the near filed of the metasurface and the
response of the unit cell. Figure 9 illustrates the measured scattering pattern can be calculated through transformation of
reflection phase of unit cell (ON state and OFF state), and the measured near field. Figure 10 shows the measured 2D, 3D
specially, the measured reflection phase difference between scattering patterns and their comparison on the principle plane.
ON and OFF states is 179.61◦. The two states correspond to In single-beam steering, the scattering patterns are measured
the 1 and 0 in the 1-bit coding scheme. based on the coding matrices obtained by back-projection and
The prototype of the programmable metasurface is fabri- deep learning approach, respectively, as shown in Figure 10a.
cated and assembled. The size of the entire metasurface is The discrepancy between back-projection-coded and deep-
768mm × 768mm, as shown in Figure 7b. The control circuits learning-coded scattering patterns is very small. Table I
of the metasurface can transfer the code matrix to the metasur- shows the detailed direction of back-projection-coded and
face and set the states of PIN diodes. Real-time coding of the deep-learning-coded single scattering beam. Figure 10b illus-
metasurface allows it to modulate the incident electromagnetic trates the measured scattering patterns corresponding to the
wave in both spatial and time domain simultaneously. coding matrices obtained by genetic algorithm and deep
The measurement setup of the metasurface is illustrated learning approach, respectively. The genetic-algorithm-coded
in Figure 7b. This programmable metasurface is measured scattering patterns agree well with the deep-learning-coded
in a near-field anechoic chamber. A linear polarized con- ones. Table II shows the detailed comparison of the direc-
ical horn antenna working at 9.37GHz illuminates the tion of genetic-algorithm-coded and deep-learning-coded dual

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
122 IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS, VOL. 10, NO. 1, MARCH 2020

Fig. 10. Measurement results. a Three measured single-beam scattering patterns of (i) back-projection and (ii) deep learning obtained coding matrices. b
Three measured dual-beam scattering patterns of (i,iii) genetic algorithm and (ii,iv) deep learning obtained coding matrices. (i,ii) are measured patterns of
beam A and (iii,iv) are measured patterns of beam B. In the legend, S1, S2, S3 represent three single-beam results; D1A, D2A, D3A represent beam A in
the three dual-beam cases and D1B, D2B, D3B represent beam B in the three dual-beam cases; BP, GA, DL denote the back-projection, genetic algorithm
and deep learning approaches, respectively.

scattering beam. The measured beam directions are slightly back-projection approach is an efficient coding scheme based
different from the desired beam directions due to the error on fast Fourier transform. However, it cannot guarantee its
in the measurement caused by feed blockage and assembly accuracy when the beams are complex with various con-
error. But the measured scattering patterns computed by the straints. Genetic algorithm, as an stochastic iterative non-
traditional approaches and the deep learning approach agree linear optimization method, can compute an optimal code
well with each other, which verifies the accuracy of the towards the beam requirements. But it cannot provide real-time
proposed approach. responses due to its large computational complexity. It usually
takes hundreds of iterations to reach the global minimum
C. Comparison of Computational Performance of the objective function. By combining advantages of both
In this above experiments, we apply three schemes to back-projection and genetic algorithm, the proposed deep
code the programmable metasurface that includes the back- learning based scheme can compute the code matrix for both
projection, genetic algorithm and deep learning approach. The simple and complex beams in real time.

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
SHAN et al.: CODING PROGRAMMABLE METASURFACES BASED ON DEEP LEARNING TECHNIQUES 123

TABLE I
S INGLE -B EAM S CATTERING PATTERNS M EASUREMENT R ESULTS

TABLE II
D UAL -B EAM S CATTERING PATTERNS M EASUREMENT R ESULTS

For a more intuitive comparison, we examine the comput- parallelization, the proposed deep network can compute the
ing time of the above three schemes to generate one code coding matrix in almost real time with a good accuracy. A 1-bit
matrix. It takes the back-projection approach 0.0025 seconds programmable metasurface with 2304 elements is fabricated
to generate one code matrix to steer a single beam. The and programmed using coding matrices computed from both
computation platform is Intel Xeon CPU E5-2667 V3 @ traditional and deep learning approaches. The results show that
3.20GHz. It can be faster on more specific computing platform deep learning approaches can compute the coding matrices
such as FPGA or digital signal processor. It takes the genetic that generate almost identical beam patterns in milliseconds.
algorithm about 38 seconds for the dual-beam case in this The preliminary results in this study shows that deep learning
paper. The computation is carried out on the same CPU and techniques can compute the coding matrices for complex beam
accelerated by one Nvidia Tesla K80 GPU. The deep learning patterns. With the help of efficient control circuits, we may
approach studied in this paper is computed on one Nvidia modulate the electromagnetic wave with a good accuracy using
P100 GPU and it takes an average of 0.0005 seconds to programmable metasurfaces in both time and spatial domain
generate one code for both single-beam and dual-beam cases. simultaneously. It may have many applications in microwave,
Although such a comparison is not fair enough because the optical, and acoustic engineering. In the next step, we will try
computing platforms are not identical, it still indicates that to extend the deep neural network to compute more complex
the deep learning approach can provide not only accurate wave scattering patterns.
but also real-time responses when coding the programmable Space-time modulation requires setting the element phase
metasurface. Furthermore, the proposed deep learning based in a pace that can keep up with the wave motion. This is
coding approach can be incorporated into hardwares, such as a challenge for large metasurfaces at high frequencies. This
FPGA that can further accelerate its computation speed and paper mainly concerns with the computation of the element
easier to communicate with the element code controller. code. However, there are more challenges for real time mod-
ulation, one of which is the transfer of the element code from
IV. C ONCLUSIONS AND D ISCUSSION the controller, such as personal computers, to every element.
In this work, we investigated the feasibility of applying deep This could be improved by moving from serial communication
learning techniques to code the programmable metasurface to parallel communication, and using hardware to do more
for multi-beam steering. A deep convolutional neural network computation. More works need to be done on this aspect.
consisting of eleven convolutional layers is built to compute This study is a start of exploration to apply deep learning
the element codes given the requirement of the scattering techniques to electromagnetic engineering, where physical
patterns. The input of the network includes the parameters laws used to dominate in research and development. With
of the desired beam pattern and the output is the coding the development of deep learning techniques, improvement
matrix to program the metasurface. This network is trained in learning capacity may allow machines to “learn” from a
with data of coding matrices computed by back-projection or large amount of physical data and “master” the physical law
nonlinear optimization approaches. With the help of massive in certain controlled boundary conditions. This study shows

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
124 IEEE JOURNAL ON EMERGING AND SELECTED TOPICS IN CIRCUITS AND SYSTEMS, VOL. 10, NO. 1, MARCH 2020

one promising example of their applications in real-time com- [20] N. K. Grady et al., “Terahertz metamaterials for linear polarization
putation of coding matrices for programmable metasurfaces. conversion and anomalous refraction,” Science, vol. 340, no. 6138,
pp. 1304–1307, Jun. 2013.
The increase of the depth and scale of the deep neural [21] X. Zhang et al., “Broadband terahertz wave deflection based on c-shape
network significantly expands its learning and generalization complex metamaterials with phase discontinuities,” Adv. Mater., vol. 25,
ability. This may allow it to approximate more complex no. 33, pp. 4567–4572, Sep. 2013.
[22] C. Pfeiffer and A. Grbic, “Metamaterial huygens’ surfaces: Tailoring
functions. The operation in each layer reduces the size of wave fronts with reflectionless sheets,” Phys. Rev. Lett., vol. 110, no. 19,
the input, but extends the number of output arrays, i.e., the 2013, Art. no. 197401.
size is reduced in one dimension but increases in another [23] N. Yu, F. Aieta, P. Genevet, M. A. Kats, Z. Gaburro, and F. Capasso,
“A broadband, background-free quarter-wave plate based on plasmonic
dimension, this computational concept is very similar to fast metasurfaces,” Nano Lett., vol. 12, no. 12, pp. 6328–6333, Dec. 2012.
Fourier transform, and many fast algorithms in numerical [24] Y. Hadad, D. L. Sounas, and A. Alù, “Space-time gradient metasurfaces,”
modeling. This may benefit the effectiveness of deep neural Phys. Rev. B, Condens. Matter, vol. 92, no. 10, 2015, Art. no. 100304.
[25] Y. Hadad, J. C. Soric, and A. Alu, “Breaking temporal symmetries for
network in working on the physical task in this paper. In the emission and absorption,” Proc. Nat. Acad. Sci. USA, vol. 113, no. 13,
long run, a hybridization of fundamental physical principles pp. 3471–3475, Mar. 2016.
with “knowledge” from big data could unleash numerous [26] T. J. Cui, M. Q. Qi, X. Wan, J. Zhao, and Q. Cheng, “Coding
engineering applications that used to be impossible due to the metamaterials, digital metamaterials and programmable metamaterials,”
Light Sci. Appl., vol. 3, no. 10, pp. e218–e218, Oct. 2014.
limit of data information and ability of computation. [27] L. Li et al., “Electromagnetic reprogrammable coding-metasurface holo-
grams,” Nature Commun., vol. 8, no. 1, p. 197, Aug. 2017.
[28] B. Xie, K. Tang, H. Cheng, Z. Liu, S. Chen, and J. Tian, “Cod-
R EFERENCES ing acoustic metasurfaces,” Adv. Mater., vol. 29, no. 6, Feb. 2017,
Art. no. 1603507.
[1] N. I. Zheludev and Y. S. Kivshar, “From metamaterials to metadevices,” [29] A. K. Bhattacharyya, Phased Array Antennas: Floquet Analysis, Synthe-
Nature Mater., vol. 11, no. 11, p. 917, 2012. sis, BFNs and Active Array Systems. Hoboken, NJ, USA: Wiley, 2006.
[2] T. J. Cui, D. Smith, and R. Liu, Metamaterials: Theory, Design, and [30] P. Nayeri, F. Yang, and A. Z. Elsherbeni, Reflectarray Antennas: Theory,
Applications. 2009. Designs, and Applications. Hoboken, NJ, USA: Wiley, 2018.
[3] D. R. Smith, W. J. Padilla, D. C. Vier, S. C. Nemat-Nasser, and [31] T. Isernia, A. Massa, A. F. Morabito, and P. Rocca, “On the optimal
S. Schultz, “Composite medium with simultaneously negative perme- synthesis of phase-only reconfigurable antenna arrays,” in Proc. 5th Eur.
ability and permittivity,” Phys. Rev. Lett., vol. 84, no. 18, pp. 4184–4187, Conf. Antennas Propag. (EUCAP), Apr. 2011, pp. 2074–2077.
Jul. 2002. [32] D. Pinchera, M. D. Migliore, and G. Panariello, “Synthesis of large
[4] J. B. Pendry, “Negative refraction makes a perfect lens,” Phys. Rev. Lett., sparse arrays using IDEA (inflating-deflating exploration algorithm),”
vol. 85, no. 18, pp. 3966–3969, Jul. 2002. IEEE Trans. Antennas Propag., vol. 66, no. 9, pp. 4658–4668, Sep. 2018.
[5] D. Schurig et al., “Metamaterial electromagnetic cloak at microwave [33] J. Johnson and Y. Rahmat-Samii, “Genetic algorithm optimization for
frequencies,” Science, vol. 314, no. 5801, pp. 977–980, Nov. 2006. aerospace electromagnetic design and analysis,” in Proc. IEEE Aerosp.
[6] U. Leonhardt, “Optical conformal mapping,” Science, vol. 312, no. 5781, Appl. Conf., vol. 1, Dec. 2002, pp. 87–102.
pp. 1777–1780, Jun. 2006. [34] D. E. Goldberg and J. H. Holland, “Genetic algorithms and machine
[7] T. Ergin, N. Stenger, P. Brenner, J. B. Pendry, and M. Wegener, “Three- learning,” Mach. Learn., vol. 3, nos. 2–3, pp. 95–99, 1988.
dimensional invisibility cloak at optical wavelengths,” Science, vol. 328, [35] J. Robinson and Y. Rahmat-Samii, “Particle swarm optimization in
no. 5976, pp. 337–339, Apr. 2010. electromagnetics,” IEEE Trans. Antennas Propag., vol. 52, no. 2,
[8] H. Yang et al., “A programmable metasurface with dynamic polarization, pp. 397–407, Feb. 2004.
scattering and focusing control,” Sci. Rep., vol. 6, p. 35692, Oct. 2016. [36] A. Lommi, A. Massa, E. Storti, and A. Trucco, “Sidelobe reduction in
[9] C. L. Holloway, E. F. Kuester, J. A. Gordon, J. O’hara, J. Booth, sparse linear arrays by genetic algorithms,” Microw. Opt. Technol. Lett.,
and D. R. Smith, “An overview of the theory and applications of vol. 32, no. 3, pp. 194–196, Feb. 2002.
metasurfaces: The two-dimensional equivalents of metamaterials,” IEEE [37] S. Caorsi, M. Donelli, A. Lommi, and A. Massa, “A real-time approach
Antennas Propag. Mag., vol. 54, no. 2, pp. 10–35, Apr. 2012. to array control based on a learned genetic algorithm,” Microw. Opt.
[10] D. Lin, P. Fan, E. Hasman, and M. L. Brongersma, “Dielectric gradient Technol. Lett., vol. 36, no. 4, pp. 235–238, Feb. 2003.
metasurface optical elements,” Science, vol. 345, no. 6194, pp. 298–302, [38] M. Donelli, S. Caorsi, F. DeNatale, M. Pastorino, and A. Massa, “Linear
Jul. 2014. antenna synthesis with a hybrid genetic algorithm,” Prog. Electromagn.
[11] L.-H. Gao et al., “Broadband diffusion of terahertz waves by multi- Res., vol. 49, pp. 1–22, 2004.
bit coding metasurfaces,” Light Sci Appl, vol. 4, no. 9, pp. e324–e324, [39] P. Rocca, L. Poli, G. Oliveri, and A. Massa, “GA-based adaptive thinning
Sep. 2015. strategy for pattern nulling in linear arrays,” in Proc. 6th Eur. Conf.
[12] K. Wang, J. Zhao, Q. Cheng, D. S. Dong, and T. J. Cui, “Broadband and Antennas Propag. (EUCAP), Mar. 2012, pp. 1526–1527.
broad-angle low-scattering metasurface based on hybrid optimization [40] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
algorithm,” Sci. Rep., vol. 4, p. 5935, Aug. 2014. no. 7553, p. 436, 2015.
[13] S. Liu et al., “Convolution operations on coding metasurface to reach [41] G. Hinton et al., “Deep neural networks for acoustic modeling in speech
flexible and continuous controls of terahertz beams,” Adv. Sci., vol. 3, recognition: The shared views of four research groups,” IEEE Signal
no. 10, Oct. 2016, Art. no. 1600156. Process. Mag., vol. 29, no. 6, pp. 82–97, Nov. 2012.
[14] L. Zhang et al., “Space-time-coding digital metasurfaces,” Nature Com- [42] T. N. Sainath, A.-R. Mohamed, B. Kingsbury, and B. Ramabhadran,
mun., vol. 9, no. 1, p. 4334, 2018. “Deep convolutional neural networks for LVCSR,” in Proc. IEEE Int.
[15] L. Cong et al., “A perfect metamaterial polarization rotator,” Appl. Phys. Conf. Acoust., Speech Signal Process., May 2013, pp. 8614–8618.
Lett., vol. 103, no. 17, Oct. 2013, Art. no. 171107. [43] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification
[16] Y. Z. Cheng et al., “Ultrabroadband reflective polarization convertor with deep convolutional neural networks,” in Proc. Adv. Neural Inf.
for terahertz waves,” Appl. Phys. Lett., vol. 105, no. 18, Nov. 2014, Process. Syst., 2012, pp. 1097–1105.
Art. no. 181111. [44] J. J. Tompson, A. Jain, Y. LeCun, and C. Bregler, “Joint training of a
[17] N. Yu et al., “Light propagation with phase discontinuities: General- convolutional network and a graphical model for human pose estima-
ized laws of reflection and refraction,” Science, vol. 334, no. 6054, tion,” in Proc. Adv. Neural Inf. Process. Syst., 2014, pp. 1799–1807.
pp. 333–337, Oct. 2011. [45] J. Yue-Hei Ng, M. Hausknecht, S. Vijayanarasimhan, O. Vinyals,
[18] S. Sun, Q. He, S. Xiao, Q. Xu, X. Li, and L. Zhou, “Gradient-index R. Monga, and G. Toderici, “Beyond short snippets: Deep networks for
meta-surfaces as a bridge linking propagating waves and surface waves,” video classification,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit.
Nature Mater., vol. 11, no. 5, pp. 426–431, May 2012. (CVPR), Jun. 2015, pp. 4694–4702.
[19] Z. Li et al., “Controlling propagation and coupling of waveguide modes [46] J. Tompson, K. Schlachter, P. Sprechmann, and K. Perlin, “Accelerating
using phase-gradient metasurfaces,” Nature Nanotech., vol. 12, no. 7, eulerian fluid simulation with convolutional networks,” in Proc. 34th Int.
pp. 675–683, Jul. 2017. Conf. Mach. Learn., vol. 17, 2017, pp. 3424–3433.

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.
SHAN et al.: CODING PROGRAMMABLE METASURFACES BASED ON DEEP LEARNING TECHNIQUES 125

[47] T. Shan, X. Dang, M. Li, F. Yang, S. Xu, and J. Wu, “Study on a 3D Shenheng Xu (Member, IEEE) received the
Possion’s equation slover based on deep learning technique,” in Proc. B.S. and M.S. degrees from Southeast University,
IEEE Int. Conf. Comput. Electromagn. (ICCEM), Mar. 2018, pp. 1–3. Nanjing, China, in 2001 and 2004, respectively, and
[48] Z. Wei and X. Chen, “Deep-learning schemes for full-wave nonlinear the Ph.D. degree in electrical engineering from the
inverse scattering problems,” IEEE Trans. Geosci. Remote Sens., vol. 57, University of California, Los Angeles (UCLA), CA,
no. 4, pp. 1849–1860, Apr. 2019. USA, in 2009.
[49] L. Li, L. G. Wang, F. L. Teixeira, C. Liu, A. Nehorai, and From 2000 to 2004, he was a Research Assistant
T. J. Cui, “DeepNIS: Deep neural network for nonlinear electromagnetic with the State Key Laboratory of Millimeter Waves,
inverse scattering,” IEEE Trans. Antennas Propag., vol. 67, no. 3, Southeast University, Nanjing. From 2004 to 2011,
pp. 1819–1825, Mar. 2019. he was a Graduate Student Researcher and later a
[50] R. Guo, X. Song, M. Li, F. Yang, S. Xu, and A. Abubakar, “Supervised Post-Doctoral Researcher at the Antenna Research,
descent learning technique for 2-D microwave imaging,” IEEE Trans. Analysis, and Measurement Laboratory, UCLA. In 2012, he joined the Depart-
Antennas Propag., vol. 67, no. 5, pp. 3550–3554, May 2019. ment of Electronic Engineering, Tsinghua University, Beijing, China, as an
[51] D. E. Golberg, Genetic Algorithms in Search, Optimization, and Machine Associate Professor. His research interests include novel designs of high-gain
Learning. Reading, MA, USA: Addison-Wesley, 1989. antennas for advanced applications, artificial electromagnetic structures, and
[52] H. Mühlenbein, M. Schomisch, and J. Born, “The parallel genetic electromagnetic and antenna theories.
algorithm as function optimizer,” Parallel Comput., vol. 17, nos. 6–7,
pp. 619–632, Sep. 1991.
[53] K. Simonyan and A. Zisserman, “Very deep convolutional networks
for large-scale image recognition,” 2014, arXiv:1409.1556. [Online].
Available: https://arxiv.org/abs/1409.1556
[54] C. M. Bishop et al., Neural Networks for Pattern Recognition. London,
U.K: Oxford Univ. Press, 1995.
[55] B. D. Ripley, Pattern Recognition and Neural Networks. Cambridge,
U.K.: Cambridge Univ. Press, 2007.
[56] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for
image recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit.
(CVPR), Jun. 2016, pp. 770–778.
[57] X. Guo, W. Li, and F. Iorio, “Convolutional neural networks for steady
flow approximation,” in Proc. 22nd ACM SIGKDD Int. Conf. Knowl.
Discovery Data Mining (KDD). New York, NY, USA: ACM, 2016,
pp. 481–490.
Tao Shan received the B.S. degree from the School
of Electronic Engineering, Xidian University, Xi’an,
China, in 2016. He is currently pursuing the Ph.D.
degree with the Department of Electronic Engineer-
ing, Tsinghua University, Beijing, China.
His current research interests include fast algo- Fan Yang (Fellow, IEEE) received the B.S. and
rithms for computational electromagnetics, opti- M.S. degrees from Tsinghua University, Beijing,
mization algorithms for programmable metasurface, China, in 1997 and 1999, respectively, and the Ph.D.
deep learning, and its applications in electromagnetic degree from the University of California at Los
modeling. Angeles (UCLA) in 2002.
From 1994 to 1999, he was a Research Assistant
Xiaotian Pan received the B.S. degree from at the State Key Laboratory of Microwave and
Zhengzhou University in 2013, and the M.S. degree Digital Communications, Tsinghua University. From
from Beihang University in 2016. He is currently 1999 to 2002, he was a Graduate Student Researcher
pursuing the Ph.D. degree with the Department of at the Antenna Laboratory, UCLA. From 2002 to
Electronic Engineering, Tsinghua University. 2004, he was a Post-Doctoral Research Engineer and
His current research interests include reconfig- Instructor at the Electrical Engineering Department, UCLA. In 2004, he joined
urable reflectarray, reconfigurable transmitarray, and the Electrical Engineering Department, The University of Mississippi, as an
the imaging applications. Assistant Professor, and was promoted to the post of an Associate Professor
in 2009. In 2011, he joined the Electronic Engineering Department, Tsinghua
University, as a Professor, and has served as the Director of the Microwave and
Antenna Institute since then. His research interests include antennas, surface
Maokun Li (Member, IEEE) received the B.S. electromagnetics, computational electromagnetics, and applied electromag-
degree in electronic engineering from Tsinghua Uni- netic systems. He has published over 300 journal articles and conference
versity, Beijing, China, in 2002, and the M.S. and papers, six book chapters, and five books Reflectarray Antennas: Theory,
Ph.D. degrees in electrical engineering from the Uni- Designs, and Applications (IEEE-Wiley, 2018), Analysis and Design of
versity of Illinois at Urbana-Champaign, in 2004 and Transmitarray Antennas (Morgan & Claypool, 2017), Scattering Analysis of
2007, respectively. Periodic Structures Using Finite-Difference Time-Domain Method (Morgan &
After graduation, he worked as a Senior Research Claypool, 2012), Electromagnetic Band Gap Structures in Antenna Engi-
Scientist with the Schlumberger-Doll Research, neering (Cambridge Univ. Press, 2009), and Electromagnetics and Antenna
Cambridge, MA, USA. Since 2014, he joined the Optimization Using Taguchi’s Method (Morgan & Claypool, 2007).
Department of Electronic Engineering, Tsinghua Dr. Yang is a fellow of the ACES. He has been a recipient of several
University, Beijing. His research interests include prestigious awards and recognitions, including the Young Scientist Award of
fast algorithms in computational electromagnetics and their applications in the 2005 URSI General Assembly and of the 2007 International Symposium
antenna modeling, electromagnetic compatibility analysis, inverse problems, on Electromagnetic Theory, the 2008 Junior Faculty Research Award of the
and so on. He has published one book chapter, 50 journal papers, 120 con- University of Mississippi, the 2009 inaugural IEEE Donald G. Dudley Jr.
ference proceedings, and three patent applications. He was also among the Undergraduate Teaching Award, and the 2011 Recipient of Global Experts
recipients of the China National 1000 Plan in 2014, and the 2017 IEEE Ulrich Program of China. He was the Technical Program Committee (TPC) Chair
L. Rohde Innovative Conference Paper Award. He also serves as an Associate of the 2014 IEEE International Symposium on Antennas and Propagation
Editor for the IEEE J OURNAL ON M ULTISCALE AND M ULTIPHYSICS C OM - and USNC-URSI Radio Science Meeting. He has served as an Associate
PUTATIONAL T ECHNIQUES , Applied Computational Electromagnetic Society Editor for the IEEE T RANSACTIONS ON A NTENNAS AND P ROPAGATION
Journal, and the Guest Editor for the special issue on “Electromagnetic Inverse (2010–2013) and an Associate Editor-in-Chief of the Applied Computational
Problems for Sensing and Imaging” in the IEEE Antennas and Propagation Electromagnetics Society (ACES) Journal (2008–2014). He is also an IEEE
Magazine. APS Distinguished Lecturer (2018–2020).

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on June 17,2020 at 15:01:47 UTC from IEEE Xplore. Restrictions apply.